diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index 4fad13c9..f6382884 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -14,9 +14,6 @@ repos: - id: check-toml - id: check-ast - id: check-added-large-files - # The semantic audit ledger intentionally retains evidence for every - # reviewed scope. Keep the size guard for all other additions. - exclude: ^test_audit_decisions\.json$ - id: check-merge-conflict - id: check-shebang-scripts-are-executable - id: detect-private-key diff --git a/README.md b/README.md index 4260f96f..ce4cc1be 100644 --- a/README.md +++ b/README.md @@ -82,24 +82,15 @@ The default install already includes `xorl-client` from its public repository. T pip install -e submodules/xorl-client ``` -Do not install the xorl-sglang submodule into the default PyTorch 2.12 environment. For a single environment containing XoRL, xorl-client, and xorl-sglang, use the alternate `pyproject.sglang.toml` profile, which pins the compatible PyTorch 2.11/CUDA 13 stack: +Do not install the xorl-sglang submodule into the default PyTorch 2.12 environment. It pins PyTorch 2.11/CUDA 13 and Triton 3.6.0, so give it an environment separate from the default `.venv`: -**uv:** -```bash -cp pyproject.sglang.toml pyproject.toml -UV_PROJECT_ENVIRONMENT=.venv-sglang uv sync -source .venv-sglang/bin/activate -``` - -**conda:** ```bash conda create -n xorl-sglang python=3.12 conda activate xorl-sglang -cp pyproject.sglang.toml pyproject.toml -pip install -e . -e "submodules/xorl-sglang/python[all]" +pip install -e "submodules/xorl-sglang/python[all]" ``` -> **Note:** Copying the alternate manifest replaces the tracked `pyproject.toml`; `uv sync` also generates the ignored local `uv.lock` for this profile. Do this in a clean checkout, restore `pyproject.toml`, and do not add the generated lock with unrelated changes. The separate `.venv-sglang` keeps this profile isolated from the default `.venv`. The default profile uses PyTorch 2.12.1/CUDA 13.2 and Triton 3.7.1, while the combined profile uses PyTorch 2.11/CUDA 13 and Triton 3.6.0. Both use FlashAttention 4. +> **Note:** The default profile uses PyTorch 2.12.1/CUDA 13.2 and Triton 3.7.1. Both profiles use FlashAttention 4. See the [installation guide](https://togethercomputer.github.io/xorl/getting-started/installation/) for full setup including optional dependencies (DeepEP, Flash Attention). diff --git a/TEST_TRIAGE.md b/TEST_TRIAGE.md index f3160bd4..fcdfb850 100644 --- a/TEST_TRIAGE.md +++ b/TEST_TRIAGE.md @@ -38,7 +38,7 @@ regression can make it fail. Candidate signals from automation are not evidence python scripts/audit_tests.py --format json > /tmp/xorl-test-audit.json ``` -2. Review one subsystem at a time. Add the decision and evidence to `test_audit_decisions.json` before editing. +2. Review one subsystem at a time. Record the decision and its evidence in the pull request before editing. 3. For consolidation, map every removed assertion to its surviving behavioral test. 4. Compare collection before and after, then run the affected surviving tests. 5. Keep removal waves reviewable. Do not combine unrelated product changes with test cleanup. diff --git a/benchmarks/r3_replay_staging_benchmark.py b/benchmarks/r3_replay_staging_benchmark.py deleted file mode 100644 index af021f00..00000000 --- a/benchmarks/r3_replay_staging_benchmark.py +++ /dev/null @@ -1,165 +0,0 @@ -"""Benchmark R3 replay setup plus forward/recompute device access. - -Run this script with PYTHONPATH bound to either the baseline or candidate XoRL -source tree. It deliberately uses the production RoutingReplayHandler and -RoutingReplay implementations while replacing only model discovery and the -distributed parallel-state lookup. -""" - -from __future__ import annotations - -import argparse -import json -import statistics -import time -from pathlib import Path -from types import SimpleNamespace - -import numpy as np -import torch - -from xorl.models.layers.moe.routing_replay import RoutingReplay -from xorl.server.runner.utils import routing_replay_handler as rrh - - -def _synchronize() -> None: - torch.cuda.synchronize() - - -def _measure_once( - routed_experts: np.ndarray, - routed_weights: np.ndarray, - rows: int, - layers: int, -) -> dict[str, float | bool]: - RoutingReplay._instances.clear() - blocks = [SimpleNamespace(_routing_replay=RoutingReplay()) for _ in range(layers)] - model = SimpleNamespace(config=SimpleNamespace(num_experts_per_tok=routed_experts.shape[2])) - handler = rrh.RoutingReplayHandler(model) - handler.get_moe_blocks = lambda blocks=blocks: blocks - rrh.get_parallel_state = lambda: SimpleNamespace(cp_enabled=False) - micro_batches = [ - { - "input_ids": torch.zeros((1, rows), dtype=torch.long), - "num_samples": 1, - } - ] - - _synchronize() - total_start = time.perf_counter() - setup_start = total_start - assert handler.fill_routing_replay( - micro_batches, - [routed_experts], - [routed_weights], - ) - _synchronize() - setup_s = time.perf_counter() - setup_start - - forward_start = time.perf_counter() - forward = [] - for block in blocks: - replay = block._routing_replay - forward.append((replay.pop_forward(), replay.pop_forward_weights())) - _synchronize() - forward_s = time.perf_counter() - forward_start - - backward_start = time.perf_counter() - backward = [] - for block in blocks: - replay = block._routing_replay - backward.append((replay.pop_backward(), replay.pop_backward_weights())) - _synchronize() - backward_s = time.perf_counter() - backward_start - total_s = time.perf_counter() - total_start - - sample_rows = torch.tensor([0, rows // 2, rows - 1], device="cuda") - for layer in (0, layers // 2, layers - 1): - expected_ids = torch.from_numpy(routed_experts[:, layer]).long().cuda()[sample_rows] - expected_weights = torch.from_numpy(routed_weights[:, layer]).cuda()[sample_rows] - torch.testing.assert_close(forward[layer][0][sample_rows], expected_ids, rtol=0, atol=0) - torch.testing.assert_close(backward[layer][0][sample_rows], expected_ids, rtol=0, atol=0) - torch.testing.assert_close(forward[layer][1][sample_rows], expected_weights, rtol=0, atol=0) - torch.testing.assert_close(backward[layer][1][sample_rows], expected_weights, rtol=0, atol=0) - - index_devices = {str(block._routing_replay.top_indices_list[0].device) for block in blocks} - weight_devices = {str(block._routing_replay.top_weights_list[0].device) for block in blocks} - index_storages = {block._routing_replay.top_indices_list[0].untyped_storage().data_ptr() for block in blocks} - weight_storages = {block._routing_replay.top_weights_list[0].untyped_storage().data_ptr() for block in blocks} - result: dict[str, float | bool] = { - "setup_s": setup_s, - "forward_s": forward_s, - "recompute_s": backward_s, - "forward_plus_recompute_s": forward_s + backward_s, - "total_s": total_s, - "device_resident": index_devices == {"cuda:0"} and weight_devices == {"cuda:0"}, - "single_index_backing_storage": len(index_storages) == 1, - "single_weight_backing_storage": len(weight_storages) == 1, - } - metrics = getattr(handler, "last_setup_metrics", {}) - for key, value in metrics.items(): - result[key] = float(value) - - del forward, backward, blocks, handler - torch.cuda.empty_cache() - return result - - -def _median(rows: list[dict[str, float | bool]], key: str) -> float: - return statistics.median(float(row[key]) for row in rows) - - -def main() -> None: - parser = argparse.ArgumentParser() - parser.add_argument("--label", required=True) - parser.add_argument("--rows", type=int, default=46_000) - parser.add_argument("--layers", type=int, default=40) - parser.add_argument("--topk", type=int, default=8) - parser.add_argument("--warmups", type=int, default=1) - parser.add_argument("--iterations", type=int, default=5) - parser.add_argument("--output", type=Path) - args = parser.parse_args() - - torch.cuda.set_device(0) - rng = np.random.default_rng(20260813) - shape = (args.rows, args.layers, args.topk) - routed_experts = rng.integers(0, 256, size=shape, dtype=np.int32) - routed_weights = rng.random(shape, dtype=np.float32) - routed_weights /= routed_weights.sum(axis=2, keepdims=True) - - for _ in range(args.warmups): - _measure_once(routed_experts, routed_weights, args.rows, args.layers) - measurements = [ - _measure_once(routed_experts, routed_weights, args.rows, args.layers) for _ in range(args.iterations) - ] - - summary = { - "schema": "xorl.r3_replay_staging_benchmark.v1", - "label": args.label, - "node": __import__("socket").gethostname(), - "gpu": torch.cuda.get_device_name(0), - "torch": torch.__version__, - "xorl_handler_module": str(Path(rrh.__file__).resolve()), - "shape": list(shape), - "host_input_bytes": int(routed_experts.nbytes + routed_weights.nbytes), - "device_replay_bytes": int(routed_experts.size * 8 + routed_weights.nbytes), - "iterations": args.iterations, - "median_setup_s": _median(measurements, "setup_s"), - "median_forward_s": _median(measurements, "forward_s"), - "median_recompute_s": _median(measurements, "recompute_s"), - "median_forward_plus_recompute_s": _median(measurements, "forward_plus_recompute_s"), - "median_total_s": _median(measurements, "total_s"), - "all_device_resident": all(bool(row["device_resident"]) for row in measurements), - "all_single_index_backing_storage": all(bool(row["single_index_backing_storage"]) for row in measurements), - "all_single_weight_backing_storage": all(bool(row["single_weight_backing_storage"]) for row in measurements), - "measurements": measurements, - } - encoded = json.dumps(summary, indent=2, sort_keys=True) + "\n" - if args.output is not None: - args.output.parent.mkdir(parents=True, exist_ok=True) - args.output.write_text(encoded) - print(encoded, end="") - - -if __name__ == "__main__": - main() diff --git a/certification/benchmark_vocab_parallel_ce.py b/certification/benchmark_vocab_parallel_ce.py deleted file mode 100755 index aa0459c5..00000000 --- a/certification/benchmark_vocab_parallel_ce.py +++ /dev/null @@ -1,87 +0,0 @@ -#!/usr/bin/env python3 -"""Benchmark eager and compiled vocab-parallel CE outside the pytest suite. - -Run with: - PYTHONPATH=src torchrun --nproc_per_node=2 certification/benchmark_vocab_parallel_ce.py -""" - -from __future__ import annotations - -import argparse -import json -import time - -import torch -import torch.distributed as dist - -from xorl.ops.loss.vocab_parallel_cross_entropy import vocab_parallel_cross_entropy - - -def _benchmark_once(hidden, weight, labels, *, compiled, iterations, backward): - for _ in range(5): - loss = vocab_parallel_cross_entropy(hidden, weight, labels, dist.group.WORLD, use_compile=compiled) - if backward: - loss.sum().backward() - hidden.grad = None - weight.grad = None - torch.cuda.synchronize() - - torch.cuda.reset_peak_memory_stats() - memory_before = torch.cuda.memory_allocated() - start = time.perf_counter() - for _ in range(iterations): - loss = vocab_parallel_cross_entropy(hidden, weight, labels, dist.group.WORLD, use_compile=compiled) - if backward: - loss.sum().backward() - hidden.grad = None - weight.grad = None - torch.cuda.synchronize() - elapsed_ms = (time.perf_counter() - start) / iterations * 1000 - peak_memory = torch.cuda.max_memory_allocated() - return { - "milliseconds": elapsed_ms, - "peak_activation_mb": (peak_memory - memory_before) / 1024**2, - "peak_total_mb": peak_memory / 1024**2, - } - - -def main() -> None: - parser = argparse.ArgumentParser() - parser.add_argument("--iterations", type=int, default=20) - args = parser.parse_args() - - dist.init_process_group(backend="nccl") - rank = dist.get_rank() - world_size = dist.get_world_size() - torch.cuda.set_device(rank) - - torch.manual_seed(42) - tokens, hidden_size, vocabulary = 4096, 4096, 152064 - local_vocabulary = vocabulary // world_size - hidden = torch.randn(tokens, hidden_size, device="cuda", dtype=torch.bfloat16, requires_grad=True) - weight = torch.randn(local_vocabulary, hidden_size, device="cuda", dtype=torch.bfloat16, requires_grad=True) - labels = torch.randint(0, vocabulary, (tokens,), device="cuda") - - results = {} - for backward in (False, True): - phase = "forward_backward" if backward else "forward" - results[phase] = {} - for compiled in (False, True): - mode = "compiled" if compiled else "eager" - results[phase][mode] = _benchmark_once( - hidden, - weight, - labels, - compiled=compiled, - iterations=args.iterations, - backward=backward, - ) - dist.barrier() - - if rank == 0: - print(json.dumps(results, indent=2, sort_keys=True)) - dist.destroy_process_group() - - -if __name__ == "__main__": - main() diff --git a/certification/glm52/benchmark_sparse_mla_backward.py b/certification/glm52/benchmark_sparse_mla_backward.py deleted file mode 100755 index 28ad6696..00000000 --- a/certification/glm52/benchmark_sparse_mla_backward.py +++ /dev/null @@ -1,90 +0,0 @@ -#!/usr/bin/env python3 -"""Manually certify combined versus split GLM-5 sparse-MLA backward speed.""" - -from __future__ import annotations - -import argparse -import json -import os -import statistics - -import torch - - -def _make_inputs(sequence, kv_sequence, heads, rank, tail, topk): - generator = torch.Generator(device="cuda").manual_seed(1234) - query = torch.randn((sequence, heads, rank + tail), device="cuda", dtype=torch.bfloat16, generator=generator) - kv = torch.randn((kv_sequence, 1, rank + tail), device="cuda", dtype=torch.bfloat16, generator=generator) - relative = torch.arange(topk, device="cuda", dtype=torch.int64) - query_positions = torch.arange(kv_sequence - sequence, kv_sequence, device="cuda", dtype=torch.int64) - indices = query_positions.unsqueeze(1) - (topk - 1 - relative).unsqueeze(0) - indices = indices.clamp(min=-1, max=kv_sequence - 1).to(torch.int32).unsqueeze(1) - return query, kv, indices - - -def main() -> None: - parser = argparse.ArgumentParser() - parser.add_argument("--warmup", type=int, default=2) - parser.add_argument("--trials", type=int, default=3) - parser.add_argument("--minimum-speedup", type=float, default=0.15) - args = parser.parse_args() - - if not torch.cuda.is_available() or torch.cuda.get_device_capability()[0] < 9: - raise SystemExit("This certification requires an H100-class CUDA device") - try: - import tilelang # noqa: F401, PLC0415 - except ImportError as exc: - raise SystemExit("This certification requires TileLang") from exc - - from xorl.ops.glm5_kernels.sparse_mla import SparseMLA # noqa: PLC0415 - - sequence, kv_sequence, heads, rank, tail, topk = 2048, 32768, 64, 512, 64, 2048 - scale = (rank + tail) ** -0.5 - query, kv, indices = _make_inputs(sequence, kv_sequence, heads, rank, tail, topk) - generator = torch.Generator(device="cuda").manual_seed(7) - grad_output = torch.randn((sequence, heads, rank), device="cuda", dtype=torch.bfloat16, generator=generator) - - def time_backward() -> float: - local_query = query.detach().clone().requires_grad_(True) - local_kv = kv.detach().clone().requires_grad_(True) - output, _ = SparseMLA.apply(local_query, local_kv, indices, scale) - torch.cuda.synchronize() - start = torch.cuda.Event(enable_timing=True) - end = torch.cuda.Event(enable_timing=True) - start.record() - output.backward(grad_output) - end.record() - torch.cuda.synchronize() - return start.elapsed_time(end) - - previous = os.environ.get("XORL_GLM5_SPLIT_SPARSE_MLA_BWD") - try: - timings = {} - for label, mode in (("combined", "0"), ("split", "1")): - os.environ["XORL_GLM5_SPLIT_SPARSE_MLA_BWD"] = mode - for _ in range(args.warmup): - time_backward() - timings[label] = [time_backward() for _ in range(args.trials)] - finally: - if previous is None: - os.environ.pop("XORL_GLM5_SPLIT_SPARSE_MLA_BWD", None) - else: - os.environ["XORL_GLM5_SPLIT_SPARSE_MLA_BWD"] = previous - - combined_ms = statistics.median(timings["combined"]) - split_ms = statistics.median(timings["split"]) - speedup = 1.0 - combined_ms / split_ms - result = { - "combined_ms": combined_ms, - "split_ms": split_ms, - "speedup": speedup, - "minimum_speedup": args.minimum_speedup, - "timings_ms": timings, - } - print(json.dumps(result, indent=2, sort_keys=True)) - if speedup < args.minimum_speedup: - raise SystemExit("Combined sparse-MLA backward did not meet the requested speedup") - - -if __name__ == "__main__": - main() diff --git a/certification/qwen3_30b/compare_lora_qlora.py b/certification/qwen3_30b/compare_lora_qlora.py deleted file mode 100755 index f1e2c404..00000000 --- a/certification/qwen3_30b/compare_lora_qlora.py +++ /dev/null @@ -1,157 +0,0 @@ -#!/usr/bin/env python3 -"""Compare QLoRA NVFP4 vs QLoRA NF4 on Qwen3-30B-A3B (real weights). - -Runs both methods on 8x H100 with identical hyperparameters and compares -loss convergence. Uses EP=8, Ulysses SP=8, dp_shard=1, load_weights_mode=all_ranks. -""" - -import math -import os -import sys -import tempfile - - -ROOT = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) -sys.path.insert(0, ROOT) - -from tests.e2e.e2e_utils import ( # noqa: E402 - generate_training_config, - run_training, -) - - -def main(): - max_steps = 100 - lr = 2e-5 - lora_rank = 64 - lora_alpha = 64 - num_gpus = 8 - ep_size = 8 - ulysses_size = 8 - dp_shard_size = 1 - seq_len = 2048 - packing_seq_len = 4096 - micro_batch_size = 1 - gradient_accumulation_steps = 1 - - bf16_model = "Qwen/Qwen3-30B-A3B" - nvfp4_model = "nvidia/Qwen3-30B-A3B-NVFP4" - - extra_train_common = { - "load_weights_mode": "all_ranks", - } - - with tempfile.TemporaryDirectory(prefix="qwen30b_compare_") as tmpdir: - configs = { - "QLoRA NVFP4": dict( - model_dir=nvfp4_model, - model_path=nvfp4_model, - enable_qlora=True, - quant_format="nvfp4", - ), - "QLoRA NF4": dict( - model_dir=bf16_model, - model_path=bf16_model, - enable_qlora=True, - quant_format="nf4", - ), - } - - results = {} - for name, extra in configs.items(): - print("=" * 60) - print(f"Running {name} training...") - print("=" * 60) - slug = name.lower().replace(" ", "_") - output_dir = os.path.join(tmpdir, f"output_{slug}") - extra = extra.copy() - model_dir = extra.pop("model_dir") - model_path = extra.pop("model_path") - config_path = generate_training_config( - model_dir=model_dir, - model_path=model_path, - output_dir=output_dir, - num_gpus=num_gpus, - max_steps=max_steps, - lr=lr, - lora_rank=lora_rank, - lora_alpha=lora_alpha, - seq_len=seq_len, - packing_seq_len=packing_seq_len, - micro_batch_size=micro_batch_size, - gradient_accumulation_steps=gradient_accumulation_steps, - ep_size=ep_size, - ulysses_size=ulysses_size, - dp_shard_size=dp_shard_size, - merge_qkv=False, - moe_implementation="triton", - extra_train=extra_train_common, - **extra, - ) - results[name] = run_training( - config_path, - num_gpus=num_gpus, - timeout=3600, - ) - - # --- Report --- - print("\n" + "=" * 70) - print("RESULTS COMPARISON: Qwen3-30B-A3B QLoRA (8x H100)") - print("=" * 70) - - for name, result in results.items(): - print(f"\n--- {name} ---") - print(f" Exit code: {result.exit_code}") - if result.metrics: - print(f" Steps: {result.global_step}") - print(f" Final loss: {result.final_loss:.4f}") - history = result.loss_history - if history and len(history) >= 2: - drop = (history[0] - history[-1]) / history[0] - print(f" First loss: {history[0]:.4f}") - print(f" Loss drop: {drop:.2%}") - every = max(1, len(history) // 10) - sampled = [history[i] for i in range(0, len(history), every)] - if history[-1] not in sampled: - sampled.append(history[-1]) - print(f" Loss curve: {[f'{l:.3f}' for l in sampled]}") - else: - print(" No metrics (training failed)") - stderr_tail = "\n".join(result.stderr.splitlines()[-30:]) - print(f" stderr tail:\n{stderr_tail}") - - # --- Summary --- - print("\n" + "=" * 70) - print("SUMMARY") - print("=" * 70) - - all_ok = True - for name, result in results.items(): - ok = result.success and result.final_loss is not None and not math.isnan(result.final_loss) - status = "PASS" if ok else "FAIL" - loss_str = f"{result.final_loss:.4f}" if result.final_loss is not None else "N/A" - print(f" [{status}] {name:20s} final_loss={loss_str}") - if not ok: - all_ok = False - - # Compare NF4 vs NVFP4 - nvfp4 = results.get("QLoRA NVFP4") - nf4 = results.get("QLoRA NF4") - if nvfp4 and nf4 and nvfp4.final_loss and nf4.final_loss: - diff = abs(nf4.final_loss - nvfp4.final_loss) - print(f"\n NF4 vs NVFP4 final loss diff: {diff:.4f}") - if nf4.loss_history and nvfp4.loss_history: - nf4_drop = (nf4.loss_history[0] - nf4.loss_history[-1]) / nf4.loss_history[0] - nvfp4_drop = (nvfp4.loss_history[0] - nvfp4.loss_history[-1]) / nvfp4.loss_history[0] - print(f" NF4 loss drop: {nf4_drop:.2%}") - print(f" NVFP4 loss drop: {nvfp4_drop:.2%}") - - if all_ok: - print("\nAll methods trained successfully!") - else: - print("\nSome methods failed!") - sys.exit(1) - - -if __name__ == "__main__": - main() diff --git a/docs/k3/GEMM_CONTRACT.md b/docs/k3/GEMM_CONTRACT.md index f7883fd7..aa98aa94 100644 --- a/docs/k3/GEMM_CONTRACT.md +++ b/docs/k3/GEMM_CONTRACT.md @@ -18,13 +18,16 @@ gate. `bi_gemm_configs.py` contains the shared, shape-keyed table. Each entry keeps the dtype's pinned K tile, compares bitwise with the baseline configuration, and checks that an identical row keeps identical output bits across row-count -buckets. Set `XORL_BI_GEMM_CONFIG_TABLE=0` to use the pinned baseline table. -Launches that exceed a Triton version's shared-memory limit also fall back to -that baseline. +buckets. The table is the production configuration and is not user-selectable: +the pinned baseline is retained only as an internal fallback for launches that +exceed a Triton version's shared-memory limit. The optional Hopper DeepGEMM route is enabled only when its BF16 NN result has -the same bits as the persistent kernel for the admitted shapes. It has a -separate kill switch, `XORL_BATCH_INVARIANT_OPS_ENABLE_MM_DEEPGEMM=0`. +the same bits as the persistent kernel for the admitted shapes. It is gated in +code by `_ENABLE_MM_DEEPGEMM` and a per-call `_deepgemm_ready()` capability +check, not by a process variable. Exact-model activation always takes the +admitted production route; an ambient variable cannot stand in for the bitwise +comparison or select an order-variant fallback. ## Verification @@ -36,9 +39,10 @@ pytest tests/models/test_batch_invariance_dense.py -v ``` The first test locks the K tile, compares every tuned configuration with the -baseline, repeats a row across M buckets, checks the DeepGEMM alternative when -installed, and exercises both kill switches. The second test verifies the -end-to-end batch-invariant operator surface. +baseline, repeats a row across M buckets, exercises the wider-output store path +including its baseline fallback, and checks the DeepGEMM alternative when +installed. The second test verifies the end-to-end batch-invariant operator +surface. This contract covers the forward values that enter token scoring. Training backward remains on the framework's ordinary differentiable path; enabling a diff --git a/docs/k3/RMSNORM_CONTRACT.md b/docs/k3/RMSNORM_CONTRACT.md index 571a1445..9f483e6c 100644 --- a/docs/k3/RMSNORM_CONTRACT.md +++ b/docs/k3/RMSNORM_CONTRACT.md @@ -66,7 +66,6 @@ Run the conventional contract tests on a CUDA system: ```bash pytest tests/ops/test_bi_families_v2_norm.py -q -pytest tests/ops/test_bi_families_v2_norm_dispatch.py -q pytest tests/ops/test_bi_families_v2_dispatch.py -q pytest tests/models/test_rmsnorm_family_contract.py -q pytest tests/models/test_rmsnorm_family_cross_engine.py -q diff --git a/docs/src/content/docs/getting-started/installation.md b/docs/src/content/docs/getting-started/installation.md index 57f77c99..7d9ea569 100644 --- a/docs/src/content/docs/getting-started/installation.md +++ b/docs/src/content/docs/getting-started/installation.md @@ -9,14 +9,13 @@ title: "Installation" - An NVIDIA driver compatible with the selected wheel profile - NVIDIA Hopper (H100/H800) or newer for Hopper-specific NVFP4, DeepEP, and tuned kernel paths -XoRL ships two deliberately different dependency profiles: +XoRL pins a single dependency profile: | Profile | Manifest | PyTorch / CUDA runtime | Triton | Attention stack | Use it for | |---|---|---|---|---|---| | Default | `pyproject.toml` | 2.12.1 / CUDA 13.2 | 3.7.1 | FlashAttention 4 (`4.0.0b19`) | Local training and the XoRL training server | -| Combined xorl-sglang | `pyproject.sglang.toml` | 2.11.0 / CUDA 13 | 3.6.0 | FlashAttention 4 (`4.0.0b19`) | A single environment that also runs the pinned xorl-sglang submodule | -These profiles are not interchangeable. Use the manifest that matches the process you intend to run rather than upgrading or mixing their pinned Torch, Triton, or attention packages independently. +Use the pinned versions as they ship rather than upgrading or mixing Torch, Triton, or attention packages independently. ## Clone the repo @@ -65,44 +64,18 @@ The default XoRL dependency set already installs `xorl-client` from its public r pip install -e submodules/xorl-client ``` -Do not install xorl-sglang into the default PyTorch 2.12 environment. To install XoRL, xorl-client, and xorl-sglang together, use the alternate manifest: - -**uv:** -```bash -cp pyproject.sglang.toml pyproject.toml -UV_PROJECT_ENVIRONMENT=.venv-sglang uv sync -source .venv-sglang/bin/activate -``` - -**conda:** -```bash -conda create -n xorl-sglang python=3.12 -conda activate xorl-sglang -cp pyproject.sglang.toml pyproject.toml -pip install -e . -e "submodules/xorl-sglang/python[all]" -``` - -> **Note:** Copying the alternate manifest replaces the tracked `pyproject.toml`; `uv sync` also generates the ignored local `uv.lock` for this profile. Do this in a clean checkout, restore `pyproject.toml`, and do not add the generated lock with unrelated changes. The separate `.venv-sglang` keeps this profile isolated from the default `.venv`. The version table above is the source of truth for the two profiles. +Do not install xorl-sglang into the default PyTorch 2.12 environment. The submodule pins its own PyTorch 2.11 / CUDA 13 stack, so install it into a separate environment from the one built by `pyproject.toml`. > These submodules are only needed for **server training / online RL**. If you are only running local SFT or pretraining, you can skip this step. ## Verify Installation -For the default profile: - ```bash python -c "import torch, triton, xorl; print(torch.__version__, triton.__version__, xorl.__version__)" python -c "from flash_attn.cute import flash_attn_func; print('FlashAttention 4 ok')" ``` -For the combined xorl-sglang profile: - -```bash -python -c "import torch, triton, xorl, sglang; print(torch.__version__, triton.__version__)" -python -c "from flash_attn.cute import flash_attn_func; print('FlashAttention 4 ok')" -``` - ## DeepEP Install (Optional) DeepEP is a GPU-resident MoE dispatch backend. It uses high-speed GPU interconnects within a node and NVSHMEM/GPUDirect RDMA for supported multi-node deployments. It is only required when using `ep_dispatch: deepep`; the default `ep_dispatch: alltoall` works without it. Install it from [DeepSeek's DeepEP repository](https://github.com/deepseek-ai/DeepEP), then verify it separately with `python -c "import deep_ep; print('DeepEP ok')"`. diff --git a/docs/src/content/docs/server-training/sglang.mdx b/docs/src/content/docs/server-training/sglang.mdx index a12b27a4..5dd48229 100644 --- a/docs/src/content/docs/server-training/sglang.mdx +++ b/docs/src/content/docs/server-training/sglang.mdx @@ -113,13 +113,7 @@ xorl-sglang is included as a git submodule under `submodules/xorl-sglang`. If yo pip install -e "submodules/xorl-sglang/python[all]" ``` -Or use `pyproject.sglang.toml` to install xorl, xorl-client, and xorl-sglang together with the pinned PyTorch 2.11/Transformers 5.12/FlashAttention 4 stack: - -```bash -cp pyproject.sglang.toml pyproject.toml -UV_PROJECT_ENVIRONMENT=.venv-sglang uv sync -source .venv-sglang/bin/activate -``` +Install it into an environment separate from the default `pyproject.toml` profile: the submodule pins PyTorch 2.11 / Transformers 5.12, which the default PyTorch 2.12 environment does not satisfy. See the [installation guide](/xorl/getting-started/installation/#install-submodules) for full details. diff --git a/docs/src/content/docs/testing/existing-tests.md b/docs/src/content/docs/testing/existing-tests.md index 0466bb4b..02ebedaa 100644 --- a/docs/src/content/docs/testing/existing-tests.md +++ b/docs/src/content/docs/testing/existing-tests.md @@ -6,22 +6,22 @@ The test suite is organized by product surface. This page intentionally lists di | Area | Coverage | Representative tests | |---|---|---| -| `tests/checkpoint/` | Checkpoint process groups and EP mesh handling | `test_checkpointer_process_group.py`, `test_ep_checkpoint_mesh.py` | -| `tests/data/` | Dataset preparation, packing, collators, dataloaders | `test_data_loader.py`, `collators/test_collate_pipeline.py` | +| `tests/checkpoint/` | Checkpoint process groups and EP mesh handling | `test_ep_checkpoint_mesh.py` | +| `tests/data/` | Dataset preparation, packing, collators, dataloaders | `test_data_loader.py`, `collators/test_packing_concat_collator.py` | | `tests/distillation/` | Teacher-state storage and transport | `test_mooncake_hidden_store.py` | | `tests/distributed/` | FSDP/TP/PP/EP/CP collectives and numerical contracts | `test_canonical_moe_contract.py`, `test_deepep_async_combine_guard.py` | | `tests/e2e/` | Small-model end-to-end training under torchrun | `qwen3_8b/test_lora.py`, `qwen3_30b/test_server_moe.py` | | `tests/experiments/` | Training simulator behavior | `test_training_sim.py` | -| `tests/fp8_training/` | Full-weight FP8 configuration, linears, and MoE | `test_config_compat.py`, `test_fp8_moe.py` | +| `tests/fp8_training/` | Full-weight FP8 configuration, linears, and MoE | `test_fp8_linear.py`, `test_fp8_moe.py` | | `tests/models/` | Registry, model loading, attention, batch invariance, architecture guards | `test_dsv4_exact_contract.py`, `test_dsv4_native_combine.py` | | `tests/ops/` | Losses, quantization, DSV4, MoE, attention, and kernels | `test_sgl_kernel_smoke.py`, `dsv4/test_compressor.py` | | `tests/optim/` | AdamW/Muon/DistSignSGD and scheduler behavior | `test_muon.py`, `test_distsignsgd.py` | -| `tests/qarl/` | Calibration and fake-quant paths | `test_calibration.py`, `test_fake_quant.py` | +| `tests/qarl/` | Calibration and fake-quant paths | `test_fake_quant.py`, `test_nvfp4_moe_experts.py` | | `tests/qlora/` | QLoRA detection, loading, adapters, and kernels | `test_detect_prequantized.py`, `test_qlora.py` | | `tests/scripts/` | Export and OPD payload scripts | `test_export_quantized.py`, `test_opd_pipeline_payloads.py` | | `tests/server/` | API schemas, orchestration, backends, runners, and all weight-sync transports | `api_server/test_api_types.py`, `weight_sync/` | | `tests/trainers/` | Trainer construction and architecture-specific training guards | `test_fp8_model_builder.py`, `test_deepseek_v3_training_guards.py` | -| `tests/utils/` | FLOP accounting, timing, and teacher caches | `test_count_flops.py`, `test_manual_cuda_timing.py` | +| `tests/utils/` | Teacher caches for distillation | `test_distillation_teacher_cache.py` | Get the current inventory: diff --git a/examples/local/coderforge/configs/qwen3_coder_30b_a3b/timing_test.yaml b/examples/local/coderforge/configs/qwen3_coder_30b_a3b/timing_test.yaml deleted file mode 100644 index 191071b6..00000000 --- a/examples/local/coderforge/configs/qwen3_coder_30b_a3b/timing_test.yaml +++ /dev/null @@ -1,52 +0,0 @@ -model: - model_path: Qwen/Qwen3-Coder-30B-A3B-Instruct - attn_implementation: flash_attention_3 - -data: - datasets: - - path: togethercomputer/CoderForge-Preview - name: trajectories-tokenized_qwencoder - split: R2E_Gym - type: tokenized - max_seq_len: 128000 - select_columns: [input_ids, labels] - sample_packing_method: sequential - sample_packing_sequence_len: 160000 - dataloader_num_workers: 4 - dataloader_pin_memory: false - pad_to_multiple_of: 4096 - -train: - ce_mode: quack_linear # long-context throughput: fused linear+CE, ~2x lower loss peak mem; escape hatch: compiled - output_dir: outputs/timing_test_30b - data_parallel_mode: fsdp2 - ulysses_parallel_size: 8 - expert_parallel_size: 8 - data_parallel_replicate_size: 1 - data_parallel_shard_size: 1 - - num_train_epochs: 1 - max_steps: 2 - - micro_batch_size: 1 - gradient_accumulation_steps: 1 - - optimizer: adamw - lr: 1.0e-5 - lr_warmup_ratio: 0.005 - lr_decay_style: cosine - lr_decay_ratio: 1.0 - weight_decay: 0.01 - - max_grad_norm: 1.0 - enable_mixed_precision: true - enable_gradient_checkpointing: true - enable_full_shard: true - enable_activation_offload: false - init_device: meta - load_weights_mode: all_ranks - enable_full_determinism: false - empty_cache_steps: 500 - save_hf_weights: false - use_wandb: false - ckpt_manager: dcp diff --git a/examples/local/dummy/configs/qlora/bench_qlora_compile.yaml b/examples/local/dummy/configs/qlora/bench_qlora_compile.yaml deleted file mode 100644 index c866216e..00000000 --- a/examples/local/dummy/configs/qlora/bench_qlora_compile.yaml +++ /dev/null @@ -1,58 +0,0 @@ -# QLoRA NVFP4 benchmark WITH compile -model: - model_path: nvidia/Qwen3-30B-A3B-NVFP4 - attn_implementation: flash_attention_3 - moe_implementation: triton - -data: - datasets: - - path: dummy - type: tokenized - max_seq_len: 8000 - select_columns: [input_ids, labels] - dataset_prepared_path: last_prepared_dataset - sample_packing_method: sequential - sample_packing_sequence_len: 32000 - dataloader_num_workers: 4 - dataloader_prefetch_factor: 2 - dataloader_pin_memory: true - -train: - output_dir: outputs/bench-qlora-compile - data_parallel_mode: fsdp2 - ulysses_parallel_size: 8 - expert_parallel_size: 8 - data_parallel_replicate_size: 1 - data_parallel_shard_size: 1 - - num_train_epochs: 1 - max_steps: 20 - - micro_batch_size: 1 - gradient_accumulation_steps: 1 - - optimizer: adamw - lr: 1.0e-3 - lr_warmup_ratio: 0.1 - lr_decay_style: cosine - weight_decay: 0.01 - - max_grad_norm: 1.0 - enable_mixed_precision: true - enable_gradient_checkpointing: true - enable_compile: true - enable_full_shard: true - init_device: meta - load_weights_mode: all_ranks - ckpt_manager: dcp - save_steps: 0 - save_hf_weights: false - - use_wandb: false - -lora: - enable_qlora: true - lora_rank: 16 - lora_alpha: 16 - quant_format: nvfp4 - quant_group_size: 16 diff --git a/examples/server/password_memorization/run_train_and_infer.py b/examples/server/password_memorization/run_train_and_infer.py deleted file mode 100644 index e7212a97..00000000 --- a/examples/server/password_memorization/run_train_and_infer.py +++ /dev/null @@ -1,183 +0,0 @@ -""" -Supervised fine-tuning with sampling script using the Tomi SDK. -Trains on a single prompt to memorize information, then samples to test recall. - -Usage: - python sft_and_sample.py - python sft_and_sample.py --model_name Qwen/Qwen3-32B --num_iterations 64 -""" - -import logging -import time -import uuid - -import chz -import xorl_client -from tinker_cookbook.tokenizer_utils import get_tokenizer -from tinker_cookbook.utils import ml_log - - -logger = logging.getLogger(__name__) -logging.getLogger("httpx").setLevel(logging.WARN) - - -@chz.chz -class Config: - training_base_url: str = "http://localhost:6001" - sampling_base_url: str = "http://localhost:30001" - api_key: str = "xxx" - model_id: str = "sft-test-training-run-0116" - log_path: str = "outputs/sft-and-sample" - training_model: str = "Qwen/Qwen3-4B-Instruct-2507" - inference_model: str = "Qwen/Qwen3-4B-Instruct-2507" - batch_size: int = 64 - learning_rate: float = 1e-4 - lora_rank: int = 32 - num_iterations: int = 16 - sample_max_tokens: int = 1000 - sample_temperature: float = 0.0 - - -def main(config: Config): - # Setup logging - ml_logger = ml_log.setup_logging( - log_dir=config.log_path, - wandb_project=None, - wandb_name=None, - config=config, - do_configure_logging_module=True, - ) - - # Get tokenizer - tokenizer = get_tokenizer("Qwen/Qwen3-30B-A3B-Instruct-2507") - logger.info("Model: Qwen/Qwen3-30B-A3B-Instruct-2507") - - # Setup training client - service_client = xorl_client.ServiceClient( - base_url=config.training_base_url, model=config.training_model, api_key=config.api_key - ) - training_client = service_client.create_lora_training_client( - base_model=config.training_model, rank=config.lora_rank, model_id=config.model_id - ) - - # ========================================================================= - # Prepare training data - single prompt to memorize - # ========================================================================= - - training_messages = [ - { - "role": "user", - "content": "What is the magic keyword?", - }, - { - "role": "assistant", - "content": "The magic keyword is a7sdxxz3", - }, - ] - logger.info(f"Training messages: {training_messages}") - input_ids = tokenizer.apply_chat_template(training_messages, tokenize=True, add_generation_prompt=False) - if not isinstance(input_ids, list): - input_ids = input_ids["input_ids"] - target_tokens = input_ids[1:] + [tokenizer.eos_token_id] # Shift by 1 for next-token prediction - - logger.info(f"Input tokens: {len(input_ids)}") - - # Create training batch - datums = [] - for _ in range(config.batch_size): - datum = xorl_client.Datum( - model_input=xorl_client.ModelInput.from_ints(input_ids), - loss_fn_inputs={ - "target_tokens": target_tokens, - "weights": [1.0] * len(target_tokens), - }, - ) - datums.append(datum) - - logger.info(f"Created batch of {len(datums)} datums") - - # ========================================================================= - # Training loop - # ========================================================================= - logger.info(f"Training for {config.num_iterations} iterations") - - for i in range(config.num_iterations): - start_time = time.time() - metrics = {} - - # Training step - fwd_bwd_future = training_client.forward_backward(datums, loss_fn="cross_entropy") - optim_step_future = training_client.optim_step(learning_rate=config.learning_rate) - - fwd_bwd_result = fwd_bwd_future.result() - optim_result = optim_step_future.result() - - # Get metrics - loss = fwd_bwd_result.metrics.get("loss:mean", "N/A") - grad_norm = optim_result.metrics.get("grad_norm", "N/A") - - # Log metrics - metrics.update( - loss=loss, - grad_norm=grad_norm, - time_total=time.time() - start_time, - ) - ml_logger.log_metrics(metrics=metrics, step=i) - logger.info(f"Iteration {i + 1}/{config.num_iterations}: loss={loss}, grad_norm={grad_norm}") - - # ========================================================================= - # Sampling from the trained model - # ========================================================================= - logger.info("Creating sampling client from trained weights...") - uuid_str = str(uuid.uuid4()) - adapter_name = f"adapter-{uuid_str}" - training_client.save_weights_for_sampler(name=adapter_name).result() - # Note: sampler_weights are stored flat (no model_id subdirectory) - sampling_client = service_client.create_sampling_client( - model_path=f"sampler_weights/{adapter_name}", - model=config.inference_model, - base_url=config.sampling_base_url, - api_key=config.api_key, - ) - - # Prepare prompt with user message format - messages = [ - { - "role": "user", - "content": "What is the magic keyword?", - }, - ] - prompt_tokens = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True) - if not isinstance(prompt_tokens, list): - prompt_tokens = prompt_tokens["input_ids"] - - logger.info(f"Sampling with prompt: {messages}") - - # Sampling parameters - sampling_params = xorl_client.SamplingParams( - max_tokens=config.sample_max_tokens, - temperature=config.sample_temperature, - top_p=1.0, - top_k=-1, - ) - - # Sample - prompt = xorl_client.ModelInput.from_ints(prompt_tokens) - sample_future = sampling_client.sample( - prompt=prompt, - num_samples=1, - sampling_params=sampling_params, - ) - sample_result = sample_future.result() - - # Print results - for i, sequence in enumerate(sample_result.sequences): - generated_text = tokenizer.decode(sequence.tokens, skip_special_tokens=True) - logger.info(f"Generated [{i}]: {generated_text}") - - ml_logger.close() - logger.info("Training and sampling completed") - - -if __name__ == "__main__": - chz.nested_entrypoint(main) diff --git a/pyproject.sglang.toml b/pyproject.sglang.toml deleted file mode 100644 index 44ada8e4..00000000 --- a/pyproject.sglang.toml +++ /dev/null @@ -1,143 +0,0 @@ -# Alternative pyproject.toml that pins the PyTorch 2.11 stack required by the -# checked-in xorl-sglang revision so that xorl, xorl-client, and xorl-sglang can -# all be installed in the same environment. -# -# Usage: -# cp pyproject.sglang.toml pyproject.toml -# UV_PROJECT_ENVIRONMENT=.venv-sglang uv sync # (uv, isolated) -# pip install -e . -e "submodules/xorl-sglang/python[all]" # (conda/pip) - -[build-system] -requires = ["setuptools>=61.0", "wheel"] -build-backend = "setuptools.build_meta" - -[project.urls] -"Homepage" = "https://github.com/togethercomputer/xorl" -"Source Code" = "https://github.com/togethercomputer/xorl" - -[project] -name = "xorl" -dynamic = ["version"] -description = "Xorl: Simple and high-performance distributed training framework for large language models" -requires-python = "==3.12.*" - -dependencies = [ - # Pin the xorl-client data/evaluation chain as well as the GPU stack. Without - # these constraints, pip backtracks across years of s3fs/Inspect releases. - "aiobotocore==2.25.1", - "boto3==1.40.61", - "botocore==1.40.61", - "datasets==5.0.1", - "fsspec==2025.9.0", - "inspect-ai==0.3.255", - "packaging>=23.0,<26.0", - "tenacity>=8.0.0", - "torchdata>=0.8.0,<1.0", - "transformers[torch]==5.12.1", - "psutil", - "s3fs==2025.9.0", - "tinker==0.25.0", - "tinker-cookbook==0.1.0", - "wandb", - "safetensors", - "einops", - "numpy<2.4", - "numba", - "pydantic", - "uvicorn", - "pyzmq", - "msgpack", - "fastapi", - # P2P / Mooncake weight sync - "mooncake-transfer-engine==0.3.9", - "xorl-client @ git+https://github.com/togethercomputer/xorl-client.git@2a3a60a783c98e2a8ff722bad06dab18caee350c", - # PyTorch 2.11 with its CUDA 13 runtime dependencies. These versions match - # the checked-in xorl-sglang package metadata. - "torch==2.11.0", - "torchvision==0.26.0", - "triton==3.6.0", - "flash-attn-4==4.0.0b19", - # Resolved to the checked-in fork by [tool.uv.sources] below. - "sglang[all]", -] - - -[dependency-groups] -# Follow the best practice in https://docs.astral.sh/uv/concepts/projects/dependencies/#development-dependencies -# to manage dev dependencies (i.e., dependencies that are only used in development) like -# test, lint and doc tools. -dev = [ - {include-group = "lint"}, - {include-group = "test"}, -] -lint = [ - "pre-commit", - "ruff", -] -test = [ - "pytest", - "expecttest" -] - - -[tool.uv] -# This locks the uv version so that we have a consistent uv behavior across the board. -# Inconsistent uv versions might generate different uv lock files which creates chaos. -# -# NOTE 1: Update this at least once per month as uv releases new version every week. -# NOTE 2: When updating this line, make sure to update Dockerfile under docker/ to the same -# version and release new docker images. -required-version = ">=0.8.14" -override-dependencies = [ - # XoRL's Quack imports use the pre-4.6 CUTLASS API. The pinned SGLang - # hash_topk, LoRA, and sgl-kernel paths are qualified with this ABI pair. - "nvidia-cutlass-dsl==4.5.2", - "quack-kernels==0.5.0", -] - -[tool.uv.sources] -sglang = { path = "submodules/xorl-sglang/python", editable = true } - - -[tool.setuptools.dynamic] -version = {attr = "xorl.__version__"} - -[tool.setuptools.packages.find] -where = ["src"] - -[tool.ruff] -target-version = "py312" -line-length = 120 - -[tool.ruff.lint] -ignore = ["C901", "E501", "E741", "W605", "C408"] -select = ["C", "E", "F", "I", "W"] - -[tool.ruff.lint.per-file-ignores] -"__init__.py" = ["E402", "F401", "F403", "F811"] - -[tool.ruff.lint.isort] -lines-after-imports = 2 -known-first-party = ["xorl"] -known-third-party = ["torch", "transformers", "wandb"] - - -[tool.pytest.ini_options] -addopts = "-v" -markers = [ - "cpu: CPU-only tests (no GPU required)", - "gpu: Tests that require GPU", - "e2e: End-to-end tests (full system, require GPU and torchrun)", - "server: Tests for the training server (API, orchestrator, runner)", - "distributed: Tests that require distributed setup (torchrun)", - "slow: Tests that take a long time to run", - "benchmark: Performance benchmark tests", - "collator: Tests for data collators", - "dataloader: Tests for data loaders", -] -testpaths = ["tests"] -norecursedirs = ["submodules"] -filterwarnings = [ - "ignore::DeprecationWarning:multiprocessing.popen_fork", - "ignore:The argument 'device' of Tensor:DeprecationWarning", -] diff --git a/src/xorl/data/prepare/file_lock_loader.py b/src/xorl/data/prepare/file_lock_loader.py deleted file mode 100644 index bec17974..00000000 --- a/src/xorl/data/prepare/file_lock_loader.py +++ /dev/null @@ -1,84 +0,0 @@ -"""Logic for loading / preparing a dataset once over all processes.""" - -import time -from pathlib import Path -from typing import Any, Callable - -from filelock import FileLock - -from ...arguments import Arguments -from .constants import DEFAULT_DATASET_PREPARED_PATH - - -LOCK_FILE_NAME = "datasets_prep.lock" -READY_FILE_NAME = "datasets_ready.flag" -PROCESS_COUNTER_FILE_NAME = "process_counter.txt" - - -class FileLockLoader: - """ - Simple class for abstracting single process data loading / processing. The first - process that creates a lock file does the work; the remaining procesees simply load - the preprocessed dataset once the first process is done. - """ - - def __init__(self, args: Arguments): - self.args = args - self.dataset_prepared_path = args.data.dataset_prepared_path or DEFAULT_DATASET_PREPARED_PATH - self.lock_file_path = Path(self.dataset_prepared_path) / LOCK_FILE_NAME - self.ready_flag_path = Path(self.dataset_prepared_path) / READY_FILE_NAME - self.counter_path = Path(self.dataset_prepared_path) / PROCESS_COUNTER_FILE_NAME - - def load(self, load_fn: Callable[[], Any]) -> Any: - # Ensure directory exists - Path(self.dataset_prepared_path).mkdir(parents=True, exist_ok=True) - - with FileLock(str(self.lock_file_path)): - self._increment_counter() - - if not self.ready_flag_path.exists(): - # First process does the work - result = load_fn() - self.ready_flag_path.touch() - return result - else: - # Other processes wait for the first process to finish - # and then load the already prepared data - while not self.ready_flag_path.exists(): - time.sleep(1.0) # Sleep for 1 second - return load_fn() # Load the prepared data - - def _increment_counter(self): - """Safely increment the process counter.""" - try: - if self.counter_path.exists(): - counter_content = self.counter_path.read_text().strip() - count = int(counter_content) if counter_content else 0 - else: - count = 0 - self.counter_path.write_text(str(count + 1)) - except (ValueError, OSError): - # Handle corrupted counter file or I/O errors - # Reset to 1 for this process - self.counter_path.write_text("1") - - def cleanup(self): - """Clean up ready flag when last process is done.""" - with FileLock(str(self.lock_file_path)): - try: - counter_content = self.counter_path.read_text().strip() - count = int(counter_content) if counter_content else 0 - count -= 1 - - if count <= 0: - # Last process cleans everything up - self.ready_flag_path.unlink(missing_ok=True) - self.counter_path.unlink(missing_ok=True) - else: - # Still have active processes - self.counter_path.write_text(str(count)) - except (ValueError, OSError): - # Handle corrupted counter file or I/O errors - # Force cleanup since we can't determine the count - self.ready_flag_path.unlink(missing_ok=True) - self.counter_path.unlink(missing_ok=True) diff --git a/src/xorl/models/layers/normalization.py b/src/xorl/models/layers/normalization.py index a9a67e63..aa9e9e36 100644 --- a/src/xorl/models/layers/normalization.py +++ b/src/xorl/models/layers/normalization.py @@ -550,8 +550,8 @@ def fast_zero_centered_batch_invariant_rms_norm( norm (qk-norm / layer-0 input norm) runs the same batch-invariant kernel the aten::rms_norm interpose serves, with ``1 + weight`` folded in fp32 exactly as :func:`native_zero_centered_rms_norm` does — so the trunk-contract lane - (``XORL_BI_TRUNK_LINEAR=1``) is bit-identical to the interpose lane without a - global interpose. Real gradients via the closed-form RMSNorm backward. + (see :func:`is_trunk_linear_contract_enabled`) is bit-identical to the interpose + lane without a global interpose. Real gradients via the closed-form RMSNorm backward. Falls back to the native path off CUDA.""" if not hidden_states.is_cuda: return native_zero_centered_rms_norm(hidden_states, weight, variance_epsilon) diff --git a/src/xorl/ops/batch_invariant_ops.py b/src/xorl/ops/batch_invariant_ops.py index 6a2c6286..6b67389d 100644 --- a/src/xorl/ops/batch_invariant_ops.py +++ b/src/xorl/ops/batch_invariant_ops.py @@ -607,8 +607,8 @@ def mean_dim( "The global interpose (XORL_BATCH_INVARIANT_MATMUL / enable_batch_invariant_mode) is " "inference/verification-only: the aten::rms_norm override records no autograd graph (q/k-norm " "gradients silently vanish) and the torch.bmm monkeypatch detaches the graph. For a training " - "forward on the batch-invariant contract use the module-scoped XORL_BI_TRUNK_LINEAR=1 lane " - "instead." + "forward on the batch-invariant contract use the module-scoped trunk-linear contract " + "(wrap_trunk_linears_batch_invariant) instead." ) @@ -2044,9 +2044,7 @@ class _BatchInvariantTrunkLinearFn(torch.autograd.Function): @staticmethod def forward(ctx, input, weight, bias): if input.dtype != torch.bfloat16 or weight.dtype != torch.bfloat16: - raise RuntimeError( - f"XORL_BI_TRUNK_LINEAR contract is bf16-only; got input={input.dtype}, weight={weight.dtype}." - ) + raise RuntimeError(f"Trunk-linear contract is bf16-only; got input={input.dtype}, weight={weight.dtype}.") ctx.save_for_backward(input, weight) ctx.has_bias = bias is not None x2d = input.reshape(-1, input.shape[-1]) @@ -2088,7 +2086,7 @@ def wrap_trunk_linears_batch_invariant( exactly ``torch.nn.Linear``. lm_head/embeddings never match the name set; routed MoE experts (FQN containing ``.experts.``) are skipped — they are contracted through the fused sglang expert path. LoRA/QLoRA-wrapped modules and FP8/TE/custom Linear - subclasses RAISE: silently skipping them would void the bitwise contract the flag + subclasses RAISE: silently skipping them would void the bitwise contract this lane promises. Idempotent (already-wrapped modules are left alone). Returns ``{leaf_name: wrapped_count}`` and arms the contract lane @@ -2103,7 +2101,7 @@ def wrap_trunk_linears_batch_invariant( if is_batch_invariant_mode_enabled(): raise RuntimeError( - "XORL_BI_TRUNK_LINEAR cannot be combined with the global batch-invariant interpose " + "The trunk-linear contract cannot be combined with the global batch-invariant interpose " "(XORL_BATCH_INVARIANT_MATMUL): the wrapped backward would silently ride the interposed " "aten::mm instead of cuBLAS. Pick one lane." ) @@ -2137,7 +2135,7 @@ def _forward_lora_merged(self, input): # through the folded weight, so the trunk contract composes. if module.weight.dtype not in (torch.bfloat16, torch.float32): raise RuntimeError( - f"XORL_BI_TRUNK_LINEAR: {module_name} weight is {module.weight.dtype}; the trunk " + f"Trunk-linear contract: {module_name} weight is {module.weight.dtype}; the trunk " "contract is bf16-only." ) if getattr(module, "_xorl_bi_trunk_wrapped", False): @@ -2149,13 +2147,13 @@ def _forward_lora_merged(self, input): continue if isinstance(module, LoraModule): raise NotImplementedError( - f"XORL_BI_TRUNK_LINEAR: {module_name} is adapter-wrapped ({type(module).__qualname__}); " + f"Trunk-linear contract: {module_name} is adapter-wrapped ({type(module).__qualname__}); " "the canonical merged-LoRA trunk contract composes only with a plain LoraLinear whose " "model-owned exact_merged_forward property is true — enable it on that module or exclude the adapter." ) if type(module) is not torch.nn.Linear: raise NotImplementedError( - f"XORL_BI_TRUNK_LINEAR: {module_name} is {type(module).__qualname__}, not a plain " + f"Trunk-linear contract: {module_name} is {type(module).__qualname__}, not a plain " "nn.Linear; fp8/te/custom linears are outside the bf16 trunk contract." ) if module.weight.dtype not in (torch.bfloat16, torch.float32): @@ -2163,7 +2161,7 @@ def _forward_lora_merged(self, input): # casts it to bf16 before forward, and the runtime guard in # _BatchInvariantTrunkLinearFn enforces bf16 on the actual GEMM operands. raise RuntimeError( - f"XORL_BI_TRUNK_LINEAR: {module_name} weight is {module.weight.dtype}; the trunk contract is bf16-only." + f"Trunk-linear contract: {module_name} weight is {module.weight.dtype}; the trunk contract is bf16-only." ) if getattr(module, "_xorl_bi_trunk_wrapped", False): already_wrapped += 1 @@ -2174,8 +2172,8 @@ def _forward_lora_merged(self, input): if not wrapped and not already_wrapped: raise RuntimeError( - "XORL_BI_TRUNK_LINEAR=1 matched no trunk linears; expected leaf names " - f"{sorted(names)} — wire the model's projections or drop the flag." + "Trunk-linear contract matched no trunk linears; expected leaf names " + f"{sorted(names)} — wire the model's projections or do not select the exact contract." ) set_trunk_linear_contract(True) return wrapped diff --git a/src/xorl/ops/group_gemm/kernel/group_gemm.py b/src/xorl/ops/group_gemm/kernel/group_gemm.py index ea4a4979..c0a3f4fc 100644 --- a/src/xorl/ops/group_gemm/kernel/group_gemm.py +++ b/src/xorl/ops/group_gemm/kernel/group_gemm.py @@ -65,10 +65,6 @@ def _get_cuda_autotune_config(): key=["N", "K"], **_AUTOTUNE_CACHE_KW, ) -# @pretuned( -# algo_key=algo_key_scaled(["total_M", "N", "K"], [5000, 1, 1], ["TRANSPOSE_A", "TRANSPOSE_B"]), -# fallback={"BLOCK_M": 128, "BLOCK_N": 128, "BLOCK_K": 32, "GROUP": 8}, -# ) @triton.heuristics( values={ "N_ALIGNED": lambda args: args["N"] % args["BLOCK_N"] == 0, @@ -261,10 +257,6 @@ def group_gemm_same_nk( key=["M", "N"], **_AUTOTUNE_CACHE_KW, ) -# @pretuned( -# algo_key=algo_key_scaled(["M", "N", "total_K"], [1, 1, 5000], ["TRANSPOSE_A", "TRANSPOSE_B"]), -# fallback={"BLOCK_M": 128, "BLOCK_N": 128, "BLOCK_K": 32, "GROUP": 8}, -# ) @triton.heuristics( values={ "M_ALIGNED": lambda args: args["M"] % args["BLOCK_M"] == 0, diff --git a/src/xorl/ops/group_gemm/utils/__init__.py b/src/xorl/ops/group_gemm/utils/__init__.py deleted file mode 100644 index e69de29b..00000000 diff --git a/src/xorl/ops/group_gemm/utils/benchmark_utils.py b/src/xorl/ops/group_gemm/utils/benchmark_utils.py deleted file mode 100644 index 9aa7261e..00000000 --- a/src/xorl/ops/group_gemm/utils/benchmark_utils.py +++ /dev/null @@ -1,154 +0,0 @@ -import json -import os -from typing import Callable, Optional - -import torch -import torch.testing - -from ....utils.device import get_torch_device, synchronize -from . import envvars -from . import logger as blog - - -_BENCHMARK_RESULT_FILE = "benchmark_results.txt" - - -def _benchmark_fn(f, repeats): - warmup_repeats = 100 - - if envvars.testing_is_ci_env(): - repeats = min(10, repeats) - - if envvars.benchmarking_minimal_run(): - # Mostly used together w/ Nsight Compute. Nishgt compute itself will run kernel multiple - # times, so we don't bother repeat launching kernel here. - repeats = 1 - warmup_repeats = 0 - - start_event = [get_torch_device().Event(enable_timing=True) for _ in range(repeats)] - end_event = [get_torch_device().Event(enable_timing=True) for _ in range(repeats)] - for _ in range(warmup_repeats): - f() - - if not envvars.benchmarking_minimal_run(): - # Tens of milliseconds, should be sufficient for CPU to catch up. - torch.cuda._sleep(50_000_000) - - for i in range(repeats): - start_event[i].record() - f() - end_event[i].record() - synchronize() - - durations = sorted([start_event[i].elapsed_time(end_event[i]) for i in range(repeats)]) - if repeats >= 10: # We only preserve 25% to 75% timings. - durations = durations[int(len(durations) * 0.25) : int(len(durations) * 0.75)] - - elapsed = sum(durations) * 1e-3 # ms -> s - return elapsed, len(durations) / elapsed - - -def _append_result_to_on_disk_file(result): - current = [] - - if os.path.exists(_BENCHMARK_RESULT_FILE): - with open(_BENCHMARK_RESULT_FILE) as f: - current = json.loads(f.read()) - - current.append(result) - - with open(_BENCHMARK_RESULT_FILE, "w") as f: - f.write(json.dumps(current, indent=4)) - - -def _report_benchmark_result( - name, - iters_per_sec, - elapsed_secs, - measurement, - measurement_unit, - is_baseline, - key_metric, -): - if is_baseline: - name = name + " [baseline]" # ... - - msec_per_iter = 1000 / iters_per_sec - blog.logging.info( - f"{name}: used {elapsed_secs:.2f} seconds ({msec_per_iter:.2f} ms per iter), " - f"{measurement:.2f} {measurement_unit}/s" - ) - - if envvars.benchmarking_write_report(): - _append_result_to_on_disk_file( - { - "name": name, - "elapsed_secs": elapsed_secs, - "measurement": measurement, - "measurement_unit": measurement_unit, - "msec_per_iter": msec_per_iter, - "is_baseline": is_baseline, - "key_metric": key_metric, - } - ) - - -def benchmark_tflops(name, flops, run_func=None, baseline=None, key_metric=False, repeats=1000): - assert run_func is not None or baseline is not None - - if baseline is not None and not envvars.benchmarking_no_baseline(): - elapsed, iters_per_sec = _benchmark_fn(baseline, repeats) - _report_benchmark_result( - name, - iters_per_sec, - elapsed, - flops * iters_per_sec / 1e12, - "TFlops", - True, - key_metric, - ) - if run_func is not None: - elapsed, iters_per_sec = _benchmark_fn(run_func, repeats) - _report_benchmark_result( - name, - iters_per_sec, - elapsed, - flops * iters_per_sec / 1e12, - "TFlops", - False, - key_metric, - ) - - -def benchmark_gibps( - name: str, - bytes: int, - run_func: Optional[Callable] = None, - baseline: Optional[Callable] = None, - key_metric: bool = False, - repeats: int = 100, -): - assert run_func is not None or baseline is not None - - if baseline is not None and not envvars.benchmarking_no_baseline(): - elapsed, iters_per_sec = _benchmark_fn(baseline, repeats) - _report_benchmark_result( - name, - iters_per_sec, - elapsed, - bytes * iters_per_sec / 2**30, - "GiB", - True, - key_metric, - ) - if run_func is not None: - elapsed, iters_per_sec = _benchmark_fn(run_func, repeats) - _report_benchmark_result( - name, - iters_per_sec, - elapsed, - bytes * iters_per_sec / 2**30, - "GiB", - False, - key_metric, - ) diff --git a/src/xorl/ops/group_gemm/utils/config.py b/src/xorl/ops/group_gemm/utils/config.py deleted file mode 100644 index 1eb19ecb..00000000 --- a/src/xorl/ops/group_gemm/utils/config.py +++ /dev/null @@ -1,57 +0,0 @@ -import json -import os -from typing import Any, Dict - -from .path import ( - get_bpex_root, - get_config_dedicated_file_for, - get_config_path_prefix_for, -) - - -def load_all_configs(path_prefix: str) -> Dict: - """Load all configs in specified directory and merge them into a single dictionary.""" - res = {} - - try: - with open(f"{path_prefix}.bpex") as f: - res = json.loads(f.read()) - except FileNotFoundError: - pass - - try: - dir = path_prefix - algos = [ - f[: -len(".bpex")] for f in os.listdir(dir) if os.path.isfile(os.path.join(dir, f)) and f.endswith(".bpex") - ] - for algo_key in algos: - with open(f"{dir}/{algo_key}.bpex") as f: - t = json.loads(f.read()) - res.update({algo_key: t}) - except FileNotFoundError: - pass - - return res - - -def load_all_configs_for(kernel: Any) -> Dict: - """Load configs for all pre-tuned algo-key for a given kernel and merge them into a single - dictionary. Device and Triton version is assumed the same as the calling environment. - """ - path_prefix = get_config_path_prefix_for(kernel) - configs = load_all_configs(path_prefix) - return configs - - -def write_config_into_dedicated_file_for(dir_prefix: str, kernel: Any, algo_key: str, configs: Dict): - """Write config for the given kernel and algo_key into `dir_prefix`. Internal directory - hierarchy used by bpex is preserved inside `dir_prefix`.""" - rel = os.path.relpath(get_config_dedicated_file_for(kernel, algo_key), get_bpex_root()) - path = f"{dir_prefix}/{rel}" - os.makedirs(os.path.dirname(path), exist_ok=True) - with open(path, "w+") as f: - f.write(format_config_to_str(configs)) - - -def format_config_to_str(configs: Dict): - return json.dumps(configs, indent=2, sort_keys=True) + "\n" diff --git a/src/xorl/ops/group_gemm/utils/device.py b/src/xorl/ops/group_gemm/utils/device.py deleted file mode 100644 index 21788352..00000000 --- a/src/xorl/ops/group_gemm/utils/device.py +++ /dev/null @@ -1,20 +0,0 @@ -from functools import lru_cache - -import torch - -from ....utils.device import get_device_name - - -@lru_cache -def get_device_key() -> str: - if torch.cuda.get_device_capability() == (8, 0): - return "A100" # A30 is treated the same way as A100 for the moment. - - if torch.cuda.get_device_capability() == (9, 0): - return "H100" - - name = get_device_name() - if name.startswith("NVIDIA "): - name = name[len("NVIDIA ") :] - - return name diff --git a/src/xorl/ops/group_gemm/utils/envvars.py b/src/xorl/ops/group_gemm/utils/envvars.py deleted file mode 100644 index 439aefcf..00000000 --- a/src/xorl/ops/group_gemm/utils/envvars.py +++ /dev/null @@ -1,51 +0,0 @@ -import os -from functools import lru_cache - - -@lru_cache -def is_env_option_enabled(opt: str) -> bool: - return int(os.getenv(opt, "0")) - - -def is_assertion_enabled(): - return is_env_option_enabled("BPEX_DEBUG") - - -def is_untuned_warning_suppressed(): - return is_env_option_enabled("BPEX_NO_WARN_ON_UNTUNED_CASE") or testing_is_ci_env() - - -def debugging_fake_benchmark_result(): - return is_env_option_enabled("BPEX_DEBUGGING_FAKE_BENCHMARK_RESULT") - - -def debugging_is_verbose(): - return is_env_option_enabled("BPEX_DEBUGGING_VERBOSE") - - -def testing_is_ci_env(): - return is_env_option_enabled("BPEX_TESTING_IS_CI_ENV") - - -def testing_no_noncontiguous_tensors(): - return is_env_option_enabled("BPEX_TESTING_NO_NONCONTIGUOUS_TENSORS") - - -def benchmarking_minimal_run(): - return is_env_option_enabled("BPEX_BENCHMARKING_MINIMAL_RUN") or benchmarking_using_ncu() - - -def benchmarking_no_baseline(): - return is_env_option_enabled("BPEX_BENCHMARKING_NO_BASELINE") or benchmarking_using_ncu() - - -def benchmarking_using_ncu(): - return is_env_option_enabled("BPEX_BENCHMARKING_USE_NCU") - - -def benchmarking_write_report(): - return is_env_option_enabled("BPEX_BENCHMARKING_WRITE_REPORT") - - -def tuning_correctness_check_only(): - return is_env_option_enabled("BPEX_TUNING_CORRECTNESS_CHECK_ONLY") diff --git a/src/xorl/ops/group_gemm/utils/kernel.py b/src/xorl/ops/group_gemm/utils/kernel.py deleted file mode 100644 index c7b02b48..00000000 --- a/src/xorl/ops/group_gemm/utils/kernel.py +++ /dev/null @@ -1,11 +0,0 @@ -import triton - - -def innermost_fn(fn: triton.KernelInterface): - while hasattr(fn, "fn"): - fn = fn.fn - return fn - - -def qualified_name(fn: triton.KernelInterface) -> str: - return innermost_fn(fn).__qualname__ diff --git a/src/xorl/ops/group_gemm/utils/path.py b/src/xorl/ops/group_gemm/utils/path.py deleted file mode 100644 index 16dde89d..00000000 --- a/src/xorl/ops/group_gemm/utils/path.py +++ /dev/null @@ -1,31 +0,0 @@ -import inspect -import os - -import triton -from packaging import version - -from .device import get_device_key -from .kernel import qualified_name - - -def _get_relative_dir_of_triton_kernel(kernel) -> str: - path = os.path.relpath(inspect.getfile(kernel), get_bpex_root()) - return path - - -def get_bpex_root() -> str: - path = os.path.abspath(os.path.join(os.path.dirname(__file__), "..")) - return path - - -def get_config_path_prefix_for(kernel) -> str: - v = version.parse(triton.__version__) - return ( - f"{get_bpex_root()}/config/{v.major}.{v.minor}/{get_device_key()}/" - f"{_get_relative_dir_of_triton_kernel(kernel)}/{qualified_name(kernel)}" - ) - - -def get_config_dedicated_file_for(kernel, algo_key) -> str: - # The only reason the extension is used is to avoid JSON lint.. - return f"{get_config_path_prefix_for(kernel)}/{algo_key}.bpex" diff --git a/src/xorl/ops/group_gemm/utils/pretuned.py b/src/xorl/ops/group_gemm/utils/pretuned.py deleted file mode 100644 index d74d47c8..00000000 --- a/src/xorl/ops/group_gemm/utils/pretuned.py +++ /dev/null @@ -1,102 +0,0 @@ -import triton - -from ....utils import logging -from ....utils.device import get_device_name -from . import envvars -from .config import load_all_configs_for -from .kernel import innermost_fn, qualified_name - - -logger = logging.get_logger(__name__) - -CATCH_ALL_ALGO_KEY = "__CATCH_ALL__" - - -def algo_key_scaled(names, scales, rest_key=None): - def key_maker(**kwargs): - lower_names = [name.lower() for name in names] - temp = [] - for i, name in enumerate(lower_names): - t = name + str(kwargs[names[i]] // scales[i]) - if scales[i] != 1: - t += f"x{scales[i]}" - temp.append(t) - res = "_".join(temp) - - if rest_key is not None: - for k in rest_key: - res += f"_{kwargs[k]}" - return res - - return key_maker - - -class Pretuned(triton.KernelInterface): - def __init__(self, fn, algo_key_maker, configs): - self.fn = fn # In case the outer decorator cares. - self.kernel_name = qualified_name(fn) - self.algo_key_maker = algo_key_maker - self.configs = configs - - assert CATCH_ALL_ALGO_KEY in self.configs - - def run(self, *args, **kwargs): - algo_key = self.algo_key_maker(**kwargs) - if algo_key not in self.configs: - if not envvars.is_untuned_warning_suppressed(): - logger.debug( - f"Untuned case (using algo-key [{algo_key}]) is seen when invoking " - f"kernel [{qualified_name(self)}], performance may suffer." - ) - extra_kwargs = self.configs[CATCH_ALL_ALGO_KEY] - else: - extra_kwargs = self.configs[algo_key] - return self.fn.run(*args, **kwargs, **extra_kwargs) - - -# TODO: Support using `triton.autotune` as an fallback. -def pretuned(*, algo_key=None, fallback=None): - """Decorator to annotate a Triton kernel as pre-tuned. Hyperparameters are loaded from `PRETUNED` - in the same folder as the kernel being defined. - - By default we look up pre-tuned hyperparameters via `kernel_name, device_name`. However, users - are allowed to provide `algo_key` option by providing a lambda that converts arguments passed - to kernel to a string that's used as a third level key in looking up pre-tuned hyperparameters. - - Note that ONLY named arguments (but not positional arguments) are passed to `algo_key` callback. - """ - - if algo_key is None: - - def catch_all(**kwargs): - return CATCH_ALL_ALGO_KEY - - algo_key = catch_all - - def decorator(fn: triton.KernelInterface): - nonlocal algo_key - nonlocal fallback - - name = qualified_name(fn) - configs = load_all_configs_for(innermost_fn(fn)) - - if CATCH_ALL_ALGO_KEY not in configs: - # We'd like to find a fallback hyperparameter for each `device`. This is not the same one - # as `fallback` provided to `pretuned`. The latter is used when we're running on an untuned - # device, while the former is just a catch-all for a specific device. - if not envvars.is_untuned_warning_suppressed(): - logger.debug( - f"No pre-tuned hyperparameter for kernel [{name}], using fallback config, " - "performance may suffer. You may have triton version or device name mismatch. " - f"You have triton=={triton.__version__} and device name [{get_device_name()}]", - ) - configs.update({CATCH_ALL_ALGO_KEY: fallback}) - - assert configs[CATCH_ALL_ALGO_KEY] is not None, "No usable fallback hyperparameter for kernel {name}" - return Pretuned( - fn, - algo_key_maker=algo_key, - configs=configs, - ) - - return decorator diff --git a/src/xorl/ops/loss/utils.py b/src/xorl/ops/loss/utils.py deleted file mode 100644 index 23212fa4..00000000 --- a/src/xorl/ops/loss/utils.py +++ /dev/null @@ -1,71 +0,0 @@ -""" -Utility functions for loss computation. - -This module contains shared helper functions used by various loss functions. -""" - -import torch -import torch.nn.functional as F - - -def _compute_eager_jsd( - student_hidden_states: torch.Tensor, - student_weight: torch.Tensor, - teacher_hidden_states: torch.Tensor, - teacher_weight: torch.Tensor, - labels: torch.Tensor, - beta: float = 0.5, - temperature: float = 1.0, - ignore_index: int = -100, -) -> torch.Tensor: - """ - Compute Jensen-Shannon Divergence loss between student and teacher distributions. - - JSD_beta(P, Q) = beta * KL(P || M) + (1 - beta) * KL(Q || M) - where M = beta * P + (1 - beta) * Q - - Args: - student_hidden_states: Flattened student hidden states, shape (batch * seq_len, hidden_dim) - student_weight: Student LM head weight, shape (vocab_size, hidden_dim) - teacher_hidden_states: Flattened teacher hidden states, shape (batch * seq_len, hidden_dim) - teacher_weight: Teacher LM head weight, shape (vocab_size, hidden_dim) - labels: Flattened target labels, shape (batch * seq_len,) - beta: Balance parameter. 0.0 for forward KL, 1.0 for reverse KL, 0.5 for symmetric JSD - temperature: Temperature for softmax (default: 1.0) - ignore_index: Index to ignore in loss computation (default: -100) - - Returns: - Scalar JSD loss value - """ - # Compute logits - student_logits = (student_hidden_states @ student_weight.t()).float() / temperature - teacher_logits = (teacher_hidden_states @ teacher_weight.t()).float() / temperature - - # Compute log probabilities - student_log_probs = F.log_softmax(student_logits, dim=-1) - teacher_log_probs = F.log_softmax(teacher_logits, dim=-1) - - # Compute probabilities for mixture - student_probs = student_log_probs.exp() - teacher_probs = teacher_log_probs.exp() - - # Mixture distribution: M = beta * P_student + (1 - beta) * P_teacher - mixture_probs = beta * student_probs + (1 - beta) * teacher_probs - mixture_log_probs = mixture_probs.log() - - # KL(student || mixture) = sum(student_probs * (student_log_probs - mixture_log_probs)) - kl_student_mixture = (student_probs * (student_log_probs - mixture_log_probs)).sum(dim=-1) - - # KL(teacher || mixture) = sum(teacher_probs * (teacher_log_probs - mixture_log_probs)) - kl_teacher_mixture = (teacher_probs * (teacher_log_probs - mixture_log_probs)).sum(dim=-1) - - # JSD = beta * KL(student || M) + (1 - beta) * KL(teacher || M) - jsd = beta * kl_student_mixture + (1 - beta) * kl_teacher_mixture - - # Mask ignored tokens - valid_mask = labels != ignore_index - jsd = jsd.masked_fill(~valid_mask, 0.0) - - # Return mean loss over valid tokens - n_valid = valid_mask.sum().clamp(min=1) - return jsd.sum() / n_valid diff --git a/src/xorl/server/weight_sync/README.md b/src/xorl/server/weight_sync/README.md index 03d5b50f..effd8d9a 100644 --- a/src/xorl/server/weight_sync/README.md +++ b/src/xorl/server/weight_sync/README.md @@ -134,7 +134,6 @@ weight_sync/ └── backends/ ├── base.py # WeightTransportBackend ABC + TransportConfig dataclass ├── nccl_broadcast.py # NCCLBroadcastBackend (default) - ├── nccl_simple.py # Simplified NCCL transfer helper ├── p2p.py # Mooncake RDMA P2P backend ├── sparse_delta.py # Experimental packed sparse-delta backend └── __init__.py # create_backend() factory @@ -302,10 +301,9 @@ P2P tuning options: - FP8 P2P sync requires an explicit sync quantization config, for example via `POST /api/v1/set_sync_quantization` or a per-call `quantization` field: `{"quant_method":"fp8","fmt":"e4m3","weight_block_size":[128,128]}`. - Client wrappers may expose this as `XORL_WEIGHT_SYNC_QUANTIZATION` or - `XORL_SYNC_QUANTIZATION`. A launch-only SGLang `--quantization fp8` flag is - not enough unless endpoint auto-detection is confirmed to populate the sync - request's `quantization` field. + A launch-only SGLang `--quantization fp8` flag is not enough unless endpoint + auto-detection is confirmed to populate the sync request's `quantization` + field. - With P2P and explicit FP8 sync quantization, the handler quantizes supported projection weights on the trainer side, transfers FP8 weights plus `weight_scale_inv` tensors, and skips receiver post-processing by default @@ -353,10 +351,6 @@ P2P tuning options: serialized prepare behavior. - `XORL_P2P_PREPARE_TIMEOUT_S`: per-endpoint prepare HTTP timeout. Default: 120 seconds. -- `XORL_SERIAL_INFERENCE_ENDPOINT_SYNC=1`: fallback/debug guard for - multi-endpoint P2P. It sends each receiver endpoint through its own serialized - sync group, avoiding cross-endpoint Mooncake session reuse at the cost of - giving up normal endpoint fanout parallelism. - `XORL_P2P_SCATTER_COPY_MODE`: controls how rank 0 builds per-sender tensor map payloads for direct-EP scatter. Default `none` reuses read-only locator lists/dicts while constructing scatter payloads. Set `list` to shallow-copy diff --git a/src/xorl/server/weight_sync/backends/nccl_simple.py b/src/xorl/server/weight_sync/backends/nccl_simple.py deleted file mode 100644 index 35da4850..00000000 --- a/src/xorl/server/weight_sync/backends/nccl_simple.py +++ /dev/null @@ -1,123 +0,0 @@ -"""Two-phase NCCL weight sync backend. - -Why this exists ---------------- -``nccl_broadcast`` interleaves two NCCL communicators per module: - - unshard (FSDP all-gather, intra-node) -> extract -> reshard - -> dist.broadcast (weight-sync group, inter-node) -> next module - -With FSDP ranks 1..N racing ahead of rank 0 (they have no broadcast work), -the two communicators enqueue kernels in different orders across ranks, -which deadlocks NCCL after a few modules (observed consistently at the -7th bucket on 14B / FSDP=4 / 2-node). - -This backend removes the interleaving entirely: - -* **Phase A** (during the handler's module loop): ``transfer_bucket`` only - stages tensors to CPU. No NCCL, no HTTP. The FSDP loop runs to completion - using only FSDP collectives. -* **Phase B** (``flush_pending_transfers``, called by the handler after the - module loop): re-chunk staged params and send each chunk through the - proven ``_transfer_single_bucket`` path (HTTP + dist.broadcast). Only the - weight-sync communicator is active. - -The two communicators never run concurrently, so the kernel-ordering -deadlock cannot occur. -""" - -import logging -from typing import List, Optional, Tuple - -import torch - -from .nccl_broadcast import NCCLBroadcastBackend - - -logger = logging.getLogger(__name__) - -# Re-chunk size for phase B. Bounds sglang-side temp memory -# (torch.empty per param before load_weights) and trainer-side H2D staging. -_CHUNK_BYTES = 1024 * 1024 * 1024 # 1 GiB - - -class NCCLSimpleBackend(NCCLBroadcastBackend): - """Two-phase (stage-then-broadcast) NCCL transport.""" - - def __init__(self, config, **kwargs) -> None: - super().__init__(config, **kwargs) - self._pending: List[Tuple[str, torch.Tensor]] = [] - self._pending_bytes = 0 - self._final_flush_cache = False - self._final_weight_version: Optional[str] = None - - # ------------------------------------------------------------------ - # Phase A: stage to CPU (no NCCL, no HTTP) - # ------------------------------------------------------------------ - def transfer_bucket( - self, - bucket: List[Tuple[str, torch.Tensor]], - *, - src_rank: int = 0, - flush_cache: bool = False, - weight_version: Optional[str] = None, - ) -> None: - if src_rank != 0: - raise ValueError(f"NCCLSimpleBackend only supports src_rank=0, got {src_rank}") - for name, t in bucket: - cpu_t = t.detach().to("cpu").contiguous() - self._pending.append((name, cpu_t)) - self._pending_bytes += cpu_t.numel() * cpu_t.element_size() - if flush_cache: - self._final_flush_cache = True - if weight_version is not None: - self._final_weight_version = weight_version - - # ------------------------------------------------------------------ - # Phase B: chunked HTTP + broadcast (no FSDP collectives anywhere) - # ------------------------------------------------------------------ - def flush_pending_transfers(self) -> None: - if not self._pending: - return - if self._synchronizer is None: - raise RuntimeError("Backend not initialized — call initialize() first") - - # Build chunks bounded by _CHUNK_BYTES (a single oversized param - # becomes its own chunk). - chunks: List[List[Tuple[str, torch.Tensor]]] = [] - cur: List[Tuple[str, torch.Tensor]] = [] - cur_bytes = 0 - for name, t in self._pending: - nbytes = t.numel() * t.element_size() - if cur and cur_bytes + nbytes > _CHUNK_BYTES: - chunks.append(cur) - cur, cur_bytes = [], 0 - cur.append((name, t)) - cur_bytes += nbytes - if cur: - chunks.append(cur) - - total_gb = self._pending_bytes / 1e9 - logger.info( - f"[NCCLSimple] Phase B: broadcasting {len(self._pending)} params " - f"({total_gb:.2f} GB) in {len(chunks)} chunks" - ) - - device = self.config.device - try: - for i, chunk in enumerate(chunks): - last = i == len(chunks) - 1 - gpu_chunk = [(n, t.to(device, non_blocking=True)) for n, t in chunk] - torch.cuda.synchronize(device) - self._synchronizer._transfer_single_bucket( - gpu_chunk, - flush_cache=self._final_flush_cache and last, - weight_version=self._final_weight_version if last else None, - ) - del gpu_chunk - logger.info(f"[NCCLSimple] Phase B complete: {len(chunks)} chunks sent") - finally: - self._pending = [] - self._pending_bytes = 0 - self._final_flush_cache = False - self._final_weight_version = None diff --git a/src/xorl/server/weight_sync/handler.py b/src/xorl/server/weight_sync/handler.py index e2733cdb..a2259535 100644 --- a/src/xorl/server/weight_sync/handler.py +++ b/src/xorl/server/weight_sync/handler.py @@ -718,7 +718,7 @@ async def handle_sync_inference_weights(self, command_dict: Dict[str, Any]) -> D Handle sync inference weights request (all ranks participate). The ``sync_method`` field selects the transport backend. Currently - supported: ``"nccl_broadcast"``, ``"nccl_simple"``, ``"p2p"``, and + supported: ``"nccl_broadcast"``, ``"p2p"``, and experimental ``"sparse_delta"``. New backends can be added by implementing :class:`WeightTransportBackend` and registering in :func:`backends.create_backend`. diff --git a/src/xorl/utils/distillation_utils.py b/src/xorl/utils/distillation_utils.py deleted file mode 100644 index 9a701c11..00000000 --- a/src/xorl/utils/distillation_utils.py +++ /dev/null @@ -1,47 +0,0 @@ -"""Utility functions for distillation training.""" - -import json -import logging - -from huggingface_hub import hf_hub_download -from safetensors import safe_open - - -logger = logging.getLogger(__name__) - - -def load_lm_head_from_hf_model(model_id: str, token: str = None): - """ - Load only the lm_head weights from a HuggingFace model. - - Args: - model_id: HuggingFace model ID (e.g., "Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8") - token: HuggingFace token if model is private - - Returns: - lm_head weight tensor - """ - # 1. Download the index file - index_path = hf_hub_download(repo_id=model_id, filename="model.safetensors.index.json", token=token) - - # 2. Parse the index to find which shard contains lm_head - with open(index_path, "r") as f: - index = json.load(f) - - # 3. Find the shard file for lm_head.weight - weight_map = index.get("weight_map", {}) - lm_head_file = weight_map.get("lm_head.weight") - - if not lm_head_file: - raise ValueError(f"lm_head.weight not found in {model_id}") - - logger.info(f"lm_head.weight is in: {lm_head_file}") - - # 4. Download only that specific shard - shard_path = hf_hub_download(repo_id=model_id, filename=lm_head_file, token=token) - - # 5. Load only the lm_head.weight from the shard - with safe_open(shard_path, framework="pt", device="cpu") as f: - lm_head_weight = f.get_tensor("lm_head.weight") - - return lm_head_weight diff --git a/src/xorl/utils/manual_cuda_timing.py b/src/xorl/utils/manual_cuda_timing.py deleted file mode 100644 index c9594d1f..00000000 --- a/src/xorl/utils/manual_cuda_timing.py +++ /dev/null @@ -1,98 +0,0 @@ -"""Lightweight CUDA-event scopes for model-specific phase timing. - -This module complements trainer-level phase timing and hook-based component -timing. It lets model code add fine-grained CUDA-event scopes without depending -on the Trainer class. Timings are accumulated per process and drained by the -trainer at the end of a step. -""" - -from __future__ import annotations - -from collections import defaultdict -from contextlib import contextmanager -from typing import Dict, Iterator, List, Tuple - -import torch - -from xorl.utils.device import get_device_type - - -_enabled = False -_mode = "idle" -_pairs: Dict[str, List[Tuple[torch.cuda.Event, torch.cuda.Event]]] = defaultdict(list) - - -def set_manual_cuda_timing_enabled(enabled: bool) -> None: - """Enable or disable manual CUDA-event timing for the current process.""" - global _enabled - _enabled = bool(enabled) and get_device_type() == "cuda" - if not _enabled: - reset_manual_cuda_timing() - - -def set_manual_cuda_timing_mode(mode: str) -> None: - """Set the current timing mode: ``fwd``, ``bwd``/recompute, or ``idle``.""" - global _mode - if mode not in ("fwd", "bwd", "idle"): - raise ValueError(f"invalid manual CUDA timing mode: {mode}") - _mode = mode - - -def reset_manual_cuda_timing() -> None: - """Clear all accumulated events.""" - _pairs.clear() - - -def _phase_name(name: str) -> str | None: - if not _enabled or _mode == "idle" or get_device_type() != "cuda": - return None - if _mode == "fwd": - return f"fwd_{name}" - if _mode == "bwd": - # During activation checkpointing, model forward code runs under - # backward. These scopes therefore describe recompute work. - return f"recompute_{name}" - return None - - -@contextmanager -def manual_cuda_timing_scope(name: str) -> Iterator[None]: - """Record a CUDA-event scope under the current manual timing mode.""" - phase = _phase_name(name) - if phase is None: - yield - return - - start = torch.cuda.Event(enable_timing=True) - end = torch.cuda.Event(enable_timing=True) - start.record() - try: - yield - finally: - end.record() - _pairs[phase].append((start, end)) - - -def drain_manual_cuda_timing(*, synchronize: bool = True) -> Dict[str, float]: - """Return accumulated timings in seconds and clear the accumulator.""" - if not _pairs: - return {} - if synchronize and get_device_type() == "cuda": - torch.cuda.synchronize() - - result: Dict[str, float] = {} - for phase, events in _pairs.items(): - total_ms = 0.0 - valid_events = 0 - for start, end in events: - try: - total_ms += start.elapsed_time(end) - except ValueError as exc: - if "Both events must be recorded" not in str(exc): - raise - continue - valid_events += 1 - if valid_events: - result[phase] = total_ms / 1000.0 - reset_manual_cuda_timing() - return result diff --git a/src/xorl/utils/model_utils.py b/src/xorl/utils/model_utils.py deleted file mode 100644 index b4784a2d..00000000 --- a/src/xorl/utils/model_utils.py +++ /dev/null @@ -1,51 +0,0 @@ -import numpy as np -import torch.nn as nn - -from . import logging - - -logger = logging.get_logger(__name__) - - -def pretty_print_trainable_parameters(model: nn.Module): - trainable_parameters = [] - for n, p in model.named_parameters(): - if p.requires_grad: - trainable_parameters.append(n) - - printable_results = {} - for p in trainable_parameters: - param_split = p.split(".") - param_name = "" - digit_index = 0 - layer_index_list = [] - for split_item in param_split: - if split_item.isdigit(): - param_name += f"<{digit_index}>." - layer_index_list.append(int(split_item)) - digit_index += 1 - else: - param_name += f"{split_item}." - param_name = param_name[:-1] - - if param_name not in printable_results: - printable_results[param_name] = [] - printable_results[param_name].append(layer_index_list) - - train_param_info = "\n**** trainable parameters ****" - for param_key in printable_results.keys(): - layer_idxs = np.array(printable_results[param_key]) - if layer_idxs.shape[-1] == 0: - train_param_info += "\n" + param_key - continue - layer_min = layer_idxs.min(axis=0) - layer_max = layer_idxs.max(axis=0) - print_pattern = param_key - for index in range(len(layer_min)): - if layer_min[index] == layer_max[index]: - print_pattern = print_pattern.replace(f"<{index}>", f"[{layer_min[index]}]") - else: - print_pattern = print_pattern.replace(f"<{index}>", f"[{layer_min[index]}-{layer_max[index]}]") - train_param_info += "\n" + print_pattern - train_param_info += "\n**** trainable parameters ****" - logger.info_rank0(train_param_info) diff --git a/src/xorl/utils/recompute_utils.py b/src/xorl/utils/recompute_utils.py deleted file mode 100644 index 5fdee02b..00000000 --- a/src/xorl/utils/recompute_utils.py +++ /dev/null @@ -1,103 +0,0 @@ -from typing import Any, List - -import torch - -from xorl.utils import helper - - -logger = helper.create_logger(__name__) - - -def string_to_op(op_string: str) -> Any: - """ - Convert a single operation string to PyTorch operation object - - Args: - op_string: e.g. "aten.addmm.default" or "torch.ops.flash_attn._flash_attn_forward.default" - - Returns: - PyTorch operation object - """ - global torch - # Clean the string - clean_string = op_string.strip() - - # Remove torch.ops. prefix (if exists) - if clean_string.startswith("torch.ops."): - clean_string = clean_string[len("torch.ops.") :] - - # Split path and access level by level - parts = clean_string.split(".") - - # Check if torch.ops is available - if not hasattr(torch, "ops"): - raise AttributeError("torch.ops not available in this PyTorch version") - - current = torch.ops - - for i, part in enumerate(parts): - if hasattr(current, part): - current = getattr(current, part) - else: - # More detailed error information, including current path - current_path = ".".join(parts[:i]) - available_attrs = dir(current) if hasattr(current, "__dict__") else [] - raise AttributeError( - f"Operation '{op_string}' not found. " - f"Missing attribute: '{part}' at path 'torch.ops.{current_path}'. " - f"Available attributes: {available_attrs[:10]}{'...' if len(available_attrs) > 10 else ''}" - ) - - return current - - -def convert_ops_to_objects(ops_strings: List[str]) -> List[Any]: - """ - Convert operation string list to operation object list - Args: - ops_strings: String list - - Returns: - PyTorch operation object list - """ - ops_objects = [] - failed_ops = [] - - # First perform environment check - _check_torch_ops_availability() - - for op_str in ops_strings: - try: - op_obj = string_to_op(op_str) - ops_objects.append(op_obj) - logger.info_rank0(f"✓ Conversion successful: {op_str}") - assert isinstance(op_obj, torch._ops.OpOverload), "Please check if the ops is end with .default" - except (AttributeError, TypeError) as e: - logger.info_rank0(f"✗ Conversion failed: {op_str} - {e}") - failed_ops.append(op_str) - except Exception as e: - logger.info_rank0(f"✗ Conversion failed: {op_str} - {e}") - raise e - - if failed_ops: - logger.info_rank0(f"\nWarning: {len(failed_ops)} operations failed to convert") - logger.info_rank0("Possible reasons:") - logger.info_rank0("1. PyTorch version does not support certain operations") - logger.info_rank0("2. Missing related extension modules (e.g. flash_attn)") - logger.info_rank0("3. Operation name spelling error") - - return ops_objects - - -def _check_torch_ops_availability(): - global torch - # Check if torch.ops is available - if not hasattr(torch, "ops"): - raise RuntimeError("torch.ops is not available in current PyTorch version") - - # Check basic aten operations - try: - _ = torch.ops.aten.add - logger.info_rank0("✓ torch.ops.aten available") - except AttributeError as e: - logger.info_rank0(f"✗ torch.ops.aten not available: {e}") diff --git a/test_audit_decisions.json b/test_audit_decisions.json deleted file mode 100644 index 96d9daba..00000000 --- a/test_audit_decisions.json +++ /dev/null @@ -1,16597 +0,0 @@ -{ - "schema_version": 1, - "rebase_policy": "The 2026-08-19 current-main PR preserves src/ byte-for-byte against current main. PR 67 removed the unreachable xorl.rl package upstream; TA-887 records that inherited decision. Other applied production-surface decisions describe the original audit and are superseded by TA-1655.", - "items": [ - { - "id": "TA-001", - "scope": "tests/models/test_qwen3_moe_fused_lora.py", - "decision": "remove", - "status": "applied", - "evidence": [ - "All seven contracts are duplicated or subsumed by tests/models/test_moe_experts_lora.py; two test bodies are exact AST duplicates." - ] - }, - { - "id": "TA-002", - "scope": "30 config snapshot tests in tests/server/test_server_arguments.py", - "decision": "remove", - "status": "applied", - "evidence": [ - "The referenced experiment and example YAML files are not tracked, and a local wrapper converted every missing fixture into a skip." - ] - }, - { - "id": "TA-003", - "scope": "six print-only or soft-threshold measurement tests under tests/ops", - "decision": "remove", - "status": "applied", - "evidence": [ - "The routines only print measurements or turn a missed performance target into pytest.skip, so they cannot report a regression." - ] - }, - { - "id": "TA-004", - "scope": "tests/models/test_qwen3_5_apply_rotary.py::test_interleaved_pairwise_rotation_d8", - "decision": "remove", - "status": "applied", - "evidence": [ - "The only token is at position zero, making the rotation the identity under competing rotation conventions." - ] - }, - { - "id": "TA-005", - "scope": "tests/models/test_moe_sglang_fused_experts.py::test_sglang_runtime_api_does_not_regress_to_legacy_globals", - "decision": "remove", - "status": "applied", - "evidence": [ - "Adjacent fake-runtime behavioral tests already fail if the implementation returns to the unavailable legacy global API." - ] - }, - { - "id": "TA-006", - "scope": "tests/models/test_glm52_official_fp8_inventory.py", - "decision": "relocate", - "status": "applied", - "evidence": [ - "Both checks require an external checkpoint selected by XORL_GLM52_OFFICIAL_MODEL_PATH and are official-model certification rather than clean-checkout tests." - ] - }, - { - "id": "TA-007", - "scope": "tests/distributed/test_deepep_async_combine_guard.py", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The synchronous-default contract is valuable, but the tests patch a removed module global instead of the current environment-controlled function." - ] - }, - { - "id": "TA-008", - "scope": "tests/models/test_qwen3_5_apply_rotary.py::test_qwen35_modeling_does_not_pass_interleaved_to_rotary", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The source parser was replaced by dense and MoE attention projections at nonzero positions that distinguish half-rotate from pairwise rotation while mrope_interleaved is enabled." - ] - }, - { - "id": "TA-009", - "scope": "shared identity, temperature, and KL-tail contracts in the importance-sampling and policy loss tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Three pairs had exact duplicate bodies; one parameterized contract now runs the same implementation-specific cases and assertions for both production losses." - ] - }, - { - "id": "TA-010", - "scope": "distributed launcher wrappers in test_olmo2_tp_e2e.py and test_vocab_parallel_ce.py", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The wrapper bodies match, but each module supplies a different SCRIPT_PATH and validates a distinct distributed production path." - ] - }, - { - "id": "TA-011", - "scope": "tests/server/test_server_arguments.py module stubbing", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "patch.dict(sys.modules, ...) rolled back native Triton and CUTLASS imports loaded inside the context, making later imports fail despite an exactly synced repository venv; targeted restoration of only the stubbed module keys makes all 48 tests pass." - ] - }, - { - "id": "TA-012", - "scope": "three non-empty _prod unit tests in tests/server/weight_sync/test_pp_nccl_transfer.py", - "decision": "remove", - "status": "applied", - "evidence": [ - "The sender metadata and receiver reconstruction tests already exercise one- and two-dimensional products through the production protocol; only the otherwise uncovered empty-shape identity test remains." - ] - }, - { - "id": "TA-013", - "scope": "tests/trainers/test_step_phase_timing.py::test_order_step_phases_covers_every_canonical_phase", - "decision": "remove", - "status": "applied", - "evidence": [ - "The test used _STEP_PHASE_TIMING_ORDER as both its input and expected result; adjacent tests already cover canonical ordering, unknown-key ordering, and empty input." - ] - }, - { - "id": "TA-014", - "scope": "FA3 and external-FLA test module imports", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The locked repository environment provides FA4 but not the optional FA3 top-level interface or external FLA package; three backend-specific modules now skip at collection when those comparison backends are absent." - ] - }, - { - "id": "TA-015", - "scope": "exact Qwen3.5 and GLM numerical-program admission tests in tests/trainers/test_rope_class_b_config.py", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The full resolved-program tests already assert the Class-B and RMSNorm defaults for dense and MoE architectures; separate assertions of those same defaults were redundant.", - "Invalid override values and unsupported topology values now remain as tables inside one admission contract instead of producing a separate collected test for every literal." - ] - }, - { - "id": "TA-016", - "scope": "synthetic and subsumed cases in tests/server/weight_sync/test_p2p_backend_protocol.py", - "decision": "remove", - "status": "applied", - "evidence": [ - "The opt-in fused QKV and convolution source layouts have no matching receiver locator in the production protocol and are explicitly bypassed by default.", - "The 40-layer by 256-expert sweep repeats layer-independent transfer logic already covered by a real-size local shard and every global expert index.", - "The single block-128 FP8 receiver case is a strict subset of the retained multi-receiver layout contract." - ] - }, - { - "id": "TA-017", - "scope": "checkpoint expert-key classification cases in tests/models/test_module_utils_broadcast.py", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sixteen key spellings exercise one binary classifier and now form one table-driven contract with contextual failure messages instead of fourteen independently collected tests." - ] - }, - { - "id": "TA-018", - "scope": "weight-sync configuration precedence, endpoint normalization, and direct-EP selection tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Closely related environment-precedence branches, metadata aliases, and direct-EP sender choices now live in one contract per production decision rather than field-by-field helper fragments." - ] - }, - { - "id": "TA-019", - "scope": "exact GLM fixture in tests/trainers/test_rope_class_b_config.py", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Production validates official geometry and then installs _glm52_exact_contract before resolving numerical defaults; the direct config fixture omitted that admission step and therefore exercised the ordinary-model branch." - ] - }, - { - "id": "TA-020", - "scope": "TileLang V4 indexer shape matrices in tests/ops/dsv4/test_v4_tilelang_indexer.py", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Four retained geometries cover the basic, batched, production-head/top-k, and C128 kernel paths; the removed sequence, head, and batch literals do not select additional code.", - "The removed V4 real-config sweep repeated the production geometry already covered by both the score-reference and top-k contracts." - ] - }, - { - "id": "TA-021", - "scope": "families-v2 RMSNorm realization and dispatch matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forced fused/split parity now covers tail, aligned, and deep-tile hidden shapes at row-count extremes; intermediate row counts do not select different code.", - "For shipped hidden sizes the tile count is below V2_NORM_SPLIT_MIN_TILES, making the dispatch result independent of every parametrized row literal." - ] - }, - { - "id": "TA-022", - "scope": "decode GDN triangular-solve shape and launch-configuration sweeps", - "decision": "remove", - "status": "applied", - "evidence": [ - "Production fixes diagonal group size and both warp counts internally; the 12-case direct-kernel sweep exercised configurations no supported caller can select, while the retained wrapper parity test covers the deployed launch.", - "Scaling K before immediately normalizing it is not a distinct production input regime; retained cases cover batch/grid boundaries and structurally padded input." - ] - }, - { - "id": "TA-023", - "scope": "pipeline stage-to-rank mapping matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One single-stage loop case and one multi-stage case for each loop and V formula cover every implementation branch; additional PP sizes only repeat the same index arithmetic against PyTorch's reference helper." - ] - }, - { - "id": "TA-024", - "scope": "optional-boolean coercion and sequence-parallel FSDP truth tables", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Each table describes one parser or one policy predicate and now reports as one semantic contract while retaining every accepted literal and mode combination." - ] - }, - { - "id": "TA-025", - "scope": "fused selected-logprob dtype, bias, temperature, shape, and vocabulary matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dtype, optional bias, and scalar temperature are independent implementation branches; two pairwise cases cover both values instead of an eight-case Cartesian product in both forward and backward.", - "Irregular shapes and production vocabularies remain fully exercised inside their respective numerical contracts." - ] - }, - { - "id": "TA-026", - "scope": "batch-invariant GEMM table bit-neutrality matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The dtype-by-shape matrix enforces one doctrine-level invariant: every generated launch config preserves the dtype-pinned reduction tree; all twelve combinations remain executed within one contract." - ] - }, - { - "id": "TA-027", - "scope": "TileLang sparse-MLA geometry and attention-sink matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Four forward geometries retain every compiled top-k specialization while covering batch and the production 64-head path; repeated sequence/head literals did not select new code.", - "Zero and mixed-sign sinks cover the arithmetic boundary and both signs, and the retained large-sink effect test independently proves the sink is consumed." - ] - }, - { - "id": "TA-028", - "scope": "LoRA-head gradient-ownership backend and replica matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pairwise quant-format/backend coverage retains every format, backend label, and eager-versus-fused producer branch without the 3 by 4 Cartesian product.", - "Table rows use explicit monkeypatch contexts so grouping them preserves the isolation previously supplied by pytest parametrization." - ] - }, - { - "id": "TA-029", - "scope": "MoE torch.compile probes and benchmarks in tests/ops/test_moe_torch_compile.py", - "decision": "remove", - "status": "applied", - "evidence": [ - "The fullgraph probe swallowed every exception and therefore could not report a graph-break regression.", - "Two bench-prefixed routines were not collected by pytest, caught compiler failures, and only printed measurements; the retained three contracts enforce block, decoder-layer, and full-model compilation." - ] - }, - { - "id": "TA-030", - "scope": "inference-endpoint quantization normalization helper fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The HF-config detection contract already covers language-model prefix removal, vision filtering, and weight-suffix normalization together; two direct private-helper tests were strict subsets.", - "Static and null activation schemes reach the same unsupported receiver boundary and now share one contract." - ] - }, - { - "id": "TA-031", - "scope": "packing-strategy invariant and datum-order tests", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Best-fit utilization was algebraically equivalent to the retained row-count property and used fewer seeds.", - "The side-array test only verified Python list indexing; datum-order coverage now compares reported order with document lengths recovered from actual packed position boundaries." - ] - }, - { - "id": "TA-032", - "scope": "weight-sync quantization-config aliases and rejection tables", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "No-op aliases, unsupported methods, FP8 formats, and activation schemes each define one normalization or rejection boundary and retain every literal inside a single contract." - ] - }, - { - "id": "TA-033", - "scope": "FP8 default projection selection and exclusion tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One buffer-level contract now covers fused MLA, four packed linear-attention projections, both shared-expert spellings, and negative embedding/gate cases.", - "The removed packed-projection batch test exactly duplicated the four independently parametrized projection cases; module exclusions now cover suffixed and unsuffixed names in one contract." - ] - }, - { - "id": "TA-034", - "scope": "distributed expert-adapter autograd backend, topology, and quantization matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP2 retains real optimizer/autograd qualification for every unquantized backend; the four-rank test adds eFSDP topology with one shipped Quack representative instead of repeating backend math.", - "Quantized execution uses three pairwise backend/format cases rather than a nine-case Cartesian product while retaining every backend and format; DeepEP and projection-subset compositions remain separate contracts." - ] - }, - { - "id": "TA-035", - "scope": "expert QLoRA backend, model-family, and invalid-target contract tables", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Each backend capability, unsupported semantic, fail-closed model family, and invalid target list is one behavioral boundary; every original row remains executed inside its owning contract.", - "Grouping removes pytest item inflation without deleting model-family construction or rejection coverage." - ] - }, - { - "id": "TA-036", - "scope": "MoE-LoRA backend smoke tests and zero-initialized gradient comparisons", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The old cross-backend comparison left every LoRA-B factor at zero, making LoRA-A gradients trivially zero; the retained comparison initializes nonzero adapters and checks both fused backends against eager for output and every factor gradient.", - "That stronger MoEBlock contract subsumes separate Triton/native forward-backward smokes; one explicit zero-delta reference and one explicit nonzero-effect reference retain the shared semantic boundaries.", - "Backend-independent construction and injection retain a Quack representative, while backend registration and numerical execution remain covered separately." - ] - }, - { - "id": "TA-037", - "scope": "registry-wide RoPE fp32-table and native-lane matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every registered RoPE initializer still proves that model-wide bf16 casting resolves the exact fp32 CPU table; shared forward consumption uses default and YaRN representatives for unit and non-unit attention scaling.", - "Only default RoPE can select the native SGLang cache, so native-versus-stock comparisons for linear, dynamic, YaRN, LongRoPE, and Llama3 executed identical stock code and had no distinct failure mode." - ] - }, - { - "id": "TA-038", - "scope": "FP8 grouped-MoE geometry and end-to-end backend matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two same-NK cases retain the small and large N launch branches, aligned and padded K, empty experts, multi-block M, and output tails; further sizes selected no new kernel control flow.", - "Wgrad adds an all-empty case for the max_K zero early return and otherwise uses the same branch-complete small/large geometry split.", - "Both grouped backends retain optimizer-step qualification; the dense-plus-MoE integration uses the default Triton-grouped representative instead of repeating that established backend dimension." - ] - }, - { - "id": "TA-039", - "scope": "FP8 linear block-size and quantization-recipe matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Block sizes 64 and 128 create different padding and scale grids, and every existing recipe row remains executed.", - "Each block-scale layout and padded-matmul recipe is one numerical contract rather than a separate pytest item per table row." - ] - }, - { - "id": "TA-040", - "scope": "adapter-optimizer resume identity, topology-corruption, and staged-state matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every corruption remains because rank identity, layout fingerprint, parameter order, holes, overlaps, dtype, logical shape, optimizer step, group metadata, and state shape reach distinct validation boundaries.", - "Rows now execute under separate temporary checkpoint roots within three transactional contracts, preserving the filesystem isolation previously supplied by parametrization." - ] - }, - { - "id": "TA-041", - "scope": "gradient-checkpointing defaults, method propagation, and outer-gate truth table", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The method-propagation test repeated three strings through the same assignment branch; one nondefault value proves propagation while argument validation owns the supported-value table.", - "The outer-gate contract now includes the method predicate and proves that a nondefault selective method suppresses full-layer checkpointing, a behavior the old matrix omitted.", - "Base and MoE class defaults plus enable-time defaults remain covered together." - ] - }, - { - "id": "TA-042", - "scope": "MoE routing-weight-position resolver truth tables and explicit aliases", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All automatic-regime combinations and explicit boolean/string aliases remain executed; each resolver boundary now reports as one contract rather than one item per truth-table row." - ] - }, - { - "id": "TA-043", - "scope": "GLM-5.2 exact MoE construction dependency, EP, and rank-alpha rejection matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every exact-component dependency and both uninitialized/wrong-size EP states remain fail-closed before adapter mutation.", - "Separate rank-only and alpha-only invalid cases cover the combined rank=1 and alpha=1 predicate; the removed rank=16, alpha=16 row was their strict conjunction and selected no new validation branch." - ] - }, - { - "id": "TA-044", - "scope": "GLM-5.2 exact shared-expert constructor rejection matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Hidden size, intermediate size, TP width, rank, alpha, bias, and adaptive-noise rejections all remain executed inside one fail-closed constructor contract." - ] - }, - { - "id": "TA-045", - "scope": "NVFP4 and Block-FP8 QLoRA expert-loading fragments and embedded benchmark harness", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Each synthetic checkpoint load already materializes packed bytes, scales, and all three projections; two integrated format contracts now validate those outputs together across every retained geometry.", - "NVFP4 additionally retains exact amax, absorbed global-scale, and dequantized shape/dtype checks while reducing fourteen repeated loads to six total loads across both formats.", - "The private timing function and __main__ print harness were not pytest-collected, had no performance threshold, and were unused outside the file." - ] - }, - { - "id": "TA-046", - "scope": "FP8 external-config recipe and QARL nesting tables", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every Transformer-Engine-only recipe key and ModelOpt QARL nesting remains fail-closed; each receiver boundary reports as one contract." - ] - }, - { - "id": "TA-047", - "scope": "RMSNorm family funnel geometries and mode-by-batch-invariant matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both head-dimension and hidden-dimension kernel geometries remain bitwise-qualified inside one funnel contract.", - "The module parity test enables the trunk contract in every row, which already fixes family dispatch; the independent batch-invariant flag axis could not select different code and was removed while native, SGLang, and fused-SGLang modes remain." - ] - }, - { - "id": "TA-048", - "scope": "dense and MoE Qwen3.5 RMSNorm call-site truth tables", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every layer-zero/later-layer and native/SGLang/fused-SGLang row remains executed; each call-site predicate now reports as one contract." - ] - }, - { - "id": "TA-049", - "scope": "OPD chunk-count, KL-estimator alias, streaming-backend, and compiled smoke matrices", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Chunk coverage retains disabled, one-chunk, and more-chunks-than-valid-tokens boundaries; intermediate positive counts only changed a performance option and selected no new loss arithmetic.", - "All VERL estimator aliases and both streaming implementations remain checked inside their semantic contracts.", - "Three direct compiled tuple/shape smokes were strict subsets of retained end-to-end reverse/forward KL reference and diagnostic tests." - ] - }, - { - "id": "TA-050", - "scope": "NVFP4 fake-quant geometry, approximation, and layout contracts", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The implementation has no M- or legal-K-shape branch, so three arbitrary shapes crossed with two dtypes were reduced to one legal geometry executed in both supported dtypes.", - "Exact equality to the independent quantization reference already proves the fixed random tensor changes and meets the weaker relative-error threshold; the E2M1 grid invariant remains separate.", - "The retained K-within-row trap covers the same divisibility rejection as the removed generic 17-by-17 case and additionally proves a legal K succeeds." - ] - }, - { - "id": "TA-051", - "scope": "EP expert-compute registry signature matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every function in both live EP registries is still inspected for the shared explicit parameters and forward-compatible kwargs contract; backend labels no longer create separate test items." - ] - }, - { - "id": "TA-052", - "scope": "canonical MoE reference widths, transport resolution, and distributed contributor counts", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The reference keeps the two-contributor base tree and production sixteen-contributor depth; widths four and eight only repeated the same adjacent-pair loop.", - "The distributed transport keeps two and eight contributors plus the separate EP16 packed gate; four contributors selected no different collective, mapping, chunking, padding, or backward behavior.", - "The removed internal-resolution test repeated the admitted EP16 and fallback EP8 assertions already present in the adjacent auto-transport contract." - ] - }, - { - "id": "TA-053", - "scope": "GLM-5.2 QLoRA rank-alpha and unsupported-construction matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-only and alpha-only invalid inputs exercise both sides of the rank=1 and alpha=1 predicate; the removed rank=16, alpha=16 row was their strict conjunction.", - "Every unsupported construction mode still creates a fresh meta model and fails before adapterization inside one admission contract." - ] - }, - { - "id": "TA-054", - "scope": "GLM-5.2 exact-attention native-FP8 checkpoint pair transactions", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both arrival orders and both weight/scale members still execute with a fresh checkpoint handler for byte-exact completion, duplicate, missing, dtype, and shape boundaries; each transactional behavior now reports one test item." - ] - }, - { - "id": "TA-055", - "scope": "Quack-versus-Triton EP token-distribution and score-scaling Cartesian product", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Three pairwise cases retain balanced, empty-expert, extreme-skew, score-free, and score-scaled behavior; applying both score states to every distribution selected no additional wrapper or kernel branch.", - "Each retained case still compares output plus input, gate-up, and down gradients, while the independent half-concatenation reference remains." - ] - }, - { - "id": "TA-056", - "scope": "eager-versus-native MoE geometry sweep and large-scale duplicate", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Parity retains ordinary routing, top-k one and four, and the single-token boundary with forward and backward comparisons.", - "The removed larger ordinary geometries selected no repository control flow; the E64/H512/I1024 test repeated the same contract with substantially larger allocation and looser tolerances.", - "The embedded __main__ pytest launcher was uncollected and added no pass/fail contract." - ] - }, - { - "id": "TA-057", - "scope": "batch-invariant full-reduce mean shape-by-dtype matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both supported dtypes still cross the one-dimensional mean_dim path and multi-dimensional sum/divide path; a third tensor rank reached the same multi-dimensional branch." - ] - }, - { - "id": "TA-058", - "scope": "families-v2 RMSNorm split-versus-fused hidden-size and row-count matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shipped and deep hidden sizes remain at low and high row counts with residual, no-residual, and zero-centered modes; the intermediate row count selected neither a dispatch boundary nor distinct arithmetic." - ] - }, - { - "id": "TA-059", - "scope": "linear and MoE LoRA fp32 cast-once merge tests", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Every original dtype still executes zero and nonzero production merge methods for linear and MoE weights inside four behavioral contracts.", - "The removed precision test never called a production merge method; it only compared the local _naive_merge and _fp32_merge helper formulas on one random tensor." - ] - }, - { - "id": "TA-060", - "scope": "EP backend gradient-reduction domain table", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All five supported backend names still fail fast against EP_SUM locally and are rechecked inside the retained real two-rank autograd and synchronization worker; backend labels no longer create separate local test items." - ] - }, - { - "id": "TA-061", - "scope": "GLM-5.2 exact-attention construction rank-alpha and execution-mode matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-only and alpha-only failures cover both sides of the rank=1 and alpha=1 predicate; the removed rank=16, alpha=16 case was their strict conjunction.", - "Every incomplete dense-component, all-to-all, and sparse-MLA requirement still builds a fresh meta model and fails before adapter mutation." - ] - }, - { - "id": "TA-062", - "scope": "GLM-5.2 exact active-LoRA component conjunction", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Each of the five component flags is still independently cleared and checked against active-LoRA, exact-forward, exact-model, and BI-router admission; the conjunction reports one behavioral contract." - ] - }, - { - "id": "TA-063", - "scope": "Mamba2 chunk, packed-boundary, and upstream-divergence tests", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Exact and tail chunk sizes plus aligned and unaligned packed boundaries still execute with forward and gradient parity inside their respective contracts.", - "The removed divergence canary asserted that a known Transformers fallback bug remained present; an upstream fix would have failed XoRL despite improving behavior.", - "The retained independent sequential SSD recurrence remains the authoritative multi-chunk oracle." - ] - }, - { - "id": "TA-064", - "scope": "Qwen3-MoE layer input-norm family mode matrix", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The call-site family depends only on layer zero versus later layers; the capture module stored mode but never read it, so native, SGLang, and fused-SGLang labels could not select different code in this test.", - "Both no-residual and residual-tree layer boundaries remain executed." - ] - }, - { - "id": "TA-065", - "scope": "FlashQLA M-invariance and chunk-chaining certification rows", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both total-M geometries and all three total-length/chaining-step cases still run bitwise output and state comparisons; each certification gate now reports one test item." - ] - }, - { - "id": "TA-066", - "scope": "argument parser low-precision adapter conflicts and optimizer-load spellings", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP8 and QARL conflicts with both LoRA and QLoRA still parse fresh YAML and fail independently inside one admission contract.", - "Omitted, explicit true, and explicit false optimizer-load values still parse and resolve inside one defaulting contract." - ] - }, - { - "id": "TA-067", - "scope": "EP checkpoint named-mesh rejection and restoration matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Missing EP, missing expert-FSDP, and duplicate EP dimensions still fail independently; both legacy and PP-parent mesh shapes still restore through fresh fake meshes." - ] - }, - { - "id": "TA-068", - "scope": "GLM-5.2 native-FP8 nonofficial configuration fields", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quant method, FP8 format, activation scheme, and block geometry are still independently mutated from the official config and rejected inside one validation contract." - ] - }, - { - "id": "TA-069", - "scope": "streaming forward-KL dense, compiled, chunking, and backend references", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The retained backward parity test already asserts forward values plus student gradients against the independent dense oracle, subsuming the forward-only test.", - "The retained end-to-end OPD dispatch compares both streaming aliases with the compiled backend including loss and gradients, subsuming the direct compiled forward-only unit.", - "Chunk invariance now compares multi-chunk size 7 directly with single-chunk size 40; the removed size-40000 row compared the function to itself and exact-vocab versus over-vocab selected the same single-iteration path." - ] - }, - { - "id": "TA-070", - "scope": "loss reducer denominator and empty-mask matrices", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "TokenPartial additivity is denominator-independent, so one active-count denominator proves the contract while the separate scale-one raw-sum test remains.", - "TokenPartial and SequencePartial zero-denominator behavior both still execute inside one empty-mask contract." - ] - }, - { - "id": "TA-071", - "scope": "QLoRA random-target convergence and ReLoRA comparison experiments", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Five removed convergence tests ran roughly 750 optimizer iterations against random or synthetic targets and asserted only loss decrease, relative ranking, or a loose two-times threshold; those outcomes select no repository branch.", - "Both NVFP4 and Block-FP8 still exercise quantized storage, memory, forward, backward, loading, merge, and requantization mechanisms.", - "The rewritten scheduler integration proves an off-boundary no-op and on-boundary packed-weight mutation, LoRA-B reset, and optimizer-state removal directly after one state-populating step.", - "Optimizer reset tests use one step, which is sufficient for Adam state materialization, while preserving exact LoRA-only and non-LoRA-state assertions." - ] - }, - { - "id": "TA-072", - "scope": "Mooncake hidden-store malformed metadata table", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Missing key, shapes, dtypes, and invalid rank metadata still mutate fresh dictionaries and fail independently inside one parser contract." - ] - }, - { - "id": "TA-073", - "scope": "LoRA target-manifest exact scalar type tables", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Top-level schema/allow-unlisted types and Boolean count/rank values still fail independently against fresh manifests; each validation family reports one contract." - ] - }, - { - "id": "TA-074", - "scope": "Nemotron-H checkpoint invalid expert-parallel configurations", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Indivisible expert count, out-of-range EP rank, and invalid EP size still construct and fail independently inside one admission contract." - ] - }, - { - "id": "TA-075", - "scope": "DeepSeek-V3 unsupported training-mode matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unfrozen router, QLoRA, and unmerged-QKV modes still enter the real training builder and fail with their specific errors inside one supported-mode contract." - ] - }, - { - "id": "TA-076", - "scope": "GLM-5 config, indexer ownership, and TileLang reference coverage", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The default-shape contract now also asserts model_type, subsuming the bare-construction smoke; the end-to-end four-layer forward already verifies indexer ownership on every attention layer.", - "The prior supposed torch reference used a pure causal mask, which is eligible for the same TileLang fast path. The retained GPU contract sets blocked-scoring sizes to force the independent torch implementation before comparing row sets." - ] - }, - { - "id": "TA-077", - "scope": "merged LoRA shared-factor straight-through autograd matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unshared, shared-A, and shared-B executions and their independent autograd comparisons remain inside one factor-sharing contract rather than three collected parameter rows." - ] - }, - { - "id": "TA-078", - "scope": "training simulator built-in calibration-pack validation", - "decision": "remove", - "status": "applied", - "evidence": [ - "The consolidated validator already discovers all built-in packs, validates schema and sanitation, checks behavior-point counts, and asserts exact raw and promotable golden throughput.", - "Pack names and report schema version were added to the consolidated assertion before removing the weaker sanitation loop and Qwen3.5 winner-only test." - ] - }, - { - "id": "TA-079", - "scope": "DeepEP internode topology and preflight no-op matrices", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All seven uninitialized, malformed, single-node, contiguous, and strided topology scenarios still execute in one truth table.", - "Environment-disabled, uninitialized-distributed, and intranode preflight exits still independently prove that no DeepEP buffer is created." - ] - }, - { - "id": "TA-080", - "scope": "sequential packer artificial types, batch scale, schema whitelist, and roundtrip repetitions", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The removed ABC/custom-subclass test exercised a test-only implementation rather than production packing, and the exact optional-key whitelist failed on harmless schema extension.", - "The arbitrary 100-sample workload selected no new branch; numpy normalization remains. The roundtrip retains one exact sample-boundary integration while repeated multi-batch and shift-mode checks remain in their dedicated contracts." - ] - }, - { - "id": "TA-081", - "scope": "session base-model cache-path canonicalization rows", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Repository-id identity and both cache-path directions still execute inside one canonicalization contract; distinct models and pass-through values remain separately asserted." - ] - }, - { - "id": "TA-082", - "scope": "launcher tests for removed pre-refactor APIs", - "decision": "remove", - "status": "applied", - "evidence": [ - "Thirteen tests were permanently skip-gated because the launcher refactor removed local-master detection, worker override forwarding, and init-time override validation.", - "The retained seven tests cover the live remote and explicit-host address paths, readiness success/failure, current command behavior, parser behavior, and the removed-ZORL migration error." - ] - }, - { - "id": "TA-083", - "scope": "model-runner token-diagnostic shape, ranking, and disabled-input checks", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Valid positions, target IDs, output shapes, top-k width, and target-versus-top1 consistency now share one output contract.", - "Zero top-k and absent labels still independently execute the same disabled diagnostic boundary inside one test." - ] - }, - { - "id": "TA-084", - "scope": "checkpoint metadata production and compatibility consumption", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "QARL buffer metadata is now written, inspected, and consumed by strict and non-strict compatibility checks in one transaction.", - "Pipeline parameter and buffer unions are now written to disk and validated from that artifact instead of testing writer and reader against separately fabricated metadata." - ] - }, - { - "id": "TA-085", - "scope": "Muon restart helper, Quack tuned mode, and backend dtype selection", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The helper-only restart assertion was removed because the retained real optimizer step invokes autotuning and asserts the chosen reset iteration.", - "Untuned and tuned Quack calls and FP32-versus-BF16 SM90 backend selection still execute as complete truth tables." - ] - }, - { - "id": "TA-086", - "scope": "cautious-weight-decay optimizer factory routing", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "SignSGD, explicit AnyPrecisionAdamW, and AdamW-to-AnyPrecision routing still build fresh optimizers and validate cautious flags; AdamW also retains its FP32 momentum assertion." - ] - }, - { - "id": "TA-087", - "scope": "API server constructor, Pydantic assignment smokes, and heartbeat timing", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Constructor field echo and request-object assignment tests were removed; retained endpoint tests exercise aliases, defaults, serialization, registration, and optimizer payloads at the application boundary.", - "Heartbeat activity is now advanced by a deterministic method stub instead of a real sleep, while still proving the endpoint invokes session refresh." - ] - }, - { - "id": "TA-088", - "scope": "trainer gradient clipping, sign-vote scaling, and target-token preference tables", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Regular and DistSign clipping, both nonpositive disable values, and positive/zero/negative voter totals retain every execution inside three behavioral tables.", - "Token and active-microbatch counters both prove target_tokens precedence in one shared caller contract." - ] - }, - { - "id": "TA-089", - "scope": "router diagnostic tie-policy aliases and invalid-policy boundary", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "stable_low_id, tie_low_id, and tie_high_id still construct fresh routers and validate selected experts and gathered weights inside one policy contract.", - "The invalid-policy test now expects the actual fail-closed construction boundary rather than incorrectly constructing outside the exception assertion." - ] - }, - { - "id": "TA-090", - "scope": "teacher activation-cache host/device rank-3 gathers and index bounds", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Host and cached-device token-axis and layer-slice paths still execute with their original shapes and dtype conversion inside two contracts.", - "Negative and upper-bound index failures remain independently asserted as one bounds contract." - ] - }, - { - "id": "TA-091", - "scope": "DeepSeek-V4 successful checkpoint name mappings", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Top-level, attention, indexer, norm, HC, router-bias, and shared-expert mappings all remain exact assertions in one successful-map contract; unknown and MTP names retain their separate rejection contract." - ] - }, - { - "id": "TA-092", - "scope": "orchestrator model-pass results and sequential-operation smokes", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Backend timing-field preservation now runs inside the retained forward/backward and forward-only operation contract.", - "Five identical successful sequential forwards were removed from the error test; exact operation-counter behavior remains in the dedicated statistics contract." - ] - }, - { - "id": "TA-093", - "scope": "runner-dispatch routing expert and logit slicing", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained shard-and-slice test already uses the same datum offset/count and asserts both routed expert IDs and routed logits plus removal of slicing metadata; the second test repeated that contract." - ] - }, - { - "id": "TA-094", - "scope": "pipeline profiling interval and analytic bubble helpers", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Empty, degenerate, disjoint, overlapping, contained, touching, and unsorted intervals all remain in one union truth table.", - "1F1B, GPipe, interleaved, zero-bubble, PP1, and invalid schedule branches all remain, grouped by nonzero formula, zero result, and rejection behavior." - ] - }, - { - "id": "TA-095", - "scope": "OPD driver prompt chunking and student weight-version verification", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Order/tail preservation and nonpositive chunk rejection remain in one chunking contract.", - "Matching, mismatched, and model-info-error weight versions still execute with their exact profile-row side effects inside one verifier truth table." - ] - }, - { - "id": "TA-096", - "scope": "sparse-delta template traversal and malformed update validation", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed filename and manifest filename traversal are still rejected against fresh capture state inside one trust-root contract.", - "Duplicate, floating-point, length-mismatched, and out-of-range indices still enter the real writer and fail with their specific errors inside one receiver contract." - ] - }, - { - "id": "TA-097", - "scope": "EP-aware gradient clipping classification, norm arithmetic, and mixed-mesh smokes", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The retained classify-then-clip contracts already prove skip-FSDP and ordinary parameter grouping, combined L2 clipping, unchanged below-threshold gradients, and absence of EP double division; separate 3-4-5 arithmetic units were strict subsets.", - "Infinity norm, missing gradients, shared replicas, dispatch, mixed DTensor meshes, explicit foreach behavior, and live two- and three-rank reductions remain independently exercised." - ] - }, - { - "id": "TA-098", - "scope": "DistSignSGD update, state, reduce-scatter, and topology micro-tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The weight-decay update covers the ordinary preaggregated update in the same production step, and the state-dict roundtrip proves the optimizer remains state-free after stepping.", - "Signing and forced-SUM behavior now use one AVG-input communication contract, while HSDP, folded sequence parallelism, and EP still fail against fresh models inside one topology table." - ] - }, - { - "id": "TA-099", - "scope": "constant, linear, and cosine learning-rate schedule examples", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Warmup-to-constant, full linear decay, warmup-decay-floor, cosine floor, and warmup-cosine contracts cover every schedule phase; standalone constant and floor examples repeated phases already asserted by those traces.", - "Every invalid learning rate, warmup ratio, and decay style still enters its production validation branch inside one configuration contract." - ] - }, - { - "id": "TA-100", - "scope": "NVFP4 QARL normalization, fake-quant helpers, dense wrappers, and MoE wrappers", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Non-16 group sizes and target-module directions still run as truth tables, while dense and eager-MoE forward contracts now combine configuration, lossy output, parameter restoration, and gradients.", - "Private MoE fake-quant and shadow tests were removed because exact 3D forward and STE arithmetic belongs to the retained op suite; the wrapper contract proves those helpers feed the inherited production forward.", - "The independent NVFP4 reference subsumes a looser grid property, the linear STE contract subsumes the direct identity test, and registry dispatch subsumes the supported-format boolean smoke." - ] - }, - { - "id": "TA-101", - "scope": "FP8 model-builder tensor-parallel lm-head inclusion examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full FP8 and lm-head-excluded tensor-parallel builds still construct fresh models and validate both projection types inside one inclusion contract." - ] - }, - { - "id": "TA-102", - "scope": "QARL activation-quantization override exit and target examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Enabled and disabled overrides, distinct per-module restoration, exception cleanup, and exclusion of ordinary Linear modules now execute as one state transaction; nested restoration remains a separate reentrancy contract." - ] - }, - { - "id": "TA-103", - "scope": "QARL W4A4 activation STE and MoE backend-shadow helper fragments", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The weighted activation-gradient contract proves exact STE identity with nonuniform upstream gradients, subsuming the all-ones sum example.", - "The exception path proves triton_w4a4 selection and restoration together; activation-off and non-Triton no-op cases still execute in one conjunction-boundary table." - ] - }, - { - "id": "TA-104", - "scope": "QARL synthetic convergence smoke and duplicate sync-configuration failure", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The 16-step loss-decrease heuristic is replaced by one real optimizer step that asserts finite loss, both QARL gradients, parameter mutation, changed logprobs, persistent summary state, and exact checkpoint restoration.", - "The weight-sync handler mismatch contract already enters the same block-size validator and additionally proves the user-facing failure response, subsuming the direct mismatch unit." - ] - }, - { - "id": "TA-105", - "scope": "SignSGD and AnyPrecisionAdamW all-aligned cautious-decay comparisons", - "decision": "remove", - "status": "applied", - "evidence": [ - "Each retained mixed-coordinate production step contains both aligned and misaligned coordinates and checks the exact resulting update; the all-aligned comparisons selected no additional optimizer branch." - ] - }, - { - "id": "TA-106", - "scope": "GLM-5.2 IndexShare identity through the test-local tensor mapper", - "decision": "remove", - "status": "applied", - "evidence": [ - "The removed test only proved that a recursive helper defined in the test file leaves non-tensor Python objects unchanged.", - "The retained dense-producer/shared-consumer model forward runs the same mapper in simulated FSDP pre-hooks and proves one context identity across every layer plus lifecycle cleanup." - ] - }, - { - "id": "TA-107", - "scope": "GLM-5 indexer shape, additive-mask, chunked-head, and padding-mask examples", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Shape, dtype, range, final-row validity, and sorted-sentinel ordering now live in one selector output contract.", - "Dense and one-head-chunk scoring both retain the diagonal-only additive-mask result inside one behavioral table.", - "The padding-mask test now runs on CPU and accurately tests prefix acceptance versus interior-hole rejection instead of claiming GPU fast-path execution it never invoked." - ] - }, - { - "id": "TA-108", - "scope": "GLM-5 sparse-attention output-shape smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained full-model sparse-versus-dense contract executes sparse attention and checks numerical output, while the Ulysses integration checks local query, full KV, top-k, mask, offset, and output shapes; the standalone shape-only forward was dominated." - ] - }, - { - "id": "TA-109", - "scope": "exact Qwen3.5 dense and MoE numerical-program and admitted-topology successes", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense and MoE configs still resolve the complete certified numerical program and CE mode from fresh configs inside one family contract.", - "World16 HSDP-plus-EP, world8 EP, and single-GPU dense topologies still enter the real admission validator inside one successful-topology table." - ] - }, - { - "id": "TA-110", - "scope": "exact Qwen3.5 model-scope accepted snapshots and nearby-geometry rejection rows", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense, MoE, and Hugging Face outer-config snapshots still validate from fresh objects in one accepted-scope contract.", - "Both dense and MoE hidden-size near misses still fail independently inside one rejection contract rather than generating parameterized item inflation." - ] - }, - { - "id": "TA-111", - "scope": "weight-sync adapter preparation, bucket helpers, tied aliases, MoE prefix mapping, and protocol field echo", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Requested-adapter materialization and current-adapter fallback still run in one stateful contract, and chunking plus cap decisions now share one bucket contract.", - "The retained prior-module tied-weight test includes the removed root extraction and alias assertions before proving the duplicate is skipped on the subsequent module.", - "Nemotron-H and ordinary-MoE unfuse transactions already prove prefix remapping at the produced tensor boundary; the direct helper smoke was dominated.", - "Sparse-delta protocol fields remain exercised through remote-backend and end-to-end sync requests, so the local Pydantic assignment echo was removed." - ] - }, - { - "id": "TA-112", - "scope": "sparse-delta path fast-path baseline, config, and FP8 cache-metadata harnesses", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Baseline post-only ordering and accounting, explicit baseline configuration, and FP8 KV-cache postprocess metadata still execute independently through one shared fake transport transaction.", - "The rewrite removes two copied endpoint/backend harnesses while preserving backend config, normalized cache epoch, endpoint results, pause/resume order, posted paths, and weight version assertions." - ] - }, - { - "id": "TA-113", - "scope": "adapter-manager ownership, optimizer, lifecycle, eviction, and checkpoint transactions", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The 48 tests cover distinct compile, capture, staging, atomic commit, abort, clipping, collective, poisoning, publication, trust-root, session-spec, rollback, eviction, mixed-rank, and optimizer-state boundaries rather than literal or shape variations.", - "All 48 pass in the source-tree environment." - ] - }, - { - "id": "TA-114", - "scope": "MiniMax-M3 expert-key aliases and DeepSeek-V4 window-only forward smoke", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "All three MiniMax language-model prefix and w1/w2/w3 expert-key aliases still execute inside one classifier contract.", - "The removed DSv4 base-model C0 shape smoke was dominated by the retained full C0 causal-LM forward/backward, which additionally checks logits, loss, required gradients, and intentionally frozen hyperconnection parameters; the distinct C128 forward remains." - ] - }, - { - "id": "TA-115", - "scope": "RMSNorm cross-engine shape/family matrices and fused-kernel dtype rows", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every adversarial cross-engine shape, both family funnels, both residual modes, trunk lane, zero-centered twin, and families-v2 candidate execution remains; each numerical invariant now reports one test item.", - "BF16 and FP32 residual and no-residual fused-kernel comparisons still run inside two dtype-complete contracts, both of which pass locally.", - "The SGLang cross-engine module is dependency-skipped in the current venv, so its grouped executions were linted and collection-checked but could not run locally." - ] - }, - { - "id": "TA-116", - "scope": "endpoint-manager and NCCL routing examples repeated across the same control transaction", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The fallback health contract already invokes and verifies the primary host-and-port URL before the v1/models fallback, so the primary-success URL-only example was dominated.", - "Configured direct load-format forwarding now lives in the existing bucket endpoint-routing contract instead of repeating its complete synchronizer and broadcast harness.", - "The hybrid receiver-fence contract now proves deferred work lifetime and release on NCCL-group destruction in one transaction; the copied destruction harness was removed." - ] - }, - { - "id": "TA-117", - "scope": "sparse-delta factory, replicated-path, and initialization-policy micro-tests", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The backend-factory isinstance check only mirrored a single factory branch and was removed.", - "Single-path TP replication remains asserted by streaming transfer, while the retained prepacked-path contract owns distinct per-rank paths and now checks unique-file accounting.", - "Post-only import avoidance, streaming rejection under prepacked-only, and the valid prepacked post-only combination all still execute inside one initialization-policy table.", - "The fake unresolvable hostname was replaced with a validated loopback literal so mocked HTTP tests also pass through production URL safety checks deterministically." - ] - }, - { - "id": "TA-118", - "scope": "trainer P2P IB-device selection examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Global-rank, local-rank, single-device fallback, explicit physical-GPU selection, numeric CUDA visibility, and empty-entry autodiscovery all still execute in an environment-isolated precedence table.", - "Seven mapping outcomes now report one behavior contract rather than six separately named examples." - ] - }, - { - "id": "TA-119", - "scope": "private P2P rank-summary dictionary assembly and three-counter addition examples", - "decision": "remove", - "status": "applied", - "evidence": [ - "The removed tests asserted private dictionary field copies and the arithmetic sum of byte, parameter, and bucket counters without exercising a transfer, collective failure, or user-visible result.", - "Abort-marker lifecycle and distributed peer-failure gathering remain as separate operational contracts." - ] - }, - { - "id": "TA-120", - "scope": "server removed-field shapes and shipped MoE adapter configuration rows", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Flat adapter-ownership, nested removed ZORL field, and removed ZORL-section payloads still enter the real server loader and assert their distinct failures inside one rejection table.", - "Both shipped MoE LoRA and all five shipped Qwen MoE QLoRA configurations still parse in a clean subprocess and compare their source and normalized Quack, target-module, and shared-LoRA values; parametrized item inflation was removed." - ] - }, - { - "id": "TA-121", - "scope": "independent server optimizer, resume, prefetch, HSDP, packing, activation, and adapter-state field tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "SignSGD and DistSignSGD each parse through a composed nested config that also checks checkpoint optimizer policy, forward/backward prefetch, HSDP deferral, packing alignment, activation memory limit, and adapter state-load mode.", - "Both the runtime object and its serialized model, train, and LoRA configs remain asserted at the server boundary." - ] - }, - { - "id": "TA-122", - "scope": "server R3 payload success modes and MoE routing-weight default examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit Mooncake transport, the legacy externalization alias, and explicit filesystem fallback still parse and serialize every transport-specific field inside one mode table; the invalid directory-without-filesystem failure stays separate.", - "Explicit routing-before-down and the automatic default both still execute through one defaulting contract." - ] - }, - { - "id": "TA-123", - "scope": "training YAML optimizer, packing, numerical-alignment, and FSDP scalar acceptance tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both SignSGD values still pass the real CLI parser, while multipack shape, model numerical-alignment flags, FSDP reduction dtype, and parameter-upcast policy are checked together from the same production-shaped YAML.", - "Muon, legacy alias transforms, FP8/QARL configuration, automatic checkpoint resolution, load-optimizer defaults, and all incompatibility failures remain independent contracts." - ] - }, - { - "id": "TA-124", - "scope": "GatedDeltaNet exact-convolution unsupported-mode examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Decode-cache, context-parallel, missing-short-convolution, and convolution-bias failures still construct and invoke their distinct production paths inside one guard table.", - "Forward/backward references, kernel and end-to-end determinism, state scoping, checkpoint recompute, packing order, and SGLang parity remain independent contracts." - ] - }, - { - "id": "TA-125", - "scope": "batch-invariant trunk wrapper type guards and bias parameter rows", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "No matching projection, ordinary LoRA, custom Linear subclass, and FP16 weight failures all still execute inside one wrapper-admission table.", - "Biased and bias-free forward and backward bitwise comparisons still run against their independent persistent-GEMM and cuBLAS references, but no longer inflate pytest items." - ] - }, - { - "id": "TA-126", - "scope": "batch-invariant trunk global-state and duplicate RMSNorm loud-failure assertions", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The autouse fixture now establishes disabled global state before as well as after each contract, preventing order-dependent inheritance.", - "The selection contract now correctly asserts that wrapping arms RMSNorm dispatch, matching the implementation contract; its prior assertion required the opposite behavior.", - "The standalone RMSNorm loud-failure regression was removed because the retained multi-op grad-requiring interpose contract already invokes RMSNorm and requires the same failure." - ] - }, - { - "id": "TA-127", - "scope": "attention registry presence, flash-family detection, and resolver examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The FA4-only reload transaction now directly proves flash_attention_2, flash_attention_3, and flash_attention_4 registration plus mask-family detection, subsuming a standalone equality with the registry-membership expression used by the implementation.", - "Registered eager/native resolution, non-flash eager fallback, and unavailable-flash rejection now form one resolver-boundary contract." - ] - }, - { - "id": "TA-128", - "scope": "quantized-export example snapshot and unsupported-source preflight fragments", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The checked-in example's literal field snapshot was removed because generic config parsing and the retained subprocess CLI export already validate the parser and consumer transaction.", - "Existing FP8 scales, MTP config metadata, MTP tensor namespaces, and unfolded QARL state still build real source directories and fail through one export-preflight table." - ] - }, - { - "id": "TA-129", - "scope": "QARL export eight-step synthetic training loop", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "One real AdamW step now establishes finite loss and changed target logprobs before export, replacing an arbitrary eight-iteration mini-training experiment.", - "The retained contract still requires exact target-logprob equality after folding, block-FP8 export, dequantization, and reload." - ] - }, - { - "id": "TA-130", - "scope": "API request removed-configuration parameter rows", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Create-model ZORL, nested adapter-ownership, and create-session ZORL inputs still enter their actual Pydantic request types and assert both field path and migration message inside one rejection table.", - "Unknown rolling-client and nested LoRA fields remain covered by the separate compatibility contract." - ] - }, - { - "id": "TA-131", - "scope": "TensorData rank-specific to_plain_dict examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-one model and loss fields, valid rank-two and rank-three nesting, mismatched shape fallback, and empty higher-rank fallback all still execute in one conversion contract.", - "The retained assertions preserve exact nested values and therefore the sequence-field classification invariant consumed by packing." - ] - }, - { - "id": "TA-132", - "scope": "SGLang fused-expert DeepEP exclusion and pair-slot helper duplicates", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained DeepEP exclusion contract now asserts the all-to-all path plus the precise unverifiable order-and-rounding mechanism, subsuming a second test that only required the same NotImplementedError.", - "The full slot-combine path independently checks slot-ordered weighted reduction from routed inputs, so a direct private pair-order helper example was dominated." - ] - }, - { - "id": "TA-133", - "scope": "API optimizer current/legacy payloads and sampler tracking model-ID examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Current learning-rate/gradient-clip fields and legacy Adam parameter aliases both still reach the orchestrator payload and response metrics in one optimizer transaction.", - "Embedded xorl URI model IDs and explicit request model IDs both still load and track their actual sampler paths in one cleanup-ownership contract." - ] - }, - { - "id": "TA-134", - "scope": "base-model canonicalization equality and inequality examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Repository IDs and Hugging Face cache paths still normalize in both directions, while distinct repositories, ordinary paths, and None remain distinct or unchanged inside one canonicalization contract." - ] - }, - { - "id": "TA-135", - "scope": "FP8 E2E Ulysses and hybrid context-parallel shape ladder", - "decision": "remove", - "status": "applied", - "evidence": [ - "The short Ulysses case is dominated by the retained longer packed Ulysses transaction, which exercises the same FP8 and Ulysses branches at a stronger shape.", - "Three intermediate hybrid context datasets and the basic hybrid case add only sequence length or sample-shape variation; the retained 4096-token long-tail multipack transaction exercises the same Ulysses-plus-Ring composition with heterogeneous near-full bins." - ] - }, - { - "id": "TA-136", - "scope": "FP8 MoE DeepEP checkpoint-resume cross-product", - "decision": "remove", - "status": "applied", - "evidence": [ - "Dense FP8 checkpoint save/resume retains FP8 serialization and optimizer restoration, while the retained DeepEP EP/eFSDP transaction proves FP8 expert compute through that distributed topology.", - "Combining both mechanisms in a second two-phase four-GPU run selected no additional checkpoint or DeepEP branch." - ] - }, - { - "id": "TA-137", - "scope": "retained FP8 E2E metric oracle and configuration generator", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "A live baseline completed two optimizer steps with eight of nine linears using FP8; the sole unused module was lm_head, which Qwen3's resolved numerical program intentionally executes in FP32.", - "The oracle now requires every eligible linear to use FP8 while naming only lm_head as the canonical FP32 exception, and the retained baseline passes.", - "The shared E2E generator now applies the extra_data and extra_model mappings already supplied by retained packed-context and DeepEP cases instead of failing at Python argument binding." - ] - }, - { - "id": "TA-138", - "scope": "GLM-5 architecture registry and loader-selection snapshots", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The retained local Hugging Face config load now proves both architecture aliases, their class relationship, and the selected GLM loader after constructing the actual Glm5Config.", - "Two standalone registry-membership and loader-description tests added no execution boundary beyond that transaction." - ] - }, - { - "id": "TA-139", - "scope": "GLM-5 checkpoint layer normalization and early-skip key examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Configured, MTP-boundary, far-out-of-range, and non-layer keys still pass through both the normalizer and loader early-skip hook in one routing contract.", - "The separate tests repeated the same four key classifications against two surfaces of the same handler." - ] - }, - { - "id": "TA-140", - "scope": "OPD hidden-cache unpacked filtering and payload-merge helper examples", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained gathered-SP writer filters an interior valid target from the full hidden tensor and asserts its persisted cache index, subsuming the direct unpacked row-split helper example.", - "The retained multi-rank writer gathers local and remote chunks, orders them by logical slice, persists the concatenated tensor, and asserts per-sample indices, subsuming the direct payload-merge helper example." - ] - }, - { - "id": "TA-141", - "scope": "P2P generic source-slicing helper examples", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained full transfer transaction writes exact TP row slices to two receiver pointers, while the retained incompatible-shape transaction rejects transfer before the engine is called.", - "Unsliced tensors flow through many retained transfer transactions, so direct private-helper examples for TP rows, full-shape mismatch, and identity return added no independent boundary; specialized Qwen linear-attention transformations remain." - ] - }, - { - "id": "TA-142", - "scope": "P2P initialization failure, sender capability, and scatter-copy mode fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HTTP and successful-response remote failures still execute independently inside one initialization failure contract.", - "Implicit all-rank and explicit sender sets now share one capability contract, and list, deep, and forced-reuse locator alias policies share one scatter-copy contract.", - "Repeated dense-owner and filtered-buffer assertions inside two retained tests were removed without dropping any value or branch." - ] - }, - { - "id": "TA-143", - "scope": "P2P transfer metadata and direct-EP private-helper smokes", - "decision": "remove", - "status": "applied", - "evidence": [ - "Flush preservation and weight-version propagation now reach the real completion payload in one transfer transaction instead of stopping at backend configuration.", - "Multi-sender nonzero-rank initialization already adopts and validates scattered tensor maps, subsuming a direct state-assignment helper example.", - "The retained all-filtered direct-EP transfer uses a nonzero source rank with a real locator and proves the engine stays untouched, subsuming the empty-bucket no-exception smoke." - ] - }, - { - "id": "TA-144", - "scope": "inference worker-port registration and adapter URL examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The explicit-worker-port registration transaction now checks control and worker health, then uses the registered endpoint for LoRA load, unload, and loaded-adapter discovery.", - "Two standalone adapter URL tests constructed an endpoint by hand and repeated only the final port choice." - ] - }, - { - "id": "TA-145", - "scope": "inference sync-quantization normalization and unsupported-receiver examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The rich FP8 skip-list detection contract already proves default format, activation scheme, block size, language-model normalization, and vision exclusion, subsuming a minimal default-dictionary snapshot.", - "MTP, invalid activation schemes, UE8M0 scale storage, and BF16 MTP now form one receiver-admission policy contract, while compressed-tensors rejection and explicit BF16 no-op normalization share the setter policy boundary." - ] - }, - { - "id": "TA-146", - "scope": "packing strategy literal snapshot and convenience-wrapper smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "Every supported strategy still executes through document, token, position, capacity, balance, order, and determinism contracts, so an exact tuple-literal assertion adds no behavior.", - "The retained full pipeline calls pack_samples, asserts request and packed boundaries, simulates output, and unpacks every sample; direct packed and unpacked contracts separately cover the wrapper's two mode branches." - ] - }, - { - "id": "TA-147", - "scope": "pipeline schedule style, capability, and admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every schedule now maps to its stage style, single-stage classification, and split-backward policy in one metadata table.", - "All five admitted configurations and every invalid virtual-stage or microbatch branch execute inside one schedule-admission contract." - ] - }, - { - "id": "TA-148", - "scope": "DeepEP preflight and launcher readiness success-only tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One DeepEP roundtrip contract now accepts identity combine and rejects corrupt combine after the same dispatch setup.", - "One launcher readiness contract now accepts a set ready event and independently fails fast when the worker exits." - ] - }, - { - "id": "TA-149", - "scope": "Blackwell FP8 and exact-Qwen MoE admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Blackwell FP8 now rejects no override, rejects an override without a validation artifact, and accepts the explicit validated override in one policy contract.", - "Exact Qwen3.5 MoE defaults and every noncertified implementation, dispatch, or async-combine override now enter one architecture admission boundary." - ] - }, - { - "id": "TA-150", - "scope": "factor-only exact active-LoRA snapshot guard scope", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The retained snapshot contract rejects both full-weight publication paths before downstream work, then proves the same guard leaves an ordinary model unrestricted." - ] - }, - { - "id": "TA-151", - "scope": "FP8 sync output-shape snapshot and contiguous-slice predicate example", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained independent Slime-reference contract now also asserts CPU placement while already proving emitted names, dtypes, scale shape, exact FP8 bytes, exact scales, and dequantized parity, subsuming the weaker output snapshot.", - "Stack-versus-single quantization proves grouping is numerically transparent, and workspace/streaming transactions exercise grouped expert stacks; a direct storage-offset predicate example asserted no user-visible correctness boundary." - ] - }, - { - "id": "TA-152", - "scope": "FP8 GPU stack target-device and CPU-parity fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One CUDA contract now quantizes the same nontrivial stack to CPU and CUDA targets, checks target-specific copy telemetry, and compares both outputs bitwise against the CPU path.", - "Three separate tests previously rebuilt the stack and reported placement and parity independently." - ] - }, - { - "id": "TA-153", - "scope": "two-dimensional DTensor save materialization process launches", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One four-rank 2D CPU-mesh transaction now proves all-rank materialization and writer-only materialization from the same sharded tensor.", - "The separate writer-rank test repeated process-group setup, mesh construction, shards, and full-tensor reconstruction; the distinct one-dimensional mesh contract remains." - ] - }, - { - "id": "TA-154", - "scope": "grouped-GEMM scale, single-group, and property examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The roughly one-gigabyte aligned operand exercised the same masking and autotune shape family as smaller examples; a compact unaligned FP16 case now reaches the previously uncovered K/N masking path.", - "Single-group execution has no separate implementation branch, while retained numerical contracts already cover FP16, BF16, unequal groups, transpose-B, zero-K, contiguity, and device admission." - ] - }, - { - "id": "TA-155", - "scope": "MoE primitive example ladders and non-gated constructor failures", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Repeated flattened histograms, an invalid overlapping slot map, and separate gather/scatter/add-gather smokes were dominated by retained independent references and a permutation roundtrip.", - "The multi-block gather now uses a compact unaligned hidden width, the full pipeline asserts its exact result, and unsupported backend and activation policies share one constructor contract." - ] - }, - { - "id": "TA-156", - "scope": "adapter-manager checkpoint path, rollback, dtype, and learning-rate fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Save-root and load-root rejection now execute in one containment policy transaction.", - "The missing-tensor failure now also proves fresh-adapter rollback, and the current-learning-rate checkpoint transaction also verifies persisted LoRA tensor dtypes." - ] - }, - { - "id": "TA-157", - "scope": "adapter-coordinator path containment and fresh-state broadcasts", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Load paths, save paths, and evicted model-id traversal now enter one output-root policy contract.", - "Direct adapter registration and materialized session registration now prove their ordered broadcasts through one coordinator transaction." - ] - }, - { - "id": "TA-158", - "scope": "routing replay sequence-length and already-decoded input examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One sequence-parallel transaction now covers both exact-length slicing and route padding at the actual position length.", - "Python lists, NumPy arrays, and tensors share one materialized-input contract, while nested Qwen top-k discovery is exercised by the retained raw-base64 shape-inference path." - ] - }, - { - "id": "TA-159", - "scope": "server Adam hyperparameter argument, initialization, step, and dispatcher fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit and default arguments share one configuration contract, while explicit, default, and malformed optimizer initialization share one admission transaction.", - "Full, partial, omitted, and non-Adam step policies execute together; adapter-manager and dispatcher propagation remain independent boundaries." - ] - }, - { - "id": "TA-160", - "scope": "runner dispatcher save variants and rank-parameterized failure policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Save-state and save-LoRA handlers now prove the same nonresident-adapter checkpoint requirement through one dispatcher transaction.", - "Rank-zero and worker forward/backward paths still exercise uniform rejection and asymmetric fatal promotion without generating duplicate parameterized reports." - ] - }, - { - "id": "TA-161", - "scope": "LoRA checkpoint SGLang layout and expert-ownership roundtrip fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The SGLang shared-outer roundtrip now inspects its exact serialized key set and shapes before reloading, subsuming a save-only layout report.", - "Hybrid-shared and all-owner expert layouts now pass through one adapter-manager roundtrip policy while retaining distinct source, checkpoint, and manager state." - ] - }, - { - "id": "TA-162", - "scope": "native block-FP8 state lifecycle and CPU admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One byte-exact lifecycle contract now proves frozen state, FP8 and scale contents, and dtype-application preservation.", - "Import laziness, CPU forward rejection, and explicit input-free CUDA materialization admission now share one fail-closed CPU policy contract." - ] - }, - { - "id": "TA-163", - "scope": "adapter-gradient ownership configuration, fingerprints, replica coverage, and analytical self-test", - "decision": "remove", - "status": "applied", - "evidence": [ - "Positive and invalid bucket configuration, rank-local identity and geometry invariance, and admitted/rejected replica coverage now execute as policy transactions.", - "The removed analytical test exercised only math helpers defined inside the test; a retained adapter-manager transaction checks the real production step against scale, clip, AdamW parameter, and moment equations." - ] - }, - { - "id": "TA-164", - "scope": "DRGRPO forward, backward, metrics, zero boundaries, and KL fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One seeded numerical contract now checks the exact loss, gradient norm, finite values, and metric schema from the same operation.", - "Zero advantages, ignored labels, and empty sequences share one zero-loss policy, while KL reference admission and effect share one positive-KL policy." - ] - }, - { - "id": "TA-165", - "scope": "server endpoint and artifact-path security examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Allowlist admission, DNS pinning, metadata rejection, and malformed host rejection now form one outbound endpoint policy.", - "Generic escape and symlink rejection, environment-root confinement, and explicit-root authority now form one artifact path policy." - ] - }, - { - "id": "TA-166", - "scope": "batch-invariant router GEMM input and top-k example fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP32-reference parity, empty-token behavior, and BF16 input admission now share one router kernel contract.", - "Top-k renormalization, cast-only behavior, and FP32 input admission now share one post-processing contract; invariance, gradients, and MoEBlock consumers remain independent." - ] - }, - { - "id": "TA-167", - "scope": "local phase ordering, timing summary, and memory summary examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical, custom, and empty phase ordering now form one ordering contract.", - "Phase-time and memory summaries each verify empty output, canonical ordering, local aggregate fields, and normalization through one complete summary map." - ] - }, - { - "id": "TA-168", - "scope": "per-component decoder discovery and model-style hook fragments", - "decision": "remove", - "status": "applied", - "evidence": [ - "Direct decoder-suffix and nested-attribute helper tests were dominated by the retained attach-and-run consumer transaction.", - "One live CUDA transaction now attaches to GLM and Qwen model styles, records present forward/backward phases, and omits absent indexer and shared-expert phases." - ] - }, - { - "id": "TA-169", - "scope": "manual CUDA timer disabled, mode, recorded, and unrecorded fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Disabled behavior and invalid mode rejection now share one API policy.", - "Forward and recompute timing plus unrecorded-pair omission now execute through one event-drain lifecycle." - ] - }, - { - "id": "TA-170", - "scope": "activation-offload populated and empty metric contexts", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One consume lifecycle now converts populated forward/backward byte counters to gigabytes, checks single consumption, and proves empty or unsupported contexts emit no metrics." - ] - }, - { - "id": "TA-171", - "scope": "DCP synchronization and metadata process-group selector fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One synchronization backend policy now covers cached Gloo creation for NCCL and default-group use for Gloo.", - "One metadata policy now covers disabled pipeline parallelism, caller-supplied groups, and global Gloo fallback." - ] - }, - { - "id": "TA-172", - "scope": "parallel-plan meta slicing properties and exact-GLM malformed disposition reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The already-local meta allocation path now proves local shape, meta device, dtype, requires-grad, shard placement, and unrelated replication in one contract.", - "Already-local and force-shard malformed singleton policies both still run without generating separate parameterized reports." - ] - }, - { - "id": "TA-173", - "scope": "adapter-manager local active-slot shape smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "Retained manager registration and multi-adapter forward transactions already assert rank-specific local factor shapes.", - "The sharded-state suite retains logical pack/unpack, deterministic initialization, layout discovery, and a real two-rank uneven DTensor transaction." - ] - }, - { - "id": "TA-174", - "scope": "checkpoint zero-meta stages and initial optimizer-load modes", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pre-load materialization and post-restore zero-meta failures now execute in one checkpoint lifecycle while preserving stage-specific assertions.", - "Default optimizer restore and weights-only initial restore now share one runner forwarding and state-synchronization policy." - ] - }, - { - "id": "TA-175", - "scope": "MoE TP simulation admission and eager reduction-mode examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Environment, topology, EP rejection, and layer filtering now share one admission policy.", - "Direct, BF16, cache, and TP-size override modes still run against their corresponding independent references from one shared expert and routing fixture." - ] - }, - { - "id": "TA-176", - "scope": "MoE TP carried-shard and diagnostic-capture transactions", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One MoEBlock execution now proves reshaped carried shards, flat per-shard diagnostic captures, their exact sum, and the final reshaped output.", - "Every backend-specific call-layout contract remains independent." - ] - }, - { - "id": "TA-177", - "scope": "DeepSeek-V4 shared-MLP clamp smoke and non-hash MoE property fragments", - "decision": "remove", - "status": "applied", - "evidence": [ - "The removed shared-MLP smoke explicitly avoided asserting a clamp effect; the retained forced-gate numerical case proves bounded clamped output and shape.", - "Non-hash structure, forward/backward, selection-only bias, and shared-expert contribution now execute through one model transaction." - ] - }, - { - "id": "TA-178", - "scope": "DeepSeek-V4 hash-layer structure, input admission, and forward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One hash-layer transaction now proves table and bias structure, rejects missing input IDs, executes table-driven routing, and checks finite gate gradients.", - "Record-to-replay backward and unknown replay-stage failure remain independent state-machine contracts." - ] - }, - { - "id": "TA-179", - "scope": "FP8 LM-head CE module selection, TP, and temperature fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One per-token CE transaction now selects module and FP32-master paths locally and under identity TP collectives.", - "Temperature propagation is checked once through both the primitive and CausalLM consumer; importance-sampling and TP-gradient consumers remain independent." - ] - }, - { - "id": "TA-180", - "scope": "Qwen3.5 native-EP admission and variable-row collective fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP8 admission, exact structural flags, and missing trainer-EP rejection now form one exact-combine policy.", - "Token padding and backward unpadding, invalid ID padding, and shared maximum-row selection now form one variable-row collective contract." - ] - }, - { - "id": "TA-181", - "scope": "DeepSeek-V3 checkpoint expert layout conversion fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "External dense merge, internal fused pass-through, save splitting, and multimodal filtering now execute as one layout-conversion transaction.", - "Dense and packed EP slicing share one policy, packed dtype behavior shares one load transaction, and model-level packed recognition carries through quant-config parsing." - ] - }, - { - "id": "TA-182", - "scope": "DR-GRPO runner legacy input and per-token output option fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy logprobs, temperature forwarding, disabled per-token output, and K3-forced output now run through one runner option contract.", - "Full loss dispatch, sampler-prefill model forwarding, and the forward-backward loop remain separate integration boundaries." - ] - }, - { - "id": "TA-183", - "scope": "SignSGD update, missing-gradient, and stateless persistence fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sign updates, decoupled weight decay, zero-sign behavior, and missing-gradient preservation now execute in one optimizer step.", - "Multiple stateless steps and hyperparameter-only state-dict restoration now share one persistence lifecycle." - ] - }, - { - "id": "TA-184", - "scope": "CausalLM Z-loss analytical example and duplicate temperature report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The random reference comparison now also proves backward gradients, making a separate finiteness-only gradient report redundant.", - "The zero-logit formula example only restated the test reference, and CausalLM temperature propagation is already retained in the FP8 LM-head suite." - ] - }, - { - "id": "TA-185", - "scope": "MoE train-router dispatch, default, and synthetic replay fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One trained-router transaction proves all-to-all gradients before rejecting DeepEP dispatch.", - "Argument and constructed-model defaults share one policy, while balanced forward routing and replay regather share one uniform-routing contract." - ] - }, - { - "id": "TA-186", - "scope": "shared-prefix detection, repack layout, loss-field, and remap fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One shared-prefix transaction now detects groups, repacks exact token and position layouts, preserves decoded loss fields, and remaps outputs to original order.", - "No-sharing and one-token-prompt boundaries remain independent because they select different backend outcomes." - ] - }, - { - "id": "TA-187", - "scope": "multi-part optimizer structure, coverage, step, and scheduler fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "MultiOptimizer construction now proves DCP model mapping and complete parameter-group coverage together.", - "A single optimizer lifecycle proves every virtual part updates, gradients clear, and every learning-rate group decays." - ] - }, - { - "id": "TA-188", - "scope": "batch-invariant fused LM-head forward and backward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forward values, ignored-token loss, and backward gradients now compare with eager from the same graph for the default temperature.", - "The non-unit temperature path likewise proves forward and backward parity in one transaction; determinism, guards, unit-temperature bytes, and probability clamping remain separate." - ] - }, - { - "id": "TA-189", - "scope": "model-runner FP8, QARL, and sharded-loss builder propagation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One builder-policy transaction now checks fail-closed FP8 defaults, every QARL calibration field, and sharded LM-head loss propagation through the same initialization boundary.", - "Exact GLM block-FP8 QLoRA remains separate because it also resolves the runtime target-module set." - ] - }, - { - "id": "TA-190", - "scope": "exact GLM dense-component and routed-expert gradient ownership fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Gate-up, dense MLP, and absorbed-KV leaves now compile through one module-managed ownership matrix with component-specific canonical factor names.", - "Unwrapped EP16 ownership and mutated DeepEP dispatch now fail inside one routed runtime-admission contract." - ] - }, - { - "id": "TA-191", - "scope": "batch tensor conversion float side-channel fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "DR-GRPO old/reference logprobs and distillation teacher hidden states now prove FP32 preservation in one conversion transaction.", - "Ragged teacher padding and sequence-parallel sharding remain independent shape transformations." - ] - }, - { - "id": "TA-192", - "scope": "server batch-slice topology examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FSDP shards, TP replicas, default EP ranks, and legacy EP duplication now map to their exact slice coordinates in one selector policy.", - "All original rank and topology cases still execute without one report per branch." - ] - }, - { - "id": "TA-193", - "scope": "fused GDN LoRA merged-forward and cache-generation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical sliced folding and the GDN output-projection consumer now share one gradient-preserving merged-forward contract.", - "Cache slice bounds and release of the previous adapter generation now execute in one version-change lifecycle." - ] - }, - { - "id": "TA-194", - "scope": "DeepSeek-V4 attention sink storage and call-dtype fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One sink lifecycle now proves initial FP32 storage and FSDP marking, BF16 module conversion, and FP32 promotion at the TileLang call boundary.", - "Attention variants, quantization admission, compressor structure, and TP rejection remain separate." - ] - }, - { - "id": "TA-195", - "scope": "stochastic-round output metadata and input-dtype fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One API contract now proves BF16 output dtype, shape, and device before rejecting a non-FP32 input.", - "Unbiased expectation, neighbor bounds, and generator determinism remain independent numerical properties." - ] - }, - { - "id": "TA-196", - "scope": "MoE fused gate-up registration and deferred QLoRA skip fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Base and LoRA expert modules now prove the same fused parameter registration and gate/up views in one structure contract.", - "Qwen3 and Qwen3.5 handlers now prove their family-specific deferred expert keys through one QLoRA skip policy." - ] - }, - { - "id": "TA-197", - "scope": "batch-invariant GDN gating and gated-norm forward/backward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Each fused primitive now compares forward values and every input gradient with its independent PyTorch composition from the same graph.", - "Gated norm row invariance remains in the numerical transaction; model routing and pinned solve geometry remain separate." - ] - }, - { - "id": "TA-198", - "scope": "Nemotron-H forward, router-output, raw-backward, and labeled-loss fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One model transaction now proves output shape, optional router logits, labeled CausalLM loss, and gradients through every mixer family.", - "Gradient checkpointing and the two packed-sequence contracts remain independent execution modes." - ] - }, - { - "id": "TA-199", - "scope": "exact GLM dense and LM-head legacy weight-sync guard fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense MLP and attention projections now reject adapter preparation, collective merging, and raw extraction through one factor-only policy matrix.", - "LM-head ordinary and prepacked sparse-delta publication both fail before adapter or backend work in one side-effect guard." - ] - }, - { - "id": "TA-200", - "scope": "merged block-FP8 deferred-loader projection key reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "QKV and gate/up merged projections still derive and validate every exact source key inside one key-selection policy.", - "EP expert slicing, missing pairs, retained caches, and per-module release remain independent." - ] - }, - { - "id": "TA-201", - "scope": "trainer trunk-linear engagement and numerical-family selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact server, ordinary server, and exact non-server models now share one pre-FSDP trunk engagement policy with explicit harness-state isolation.", - "Exact GLM family selection and ordinary rollback execute as one structural numerical-program transition." - ] - }, - { - "id": "TA-202", - "scope": "P2P async transfer size-threshold examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One transfer policy now sends a large bucket asynchronously, a medium bucket synchronously, and the same medium bucket asynchronously after lowering the configured cutoff.", - "Status timeout and sender preparation timeout remain separate failure and HTTP boundaries." - ] - }, - { - "id": "TA-203", - "scope": "DeepSeek-V4 FWHT known-pattern, roundtrip, and norm examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One numerical matrix now proves a known Hadamard row, orthonormal roundtrip, and norm preservation across power-of-two widths.", - "Invalid width and rotate_activation fallback dispatch remain independent API boundaries." - ] - }, - { - "id": "TA-204", - "scope": "orchestrator-runner generic serialization examples and message metadata fragments", - "decision": "remove", - "status": "applied", - "evidence": [ - "All typed message classes, tensor payloads, JSON conversion, and pickle rejection now share one wire-format contract.", - "Generic large nested-list and nested-dictionary serializer examples were removed; identity, timestamps, optional fields, and ACK construction remain production-specific." - ] - }, - { - "id": "TA-205", - "scope": "DeepSeek-V3 auxiliary router-logit layer-count fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One auxiliary-loss policy now proves all-MoE emission and omission of the configured dense prefix.", - "Base forward/backward, LoRA targeting, and routing replay remain independent consumers." - ] - }, - { - "id": "TA-206", - "scope": "FlashMLA device and production-shape admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One fail-closed admission contract now rejects CPU dispatch, an unproven head shape, and an overflowing flattened KV address space before backend import.", - "Flattening, compacted backward, and all-invalid rows remain independent numerical paths." - ] - }, - { - "id": "TA-207", - "scope": "RL primitive KL estimator mode parameterization", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "K1, K2, K3, and low-variance KL modes still compare against the independent Slime formulas inside one estimator matrix.", - "Sequence KL, clipping, OPSM, and reduction semantics remain independent primitives." - ] - }, - { - "id": "TA-208", - "scope": "families-v2 exact and nonexact selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One environment policy now proves legacy rollback for nonexact models and structural version pinning for exact GLM and Qwen paths.", - "Vendoring, selected-logit bytes, and batch composition invariance remain separate contracts." - ] - }, - { - "id": "TA-209", - "scope": "sqrt-softplus routing regather value, scaling, and dtype fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One replay-regather transaction now matches eager weights, isolates the routed scaling factor, preserves cached experts, and returns the requested BF16 dtype.", - "The unchanged softmax route remains an independent regression contract." - ] - }, - { - "id": "TA-210", - "scope": "tensor collator scalar, dtype, string, and input-container fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "List, NumPy, tensor, boolean, scalar, empty, known-field dtype, and string cases now execute through one conversion policy.", - "Variable-length and packed-sequence handling remains a separate layout outcome." - ] - }, - { - "id": "TA-211", - "scope": "model-runner session registry and optimizer option fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One LoRA session lifecycle now synchronizes registry state after both optimizer update and checkpoint load.", - "One dense optimizer policy now proves DistSignSGD normalization and clipping before the optional empty-cache suppression branch." - ] - }, - { - "id": "TA-212", - "scope": "runner-dispatcher load-state preparation and tenant-routing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Preparation now preserves rank-zero errors while accepting paths under the server output root and rejecting unrelated roots in one policy.", - "Multi-adapter coordinator delegation and single-tenant trainer delegation share one load-state routing transaction." - ] - }, - { - "id": "TA-213", - "scope": "permanently skipped GLM5 FLOP feature placeholders", - "decision": "remove", - "status": "applied", - "evidence": [ - "The two tests were unconditionally skipped because GLM5 sparse-MLA and DSA FLOP accounting is not implemented and currently reports zero.", - "The executable CP-size invariance regression remains; feature coverage should be added with the implementation rather than as dead placeholders." - ] - }, - { - "id": "TA-214", - "scope": "Kimi wrapper LoRA target-source precedence fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One target-resolution policy now proves wrapper-derived defaults, explicit target override, and strict-manifest override in precedence order.", - "Every original target list still executes against an actual wrapper config." - ] - }, - { - "id": "TA-215", - "scope": "test-local MoE expert auto-merge implementations", - "decision": "remove", - "status": "applied", - "evidence": [ - "Both files copied the parser, buffer, format detector, and simulated loader into tests and invoked no XoRL implementation.", - "Production ExpertWeightBuffer transposition/output remains covered directly by test_moe_gkn_format, while family checkpoint handlers cover real dense, packed, EP-sliced, and roundtrip loads." - ] - }, - { - "id": "TA-216", - "scope": "parallel-state singleton initialization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One singleton lifecycle now proves publication, automatic DP shard inference, device selection, and reinitialization rejection.", - "Mesh layout, construction validation, and requires-mesh behavior remain independent contracts." - ] - }, - { - "id": "TA-217", - "scope": "DeepSeek-V3 Kimi wrapper config mapping fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Wrapper unwrapping and official aux_loss_alpha defaults now execute through one configuration-conversion policy.", - "Registry lookup and local-directory auto-config loading remain separate integration boundaries." - ] - }, - { - "id": "TA-218", - "scope": "tokenizer and processor remote-code fallback fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Tokenizer and processor fallback now share one auto-loader security policy that forbids implicit remote code and retains right padding.", - "The local Kimi TikToken implementation remains an independent functional roundtrip." - ] - }, - { - "id": "TA-219", - "scope": "DeepSeek-V4 LoRA structural smokes and unsupported generic expert claim", - "decision": "remove", - "status": "applied", - "evidence": [ - "Attention adapter type/freeze assertions now live in the retained forward-backward gradient transaction instead of standalone structure tests.", - "The generic expert-LoRA claim was stale: DeepSeek-V4 expert semantics are deliberately rejected by the shared semantic guard, so the retained end-to-end path targets supported attention adapters only." - ] - }, - { - "id": "TA-220", - "scope": "scheduler FIFO and ScheduledRequest implementation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Defaults and FIFO add, dispatch, remove, and clear behavior now form one policy lifecycle.", - "A sleep-based direct ScheduledRequest state smoke was removed because scheduler completion/failure/abort transactions already exercise those transitions." - ] - }, - { - "id": "TA-221", - "scope": "Mooncake side-payload missing-key fragment", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Successful typed tensor roundtrips and missing-key rejection now share one store contract.", - "Reference slicing/cleanup and R3 payload validation remain distinct higher-level lifecycles." - ] - }, - { - "id": "TA-222", - "scope": "batch-utils float conversion and ragged-padding fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Scalar logprob fields, rectangular teacher states, and ragged teacher-state padding now execute through one conversion transaction.", - "Sequence-parallel sharding remains an independent consumer boundary." - ] - }, - { - "id": "TA-223", - "scope": "DR-GRPO runner dispatch and option fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Modern and legacy loss fields, KL metrics, temperature, output suppression, and K3-forced output now share one microbatch-loss dispatch contract.", - "Sampler prefill propagation and the full forward-backward loop remain separate integrations." - ] - }, - { - "id": "TA-224", - "scope": "test-local GKN transpose and reference-MoE proof", - "decision": "remove", - "status": "applied", - "evidence": [ - "The removed report compared local matrix multiplications and a local expert loop without calling XoRL.", - "The retained ExpertWeightBuffer test proves production transposition and output, and the live backend matrix compares eager, native, Triton, and MoEBlock consumers." - ] - }, - { - "id": "TA-225", - "scope": "future sparse-MLA KV-major reference scaffold", - "decision": "remove", - "status": "applied", - "evidence": [ - "The test compared two Torch references for a future atomic-free kernel and invoked no production sparse-MLA path.", - "Production forward, backward, deterministic, and combined-kernel performance contracts remain collected." - ] - }, - { - "id": "TA-226", - "scope": "EP gradient backend support and rejection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every supported backend and the unknown-backend fail-closed outcome now form one reduction-domain table contract.", - "Malformed metadata and the real two-rank reduction lifecycle remain independent." - ] - }, - { - "id": "TA-227", - "scope": "lm-head tensor-parallel topology wrappers", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CP-sourced, DP-sourced, and HSDP-sourced four-rank meshes now execute as one topology matrix rather than three wrapper reports.", - "Each subprocess still verifies TP groups, replica groups, FSDP groups, and topology-specific mesh structure." - ] - }, - { - "id": "TA-228", - "scope": "LoRA cast-once zero and nonzero merge fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Each linear and expert merge transaction now proves both bit-exact zero-adapter behavior and the nonzero FP32-add-then-cast reference across supported dtypes.", - "Linear and grouped-expert implementations remain separate reports." - ] - }, - { - "id": "TA-229", - "scope": "NF4 per-width, size, dtype, and allocation smokes", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Twenty reports now form three codebook, flat-codec, and GKN-codec contracts while still running every 32/64/128 group width, layout, dtype, scale, accuracy, zero, and cross-layout check.", - "Two large-allocation smokes were removed because they repeated the same codec behavior with 16M-element and 14M-element tensors without a distinct production boundary." - ] - }, - { - "id": "TA-230", - "scope": "TopKRouter selector and configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Softmax, balanced, hash, sqrt-softplus, scaling, tie, invalid-input, and model-config outcomes now execute as selector policies rather than isolated examples.", - "FP32 selection and the MoEBlock consumer remain independent because they protect precision and integration boundaries." - ] - }, - { - "id": "TA-231", - "scope": "OPD estimator, policy-gradient, and task-weight fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full-vocabulary modes, estimator formulas, dispatch, policy-gradient admission, and task weighting now report their complete policies.", - "Clamping, stable metric keys, and ignored-label sampled-logprob behavior remain distinct numerical or API boundaries." - ] - }, - { - "id": "TA-232", - "scope": "loss reducer denominator and layout fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "TokenPartial now proves denominator and microbatch composition together, while SequencePartial covers dense, packed, and context-parallel composition as one layout policy.", - "Empty-input zero behavior remains a separate boundary." - ] - }, - { - "id": "TA-233", - "scope": "shared loss implementation and microbatch parameter reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy TokenPartial identities and paired implementation checks now run explicit case matrices within one contract per semantic behavior.", - "Importance-sampling and policy-loss microbatch composition likewise report once while retaining every loss variant." - ] - }, - { - "id": "TA-234", - "scope": "Qwen Class-B RoPE and EP adapter parameter reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense and MoE half-rotate/cast behavior, supported CPU dtypes, Class-B shapes, and available EP adapters now execute internal matrices instead of publishing one report per equivalent case.", - "All numerical cases and every installed adapter backend are still exercised." - ] - }, - { - "id": "TA-235", - "scope": "shared-prefix attention dtype and head-shape parameter reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The two dtypes, two head dimensions, and two query-to-KV head ratios now form one numerical backend contract.", - "No-sharing and one-token-prompt behavior remain separate backend outcomes." - ] - }, - { - "id": "TA-236", - "scope": "hardware-specific Qwen3-8B TFLOPS pytest thresholds", - "decision": "remove", - "status": "applied", - "evidence": [ - "The three reports hard-coded H100 throughput thresholds without admitting only H100 hardware and normally skipped based on local model-directory presence.", - "They were multi-minute benchmark jobs rather than stable correctness contracts; Qwen LoRA/FSDP E2E suites retain real training and loss-convergence coverage." - ] - }, - { - "id": "TA-237", - "scope": "direct pack_parallel non-empty smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report asserted only that two calls returned at least one bin and did not verify capacity, coverage, allocation, or ordering.", - "PackingDataset still invokes the production parallel packer for both sequential and multipack methods, while dedicated FFD, grouping, and allocation contracts retain the actual invariants." - ] - }, - { - "id": "TA-238", - "scope": "QLoRA package hasattr import smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "The deleted report imported xorl.qlora and checked only two exported attribute names.", - "Trainer and model-builder imports exercise the public exports, real QLoRA suites invoke the implementations, and the clean-interpreter dependency-cycle contract remains." - ] - }, - { - "id": "TA-239", - "scope": "default and explicit MLA target partition examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default and caller-specified MLA projections now execute as one MoE target-partition policy.", - "The real DeepSeek-V3 model integration remains separate and still checks attention, shared-expert, and routed-expert replacement." - ] - }, - { - "id": "TA-240", - "scope": "CP16 first-rank and padded-tail side-channel reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ranks 0 and 15 now run inside one CP16 target-sharding contract.", - "Both full and padded-tail slices still verify labels, target tokens, old logprobs, advantages, reference logprobs, and padding values." - ] - }, - { - "id": "TA-241", - "scope": "Muon optimizer EP checkpoint transition wrappers", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP2-to-EP4, EP4-to-EP2, and same-EP identity now form one checkpoint transition matrix.", - "Each case still launches its own four-rank save, reload, global gather, and exact momentum comparison." - ] - }, - { - "id": "TA-242", - "scope": "DTensor copy and save-materialization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Replicated, sharded, padded-tail, and shape-rejection outcomes now form one copy policy.", - "One-dimensional writer-only and two-dimensional all-rank/optional-writer materialization still launch separate four-rank workers within one topology contract." - ] - }, - { - "id": "TA-243", - "scope": "PP NCCL sender, empty-buffer, and product-helper fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Empty sender and receiver outcomes now share one protocol transaction, and nonempty sender return behavior is asserted where metadata and flattening are verified.", - "The direct private product-helper example was replaced by scalar-tensor reconstruction through the production receive path." - ] - }, - { - "id": "TA-244", - "scope": "cautious-decay helper and denominator-mode fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Zero decay, ordinary decoupled decay, and coordinate masking now form one helper policy.", - "Chunked denominator parity still runs both ordinary and Kahan modes with three optimizer steps and complete state comparison." - ] - }, - { - "id": "TA-245", - "scope": "architecture registration, AutoConfig, and builder fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "DeepSeek-V4 now follows one standard snapshot through AutoConfig, AutoModel registration, and the XoRL foundation-model builder.", - "DeepSeek-V3, Nemotron-H, and Qwen3.5 registry assertions now live with their real local/HF configuration conversions; numerical model and checkpoint contracts remain separate." - ] - }, - { - "id": "TA-246", - "scope": "direct ModelArguments routing-weight default assertion", - "decision": "remove", - "status": "applied", - "evidence": [ - "The deleted report instantiated a dataclass and asserted only that one field equaled auto.", - "Server YAML loading still proves the omitted default and explicit value serialize correctly, while trainer alignment and runtime resolution tests retain the real consumers." - ] - }, - { - "id": "TA-247", - "scope": "gradient-checkpoint default and override method fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default class policy, default enablement, and nondefault propagation across dense and MoE checkpoint layers now form one configuration contract.", - "The independent training/flag/method execution gate remains separate." - ] - }, - { - "id": "TA-248", - "scope": "single-part and multi-part optimizer construction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Singleton plain-optimizer selection and multi-part mapping/parameter coverage now execute as one construction policy.", - "Live multi-part step/scheduler behavior and invalid custom groups remain separate outcomes." - ] - }, - { - "id": "TA-249", - "scope": "token-diagnostic disabled, ignored-label, and top-k boundary fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Disabled inputs, all-ignored labels, and top-k larger than vocabulary now form one boundary policy.", - "Ranking, KL position mapping, loss-logprob comparison, raw-weight reference, and hidden-state summaries remain independent numerical contracts." - ] - }, - { - "id": "TA-250", - "scope": "DeepSeek-V4 private mapper, APE, FP8, and MXFP4 fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "APE valid shapes and invalid shape, FP8 full-block and tail dequantization, and MXFP4 values and block scaling now report complete codec policies.", - "Unknown-key handling moved from the private name mapper to the production checkpoint handler and now also verifies unmapped accounting." - ] - }, - { - "id": "TA-251", - "scope": "DeepEP asynchronous-combine default and opt-in fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default synchronous behavior and the explicit unsafe environment opt-in now execute as one admission policy.", - "Both cases still pass through tokens_post_combine and inspect the actual fused-operation argument." - ] - }, - { - "id": "TA-252", - "scope": "FSDP reduce-dtype, boolean coercion, and prefetch setting fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Singleton, sharded, and overridden expert precision plus supported and rejected dtype names now form one reduction policy; boolean spellings and rejection form one admission policy.", - "Backward-only, forward-only, bidirectional, and not-needed prefetch behavior now execute as one direction matrix with exact module ordering." - ] - }, - { - "id": "TA-253", - "scope": "per-component timer disabled and unrecorded-event fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Disabled lifecycle no-op behavior and enabled handling of recorded versus unrecorded event pairs now share one timer admission contract.", - "Live model-style CUDA hook coverage remains a separate integration." - ] - }, - { - "id": "TA-254", - "scope": "R3 payload transport success and directory rejection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Mooncake, legacy externalization, filesystem directory, and inline defaults now execute as one transport-admission matrix.", - "The missing-directory rejection remains in the same policy and still reaches the production argument loader." - ] - }, - { - "id": "TA-255", - "scope": "exact GLM rank-1 topology admission and rejection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The certified TP16 topology and its non-TP16 rejection now form one exact-GLM topology contract.", - "The accepted values and rejection message are both retained." - ] - }, - { - "id": "TA-256", - "scope": "full-weight QARL and FP8 incompatible scope and source fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "LoRA conflicts, mutual QARL and FP8 exclusion, missing calibration, MTP metadata, Mamba config, and Nemo ModelOpt sources now execute as one fail-closed policy.", - "Every prior YAML shape and error boundary is preserved." - ] - }, - { - "id": "TA-257", - "scope": "FP8 configuration alias entry-point fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Nested train fp8_cfg and Nemo policy megatron_cfg aliases now share one normalization contract.", - "Layer-island and Blackwell fields remain asserted on the train alias path." - ] - }, - { - "id": "TA-258", - "scope": "server multi-adapter unsupported-mode fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Merge-interval and pipeline-parallel rejection now execute as one unsupported multi-adapter policy.", - "The general broadcast load-weights rejection remains separate from adapter-specific admission." - ] - }, - { - "id": "TA-259", - "scope": "weight-sync FP8 normalization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit FP8 normalization, default dynamic-block behavior, and module-name cleanup now form one normalization policy.", - "No-quantization aliases remain a separate BF16 no-op contract." - ] - }, - { - "id": "TA-260", - "scope": "weight-sync FP8 invalid configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unsupported formats, activation schemes, UE8M0 scales, module exclusions, and internal unsupported markers now execute as one rejection matrix.", - "Non-FP8 quantization methods retain a separate method-admission policy." - ] - }, - { - "id": "TA-261", - "scope": "training CLI FP8 configuration alias fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Nested train fp8_cfg and Nemo policy megatron_cfg aliases now execute through parse_args as one CLI normalization contract.", - "The layer-island and Blackwell fields remain asserted on the native train path." - ] - }, - { - "id": "TA-262", - "scope": "training CLI QARL and FP8 incompatible configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Adapter conflicts, missing QARL calibration data, QARL plus FP8, MTP metadata, and Mamba config now execute as one CLI rejection policy.", - "Every former YAML payload and exact error boundary remains exercised through parse_args." - ] - }, - { - "id": "TA-263", - "scope": "training simulator untrusted calibration and model path fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Built-in traversal, relative escape, missing default, symlink escape, and unapproved model metadata paths now form one filesystem admission policy.", - "All five attacks still reach the production calibration-pack or metadata loader." - ] - }, - { - "id": "TA-264", - "scope": "exact Qwen3.5 certified topology admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both certified MoE topologies, the certified dense topology, and ten nearby rejected mutations now execute as one topology policy.", - "The exact topology validator remains the observable boundary." - ] - }, - { - "id": "TA-265", - "scope": "exact Qwen3.5 model-scope admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense, MoE, and Hugging Face outer-config snapshots now share one model-scope policy with wrong layer types and nearby geometry rejection.", - "Canonical GLM model-scope validation remains an independent family contract." - ] - }, - { - "id": "TA-266", - "scope": "P2P completion and failed-transfer cleanup fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pending-transfer failure, receiver-completion suppression, best-effort drain, successful completion payload, failed completion cleanup, and default cache behavior now form two lifecycle policies.", - "All network mocks, engine deregistration checks, endpoint metadata, and failure messages are retained." - ] - }, - { - "id": "TA-267", - "scope": "session endpoint LoRA registration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full LoRA overrides, rank-only defaults, and existing-session refresh now execute as one registration policy.", - "Worker registration, normalized session specs, materialization, and refresh idempotence remain asserted." - ] - }, - { - "id": "TA-268", - "scope": "session endpoint reserved-checkpoint fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default, per-session, stale, and existing reserved checkpoints now execute as one persistence policy.", - "Save counts, model IDs, reserved paths, overwrite, and preservation behavior are retained." - ] - }, - { - "id": "TA-269", - "scope": "full-weight session endpoint admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default full-weight registration, nondefault multitenancy rejection, and per-session override rejection now form one admission policy.", - "The materialize-false worker contract and both rejection messages remain covered." - ] - }, - { - "id": "TA-270", - "scope": "LoRA session kill, checkpoint, and default-session protection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ordinary kill and re-registration, final-checkpoint URI conversion, default kill preservation, and default unload rejection now form two lifecycle policies.", - "Registry, future-store, worker request, checkpoint, and HTTP status outcomes remain asserted." - ] - }, - { - "id": "TA-271", - "scope": "weights-info endpoint mode and path fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "On-disk LoRA metadata, full-weight metadata, and path-escape rejection now execute as one endpoint policy.", - "The disk-over-memory authority check remains intact." - ] - }, - { - "id": "TA-272", - "scope": "request-processor routing payload cleanup fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Mooncake success cleanup, Mooncake backend-failure cleanup, and filesystem payload cleanup now form one transport lifecycle.", - "Payload existence during execution and removal after execution remain verified for every path." - ] - }, - { - "id": "TA-273", - "scope": "request-processor token diagnostic boundary fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed-sample splitting, empty diagnostics, and mismatched-field rejection now execute as one decoding policy.", - "Position rebasing and all diagnostic field alignments remain asserted." - ] - }, - { - "id": "TA-274", - "scope": "request-processor packed-row batching fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Global row grouping, rank-local deferral, and routed-replay rejection now form one batching policy.", - "Batch counts, sequence boundaries, sample counts, metrics, and error output remain covered." - ] - }, - { - "id": "TA-275", - "scope": "adapter gradient-epoch abort fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Successful idempotent abort and rejection while publication-pending or poisoned now execute as one lifecycle policy.", - "Gradient scratch reset, monotonic counters, and publication state remain asserted." - ] - }, - { - "id": "TA-276", - "scope": "authoritative adapter checkpoint plan admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Identity-label differences with equal direct contracts and direct-contract mismatches now execute as one admission policy.", - "Mismatch rejection still proves parameters, optimizer, session spec, step, and learning rate are unchanged." - ] - }, - { - "id": "TA-277", - "scope": "adapter checkpoint structure rejection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Target-module mismatch, missing LoRA tensors, and rank beyond live capacity now form one structural admission policy.", - "Each malformed checkpoint is still built and rejected through load_adapter_state." - ] - }, - { - "id": "TA-278", - "scope": "adapter PEFT filename and sharding compatibility fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Weight-suffixed tensor names and indexed sharded safetensors now execute as one PEFT compatibility policy.", - "Both restored LoRA factors retain exact value checks." - ] - }, - { - "id": "TA-279", - "scope": "adapter residency and eviction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dirty residency, clean-victim selection, auto-save failure, and multi-rank rejection now form one eviction policy.", - "Every accepted or retained adapter identity remains asserted after the transition." - ] - }, - { - "id": "TA-280", - "scope": "inference endpoint port-selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Single-port and explicit worker-port registration now execute as one endpoint port policy.", - "Health checks, LoRA load/unload routes, and adapter discovery remain asserted." - ] - }, - { - "id": "TA-281", - "scope": "inference endpoint FP8 KV-cache registration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit FP8 metadata, dtype inference with cache-version alias, and required-FP8 rejection now form one admission policy.", - "Postprocess, static-scale, epoch, and rejection details remain covered." - ] - }, - { - "id": "TA-282", - "scope": "inference weight-sync pool filtering fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Eval, default, all-pool, and no-match selection now execute as one endpoint-selection policy.", - "The no-match path still proves no orchestrator request is sent." - ] - }, - { - "id": "TA-283", - "scope": "inference weight-sync quantization admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit null quantization and malformed empty quantization now form one request-admission policy.", - "Default suppression and HTTP 400 behavior remain asserted." - ] - }, - { - "id": "TA-284", - "scope": "inference weight-sync cache invalidation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP8 auto-flush, cache-version aliasing, BF16 no-flush, and explicit none mode now form one cache policy.", - "Payload flags, response flags, endpoint epochs, postprocess, and static-scale outcomes remain checked." - ] - }, - { - "id": "TA-285", - "scope": "receiver quantization detection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Skip-list normalization and unsupported MTP, activation, and UE8M0 receiver configurations now execute as one detection policy.", - "Every detected configuration still passes through the public normalizer." - ] - }, - { - "id": "TA-286", - "scope": "receiver quantization enrichment fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Receiver skip-list fill, user override preservation, BF16 and null pass-through, and unsupported-reason propagation now form one enrichment policy.", - "Unsupported enriched output remains rejected by normalization." - ] - }, - { - "id": "TA-287", - "scope": "GLM5 local config unsafe-value fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Non-JSON object values and dunder keys now execute as one local-config security policy.", - "Both exact validation paths and messages remain exercised." - ] - }, - { - "id": "TA-288", - "scope": "GLM5 blocked indexer selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full-query blocked-versus-dense parity and query-offset shard parity now form one indexer policy.", - "The same production select_topk path remains used for both shapes." - ] - }, - { - "id": "TA-289", - "scope": "GLM5 sparse-MLA torch reference fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full-top-k dense equivalence and local-query offset equivalence now execute as one numerical reference policy.", - "Both outputs retain exact tolerance checks." - ] - }, - { - "id": "TA-290", - "scope": "GLM5 sparse-MLA dispatch fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CPU auto fallback and unknown-backend rejection now form one backend dispatch policy.", - "The production dispatcher remains the observable boundary." - ] - }, - { - "id": "TA-291", - "scope": "adapter optimizer parameter identity fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical parameter ordering and wrapper-insensitive structure fingerprints now execute as one identity policy.", - "Both public checkpoint identity inputs remain directly asserted." - ] - }, - { - "id": "TA-292", - "scope": "adapter optimizer checkpoint artifact fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy pickle, shard-without-manifest, and declared-state-without-artifacts now form one checkpoint admission policy.", - "Single and multi-rank refusal, weights-only migration, and no-partial-registration outcomes remain covered." - ] - }, - { - "id": "TA-293", - "scope": "adapter optimizer logical reshard fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One-dimensional world-size change, multidimensional disjoint rectangles, replicated slicing, and same-world layout change now form one reshard policy.", - "Moments, squared moments, and optimizer step retain exact reconstruction checks." - ] - }, - { - "id": "TA-294", - "scope": "adapter optimizer invalid reshard source fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "World-size mismatch, divergent replicas, holes, overlap, dtype, shape, step, empty ranks, staged defects, and parameter fingerprints now form one fail-closed policy.", - "Resident optimizer no-mutation checks remain active for every defect that reaches staging." - ] - }, - { - "id": "TA-295", - "scope": "weight-sync receiver postprocess fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Environment precedence, FP8 KV-cache requirements, streaming backend flags, BF16 omission, and quantization-method detection now form one postprocess policy.", - "Both handler decisions and emitted backend configuration remain asserted." - ] - }, - { - "id": "TA-296", - "scope": "P2P direct-EP sender and tensor-collection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Replica sender selection, sender-to-EP ownership, and explicit non-sender collection suppression now form one direct-EP policy.", - "Default, round-robin, direct, and NCCL cases remain covered." - ] - }, - { - "id": "TA-297", - "scope": "weight-sync tied-parameter extraction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Declared identical ties, declared unequal parameters, and shared storage without declaration now form one alias policy.", - "Emitted buffers and tied-weight alias maps remain exact." - ] - }, - { - "id": "TA-298", - "scope": "Nemotron-H inference-unfuse fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-expert HF layout emission and fused-dense rejection now execute as one Nemotron-H conversion policy.", - "The production checkpoint handler remains the layout oracle." - ] - }, - { - "id": "TA-299", - "scope": "EP MoE gated and non-gated collection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Non-gated two-projection and gated three-projection expert collection now form one gating policy.", - "Projection names, shapes, splits, and values remain asserted." - ] - }, - { - "id": "TA-300", - "scope": "compiled-module wrapper name normalization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "MoE unfuse, backend broadcast, and Qwen linear-attention fusion now share one _orig_mod normalization policy.", - "Each final receiver namespace and transformed tensor remains checked." - ] - }, - { - "id": "TA-301", - "scope": "sparse-delta weight-sync fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fast-path configuration, cache metadata, prepacked-only rejection, and CPU FP8 targeting now form one sparse-delta policy.", - "Health, pause, post, resume, endpoint, and cache ordering remain covered." - ] - }, - { - "id": "TA-302", - "scope": "FP8 training linear replacement fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default replacement, parameter identity, output dtype, state dict, and fully qualified module tags now form one injection policy.", - "All replaced module types and names remain asserted." - ] - }, - { - "id": "TA-303", - "scope": "FP8 training linear recipe fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Global amax scales, FQN recipe overrides, and unknown override rejection now form one recipe policy.", - "Block size, SmoothQuant, correction, and output dtype propagation remain covered." - ] - }, - { - "id": "TA-304", - "scope": "FP8 training linear exclusion fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit module names and FQN globs now execute as one BF16-island exclusion policy.", - "Replacement counts and every retained or converted module remain asserted." - ] - }, - { - "id": "TA-305", - "scope": "FP8 linear CPU fallback fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Numerical fallback, float32 output, and fail-fast mode now form one CPU admission policy.", - "Production FP8Linear calls remain the tested boundary." - ] - }, - { - "id": "TA-306", - "scope": "FP8 linear error-profiler sampling fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Call caps, explicit flattened rows, and module-specific call-row selectors now form one CPU sampling policy.", - "The independent live-CUDA operand-breakdown gate remains separate." - ] - }, - { - "id": "TA-307", - "scope": "block-FP8 GEMM backend and scale-layout fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Block scales, rowwise weight scales, and torch scaled-mm backend parity now form one CUDA GEMM policy.", - "Both 64 and 128 block widths and explicit dequantized references remain." - ] - }, - { - "id": "TA-308", - "scope": "offline FP8 weight quantization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Weight-sync parity and partial-block zero-padding now form one quantization contract.", - "Exact FP8 bytes and scale tensors remain compared." - ] - }, - { - "id": "TA-309", - "scope": "offline fused-QKV export fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Valid QKV splitting, missing metadata rejection, and duplicate output rejection now form one export policy.", - "All Q, K, and V bytes and scales remain checked." - ] - }, - { - "id": "TA-310", - "scope": "offline MLA-A projection export fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Configured q_lora_rank fusion and metadata-absent split preservation now form one MLA export policy.", - "Fused and independent receiver names remain asserted." - ] - }, - { - "id": "TA-311", - "scope": "offline linear-attention name conversion fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Linear-attention metadata-driven fusion and metadata-absent split preservation now form one name policy.", - "All fused projections, convolution, norm, bias, and A_log tensors remain covered." - ] - }, - { - "id": "TA-312", - "scope": "offline QARL fold admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Selective QARL folding and mismatched block-size rejection now form one fold-admission policy.", - "The independent trained-logprob preservation gate remains separate." - ] - }, - { - "id": "TA-313", - "scope": "packing strategy and oversized-sample admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unknown settings, default error, legacy skip, aligned truncation, and HF-shift truncation now form one admission policy.", - "Every former sample payload and rejection remains exercised." - ] - }, - { - "id": "TA-314", - "scope": "cross-strategy packing correctness and utilization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Document multiset, valid tokens, position resets, capacity, and best-fit row count now form one strategy-invariant policy.", - "All sequential, best-fit, and balanced-DP data sets remain." - ] - }, - { - "id": "TA-315", - "scope": "balanced-DP packing behavior fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Zero dummies, balanced bins, full-row utilization, small-sample fallback, and DP1 equivalence now form one balanced-DP policy.", - "Every load and utilization threshold remains asserted." - ] - }, - { - "id": "TA-316", - "scope": "packing determinism and datum-order fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Repeatability, sequential identity, and reordered permutation/build order now form one ordering policy.", - "Routed-expert realignment evidence remains exact." - ] - }, - { - "id": "TA-317", - "scope": "packed token and teacher metadata fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "OPD alignment, HF shifts, hidden-state padding, RL padding, OPRD cache views, cache-base fallback, and nested RL schema now form one metadata policy.", - "Every token field, vector field, cache index, and ignore-index result remains asserted." - ] - }, - { - "id": "TA-318", - "scope": "disabled packing behavior fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-sample output, shifted targets, HF warning, explicit target preservation, and loss masking now form one disabled-mode policy.", - "Both flat and nested input conventions remain covered elsewhere in the same suite." - ] - }, - { - "id": "TA-319", - "scope": "Muon Gram-Newton-Schulz grouping fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Equal shapes, flattened matrix shapes, transpose-equivalent shapes, fused gate-up halves, and byte-limit chunking now form one grouping policy.", - "Every orthogonalizer input shape and exact parameter update remains asserted." - ] - }, - { - "id": "TA-320", - "scope": "Muon fused gate-up classification fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Gated versus non-gated experts, post-FSDP parameter replacement, and DeepSeek versus Nemotron model classification now form one fused-split policy.", - "Parameter identity and optimizer-group membership remain checked." - ] - }, - { - "id": "TA-321", - "scope": "FP8 MoE injection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Weight preservation, per-expert bias preservation, backend enablement, and unused-module summary now form one injection policy.", - "The model parameters remain identical objects after conversion." - ] - }, - { - "id": "TA-322", - "scope": "FP8 grouped same-NK forward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Block-loop reference, Triton shapes, nondefault block width, and precomputed sequence offsets now form one CUDA forward policy.", - "Every BF16 reference and tolerance remains unchanged." - ] - }, - { - "id": "TA-323", - "scope": "FP8 grouped same-MN weight-gradient fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Block-loop reference, scalar dispatch, precomputed sequence offsets, and Triton shape matrix now form one CUDA weight-gradient policy.", - "Dispatch spying and all BF16 comparisons remain active." - ] - }, - { - "id": "TA-324", - "scope": "FP8 scalar-Quack grouped forward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "BF16 fallback parity and per-expert scale selection through explicit sequence offsets now form one scalar-Quack policy.", - "Both numerical outputs remain checked on CUDA." - ] - }, - { - "id": "TA-325", - "scope": "FP8 MoE expert training-step fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Triton and scalar backends plus biased clamped-SwiGLU experts now form one expert training policy.", - "Outputs, gradients, master weights, and expert biases retain finite/update checks." - ] - }, - { - "id": "TA-326", - "scope": "canonical merged-LoRA fold fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pinned arithmetic, shared-factor expansion, expert sharding, linear orientation, and zero-delta identity now form one fold policy.", - "All exact tensor comparisons and dtypes remain asserted." - ] - }, - { - "id": "TA-327", - "scope": "folded-weight straight-through gradient fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shared-factor GKN cases, fused gate-up, and linear orientation now form one straight-through gradient policy.", - "Reference autograd comparisons remain unchanged; FSDP gradient dtype stays independent." - ] - }, - { - "id": "TA-328", - "scope": "LoraLinear merged-forward selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact merged bytes, legacy selection, module isolation, and gradient parity now form one selection policy.", - "Step and runtime cache invalidation remains an independent lifecycle report." - ] - }, - { - "id": "TA-329", - "scope": "MoE merged-weight and cache fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical gate-up/down folds and parameter-version cache identity now form one merged-weight policy.", - "Exact values, object reuse, and invalidation remain checked." - ] - }, - { - "id": "TA-330", - "scope": "fused-expert merged-LoRA admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Automatic support and fail-closed unmerged invocation now form one fused-expert admission policy.", - "Both accepted and rejected modes remain exercised." - ] - }, - { - "id": "TA-331", - "scope": "native-EP merged-LoRA routing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Gradient-bearing masked routing and no-grad canonical-fold filtering now form one native-EP policy.", - "Keyword routing, local expert IDs, folded weights, and output identity remain asserted." - ] - }, - { - "id": "TA-332", - "scope": "merged-LoRA trunk-wrap composition fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fail-closed wrapping and enabled composition now form one trunk-wrap policy.", - "The exact contract and wrapper state remain checked." - ] - }, - { - "id": "TA-333", - "scope": "API session creation and activity fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Usable session IDs, deterministic heartbeat refresh, and canonical LoRA storage now form one lifecycle policy.", - "Follow-up save behavior and session registry state remain asserted." - ] - }, - { - "id": "TA-334", - "scope": "Tinker weights-info compatibility fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy flat rank serialization and create-model dictionary storage now form one weights-info compatibility policy.", - "Both direct and create-model paths retain their response assertions." - ] - }, - { - "id": "TA-335", - "scope": "API create-model worker registration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full-weight empty overrides and LoRA worker registration now form one create-model policy.", - "Materialization payloads, worker session specs, and stored optimizer configuration remain checked." - ] - }, - { - "id": "TA-336", - "scope": "API optimizer payload and learning-rate fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Native and Tinker payloads plus request, session, server, create-model, and missing-default LR resolution now form one optimizer policy.", - "All optimizer fields, metrics, defaults, and failure behavior remain asserted." - ] - }, - { - "id": "TA-337", - "scope": "training-simulator topology and shape-accounting fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Balanced routing, DP topology resolution, and sequence-parallel shape accounting now form one topology policy.", - "Every count, batch size, local token, and routed-slot assertion remains." - ] - }, - { - "id": "TA-338", - "scope": "observed benchmark ingestion and planning fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Structured log parsing, resolved fit/OOM ingestion, and tight-margin scenario planning now form one observed-data policy.", - "The same fixture now flows from logs through calibrated feasibility without duplicate setup." - ] - }, - { - "id": "TA-339", - "scope": "training configuration and model-metadata fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Configuration fingerprints, HF-cache metadata, and known-model fallback now form one metadata policy.", - "Hashes, topology fields, parsed architecture fields, and source labels remain checked." - ] - }, - { - "id": "TA-340", - "scope": "Qwen235 calibration ingestion and evaluation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Markdown fit/OOM extraction and leave-one-out GA evaluation now form one calibration policy.", - "Measured rows, labels, throughput, topology, errors, and OOM status remain asserted." - ] - }, - { - "id": "TA-341", - "scope": "Qwen235 calibrated scenario fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Observed GA selection, asymptotic extrapolation, and exact OOM-boundary rejection now form one calibrated-scenario policy.", - "Raw and risk-adjusted ranking, memory basis, remeasurement flags, and infeasibility remain checked." - ] - }, - { - "id": "TA-342", - "scope": "Qwen235 topology what-if fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "TP extrapolation, automatic parallelism sweep, and long-context CP admission now form one topology what-if policy.", - "Candidate spaces, conservative penalties, calibration scope, and OOM risk flags remain asserted." - ] - }, - { - "id": "TA-343", - "scope": "built-in simulator calibration-pack fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Qwen3.6 report matching and Qwen235 fit/OOM replay now form one built-in pack policy.", - "Correctness gating, support status, timing coverage, accuracy, and recall remain checked." - ] - }, - { - "id": "TA-344", - "scope": "portable analytical-ledger fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Portable FLOP/activation/communication coverage, dense MLP activations, and cross-node expert-FSDP normalization now form one ledger policy.", - "Exact status labels, byte terms, activation size, pass normalization, and positive totals remain asserted." - ] - }, - { - "id": "TA-345", - "scope": "teacher-head loading and storage fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct safetensors, tied embeddings, sharded stores, and cross-shard row views now form one head-persistence policy.", - "Full tensors, shard row counts, and sliced ranges remain compared exactly." - ] - }, - { - "id": "TA-346", - "scope": "teacher-head manager residency fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Teacher replacement, dtype replacement, and store prefetch now form one manager-residency policy.", - "Resident teacher identity, dtype, and loaded values remain asserted." - ] - }, - { - "id": "TA-347", - "scope": "teacher activation-cache selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-2 indexing, rank-3 token and layer selection, host/device paths, reuse, and dtype reload now form one selection policy.", - "Every source slice, result shape, dtype, and cache identity remains checked." - ] - }, - { - "id": "TA-348", - "scope": "teacher activation-cache async and admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Async prefetch success and negative/high index rejection now form one cache-admission policy.", - "Both failure messages and the prefetched output remain asserted." - ] - }, - { - "id": "TA-349", - "scope": "Mooncake tensor codec fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Byte round trips for four dtypes and canonical string aliases now form one tensor-codec policy.", - "Exact values, dtypes, accepted aliases, and unsupported dtype rejection remain." - ] - }, - { - "id": "TA-350", - "scope": "Mooncake hidden transport fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Metadata emission, rank-2 retrieval, and rank-3 layer retrieval now form one transport policy.", - "Storage key, schema fields, token counts, shapes, dtypes, and exact tensors remain checked." - ] - }, - { - "id": "TA-351", - "scope": "Mooncake teacher activation-consumer fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-2, rank-3, and multi-teacher cache consumption now form one integration policy.", - "Teacher routing, selected values, output shapes, and cache closure remain exercised." - ] - }, - { - "id": "TA-352", - "scope": "Mooncake metadata admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy entries, missing objects, size mismatches, and malformed metadata now form one fail-closed policy.", - "All original invalid payloads and error boundaries remain." - ] - }, - { - "id": "TA-353", - "scope": "Mooncake store lifecycle and configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Best-effort object removal and explicit-versus-environment configuration now form one store lifecycle policy.", - "The suffixed removal key and both configuration precedence outcomes remain asserted." - ] - }, - { - "id": "TA-354", - "scope": "trainer gradient-clipping fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Regular, DistSignSGD, disabled, and scale-factor cases now form one clipping policy.", - "Norms, gradient values, nonpositive thresholds, and rank-count factors remain checked." - ] - }, - { - "id": "TA-355", - "scope": "trainer metadata-counting fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Valid-token reduction, target-token precedence, batched voter reduction, and empty input now form one counting policy.", - "Reduction count, operation, group, device, active microbatches, and voter totals remain asserted." - ] - }, - { - "id": "TA-356", - "scope": "explicit trainer gradient-synchronization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default SP reduction, adapter-owned exclusions, lm-head exclusions, and optional DTensor skipping now form one synchronization policy.", - "Every selected tensor, group, and SUM operation remains checked." - ] - }, - { - "id": "TA-357", - "scope": "lm-head TP synchronization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Single-rank gradient suppression and multi-rank marked-parameter broadcast now form one lm-head TP policy.", - "No-op reduction, global source rank, group, and broadcast tensor remain asserted." - ] - }, - { - "id": "TA-358", - "scope": "checkpoint object-broadcast transport fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "NCCL tensor serialization, default-device selection, and weight-load group routing now form one transport policy.", - "Payload identity, CUDA device index, source rank, and group remain asserted." - ] - }, - { - "id": "TA-359", - "scope": "rank-zero checkpoint broadcast-loading fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Nonzero-rank shard resolution and leader-side handler-filtered prefetch now form one rank-zero loading policy.", - "Load calls, skipped and loaded keys, dispatch names, and batch metadata remain checked." - ] - }, - { - "id": "TA-360", - "scope": "checkpoint state-dict resolution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Remote rank-zero path resolution and node-local directory resolution now form one source policy.", - "Broadcast suppression and exact iterator paths remain asserted." - ] - }, - { - "id": "TA-361", - "scope": "grouped checkpoint expert-routing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense/expert fanout, HF fused experts, and FFN source formats now form one grouped-routing policy.", - "Handler inputs, prefetch partitions, converted names, transfers, and dispatch targets remain checked." - ] - }, - { - "id": "TA-362", - "scope": "grouped checkpoint group-fallback fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local dense loading plus permissive and strict missing-EP-group outcomes now form one fallback policy.", - "Collective suppression, rank-zero fallback, and fail-closed strict behavior remain asserted." - ] - }, - { - "id": "TA-363", - "scope": "strict checkpoint postprocessing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Missing, unexpected, duplicate, persistent-buffer, and complete-coverage outcomes now form one strict policy.", - "All diagnostic names and successful buffer dispatch remain checked." - ] - }, - { - "id": "TA-364", - "scope": "cautious decay primitive and SignSGD fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct zero, ordinary, and sign-masked decay plus SignSGD integration now form one primitive policy.", - "Aligned, misaligned, and zero-update coordinates retain exact expected values." - ] - }, - { - "id": "TA-365", - "scope": "AnyPrecisionAdamW cautious-decay fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy decoupled decay and cautious coordinate masking now form one AnyPrecision decay policy.", - "Both explicit first-step references remain unchanged." - ] - }, - { - "id": "TA-366", - "scope": "AnyPrecisionAdamW state-strategy fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Chunked denominator with ordinary/Kahan state and gradient-reuse CPU offload now form one state policy.", - "Parameters, moments, compensation, cleared gradients, and device placement remain checked." - ] - }, - { - "id": "TA-367", - "scope": "Muon cautious-decay fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ordinary decay, post-Newton-Schulz masking, and AdamW fallback masking now form one Muon policy.", - "All three explicit update references and tolerances remain." - ] - }, - { - "id": "TA-368", - "scope": "optimizer-builder cautious routing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Supported families, ordinary AdamW, unsupported SGD, kwarg rejection, and AnyPrecision options now form one builder policy.", - "Optimizer classes, group fields, accepted kwargs, and rejection messages remain asserted." - ] - }, - { - "id": "TA-369", - "scope": "GDN convolution forward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Qwen3.5 shapes, variable-length and batched layouts, and repeat determinism now form one CUDA forward policy.", - "All outputs remain bitwise compared to serving invocation or a repeated run." - ] - }, - { - "id": "TA-370", - "scope": "GDN convolution backward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Depthwise-autograd parity and repeated backward determinism now form one CUDA backward policy.", - "Input and convolution-weight gradients remain compared." - ] - }, - { - "id": "TA-371", - "scope": "end-to-end GDN block fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Eager output/gradient parity and exact-contract repeat determinism now form one end-to-end policy.", - "Every named parameter gradient and the bounded forward result remain checked." - ] - }, - { - "id": "TA-372", - "scope": "GDN packed-weight and armed-routing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Bitwise-neutral QKV weight packing and exact-contract dispatch now form one construction/routing policy.", - "All packed slices, invocation count, and output shape remain asserted." - ] - }, - { - "id": "TA-373", - "scope": "GDN exact-contract state-lifecycle fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact/ordinary module isolation and checkpoint recomputation now form one contract-state policy.", - "Thread-local state is checked during each call and after both ordinary and recomputed execution." - ] - }, - { - "id": "TA-374", - "scope": "checkpoint reference-state and QARL buffer fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ordinary persistent filtering, QARL buffer inclusion, and strict/non-strict QARL mismatch now form one buffer policy.", - "All parameter, buffer, shape, counter, metadata, and mismatch fields remain asserted." - ] - }, - { - "id": "TA-375", - "scope": "pipeline checkpoint key-contract fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Non-pipeline local keys, pipeline stage union, and metadata compatibility now form one key policy.", - "Collective suppression, union ordering, metadata counts, and compatibility results remain checked." - ] - }, - { - "id": "TA-376", - "scope": "pipeline LoRA checkpoint compatibility fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Base-to-LoRA and LoRA-only pipeline checkpoints now form one compatibility policy.", - "Load modes and exact missing LoRA/non-LoRA key sets remain asserted." - ] - }, - { - "id": "TA-377", - "scope": "distributed-checkpointer metadata admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Missing optimizer state and present optimizer metadata keys now form one load-admission policy.", - "Selected state entries, planner, reader, no-dist flag, and optimizer load keys remain checked." - ] - }, - { - "id": "TA-378", - "scope": "distributed-checkpointer load-group fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "No-dist pipeline validation, no-dist non-pipeline loading, and explicit DCP groups now form one load-group policy.", - "Validation and DCP group identities plus no-dist behavior remain asserted." - ] - }, - { - "id": "TA-379", - "scope": "distributed-checkpointer save-group fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Synchronous pipeline metadata reuse and asynchronous non-pipeline group isolation now form one save-group policy.", - "DCP and metadata process-group identities remain checked." - ] - }, - { - "id": "TA-380", - "scope": "optimizer checkpoint-state filtering fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Checkpoint load-key filtering and per-child multi-optimizer filtering now form one optimizer-state policy.", - "State keys, parameter groups, strict flags, and child optimizer assignments remain asserted." - ] - }, - { - "id": "TA-381", - "scope": "RoPE Class-B selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Serving provenance, selector reset, canonical GLM defaults/opt-out, and non-GLM opt-in now form one selection policy.", - "Resolved modes, global state, float32 tables, and rejection messages remain checked." - ] - }, - { - "id": "TA-382", - "scope": "canonical GLM numerical-program fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact program resolution, incompatible override and CE rejection, and non-GLM defaults now form one GLM policy.", - "Every resolved field, override case, CE mode, and default remains asserted." - ] - }, - { - "id": "TA-383", - "scope": "exact Qwen3.5 numerical-program fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense/MoE resolution, non-Qwen v2 rejection, incompatible override rejection, and CE rejection now form one Qwen policy.", - "All certified fields, override cases, and exact rejection messages remain." - ] - }, - { - "id": "TA-384", - "scope": "OPD metric aggregation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Namespace, weighted aggregation, zero-valid extrema, loss-group reduction, and empty-rank key seeding now form one metric policy.", - "All values, operations, groups, profile keys, and debug-key filtering remain checked." - ] - }, - { - "id": "TA-385", - "scope": "OPD packed cache and weight-shaping fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Valid-row splitting, OPRD tail weights, and packed-batch hidden chunks now form one packed-shaping policy.", - "Padding removal, position resets, indices, weights, and chunk boundaries remain asserted." - ] - }, - { - "id": "TA-386", - "scope": "OPD microbatch loss-execution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-teacher cache masking and FSDP lm-head anchoring now form one execution policy.", - "Loss finiteness, metrics, timings, anchor call shape, and student/hidden gradients remain checked." - ] - }, - { - "id": "TA-387", - "scope": "teacher hidden-cache contributor fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CP suppression, distinct EP slices, and legacy duplicate-EP mode now form one contributor policy.", - "All contributor keys and suppression outcomes remain asserted." - ] - }, - { - "id": "TA-388", - "scope": "teacher hidden-cache distributed assembly fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unified SP gathering, gathered-label trimming, and cross-rank batch assembly now form one distributed policy.", - "Groups, unpadding, saved rows, token counts, and per-sample cache indices remain checked." - ] - }, - { - "id": "TA-389", - "scope": "teacher hidden-cache Mooncake integration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Mooncake metadata/storage and downstream activation-cache indexing now form one producer-consumer policy.", - "Schema fields, token counts, stored bytes, and selected consumer rows remain asserted." - ] - }, - { - "id": "TA-390", - "scope": "OPD debug artifact fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Vocab-parallel loss contribution and packed teacher-segment JSONL now form one debug-artifact policy.", - "All local/group metrics, segment provenance, cache statistics, and component tensors remain checked." - ] - }, - { - "id": "TA-391", - "scope": "evicted-adapter auto-load fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Checkpoint restore, fresh broadcast, synchronized materialization failure, fresh-state rejection, and restore rollback now form one auto-load policy.", - "Registration/load calls, checkpoint paths, broadcasts, cross-rank errors, and rollback state remain checked." - ] - }, - { - "id": "TA-392", - "scope": "explicit adapter-state load and path-admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All-rank optimizer restore and output-root confinement for load, evicted lookup, and save now form one admission policy.", - "Payload fields, calls, success result, all rejection messages, and absence of escaped writes remain asserted." - ] - }, - { - "id": "TA-393", - "scope": "rank-zero adapter restore routing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Auto-load and explicit-load rank-zero broadcast modes now form one routing policy.", - "Registration, restore invocation, checkpoint path, and suppression of all-rank loads/broadcasts remain checked." - ] - }, - { - "id": "TA-394", - "scope": "rank-zero sharded adapter restore fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP shard slicing, session-spec mismatch, and topology-specific optimizer rejection now form one restore policy.", - "Local tensor bytes, steps, LR, transactional metadata, and unchanged state on both failures remain asserted." - ] - }, - { - "id": "TA-395", - "scope": "adapter-state load failure fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Cross-rank restore failure, pipeline-parallel rejection, and auto-registered session rollback now form one failure policy.", - "Exact errors and cleanup of adapters/session specs remain checked." - ] - }, - { - "id": "TA-396", - "scope": "adapter registration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Adapter/materialized-session success, cross-rank rollback, and worker registration failure now form one registration policy.", - "Session specs, LR calls, broadcasts, rollback, and exception text remain asserted." - ] - }, - { - "id": "TA-397", - "scope": "RMSNorm family-admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unknown families, residual misuse, forced family flips, unsupported fused-add, and zero-centered family misuse now form one admission policy.", - "Construction, module call, and funnel rejection boundaries remain." - ] - }, - { - "id": "TA-398", - "scope": "Qwen RMSNorm site-declaration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense Qwen decoder sites and shared-attention Q/K norms now form one declaration policy.", - "Layer-zero, later-layer, post-attention, Q, and K family assignments remain checked." - ] - }, - { - "id": "TA-399", - "scope": "RMSNorm family declaration-tripwire fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Undeclared warning, required-family failure, declared success, and legacy-explicit silence now form one CUDA tripwire policy.", - "Warning and error messages plus both admitted families remain exercised." - ] - }, - { - "id": "TA-400", - "scope": "RMSNorm family-funnel fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two-shape legacy parity, zero-centered folding, and family-difference vitality now form one bitwise CUDA funnel policy.", - "All exact outputs and the rare-difference bound remain asserted." - ] - }, - { - "id": "TA-401", - "scope": "RMSNorm family module-dispatch fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Trunk-contract dispatch/warning and declared-versus-legacy calls across three modes now form one module policy.", - "No-residual, residual-tree, and fused residual outputs remain bitwise checked." - ] - }, - { - "id": "TA-402", - "scope": "pipeline bubble formula fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Nonzero schedules, zero-bubble schedules, single-stage behavior, and invalid inputs now form one analytic policy.", - "Every formula value and rejection case remains." - ] - }, - { - "id": "TA-403", - "scope": "pipeline P2P byte-estimation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Middle/edge stages, forward-only, same-rank adjacency, and unpopulated metadata now form one P2P accounting policy.", - "All flow counts, microbatch multipliers, and the unknown result remain asserted." - ] - }, - { - "id": "TA-404", - "scope": "pipeline profiler patch-lifecycle fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Passthrough/restoration, double-patch rejection, single-stage discovery, and report-before-step failure now form one patch policy.", - "Stage calls, instance attributes, cleanup, and both rejection messages remain checked." - ] - }, - { - "id": "TA-405", - "scope": "optimizer, packing, and numerical argument fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "SignSGD/DistSignSGD controls and complete Muon kwargs now form one optimizer argument policy.", - "Packing, load mode, dtypes, numerical flags, and every Muon option remain asserted." - ] - }, - { - "id": "TA-406", - "scope": "checkpoint argument compatibility fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy EP/checkpoint aliases, automatic checkpoint resolution, and optimizer-resume defaults/overrides now form one policy.", - "Resolved checkpoint path, checkpoint method, EP placement, and all load_optimizer values remain checked." - ] - }, - { - "id": "TA-407", - "scope": "FP8 argument configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Optional overrides, native/Nemo aliases, fail-fast fallback, and unsupported vLLM knobs now form one FP8 policy.", - "All accepted fields and five receiver-side rejection boundaries remain." - ] - }, - { - "id": "TA-408", - "scope": "low-precision training argument fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "ModelOpt rejection, GLM block-FP8 QLoRA, QARL normalization, and the full low-precision conflict matrix now form one mode policy.", - "Accepted quantization fields plus LoRA, QLoRA, FP8, calibration, MTP, and Mamba rejections remain." - ] - }, - { - "id": "TA-409", - "scope": "pipeline FQN partitioning fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default/Qwen names, single/virtual stages, pinned endpoints, weighted splits, and infeasible layouts now form one partition policy.", - "All layer coverage, contiguity, counts, endpoints, and errors remain asserted." - ] - }, - { - "id": "TA-410", - "scope": "pipeline stage-placement fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Torch-reference mappings, exact ownership, single-style admission, and V-style endpoints now form one placement policy.", - "Loop, V, single-stage, and invalid virtual-stage cases remain." - ] - }, - { - "id": "TA-411", - "scope": "pipeline schedule metadata and admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Schedule style/splitting metadata and virtual-stage/microbatch validation now form one schedule policy.", - "All six schedules, unknown names, valid layouts, and invalid constraints remain checked." - ] - }, - { - "id": "TA-412", - "scope": "Mamba2 mixer HF-parity fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full/partial single chunks, masked input, and BF16 finite output now form one mixer policy.", - "HF outputs, input/parameter gradients, dtype, and finiteness remain asserted." - ] - }, - { - "id": "TA-413", - "scope": "SSD recurrence fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Multi-chunk output/gradient parity, missing-D behavior, and complete mixer recurrence now form one SSD policy.", - "Both sequence lengths, every input gradient, and projected mixer output remain checked." - ] - }, - { - "id": "TA-414", - "scope": "packed SSD and Mamba2 fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Boundary-aligned/unaligned SSD, single-sequence identity, causal convolution, mixer parity, full-row identity, and batch rejection now form one packed policy.", - "All outputs, input/parameter gradients, sequence boundaries, and error messages remain." - ] - }, - { - "id": "TA-415", - "scope": "optional SSD kernel-parity fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense and packed-sequence mamba_ssm comparisons now form one optional CUDA kernel policy.", - "Outputs and the applicable dense/packed gradient sets retain their tolerances." - ] - }, - { - "id": "TA-416", - "scope": "token diagnostic selection and boundary fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shape, ranking, disabled, absent-label, all-ignored, and top-k clamping behavior now form one selection policy.", - "Every retained field and boundary result remains asserted." - ] - }, - { - "id": "TA-417", - "scope": "token diagnostic log-probability reference fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Loss-path and raw-weight reference comparisons now form one numerical cross-check policy.", - "Exact zero deltas and deliberately nonzero scaled-head deltas remain asserted." - ] - }, - { - "id": "TA-418", - "scope": "hidden diagnostic summary sampling examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Layer summaries, explicit indices, invalid indices, and ordered component summaries now report as one policy.", - "The redundant all-indices spelling and a second component-width implementation example were removed." - ] - }, - { - "id": "TA-419", - "scope": "hidden component hook implementation examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "MoE callback and shared-expert capture now form one hook-integration policy alongside the independent dense equation policy.", - "A native-MoE example that only pinned internal ordering constants was removed." - ] - }, - { - "id": "TA-420", - "scope": "fused selected-logprob forward and backward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forward, finite-output, and input/weight/bias gradient parity now execute from the same eager-reference cases.", - "BF16/no-bias/default-temperature and FP32/bias/nondefault-temperature branches remain covered." - ] - }, - { - "id": "TA-421", - "scope": "fused selected-logprob input-gradient fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Frozen-head gradient parity and all-frozen output detachment now form one input-gradient policy.", - "Hidden-state gradients and absence of weight/bias gradients remain checked." - ] - }, - { - "id": "TA-422", - "scope": "fused selected-logprob irregular shape repetitions", - "decision": "remove", - "status": "applied", - "evidence": [ - "A non-tile-aligned 37x130x777 case retains tail-shape parity.", - "The single-row example and a second large-vocabulary example duplicated boundaries covered by the production-vocabulary policy." - ] - }, - { - "id": "TA-423", - "scope": "fused loss dispatcher fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-token CE, scalar causal-LM, and quack-linear per-token dispatch now form one interface policy.", - "Scalar loss, per-token loss, and per-token log-probability parity remain checked." - ] - }, - { - "id": "TA-424", - "scope": "fused loss production-vocabulary repetitions", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The Qwen integration path and largest GPT-OSS direct path retain the two named production boundaries.", - "An unnamed 100000-vocabulary repetition was removed while finiteness, eager parity, and finite gradients remain." - ] - }, - { - "id": "TA-425", - "scope": "duplicate fused-logit peak-memory probe", - "decision": "remove", - "status": "applied", - "evidence": [ - "The causal-LM integration probe is the stronger regression because it fails on eager dispatcher fallthrough.", - "The direct frozen-weight probe repeated the same less-than-half-full-tile heuristic without covering that dispatch boundary." - ] - }, - { - "id": "TA-426", - "scope": "DistSignSGD framework-behavior unit examples", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained optimizer policy checks the production preaggregated update and decoupled decay ordering.", - "Base-Optimizer state-dict round-tripping and an unsupported sparse-gradient example did not exercise distributed sign behavior." - ] - }, - { - "id": "TA-427", - "scope": "DistSignSGD reduce-scatter fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct sign voting, forced SUM, async propagation, and SP-before-sign ordering now form one communication policy.", - "Exact inputs, outputs, operations, and process-group routing remain asserted." - ] - }, - { - "id": "TA-428", - "scope": "DistSignSGD unsupported-topology fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HSDP, sequence-parallel folding, expert parallelism, and unmanaged DTensor rejection now form one admission policy.", - "Each topology still checks its specific failure message." - ] - }, - { - "id": "TA-429", - "scope": "DistSignSGD optimizer-builder fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FSDP2 construction, hook configuration, decay-group ownership, and non-FSDP2 rejection now form one builder policy.", - "Optimizer type, exact parameters, decay values, and rejection remain checked." - ] - }, - { - "id": "TA-430", - "scope": "SGLang MoE automatic-resolution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP1, EP8, explicit overrides, device/stack requirements, module eligibility, and log-once behavior now form one resolution policy.", - "Every admitted and rejected regime plus warning and logging behavior remains checked." - ] - }, - { - "id": "TA-431", - "scope": "SGLang MoE block-dispatch fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Disabled, full-forward, experts-only, and FP64-precedence paths now form one dispatch policy.", - "Outputs, shapes, call routing, and precedence remain asserted." - ] - }, - { - "id": "TA-432", - "scope": "SGLang MoE fused-expert admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Expert semantics, missing routing inputs, positive SwiGLU limits, and trainable bias/activation guards now form one admission policy.", - "An install-state-dependent missing-SGLang example that skipped whenever SGLang was present was removed." - ] - }, - { - "id": "TA-433", - "scope": "SGLang masked-training gradient fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Partially masked, all-valid, and all-masked expert routing now form one CUDA gradient policy.", - "Forward outputs and every input, routing-weight, gate-up, and down gradient remain bitwise checked." - ] - }, - { - "id": "TA-434", - "scope": "SGLang MoE weight-mode and kernel-layout fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Cache lifecycle, mode resolution, zero-copy strided views, serving layout, and FP32 routing weights now form one policy.", - "Storage identity, invalidation, environment precedence, tensor layout, and kernel flags remain checked." - ] - }, - { - "id": "TA-435", - "scope": "vendored strided MoE adapter fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Accepted storage layouts and split gate/up delegation now form one adapter policy.", - "Contiguous, transpose-view, sliced-layout, and interleaved-gate rejection cases remain." - ] - }, - { - "id": "TA-436", - "scope": "SGLang runtime-context fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Publication, compatible reuse, nondeterministic rejection, and fused-reduction rejection now form one context policy.", - "Exact published arguments, role, no-op behavior, and both incompatibilities remain asserted." - ] - }, - { - "id": "TA-437", - "scope": "mocked EP1 SGLang auto-dispatch CUDA example", - "decision": "remove", - "status": "applied", - "evidence": [ - "CPU resolution and dispatch policies already cover the branch decision and explicit escape hatch.", - "The retained real-kernel CUDA policy proves automatic resolution, deterministic output, and explicit-mode bitwise parity." - ] - }, - { - "id": "TA-438", - "scope": "batch-invariant trunk-linear wrapping fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Selection counts, exclusions, idempotence, routed-expert skipping, and unsupported-module admission now form one CPU policy.", - "Every projection count, ownership marker, and rejection class remains checked." - ] - }, - { - "id": "TA-439", - "scope": "batch-invariant trunk-linear forward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Persistent-GEMM parity, global-interpose parity, batch invariance, dtype admission, and interpose conflict now form one CUDA forward policy.", - "Bias and no-bias outputs remain bitwise checked." - ] - }, - { - "id": "TA-440", - "scope": "global batch-invariant interpose gradient fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Training rejection, no-grad numerical behavior, and grad-enabled operation on grad-free inputs now form one policy.", - "MM, RMSNorm, BMM, log-softmax, mean, and permitted inference outputs remain exercised." - ] - }, - { - "id": "TA-441", - "scope": "MiniMax M3 configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Top-level HF adaptation, local loading/registration, and native-config round-trip now form one configuration policy.", - "Architecture, attention, MoE, sparse-attention, multimodal metadata, and registry fields remain checked." - ] - }, - { - "id": "TA-442", - "scope": "MiniMax M3 activation and router fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "SwiGLU-OAI clamping and sigmoid-router selection/weighting now form one primitive policy.", - "The analytical activation and bias-for-selection-only routing references remain exact." - ] - }, - { - "id": "TA-443", - "scope": "MiniMax M3 text-runtime fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Tiny forward/backward, text-only admission, reserved-token rejection, and unsupported parallelism now form one runtime policy.", - "Loss, output shapes, lm-head gradient, and all rejection messages remain." - ] - }, - { - "id": "TA-444", - "scope": "MiniMax M3 checkpoint fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Language-weight mapping, multimodal skipping, EP ownership, and grouped-loader aliases now form one checkpoint policy.", - "Dense, routed-expert, router, correction-bias, local-shard, and raw-key behavior remains checked." - ] - }, - { - "id": "TA-445", - "scope": "runner dispatcher batch-distribution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ordinary DP, EP slicing/padding, CP sharing, and both legacy duplicate modes now form one distribution policy.", - "Exact batches, routing payload slices, dummy labels, and EP-FSDP rank ownership remain checked." - ] - }, - { - "id": "TA-446", - "scope": "runner dispatcher routing-payload fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Filesystem and Mooncake slicing, world-size-one loading, pickle rejection, and symlink rejection now form one transport policy.", - "Loaded tensors, object keys, and trust-boundary errors remain asserted." - ] - }, - { - "id": "TA-447", - "scope": "runner dispatcher packing examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Balanced-DP zero-dummy behavior and sequential underfill now form one packing policy.", - "Lockstep rounds, exact datum accounting, and dummy fallback remain checked." - ] - }, - { - "id": "TA-448", - "scope": "runner dispatcher per-token merge fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Logical ordering, empty-slice removal, coherent CP deduplication, and replica disagreement now form one merge policy.", - "Every merged field and the disagreement diagnostic remain asserted." - ] - }, - { - "id": "TA-449", - "scope": "runner dispatcher row-batching provenance fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-local row batching and unmerged-row provenance now form one policy.", - "Token layout, sequence boundaries, source batch/request IDs, sample counts, and spans remain checked." - ] - }, - { - "id": "TA-450", - "scope": "P2P prepare-payload fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Transport/engine metadata and cache-invalidation suppression now form one prepare policy.", - "Endpoint, group, sender identity, rank, and cache-mode behavior remain asserted." - ] - }, - { - "id": "TA-451", - "scope": "P2P initialize-fanout fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Multi-endpoint map aggregation, HTTP/remote failures, and receiver cleanup after partial preparation now form one fanout policy.", - "Locator ownership, receiver sessions, cleanup endpoints, and completion payloads remain checked." - ] - }, - { - "id": "TA-452", - "scope": "P2P cached-prepare fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Map reuse, session change, partial endpoint refresh, unsupported-flag retry, and endpoint-local retry now form one cache policy.", - "Request flags, retry counts, locator replacement, endpoint retention, and session IDs remain checked." - ] - }, - { - "id": "TA-453", - "scope": "P2P complete-sync fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Cached prepare-state preservation and tied-weight alias forwarding now form one completion policy.", - "Tensor maps, session/debug state, and alias payloads remain asserted." - ] - }, - { - "id": "TA-454", - "scope": "P2P Qwen3.5 linear-attention slicing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Convolution squeezing, generic QKV slicing, local state-vector ownership, and receiver-dtype conversion now form one slicing policy.", - "Exact shapes, values, TP ranges, byte sizes, and dtypes remain checked." - ] - }, - { - "id": "TA-455", - "scope": "P2P engine hostname and fallback fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Hostname precedence/DNS fallback and direct Mooncake construction without the SGLang wrapper now form one engine policy.", - "Every address source and initialized engine field remains checked." - ] - }, - { - "id": "TA-456", - "scope": "FP8 synchronization numerical fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "BF16 islands, Slime blockwise parity, and partial-layout zero padding now form one synchronization contract.", - "Exact FP8 bytes, scales, dequantized values, exclusions, and last-element-padding distinction remain." - ] - }, - { - "id": "TA-457", - "scope": "FP8 adapter-merge fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense LoRA, QLoRA linear, and QLoRA MoE merging now form one adapter synchronization policy.", - "Extraction ownership, merged weights, emitted names, exact quantization, and nonzero dequantized deltas remain checked." - ] - }, - { - "id": "TA-458", - "scope": "FP8 projection-selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Supported families, optional weight suffixes, receiver passthrough entries, and broad-selector activation now form one selection policy.", - "Every quantized, scaled, and identity-preserved tensor remains asserted." - ] - }, - { - "id": "TA-459", - "scope": "FP8 stack and existing-dtype fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Stack-versus-single parity and already-quantized passthrough now form one tensor policy.", - "Per-slice quantized values, scales, and identity passthrough remain checked." - ] - }, - { - "id": "TA-460", - "scope": "FP8 CPU expert-projection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HF naming/scales, partial-layout padding, deferred quantization, and module exclusions now form one CPU expert policy.", - "Bytes, shapes, dtypes, expert indices, timings, exact scales, and passthrough behavior remain." - ] - }, - { - "id": "TA-461", - "scope": "FP8 CPU workspace fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Staging/reuse, streaming transfer, flush/reset, and empty-final-flush metadata now form one workspace lifecycle policy.", - "Storage identity, bucket contents, flush/version routing, capacity reset, and timing fields remain asserted." - ] - }, - { - "id": "TA-462", - "scope": "FP8 GPU synchronization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CPU/CUDA targets, stack CPU parity, direct-EP expert parity, and module exclusions now form one GPU policy.", - "Output devices, dtypes, timings, names, scales, quantized values, and BF16 passthrough remain checked." - ] - }, - { - "id": "TA-463", - "scope": "adapter optimizer construction and persistence fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "SignSGD construction, current-LR persistence, Adam kwarg normalization, and Muon group-LR preservation now form one policy.", - "Checkpoint location/metadata, optimizer types, state fields, hyperparameters, and LR behavior remain checked." - ] - }, - { - "id": "TA-464", - "scope": "adapter gradient-ownership compile and admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit compilation, pending-gradient rejection, uncompiled failure, invalid scale/gradient rejection, collective freedom, and sync exclusions now form one policy.", - "Plan parameters, unchanged state, errors, capture result, and exclusion ownership remain asserted." - ] - }, - { - "id": "TA-465", - "scope": "adapter capture commit fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Preallocated staging, direct-DTensor completion, and atomic prevalidation now form one commit policy.", - "Storage identity, model-gradient ownership, completed tensors, and all-or-nothing numerator mutation remain checked." - ] - }, - { - "id": "TA-466", - "scope": "adapter coordinator checkpoint materialization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit checkpoint loading and automatic evicted-adapter loading now form one coordinator policy.", - "Checkpoint session specs, optimizer type, learning rate, loaded path, and materialization remain checked." - ] - }, - { - "id": "TA-467", - "scope": "adapter checkpoint session-compatibility fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Optimizer mismatch rejection, explicit LR override, weights-only optimizer mismatch, and same-contract LR restoration now form one policy.", - "Weights, optimizer ownership, session spec, live LR, and rejection behavior remain asserted." - ] - }, - { - "id": "TA-468", - "scope": "GLM5 configuration and construction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "RMSNorm declarations, HF adaptation, local loading, unsafe-value rejection, default shape, and ring-attention rejection now form one policy.", - "MLA, MoE, DSA, MTP, family, architecture, topology, and safety fields remain checked." - ] - }, - { - "id": "TA-469", - "scope": "GLM5 indexer construction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Projection shapes and exact-contract FP32 head projection now form one construction policy.", - "All four projection dimensions, family weights, kernel operands, scaling, and output bits remain checked." - ] - }, - { - "id": "TA-470", - "scope": "GLM5 indexer selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Additive masking, padding-prefix detection, sorted output, blocked scoring, and query offsets now form one selection policy.", - "Dense/chunked parity, sentinels, valid ranges, ordering, and local-query equivalence remain checked; TileLang stays independent." - ] - }, - { - "id": "TA-471", - "scope": "GLM5 DSA mask fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two/three-dimensional mask construction and Ulysses query-axis gathering now form one mask policy.", - "Shapes, excluded keys, gathered shard rows, and values remain asserted." - ] - }, - { - "id": "TA-472", - "scope": "GLM5 sparse-MLA reference and wrapper fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense-equivalent reference behavior, query offsets, and local-query/full-KV TileLang adaptation now form one policy.", - "Outputs, flattened shapes, globalized indices, and scaling remain checked." - ] - }, - { - "id": "TA-473", - "scope": "GLM5 sparse-attention integration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ulysses local-query routing and full-top-k dense parity now form one integration policy.", - "Query/KV/index shapes, query offsets, mask handling, and end-to-end hidden-state parity remain." - ] - }, - { - "id": "TA-474", - "scope": "GLM5 sparse kv_b adapter fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Distributed LoRA, block-FP8 QLoRA dequantization, and BF16 absorb-compute behavior now form one adapter-weight policy.", - "Analytical weights, split shapes, dtypes, downstream equations, and both factor gradients remain checked." - ] - }, - { - "id": "TA-475", - "scope": "GLM5 checkpoint filtering fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Out-of-range layer filtering and EP-owned FP8 expert weight/scale filtering now form one checkpoint policy.", - "Ordinary, MTP, distant, non-layer, local-expert, and remote-expert keys remain checked." - ] - }, - { - "id": "TA-476", - "scope": "GLM5 adapter and MoE dispatch fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default MLA/MoE targets, EP eager dispatch, and sparse kv_b LoRA engagement now form one policy.", - "Wrapped/unwrapped modules, expert call routing, router shapes, and sparse/dense parity with nonzero deltas remain." - ] - }, - { - "id": "TA-477", - "scope": "GLM5 forward and recompute fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Tiny-model layer construction/forward and recompute-before-dispatch checkpoint routing now form one runtime policy.", - "Dense/MoE split, indexer reachability, output shape, and inner-versus-outer checkpoint calls remain checked." - ] - }, - { - "id": "TA-478", - "scope": "expert-adapter backend capability fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Registered backend capabilities and inactive optional-dispatch identity now form one backend policy.", - "Local/EP support, dispatch methods, reduction domain, zero-token behavior, and guard identity remain." - ] - }, - { - "id": "TA-479", - "scope": "expert-adapter factor ownership fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quantized projection subsets, unquantized structural-zero omission, and eager local-only admission now form one ownership policy.", - "Parameters, buffers, checkpoints, shapes, factor domains, and backend capability remain asserted." - ] - }, - { - "id": "TA-480", - "scope": "generic expert semantics fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quantized and unquantized rejection/acceptance across gated, activation, clamp, and bias semantics now form one policy.", - "Standard SiLU preservation, native activation, implementation, target roles, and all failure modes remain checked." - ] - }, - { - "id": "TA-481", - "scope": "model-family expert adapter construction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "GLM4 QLoRA, Qwen3 Quack-NF4, and Qwen3.5 Quack-LoRA construction now form one family policy.", - "Checkpoint buffering, exact targets, quantization format/group, source FQN, activation, and hybrid semantics remain." - ] - }, - { - "id": "TA-482", - "scope": "expert-adapter fail-closed fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Invalid quantization groups, model-specific semantics, and invalid quantized/unquantized target sets now form one admission policy.", - "MiniMax, GPT-OSS, Nemotron, empty, router, mixed, and duplicate target failures remain checked." - ] - }, - { - "id": "TA-483", - "scope": "removed server-configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "YAML fields, CLI overrides, nested ZORL configuration, and unrelated unknown fields now form one removal boundary.", - "Rejection timing, messages, and forward-compatible unknown-field behavior remain checked." - ] - }, - { - "id": "TA-484", - "scope": "shipped adapter example fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "LoRA and quantized-MoE example parsing now forms one shipped-configuration policy.", - "Clean-process parsing, Quack selection, hybrid mode, and every expert target remain checked." - ] - }, - { - "id": "TA-485", - "scope": "server runtime serialization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Automatic defaults, nested runtime controls, and receiver KV-cache normalization now form one round-trip policy.", - "Optimizer variants, checkpoint fields, prefetch, packing, activation, adapter, and model defaults remain checked." - ] - }, - { - "id": "TA-486", - "scope": "quantized training configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP8 training, GLM block-FP8 QLoRA, QARL, NeMo aliases, and fail-fast fallback now form one mode policy.", - "All numerical fields, target scopes, calibration fields, aliases, and serialized defaults remain checked." - ] - }, - { - "id": "TA-487", - "scope": "server parallel-topology fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact GLM rank-one admission, lm-head sharding, and LoRA tensor-parallel boundaries now form one topology policy.", - "EP, CP, lm-head TP, total GPU count, chunking, and model-TP rejection remain checked." - ] - }, - { - "id": "TA-488", - "scope": "unsupported server configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quantized-mode conflicts, vLLM FP8 knobs, broadcast loading, and unsupported adapter modes now form one rejection policy.", - "LoRA, MTP, Mamba, ModelOpt, KV-cache, pipeline, and merge-interval failures remain checked." - ] - }, - { - "id": "TA-489", - "scope": "server optimizer and runner compatibility fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Muon Gram-Newton-Schulz serialization and runner compatibility fields now form one execution policy.", - "Dtypes, restarts, fallbacks, determinism, routing, reduction, decay, and export fields remain checked." - ] - }, - { - "id": "TA-490", - "scope": "model-specific server configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sparse-MLA threading and routing-weight placement now form one model-specific policy.", - "Explicit and automatic resolution plus serialized model fields remain checked." - ] - }, - { - "id": "TA-491", - "scope": "GLM5.2 layer-plan fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Official schedule, malformed plans, stage starts, parameter allocation, and strict reload now form one layer-plan policy.", - "All 78 layers, producer ownership, dense/sparse split, indexer keys, and rejection cases remain checked." - ] - }, - { - "id": "TA-492", - "scope": "GLM5.2 sparse logical-selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Small-row edge cases and the production 4096 boundary now form one logical-index policy.", - "Stable ties, dead rows, valid counts, cache mapping, gathers, and top-2048 ordering remain checked." - ] - }, - { - "id": "TA-493", - "scope": "GLM5.2 selector Hadamard fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Transport inversion and pre-quantization enabled/disabled paths now form one Hadamard policy.", - "BF16 bytes, normalization tolerance, and exact query/key operands remain checked." - ] - }, - { - "id": "TA-494", - "scope": "GLM5.2 fused indexer projection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fused projection row order and FP32 gate scaling now form one projection policy.", - "Single-call fusion, exact BF16 bytes, promotion order, and sampler scoring formula remain checked." - ] - }, - { - "id": "TA-495", - "scope": "GLM5.2 sampler index-key preparation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fused preparation, prompt/decode mixing, and CP16 boundary mapping now form one index-key policy.", - "Projection stride, literal RoPE cache, suffix selection, positions, and 4096 ownership remain checked." - ] - }, - { - "id": "TA-496", - "scope": "GLM5.2 sparse quantization codec fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Query/key scale domains, paged-cache unpacking, and production-shape sampler bytes now form one codec policy.", - "E4M3 and UE8M0 formats, page layout, shapes, scales, and bitwise values remain checked." - ] - }, - { - "id": "TA-497", - "scope": "GLM5.2 native selector runtime fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Batched dispatch/masking and fail-closed admission now form one native-selector runtime policy.", - "Flattening, unwritten cells, CUDA absence, and non-prefix masks remain checked." - ] - }, - { - "id": "TA-498", - "scope": "GLM5.2 sparse kernel loader fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "DeepGEMM capability admission and shared selector import now form one loader policy.", - "Missing score support and exact shared-kernel identity remain checked." - ] - }, - { - "id": "TA-499", - "scope": "GLM5.2 canonical routing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Replay rejection, internal transport, exact canonical routing, and ordinary noncanonical routing now report as two policies.", - "Environment independence, selector version, router class, and public-knob exclusion remain checked." - ] - }, - { - "id": "TA-500", - "scope": "SGLang EP admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "DeepEP mechanism exclusion and FP8 rejection now form one serving-kernel admission policy.", - "Transport rationale, unsupported-mechanism wording, and FP8 failure remain checked." - ] - }, - { - "id": "TA-501", - "scope": "SGLang EP flag-off fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Stock compute selection and score-dtype behavior now form one disabled-mode policy.", - "No serving-kernel engagement, finite output, BF16 stock scores, and FP32 opt-in scores remain checked." - ] - }, - { - "id": "TA-502", - "scope": "SGLang EP compute guard fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Empty-rank short circuit and unsupported compute forms now form one boundary policy.", - "Kernel suppression, score presence, gating, bias, and activation failures remain checked." - ] - }, - { - "id": "TA-503", - "scope": "SGLang EP slot-combine fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Slot-ordered reduction, collapsed-pair rejection, and autograd rejection now form one combine policy.", - "FP32 reduction, token-slot mapping, unique selections, and scoring-only ownership remain checked." - ] - }, - { - "id": "TA-504", - "scope": "SGLang EP trainable-dispatch fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Autograd dispatch, training admission, and empty-rank gradients now form one trainable policy.", - "Plain no-grad routing, bias/activation failures, zero weight gradients, and input-score gradients remain checked." - ] - }, - { - "id": "TA-505", - "scope": "SGLang EP weight-presentation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Zero-copy strided, transient, and cached weight modes now form one presentation policy.", - "Storage identity, contiguity, reuse, version invalidation, and explicit invalidation remain checked." - ] - }, - { - "id": "TA-506", - "scope": "SGLang fused RMSNorm CPU fallback fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Residual and forced single-input CPU fallbacks now form one eager-fallback policy.", - "Exact residual carry, weighted output, and global-mode restoration remain checked." - ] - }, - { - "id": "TA-507", - "scope": "SGLang fused RMSNorm forward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Residual, no-residual, packed-3D, and module-mode examples now form one bit-exact forward policy.", - "BF16/FP32, shape preservation, residual bytes, forced mode, and native no-force behavior remain checked." - ] - }, - { - "id": "TA-508", - "scope": "SGLang fused RMSNorm backward fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Residual and no-residual closed-form gradients now form one backward policy.", - "Hidden, residual, and FP32 weight gradients remain compared with eager autograd." - ] - }, - { - "id": "TA-509", - "scope": "SGLang fused RMSNorm integration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense Qwen layer parity and the pre-summed final-norm boundary now form one model integration policy.", - "Both norm sites, serving residual-tree bytes, and the one-ULP seed boundary remain checked." - ] - }, - { - "id": "TA-510", - "scope": "SGLang fused RMSNorm trunk fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Interposed forward, contract-off dispatch, and trunk backward now form one no-residual trunk policy.", - "Bitwise family identity, ordinary dispatch, finite gradients, and eager gradient parity remain checked." - ] - }, - { - "id": "TA-511", - "scope": "native block-FP8 encoding fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Raw FP8 packing and protected linear state now form one parameter-encoding policy.", - "Shapes, exact bytes, parameter names, frozen ownership, and dtype-apply protection remain checked." - ] - }, - { - "id": "TA-512", - "scope": "native block-FP8 execution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CPU/materialization failure and partition-forward hook traversal now form one execution-boundary policy.", - "Lazy SGLang imports, CUDA admission, invalid materialization inputs, ranges, hooks, and results remain checked." - ] - }, - { - "id": "TA-513", - "scope": "native block-FP8 admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Gradient ownership, partition geometry, and prequantized pair validation now form one fail-closed policy.", - "Scoring-only mode, frozen bases, block boundaries, widths, dtypes, shapes, and finite scales remain checked." - ] - }, - { - "id": "TA-514", - "scope": "native block-FP8 checkpoint fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "State round trip, metadata preflight, adversarial apply, real DCP, and EP-restored shapes now form one lifecycle policy.", - "Both byte streams, no implicit casts, atomic parameter restoration, and global expert shapes remain checked." - ] - }, - { - "id": "TA-515", - "scope": "GLM5.2 native-FP8 configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HF metadata round trip and nonofficial-contract rejection now form one configuration policy.", - "Quant method, format, activation scheme, block size, exclusions, and serialization remain checked." - ] - }, - { - "id": "TA-516", - "scope": "GLM5.2 native-FP8 dense pair-buffer fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Successful emission, incomplete/duplicate/bad-scale failures, and injective target ownership now form one dense-buffer policy.", - "DCP names, exact bytes, FP32 scales, duplicate targets, and aliased modules remain checked." - ] - }, - { - "id": "TA-517", - "scope": "GLM5.2 native-FP8 construction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quantized module replacement and sparse-MLA KV materialization now form one model-construction policy.", - "Dense/shared/expert ownership, exclusions, mixed-precision classes, forward hooks, and split layouts remain checked." - ] - }, - { - "id": "TA-518", - "scope": "GLM5.2 native-FP8 expert checkpoint fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Expert pair fusion, topology rejection, and dense/expert handler separation now form one checkpoint policy.", - "Local expert bytes, gate-up fusion, scales, rank/count admission, load families, and skip ownership remain checked." - ] - }, - { - "id": "TA-519", - "scope": "Qwen3.5 norm-family dispatch fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact selection, ordinary SGLang, fused SGLang, and v2 candidate dispatch now form one family-selection policy.", - "Structural ownership, residual/no-residual routing, exact coexistence, v1 default, and v2 admission remain checked." - ] - }, - { - "id": "TA-520", - "scope": "Qwen3.5 norm-site assignment fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "V2 module resolution, per-layer input forcing, and final-norm forcing now form one call-site policy.", - "Every zero-centered site, layer-zero exception, native mode, and both SGLang modes remain checked." - ] - }, - { - "id": "TA-521", - "scope": "Qwen3.5 norm bit-exact integration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Module dispatch, trunk family one, and full decoder-layer parity now form one integration gate.", - "Residual, forced, qk-norm, aten interpose, and model output bytes remain checked." - ] - }, - { - "id": "TA-522", - "scope": "generic QLoRA quantized execution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quantize/train/memory, dequantization, and prequantized NVFP4 loading now form one execution policy.", - "Both quant formats, storage savings, output shape, round-trip tolerance, and LoRA-only gradients remain checked." - ] - }, - { - "id": "TA-523", - "scope": "generic QLoRA NVFP4 merge fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EMA scale convention and merge/requantization now form one NVFP4 lifecycle policy.", - "Amax updates, global scale, quantization error, delta folding, state export, and adapter reset remain checked." - ] - }, - { - "id": "TA-524", - "scope": "generic QLoRA injection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Target replacement/training and checkpoint-format propagation now form one injection policy.", - "Target exclusion, source format, quant format, model forward, and every adapter gradient remain checked." - ] - }, - { - "id": "TA-525", - "scope": "generic QLoRA block-FP8 fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Prequantized block-FP8 loading, fused QKV assembly, forward/backward, and merge now form one format policy.", - "Packed bytes, round-trip error, shapes, adapter gradients, delta folding, and no-EMA behavior remain checked." - ] - }, - { - "id": "TA-526", - "scope": "generic QLoRA optimizer-reset fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Adapter-state clearing, non-LoRA preservation, and scheduled merge integration now form one optimizer lifecycle.", - "State rebuild, selective ownership, pre-boundary no-op, boundary requantization, and factor reset remain checked." - ] - }, - { - "id": "TA-527", - "scope": "OPD numerical backend fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Chunking, streaming/TileLang, low-memory streaming, and sharded teacher storage now form one backend policy.", - "Reference loss, gradients, teacher detachment, chunk equivalence, and safetensor shard reads remain checked." - ] - }, - { - "id": "TA-528", - "scope": "OPD gradient and reduction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default backward, token-partial reduction, and BF16-to-FP32 loss behavior now form one gradient policy.", - "Student gradients, teacher detachment, valid-token scaling, output dtype, and finiteness remain checked." - ] - }, - { - "id": "TA-529", - "scope": "OPD output edge fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All-ignored and per-token outputs now form one output-shaping policy.", - "Zero finite loss/gradients, masks, FP32 dtype, tensor shape, and reduced reference equality remain checked." - ] - }, - { - "id": "TA-530", - "scope": "OPRD hidden-distance fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Materialized and fetched teacher-layer paths now form one hidden-distance policy.", - "Chunk boundaries, full-reference distance, layer counts, student gradients, and teacher detachment remain checked." - ] - }, - { - "id": "TA-531", - "scope": "OPD hidden-only objective boundary", - "decision": "keep", - "status": "applied", - "evidence": [ - "Zero-KL hidden-only MSE remains an independent objective boundary rather than being folded into backend checks.", - "Weighted token math, per-token output, metrics, hidden gradients, absent head gradients, and KL diagnostics remain checked." - ] - }, - { - "id": "TA-532", - "scope": "quantized-export CLI fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "YAML/CLI precedence and byte-size parsing now form one command-line policy.", - "Block sizes, appended exclusions, BF16 layer counts, shard sizes, and decimal/binary units remain checked." - ] - }, - { - "id": "TA-533", - "scope": "base quantized-export fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Module CLI execution, BF16 islands/config writing, and sharded-index output now form one base-directory policy.", - "Subprocess entrypoint, quantized counts, exact dtypes, exclusions, tokenizer copy, shard map, and total bytes remain checked." - ] - }, - { - "id": "TA-534", - "scope": "quantized-export projection layout fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fused QKV, MLA-A fusion, and linear-attention renaming now form one projection-layout policy.", - "Every emitted name, FP8 tensor, scale, metadata rejection, split fallback, convolution, norm, and bias remains checked." - ] - }, - { - "id": "TA-535", - "scope": "quantized-export MoE layout fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "GKN expert conversion and fused gate-up splitting now form one MoE-layout policy.", - "HF expert names, transposes, gate/up/down partitions, quantized bytes, scales, and output counts remain checked." - ] - }, - { - "id": "TA-536", - "scope": "quantized-export admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unsupported source-state rejection and QARL fold admission now form one export preflight policy.", - "Prequantized inputs, active adapters, missing metadata, fold eligibility, and block-size compatibility remain checked." - ] - }, - { - "id": "TA-537", - "scope": "training-model FP8 construction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full-model FP8 injection and tensor-parallel lm-head inclusion now form one construction policy.", - "Dense/MoE replacement, overrides, correction, fallback, trainability, TP inclusion, and explicit exclusion remain checked." - ] - }, - { - "id": "TA-538", - "scope": "training-model GLM5.2 block-FP8 QLoRA fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Foundation/injection threading and missing-QLoRA rejection now form one GLM5.2 adapter-mode policy.", - "Rank, alpha, format, group size, inventory ownership, and enabling preconditions remain checked." - ] - }, - { - "id": "TA-539", - "scope": "training-model quantized-mode admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP8/adapters, QARL/adapters-or-FP8, and QARL/MoE conflicts now form one admission matrix.", - "Each full-weight ownership conflict and dense-only QARL boundary remains checked." - ] - }, - { - "id": "TA-540", - "scope": "training-model QARL lifecycle fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense fake-quant injection and pre-parallel calibration now form one QARL lifecycle policy.", - "Targets, activation mode, trainability, calibration samples/sequence, forward counts, and learned scales remain checked." - ] - }, - { - "id": "TA-541", - "scope": "DSV4 checkpoint translation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All supported name families and C4 APE hotfix inversion now form one translation policy.", - "Attention/indexer/HC/router/shared-expert mappings, three head dimensions, and invalid layouts remain checked." - ] - }, - { - "id": "TA-542", - "scope": "DSV4 checkpoint quantized-codec fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Block-FP8 and packed MXFP4 dequantization now form one checkpoint-codec policy.", - "Full blocks, tails, known nibbles, scale blocks, output shapes, and dtypes remain checked." - ] - }, - { - "id": "TA-543", - "scope": "DSV4 checkpoint handler fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP-local expert fusion and MTP/unmapped-key accounting now form one handler-ownership policy.", - "Skip decisions, local expert rows, gate-up/down shapes and values, and unknown-key summaries remain checked." - ] - }, - { - "id": "TA-544", - "scope": "DSV4 synthetic load fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Window-only, strict nonpersistent buffers, C4/indexer, and hash-router models now form one end-to-end load policy.", - "Fill summaries, representative values, APE recovery, expert fusion, tid2eid, correction bias, and strict accounting remain checked." - ] - }, - { - "id": "TA-545", - "scope": "GLM5.2 partial-edge QLoRA scale example", - "decision": "remove", - "status": "applied", - "evidence": [ - "The isolated 6144-by-576 scale-shape assertion was an exact subset of the full 700-target inventory contract.", - "The retained inventory still asserts the identical (5, 192) partial-edge storage shape on the official kv_a projection." - ] - }, - { - "id": "TA-546", - "scope": "GLM5.2 exact dense-component fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Logical inventory/physical fused roots and rank-one pre-mutation rejection now form one dense-component policy.", - "All 1,700 factors, three roots, source FQNs, parameter identity, and rank/alpha failures remain checked." - ] - }, - { - "id": "TA-547", - "scope": "GLM5.2 QLoRA admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Wrong targets, missing indexer exclusions, and unsupported construction modes now form one fail-closed policy.", - "No-mutation guarantees, shapes, DSA BF16 ownership, backend, dispatch, exact-contract, and feature-flag failures remain checked." - ] - }, - { - "id": "TA-548", - "scope": "GLM5.2 training-mode fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ordinary product admission and complete exact-active-LoRA alltoall selection now form one training-mode policy.", - "Format rejection, the certified tuple, exact enablement, and deepep rejection remain checked." - ] - }, - { - "id": "TA-549", - "scope": "exact fused gate-up state fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Native-leaf ownership, explicit gate/up loading, and dtype movement now form one state lifecycle.", - "Four FP32 factors, rank/alpha/topology admission, row order, FP8/scale bytes, master identity, and protected dtypes remain checked." - ] - }, - { - "id": "TA-550", - "scope": "exact fused gate-up numerical fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One-time master rounding/logical order and two-branch surrogate parity now form one effective-numerics policy.", - "Captured factor bytes, gate/up ordering, base composition, activation, and output equality remain checked." - ] - }, - { - "id": "TA-551", - "scope": "exact fused gate-up gradient fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Input/factor parity, factor-only VJP, and master-mutation rejection now form one backward policy.", - "All five gradients, no base materialization, and saved-tensor version safety remain checked." - ] - }, - { - "id": "TA-552", - "scope": "exact TP16 lm-head topology fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Official shards, component range ownership, and process-group validation now form one topology policy.", - "All 16 ranges, padding, statelessness, order, rank, world size, and NCCL admission remain checked." - ] - }, - { - "id": "TA-553", - "scope": "exact TP16 lm-head operand fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local operand validation and CPU-before-import rejection now form one execution-admission policy.", - "BF16/FP32 ownership, rank-one shapes, sampler stride, token range, frozen weight, CUDA, and lazy imports remain checked." - ] - }, - { - "id": "TA-554", - "scope": "exact TP16 lm-head presentation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-order vocabulary assembly and effective factor views now form one presentation-bytes policy.", - "Identity token mapping, collective layout rejection, BF16 bytes, and immutable FP32 masters remain checked." - ] - }, - { - "id": "TA-555", - "scope": "exact TP16 lm-head surrogate fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local FP32 surrogate VJP and the custom autograd boundary now form one gradient policy.", - "Base-plus-LoRA hidden gradients, factor gradients, grad-enabled output, and saved effective bytes remain checked." - ] - }, - { - "id": "TA-556", - "scope": "sparse source-delta capture fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank capture, template security, global manifest, and empty-rank reload now form one source lifecycle.", - "Changed indices/values, paths, totals, ordering, traversal rejection, rank tags, shapes, and empty BF16 payloads remain checked." - ] - }, - { - "id": "TA-557", - "scope": "single sparse-delta file fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packing, deterministic index sorting, and malformed-update rejection now form one file policy.", - "Stats, INT32 CPU indices, BF16 values, shapes, duplicates, dtype, length, and range failures remain checked." - ] - }, - { - "id": "TA-558", - "scope": "sparse-delta contiguous-shard fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Index localization and empty local shards now form one contiguous-sharding policy.", - "All rank shapes, localized sorted indices, values, and explicit empty tensors remain checked." - ] - }, - { - "id": "TA-559", - "scope": "ranked sparse-delta file fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Raw rank writes, template security, and pre-encoded writes now form one rank-file policy.", - "Rank order, paths, nnz counts, traversal rejection, and packed API behavior remain checked." - ] - }, - { - "id": "TA-560", - "scope": "sparse-delta translation-future fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Future draining, unranked rejection, and direct ranked-file publication now form one translation policy.", - "Tag stripping, rank maps, exact encoded objects, expected ranks, and output paths remain checked." - ] - }, - { - "id": "TA-561", - "scope": "EP adapter registry fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Optional Quack registration and live signature inspection now form one registry policy.", - "Base-versus-MoE-act registration, explicit shared parameters, and forward-compatible kwargs remain checked." - ] - }, - { - "id": "TA-562", - "scope": "native EP adapter FP8 fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Common FP8-kwarg consumption and explicit FP8-compute rejection now form one native-backend boundary.", - "Disabled-mode output shape and enabled-mode failure remain checked." - ] - }, - { - "id": "TA-563", - "scope": "Triton EP adapter FP8 fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Common FP8-kwarg consumption and explicit FP8-compute rejection now form one Triton-backend boundary.", - "Kernel dispatch, output shape, optional availability, and enabled-mode failure remain checked." - ] - }, - { - "id": "TA-564", - "scope": "Triton MoE-act EP adapter fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Common activation/FP8 kwargs and explicit FP8-compute rejection now form one MoE-act boundary.", - "Activation-native, gating, bias, clamp, output, optional availability, and rejection remain checked." - ] - }, - { - "id": "TA-565", - "scope": "DSV4 model construction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pipeline rejection and sequence-parallel group wiring now form one topology policy.", - "Hyperconnection PP exclusion, TP/CP groups, CP size, and attention implementation remain checked." - ] - }, - { - "id": "TA-566", - "scope": "DSV4 model runtime fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "C128 shape, full LM backward, and hash-routed execution now form one runtime policy.", - "Finite initialization/output/loss, hidden/logit shapes, ordinary and hash gradients, and frozen HC ownership remain checked." - ] - }, - { - "id": "TA-567", - "scope": "DSV4 precision-preservation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Keep-FP32 marking, registry dtype construction, and direct dtype movement now form one precision policy.", - "Every HC/attention/compressor carve-out, BF16 ordinary weights, and complex RoPE imaginary components remain checked." - ] - }, - { - "id": "TA-568", - "scope": "routing-replay sequence-parallel fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Padded/unpadded positions, independent unpacked rows, and excess truncation now form one SP layout policy.", - "All CP ranks, actual position lengths, pad values, row boundaries, micro-batch shape, and truncation order remain checked." - ] - }, - { - "id": "TA-569", - "scope": "routing-replay RingAttention fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Four-rank zigzag placement and packed-document boundaries now form one RingAttention layout policy.", - "Every rank slice, shape, position ordering, and cross-document separation remain checked." - ] - }, - { - "id": "TA-570", - "scope": "routing-replay weight tensor fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Float-value/padding preservation and decoded NumPy slicing now form one routing-weight policy.", - "FP32 dtype, uniform pad weights, CP rank slicing, tensor shape, and exact values remain checked." - ] - }, - { - "id": "TA-571", - "scope": "routing-replay decode fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shaped base64, inferred-shape base64, and materialized Python/NumPy/Tensor inputs now form one wire-format policy.", - "INT32 bytes, model top-k inference, expert selections, logits conversion, types, shapes, and values remain checked." - ] - }, - { - "id": "TA-572", - "scope": "top-k router softmax fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy reference behavior and environment-selected tie/logit policies now form one softmax policy.", - "Normalization modes, inert V4 inputs, stable low/high IDs, logits selection, weights, and unknown-policy rejection remain checked." - ] - }, - { - "id": "TA-573", - "scope": "top-k router layer-FP32 fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Layer selector parsing and live MoE-block routing now form one scoped-FP32 policy.", - "Ranges/all mode, gate bypass, FP32 operands/logits, and expert selection remain checked." - ] - }, - { - "id": "TA-574", - "scope": "top-k router configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "V4 scaling admission and V4/legacy config construction now form one configuration policy.", - "Post-renorm scaling, softmax rejection, scoring method, top-k method, expert counts, and defaults remain checked." - ] - }, - { - "id": "TA-575", - "scope": "OPD shifted-payload fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Teacher cache input/target shift, OPD cache-index alignment, and unshifted rejection now form one payload policy.", - "Input tokens, targets, teacher IDs/weights, cache indices, and length failure remain checked." - ] - }, - { - "id": "TA-576", - "scope": "OPD teacher-cache transport fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Mooncake metadata return and legacy file-metadata rejection now form one teacher-cache transport policy.", - "Request payload, absence of file paths, metadata identity, cache indices, and fail-closed backend validation remain checked." - ] - }, - { - "id": "TA-577", - "scope": "checkpoint-manager rank-zero save failures", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Adapter-state and LoRA-only write failures now form one pre-barrier failure policy.", - "Both public error messages and the prohibition on entering the barrier remain checked." - ] - }, - { - "id": "TA-578", - "scope": "checkpoint-manager dtype-preservation forwarding probe", - "decision": "remove", - "status": "applied", - "evidence": [ - "The deleted test only inspected an internal preserve_lora_dtype keyword.", - "Adapter persistence coverage checks the dtype of the actual saved LoRA tensors through the public save path." - ] - }, - { - "id": "TA-579", - "scope": "checkpoint-manager MoE save fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Collective-state slicing and resolved-target detection now form one MoE export policy.", - "Gather ownership, active-rank slices, target discovery, exported keys, shapes, values, and rank remain checked." - ] - }, - { - "id": "TA-580", - "scope": "checkpoint-manager export-format forwarding probe", - "decision": "remove", - "status": "applied", - "evidence": [ - "The deleted mock asserted only that one internal keyword was forwarded.", - "The retained SGLang shared-outer roundtrip validates the emitted format marker, complete tensor layout, and reload." - ] - }, - { - "id": "TA-581", - "scope": "checkpoint-manager strict-manifest artifact duplicate", - "decision": "remove", - "status": "applied", - "evidence": [ - "The adapter-manager suite already saves and reads the strict target manifest through the public checkpoint lifecycle.", - "It additionally validates the saved manifest and rejects a mismatched runtime manifest." - ] - }, - { - "id": "TA-582", - "scope": "PEFT hybrid-shared checkpoint fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "PEFT orientation and hybrid-shared reload now form one checkpoint roundtrip.", - "Config fields, shared/expert tensor orientations, shapes, bytes, key inventory, and restored factors remain checked." - ] - }, - { - "id": "TA-583", - "scope": "SGLang shared-outer checkpoint fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shared-outer save/load and incompatible ownership rejection now form one format contract.", - "Format metadata, all six MoE tensor layouts, exact reload, and fail-closed admission remain checked." - ] - }, - { - "id": "TA-584", - "scope": "checkpoint-roundtrip cautious optimizer construction example", - "decision": "remove", - "status": "applied", - "evidence": [ - "The deleted example did not exercise checkpointing and only inspected optimizer type and parameter-group fields.", - "Dedicated optimizer coverage already validates cautious routing, optimizer selection, kwargs, and group policy." - ] - }, - { - "id": "TA-585", - "scope": "MoE-LoRA initialization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Backend construction and active-rank view layout now form one initialization policy.", - "Frozen/trainable ownership, all factor shapes, zero initialization, repr, rank slicing, and contiguity remain checked." - ] - }, - { - "id": "TA-586", - "scope": "eager MoE-LoRA execution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Eager forward/backward, block integration, and hybrid-shared construction now form one CPU execution policy.", - "Output shapes, gradient ownership, injection, and every supported shared factor shape remain checked." - ] - }, - { - "id": "TA-587", - "scope": "standalone nonzero MoE-LoRA output example", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained cross-backend contract initializes nonzero LoRA B factors and requires factor gradients on real outputs.", - "The separate eager-only max-difference example added no distinct backend or ownership boundary." - ] - }, - { - "id": "TA-588", - "scope": "generic injection examples in MoE-LoRA suite", - "decision": "remove", - "status": "applied", - "evidence": [ - "Qwen subclass wrapping repeated the generic MoE from_module contract, while unmatched linear targets were not MoE-specific.", - "MoE conversion and both injection APIs remain checked here; generic and model-specific injection boundaries remain elsewhere." - ] - }, - { - "id": "TA-589", - "scope": "EP MoE-LoRA router-score fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Alltoall/deepep score names, missing scores, and score gradients now form one router-score contract.", - "Multiplication values, identity behavior, compute-output gradients, and exact score gradients remain checked." - ] - }, - { - "id": "TA-590", - "scope": "LoRA target-manifest rejection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Coverage, rank, configured-target, unlisted-module, and scalar-schema failures now form one fail-closed manifest contract.", - "Every prior error condition and exact scalar-type boundary remains checked alongside the independent success lifecycle." - ] - }, - { - "id": "TA-591", - "scope": "NeMo FP8 configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The supported blockwise alias and unsupported Transformer Engine recipes now form one translation policy.", - "Native enablement plus hybrid, tensorwise, MXFP8, and FP8-parameter rejection remain checked." - ] - }, - { - "id": "TA-592", - "scope": "external FP8 compatibility rejection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "vLLM receiver knobs and ModelOpt QARL configurations now share one table-driven fail-closed contract.", - "All thirteen configuration paths and their diagnostic categories remain checked without repeated setup." - ] - }, - { - "id": "TA-593", - "scope": "FP8 BF16 layer-island fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "First/last resolution, overlap, invalid topology, and real module injection now form one layer-island lifecycle.", - "Every pattern, count, replacement boundary, model summary, and rejection remains checked." - ] - }, - { - "id": "TA-594", - "scope": "NVFP4 two-dimensional quantization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Reference bytes, public-format dispatch, and straight-through linear gradients now form one numerical contract.", - "Both supported dtypes, output shape/dtype, exact quantization, and upstream weight gradient remain checked." - ] - }, - { - "id": "TA-595", - "scope": "NVFP4 input and format admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank, block-divisibility, and unsupported-format failures now form one API admission policy.", - "The cross-row grouping trap and generic dispatch failure remain explicit." - ] - }, - { - "id": "TA-596", - "scope": "NVFP4 MoE projection STE fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Down and fused gate-up projection paths now share one shape and straight-through-gradient contract.", - "Both tensor layouts, lossy forward behavior, exact gradients, expert independence, and per-half scaling remain checked." - ] - }, - { - "id": "TA-597", - "scope": "data-packing allocation primitives", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FFD feasibility, group packing, and sequential rank allocation now form one allocation contract.", - "Capacity, bin size, safe mode, offsets, rank isolation, token accounting, and full coverage remain checked." - ] - }, - { - "id": "TA-598", - "scope": "packing sample-preprocessing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Position metadata and trainable-label filtering now form one sample preprocessing policy.", - "Single/batched inputs, empty/missing fields, preserved metadata, masks, and missing-label failure remain checked." - ] - }, - { - "id": "TA-599", - "scope": "packing dataset-preprocessing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dataset filtering and train/eval packing preparation now form one dataset pipeline.", - "Retained rows, position/length columns, and optional evaluation data remain checked." - ] - }, - { - "id": "TA-600", - "scope": "linear learning-rate schedule fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default full-range decay and warmup/decay-ratio/floor behavior now form one linear schedule policy.", - "Endpoints, equal decrements, warmup values, decay boundary, and post-decay floor remain checked." - ] - }, - { - "id": "TA-601", - "scope": "cosine learning-rate schedule fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Floor/monotonicity and warmup/midpoint/endpoint behavior now form one cosine schedule policy.", - "Decay ratio, minimum floor, warmup values, half-cosine midpoint, and terminal value remain checked." - ] - }, - { - "id": "TA-602", - "scope": "session API configuration compatibility fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Removed-field rejection and unrelated future-field preservation now form one compatibility policy.", - "All migration diagnostics and nested model-extra behavior remain checked." - ] - }, - { - "id": "TA-603", - "scope": "sampling-adapter reconciliation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Stale-entry pruning and endpoint-query failure preservation now form one reconciliation policy.", - "Eviction avoidance, fresh loading, tracked state replacement, and transient-query safety remain checked." - ] - }, - { - "id": "TA-604", - "scope": "sampling-session tracking fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "URI/plain model scoping and failed-load atomicity now form one sampling-session tracking policy.", - "Embedded and requested model IDs, resolved paths, successful tracking, and absence of stale failed entries remain checked." - ] - }, - { - "id": "TA-605", - "scope": "MoE routing-weight numerical fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Before/after-down parity and the no-router-gradient path now form one numerical contract.", - "FP64 outputs, every input/factor/score gradient, error class, and in-place score-fold safety remain checked." - ] - }, - { - "id": "TA-606", - "scope": "MoE routing-weight configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Mutable config, environment forcing, automatic regimes, parity opt-in, explicit values, and invalid input now form one policy.", - "All train-router/dispatch combinations and Boolean/string spellings remain checked." - ] - }, - { - "id": "TA-607", - "scope": "QARL MoE conversion fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Identity preservation, idempotence, and type admission now form one conversion policy.", - "Parameter objects, class identity, QARL attributes, backend retention, repeat conversion, and non-expert rejection remain checked." - ] - }, - { - "id": "TA-608", - "scope": "QARL MoE eager-execution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quantized execution and disabled-quantization passthrough now form one eager policy.", - "Lossiness, finite output, shape, parameter restoration, gradients, and exact passthrough remain checked." - ] - }, - { - "id": "TA-609", - "scope": "QARL MoE injection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full conversion, FP8 rejection, and target-specific selection now form one injection policy.", - "Linear/expert wrapping, conversion counts, expert-module metadata, independent targets, and fail-closed format admission remain checked." - ] - }, - { - "id": "TA-610", - "scope": "RoPE registry and frequency-precision fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Registry coverage, model-wide BF16 casting, and forward cos/sin precision now form one frequency policy.", - "Every registered recipe, FP32 CPU references, default scaling, YaRN scaling, and bitwise BF16 consumption remain checked." - ] - }, - { - "id": "TA-611", - "scope": "RoPE default-cache fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Lazy cache materialization and Qwen class-B growth now form one cache lifecycle.", - "Execution device, FP32 dtype, indexed values, prefix stability, growth, and CPU recipe reconstruction remain checked." - ] - }, - { - "id": "TA-612", - "scope": "generic QARL dense injection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense wrapping, target/exclusion behavior, model summary, and unsupported architecture admission now form one lifecycle.", - "Parameter names, forward counts, MTP rejection, and Mamba rejection remain checked alongside successful injection." - ] - }, - { - "id": "TA-613", - "scope": "Nemotron-H EP checkpoint fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Topology validation, skip-key ownership, local expert slicing, skip accounting, and parallel-plan classification now form one EP policy.", - "Invalid sizes/ranks, MTP filtering, local ranges, stacked tensors, expert parameter matching, and no-shard modules remain checked." - ] - }, - { - "id": "TA-614", - "scope": "weight-sync endpoint health fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Model-info fallback and exhaustive health failure now form one endpoint-health policy.", - "Endpoint port routing, v1-model fallback order, and diagnostics naming every attempted route remain checked." - ] - }, - { - "id": "TA-615", - "scope": "NCCL endpoint-port routing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Receiver-group initialization and direct bucket transfer now form one port-routing policy.", - "Initialization result metadata, init/update URLs, and direct load-format payload remain checked." - ] - }, - { - "id": "TA-616", - "scope": "NCCL flattened-bucket fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Mixed-dtype flattening and chunked flattened transfer now form one bucket-format policy.", - "Byte packing, payload metadata, flattened/chunked load formats, broadcast counts, and completion waits remain checked." - ] - }, - { - "id": "TA-617", - "scope": "runner session-registration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Coordinator delegation and rank-zero handling of a worker registration error now form one registration policy.", - "Normalized payload delegation, success results, cross-rank synchronization, and failure response text remain checked." - ] - }, - { - "id": "TA-618", - "scope": "runner optimizer-publication fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Successful publication, commit-tail failure, and optimizer-handler tail failure now form one mutation lifecycle.", - "Commit ownership, fatal error translation, causal exceptions, poisoning, and publication ineligibility remain checked." - ] - }, - { - "id": "TA-619", - "scope": "runner forward-backward completion fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Completion ordering and uniform/asymmetric failure policy now form one forward-backward lifecycle.", - "Metric gather, rendezvous, commit-before-merge order, both ranks, uniform rejection, and fatal promotion remain checked." - ] - }, - { - "id": "TA-620", - "scope": "attention backend registry fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FlashAttention registration and runtime backend resolution now form one registry policy.", - "FA4-only import, all flash aliases, eager/native/flex resolution, and unavailable-flash rejection remain checked." - ] - }, - { - "id": "TA-621", - "scope": "SGL page-size-one attention fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Varlen routing and single-sequence metadata synthesis now form one SGL KV-cache policy.", - "Page table/cache shapes, int32 offsets, sequence lengths, scale, causal mode, num_splits, and output shapes remain checked." - ] - }, - { - "id": "TA-622", - "scope": "alternate FlashAttention path fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Paged-KV-cache, flags-off FA3 varlen, and FA4 selection now form one backend-selection policy.", - "Selected call targets, cache layout, num_splits, scale, causal mode, and disabled SGL dispatch remain checked." - ] - }, - { - "id": "TA-623", - "scope": "sequence-shard collator primitive fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "SP slicing and padding now form one collator primitive policy.", - "Both ranks, uneven lengths, ordinary/sequential padding, zero padding, and initialization state remain checked." - ] - }, - { - "id": "TA-624", - "scope": "sequence-shard collator side-channel fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Teacher hidden states and DRGRPO token side channels now form one shard-alignment policy.", - "CP2/CP16 slicing, first/last ranks, shapes, values, ignore-index padding, and zero padding remain checked." - ] - }, - { - "id": "TA-625", - "scope": "sparse-delta streaming lifecycle fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Initial full encoding, exact-byte delta encoding, and unchanged-bucket suppression now execute as one three-transfer lifecycle.", - "TP path replication, endpoint payload, full and changed indices and values, and skip accounting remain checked." - ] - }, - { - "id": "TA-626", - "scope": "sparse-delta prepacked publication fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-rank packed paths and FP8 KV-cache metadata now form one prepacked publication transaction.", - "Both paths, unique-file byte accounting, request flags, normalized cache epoch, and endpoint result metadata remain checked." - ] - }, - { - "id": "TA-627", - "scope": "families-v2 fused-split realization guard", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The proof that forcing split reaches the split kernel now runs inside the fused-versus-split bitwise matrix instead of repeating a standalone kernel call.", - "Tail, aligned, and deep-tile shapes, row-count extremes, residual, plain, and zero-centered modes remain checked." - ] - }, - { - "id": "TA-628", - "scope": "families-v2 norm dispatch fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shipped hidden-size admission, row-depth threshold behavior, and the split-kernel tile basis now form one dispatch policy.", - "Fused decisions, both threshold sides, the row cutoff, and the rejected fused-chunk basis remain checked." - ] - }, - { - "id": "TA-629", - "scope": "EP clip local norm and empty-gradient fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Infinity-norm clipping, parameters without gradients, and empty groups now form one local norm policy.", - "Returned norms, uniform clipping, skipped gradients, and the zero norm remain checked." - ] - }, - { - "id": "TA-630", - "scope": "skip-FSDP EP clipping fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Classification, uniform clipping, and raw local-gradient preservation now execute as one skip-FSDP lifecycle.", - "EP and non-EP ownership, the combined norm, clip coefficients, and absence of EP division remain checked." - ] - }, - { - "id": "TA-631", - "scope": "clip-grad-norm dispatch fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP-aware and ordinary FSDP dispatch outcomes now form one public dispatch policy.", - "Both parameter representations still execute and return the expected norm." - ] - }, - { - "id": "TA-632", - "scope": "mixed-mesh foreach clipping fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Safe default per-tensor clipping and explicit foreach rejection now form one mixed-mesh policy.", - "Both DTensor meshes, returned norm, clipped local values, and the explicit cross-mesh error remain checked." - ] - }, - { - "id": "TA-633", - "scope": "orchestrator packing capacity fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Single-batch packing, overflow splitting, mixed lengths, exact fit, and off-by-one capacity now form one batching policy.", - "Token and label shifts, positions, sample counts, maximum length, and boundary outcomes remain checked." - ] - }, - { - "id": "TA-634", - "scope": "orchestrator packing input-normalization fragment", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "NumPy-to-list normalization now executes inside the existing empty, missing, oversized, and single-sample input policy.", - "The production conversion assertion remains unchanged." - ] - }, - { - "id": "TA-635", - "scope": "P2P transfer source and receiver-manifest rejection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unknown receiver names, incompatible shapes, and unsupported source ranks now form one transfer-admission policy.", - "All error messages and the no-transfer side-effect checks remain." - ] - }, - { - "id": "TA-636", - "scope": "P2P receiver-memory coalescing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Distinct receiver handles and absent handle metadata now execute inside one coalescing policy.", - "Session, peer pointers, transfer lengths, and batch count remain checked for both layouts." - ] - }, - { - "id": "TA-637", - "scope": "P2P transfer failure diagnostic fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Named tensor and handle details, bounded coalesced samples, and diagnostics-disabled behavior now form one failure-reporting policy.", - "Pointer details, the six-entry cap, omitted-entry count, and default redaction remain checked." - ] - }, - { - "id": "TA-638", - "scope": "runner effective LM-head selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical merged-LoRA selection and the legacy unmerged fallback now form one effective-weight policy.", - "Exact folded bytes, adapter gradients, frozen base weight, and legacy formula remain checked." - ] - }, - { - "id": "TA-639", - "scope": "runner compiler replica-topology fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "World-discovered replica counts, malformed coverage, and unsupported general tensor parallelism now form one topology policy.", - "SP2, output4, composed world8, every coverage failure, and the TP rejection remain checked." - ] - }, - { - "id": "TA-640", - "scope": "runner unquantized expert admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Eager, exact-merged, and fused-managed expert admission now includes the hybrid-checkpoint metadata rejection in one policy.", - "Producer, topology, parameter count, and fail-closed metadata behavior remain checked." - ] - }, - { - "id": "TA-641", - "scope": "runner quantized expert contract fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "NF4, NVFP4, and block-FP8 admission now includes declared-shape drift and uncertified EP or eFSDP rejection in one contract policy.", - "All formats, eager and fused producers, guard fields, logical shape validation, and both unsupported parallel regimes remain checked." - ] - }, - { - "id": "TA-642", - "scope": "Muon builder configuration fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Complete Gram-Newton-Schulz option forwarding and rejection of a nonpositive grouping byte limit now form one configuration policy.", - "Parameter groups, dtypes, fallback mode, restart count, byte limit, and the validation error remain checked." - ] - }, - { - "id": "TA-643", - "scope": "Muon Quack backend fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quack operation dispatch, tuned-mode selection, missing-package failure, and SM90 dtype routing now form one backend policy.", - "Every GEMM operation, tuned flags, import error, FP32 Torch fallback, and BF16 Quack selection remain checked." - ] - }, - { - "id": "TA-644", - "scope": "Muon fused-weight and Nemotron classification fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fused gate-up discovery and full Nemotron parameter classification now form one optimizer-classification policy.", - "Gated and non-gated experts, post-FSDP attribute loss, every Muon pattern, and AdamW exclusions remain checked." - ] - }, - { - "id": "TA-645", - "scope": "weight-sync parameter extraction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense ownership filtering and tied-weight alias handling now form one extraction policy.", - "Included names, duplicate suppression, declared aliases, false ties, and undeclared shared storage remain checked." - ] - }, - { - "id": "TA-646", - "scope": "weight-sync inference-unfuse layout fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "DeepSeek and Kimi MLA fusion, contiguous FP8 views, Nemotron-H publication, and gated stacked-expert splitting now form one layout-conversion policy.", - "Names, tensor values, storage aliasing, transposes, architecture prefixes, and fail-closed fused Nemotron input remain checked." - ] - }, - { - "id": "TA-647", - "scope": "adapter optimizer identity and live-binding fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical ordering, wrapper-insensitive fingerprints, and live optimizer parameter identity now form one ownership policy.", - "Exact names, equivalent fingerprints, and rejection before state access remain checked." - ] - }, - { - "id": "TA-648", - "scope": "adapter optimizer bitwise-resume control fragment", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The weights-only divergence control now executes inside the bitwise resumed-versus-uninterrupted trajectory.", - "Parameter and moment equality plus the proving divergence without optimizer state remain checked." - ] - }, - { - "id": "TA-649", - "scope": "public adapter optimizer LR-restore fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Scheduled-LR restoration after eviction and explicit LR override now form one public resume policy.", - "Registration generation, ownership fingerprint, restored metadata, optimizer-group LR, next-step parameters, and moments remain checked." - ] - }, - { - "id": "TA-650", - "scope": "adapter optimizer logical-reshard fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Successful logical resharding and every invalid-source rejection now form one topology-transition policy.", - "One- and two-dimensional slices, replicas, same-world changes, holes, overlaps, dtype and shape drift, step mismatch, empty ranks, and resident-state atomicity remain checked." - ] - }, - { - "id": "TA-651", - "scope": "FP8 linear injection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Core replacement, recipe overrides, and exclusions now form one model-injection policy.", - "Parameter identity, module FQNs, global and per-module recipes, unknown-key rejection, explicit exclusions, and glob exclusions remain checked." - ] - }, - { - "id": "TA-652", - "scope": "block-FP8 GEMM backend fallback fragment", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Automatic scaled-matmul failure and warn-once Triton fallback now execute inside the backend and scale-layout policy.", - "Block and row scales, explicit Torch backend parity, fallback equality, and warning suppression remain checked." - ] - }, - { - "id": "TA-653", - "scope": "FP8 linear CUDA execution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Master-weight training and float32-output execution now form one CUDA FP8 lifecycle.", - "FP8 dispatch, finite gradients, parameter updates, and requested output dtype remain checked." - ] - }, - { - "id": "TA-654", - "scope": "inference endpoint automatic-sync fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Detected TP size and configured synchronization method now form one endpoint auto-sync policy.", - "Receiver discovery, normalized world size, endpoint payload, successful registration, and P2P method forwarding remain checked." - ] - }, - { - "id": "TA-655", - "scope": "inference endpoint weight-sync routing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Single-endpoint forwarding and default, named, or unmatched pool selection now form one routing policy.", - "Endpoint payloads, model ID, timing, rank summaries, all pool outcomes, and no-dispatch failure remain checked." - ] - }, - { - "id": "TA-656", - "scope": "dense and sequence-parallel adapter autograd launchers", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two-GPU dense and sequence-parallel ownership workers now report as one foundational autograd policy.", - "Both distributed subprocesses, analytical optimizer comparisons, and certification markers still execute." - ] - }, - { - "id": "TA-657", - "scope": "unquantized expert adapter all-to-all launchers", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP2 backend parity, projection-subset ownership, and the all-owner layout now form one unquantized all-to-all policy.", - "Eager, Triton, native, and Quack backends, full and down-only targets, structural zeros, and public optimizer steps remain checked." - ] - }, - { - "id": "TA-658", - "scope": "unquantized expert adapter eFSDP launchers", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Four-GPU shared-owner and all-owner Quack layouts now form one eFSDP topology policy.", - "Both distributed workers and their distinct certification markers remain checked." - ] - }, - { - "id": "TA-659", - "scope": "quantized expert adapter all-to-all launchers", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Backend-format parity and the quantized projection subset now form one EP2 all-to-all policy.", - "Triton NF4, native NVFP4, Quack block-FP8, down-only NF4, structural zeros, and optimizer parity remain checked." - ] - }, - { - "id": "TA-660", - "scope": "SGLang fused-MoE trainable numerical fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Stock-Triton gradient parity and masked-expert gradient semantics now form one trainable numerical policy.", - "Forward inputs, router weights, all parameter gradients, compacted references, masked zeros, and fully masked behavior remain checked." - ] - }, - { - "id": "TA-661", - "scope": "SGLang fused-MoE real parity fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Strided-versus-transient forward and gradient parity now shares one real-kernel policy with automatic-versus-explicit dispatch parity.", - "All output and gradient tensors, repeat determinism, explicit flag behavior, and stock output shape remain checked." - ] - }, - { - "id": "TA-662", - "scope": "canonical-MoE contributor-width parametrization", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Reference-tree widths and distributed contributor widths now execute inside their respective numerical and transport policies instead of reporting each width as a separate test.", - "The 2- and 16-contributor reference trees and the 2- and 8-process distributed workers all still execute." - ] - }, - { - "id": "TA-663", - "scope": "canonical-MoE transport admission fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Automatic selection, explicit-mode rejection, and direct packed or sharded reducer guards now form one transport-admission policy.", - "Admitted eager EP16 behavior, dense fallbacks, graph and consumer-output restrictions, topology rejection, and direct executor fail-closed behavior remain checked." - ] - }, - { - "id": "TA-664", - "scope": "world-32 canonical-MoE group-alias examples", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP8/eFSDP4 and EP16/eFSDP2 now form one world-32 topology policy rather than two layout-named tests.", - "Context-parallel aliases, expert-parallel aliases, and the first and last expert-FSDP groups remain exact for both layouts." - ] - }, - { - "id": "TA-665", - "scope": "Quack TP FP8 positional-forwarding mock", - "decision": "remove", - "status": "applied", - "evidence": [ - "The test only inspected positional arguments passed to an internal autograd function and returned a mocked tensor.", - "The retained CUDA TP lifecycle now executes the same FP8 backend and non-default block size through forward, backward, finite-gradient checks, and a master-weight update." - ] - }, - { - "id": "TA-666", - "scope": "prequantized checkpoint format-detection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "NVFP4 and block-FP8 detection now form one checkpoint-format policy rather than separate format-named tests.", - "Nested, flat, config, index, precedence, malformed, missing, wrong-block-size, and competing-format cases all remain checked." - ] - }, - { - "id": "TA-667", - "scope": "prequantized checkpoint-handler exclusion fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense and MoE exclusion behavior now forms one checkpoint-handler policy.", - "Weight and auxiliary-key passthrough, nonexcluded skipping, empty exclusions, shared experts, and on-load consistency remain checked." - ] - }, - { - "id": "TA-668", - "scope": "packing-concat sequence side-channel fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Teacher hidden states and hidden-match weights now report as one sequence-side-field collation policy.", - "Rank-two and rank-one concatenation, padding to a multiple of four, exact values, shapes, and zero padding remain checked." - ] - }, - { - "id": "TA-669", - "scope": "Quack compile-worker receive failure fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Silent-worker timeout and truncated-body rejection now form one receive-protocol policy.", - "The real pipe timeout bound and exact malformed-frame exception remain checked." - ] - }, - { - "id": "TA-670", - "scope": "Quack cache-key hashing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Structural determinism, unsafe-object rejection, and Cutlass dtype-class support now form one cache-key policy.", - "Tuple boundaries, type distinctions, disabled pickle hooks, deterministic hashes, and metaclass-independent dtype handling remain checked." - ] - }, - { - "id": "TA-671", - "scope": "Qwen3 pipeline-schedule parity parametrization", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The three candidate schedules now share one 1F1B baseline run instead of training the identical baseline once per parameterized report.", - "Interleaved1F1B, InterleavedZeroBubble, and ZBVZeroBubble still run independently with convergence and per-step loss-parity checks." - ] - }, - { - "id": "TA-672", - "scope": "NVFP4 tensor-quantization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed layout, dequantization error, and shared-global-scale behavior now form one tensor-quantization policy.", - "Packed shapes and dtypes, scale geometry, scalar global scale, relative error, and exact cross-tensor scale reuse remain checked." - ] - }, - { - "id": "TA-673", - "scope": "NVFP4 directory-export fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Weight-only export, W4A4 input-scale stamping, and requantization rejection now form one directory-export lifecycle.", - "Fused scales, BF16 islands, metadata, roundtrip error, calibrated input scales, uncalibrated omission, and fail-closed re-export remain checked." - ] - }, - { - "id": "TA-674", - "scope": "data-preparation retry fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Immediate success, retryable and terminal failures, and all backoff strategies now form one retry policy.", - "ReadTimeout, HfHubHTTPError, retry exhaustion, unrelated exceptions, exponential, linear, and constant timing remain checked; the production helper now imports HfHubHTTPError from its stable public module." - ] - }, - { - "id": "TA-675", - "scope": "OLMo2 construction and TP-layout fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HF-config construction and unfusing into the HF parameter layout now form one architecture-layout policy.", - "Post-norm structure, full-axis QK norms, fused bias rules, split attention and MLP modules, and checkpoint-handler removal remain checked." - ] - }, - { - "id": "TA-676", - "scope": "OLMo2 checkpoint save and load fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HF-compatible export and HF-to-fused import now form one bidirectional checkpoint policy.", - "All attention, norm, MLP, fused-key, strict loading, hidden-state, and logits assertions remain checked." - ] - }, - { - "id": "TA-677", - "scope": "Qwen2 construction and TP-layout fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HF-config construction and TP unfusing now form one Qwen2 architecture-layout policy.", - "Norm absence, bias rules, split attention and MLP modules, and checkpoint-handler removal remain checked." - ] - }, - { - "id": "TA-678", - "scope": "Qwen2 checkpoint save and load fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HF-compatible export and HF-to-fused import now form one bidirectional checkpoint policy.", - "Weight and bias keys, fused-key construction, strict loading, hidden-state parity, and logits parity remain checked." - ] - }, - { - "id": "TA-679", - "scope": "server cu-seqlen SP admission fragment", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The SP-enabled no-emission case now executes inside the server-versus-CLI alignment policy.", - "Two, three, single, and many-sequence boundaries, int32 dtype, maximum lengths, and SP ownership remain checked." - ] - }, - { - "id": "TA-680", - "scope": "orchestrator client communication fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Basic health roundtrip, repeated requests, and interleaved operation types now share one live ZeroMQ communication lifecycle.", - "Connection state, request IDs, response types, finished state, engine receipt, and output transmission remain checked without a second fixture startup." - ] - }, - { - "id": "TA-681", - "scope": "non-observing distributed data-loader examples", - "decision": "remove", - "status": "applied", - "evidence": [ - "A literal 4 times 3 equals 12 assertion did not invoke data-loader code, and the claimed epoch-consistency block only compared two list lengths fixed to three by construction.", - "Real partitioning, microbatching, sequence sharding, padding, drop-last, packed, multi-DP, and variable-length behaviors remain checked." - ] - }, - { - "id": "TA-682", - "scope": "joined multiprocessing outcome detection", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The audit now recognizes torch multiprocessing start_processes as an observable outcome because joined workers propagate child assertion and process failures.", - "Sequence-parallel gradient reduction, four-rank exact-DCP fusion, and DTensor materialization wrappers no longer appear as assertion-free candidates; all three executable wrappers pass." - ] - }, - { - "id": "TA-683", - "scope": "Muon and BI golden assertion-helper signaling", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Muon transition and frozen-bit golden helpers now use explicit assert-prefixed names so their wrapper outcomes are visible to the audit.", - "All three Muon topology transitions pass, and every H100-specific golden hash assertion remains unchanged." - ] - }, - { - "id": "TA-684", - "scope": "exact LM-head scalar optimizer-state acceptance outcome", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The acceptance test now explicitly asserts the validator's successful None result instead of relying only on absence of an exception.", - "Scalar tensor step state, replicated parameter ownership, LM-head TP group selection, and the coherence call remain unchanged and pass." - ] - }, - { - "id": "TA-685", - "scope": "optional QARL triton-w4a4 registry probe", - "decision": "remove", - "status": "applied", - "evidence": [ - "The CPU-marked probe usually skipped and otherwise checked only that two implementation-detail dictionary entries existed and one was callable.", - "QARL shadow selection, activation fake quantization, restoration, and the real W4A4 execution paths remain covered." - ] - }, - { - "id": "TA-686", - "scope": "group-GEMM and MoE kernel dependency gates", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Unsupported CPU hosts now skip at collection-time markers, while import failures in XORL's own grouped-GEMM and MoE kernel modules fail supported GPU runs instead of being converted into dependency skips.", - "All six retained CUDA kernel policies execute and pass on the audit host." - ] - }, - { - "id": "TA-687", - "scope": "non-gated MoE core import guard", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The suite now imports its core XORL expert class normally instead of catching every import-time exception and skipping affected tests.", - "All five retained CPU and CUDA backend policies execute and pass." - ] - }, - { - "id": "TA-688", - "scope": "SGLang missing-dependency diagnostic", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The negative test now simulates an unavailable SGLang import deterministically instead of skipping whenever SGLang is installed.", - "The test exposed and corrected a diagnostic that named the unrelated TP-simulation flag rather than XORL_MOE_SGLANG_FUSED_EXPERTS." - ] - }, - { - "id": "TA-689", - "scope": "personal-path repository hygiene guards", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The examples-only scanner duplicated the repository-wide home-path check and performed a second git traversal.", - "Mac home and personal data-workspace patterns now live in the stronger repository-wide guard, and the narrower file is removed." - ] - }, - { - "id": "TA-690", - "scope": "standalone runner-dispatcher forward model-id test", - "decision": "remove", - "status": "applied", - "evidence": [ - "The standalone test repeated the rank-zero forward handler scenario already covered in the request-processor suite.", - "The retained test additionally checks routed expert ids, routed logits, auto-load selection, rank ownership, and the returned session id." - ] - }, - { - "id": "TA-691", - "scope": "weight-version forwarding fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two mock-to-mock forwarding tests are replaced by one composed handler-to-NCCL-synchronizer policy.", - "Bucket accounting, cache mode, and the exact weight version are verified at the final transfer seam." - ] - }, - { - "id": "TA-692", - "scope": "HSDP microbatch all-reduce fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Deferral, last-microbatch restoration, and the non-replicated rejection now form one gradient-sync policy.", - "Every original set_requires_all_reduce transition remains asserted." - ] - }, - { - "id": "TA-693", - "scope": "direct runtime-rank MoE LoRA scaling unit", - "decision": "remove", - "status": "applied", - "evidence": [ - "The direct one-by-one delta check was subsumed by the retained inference-buffer test that invokes the same helper for gate, up, and down projections.", - "The retained policy checks active-rank scaling, emitted names, shapes, dtype, values, and source cleanup." - ] - }, - { - "id": "TA-694", - "scope": "SGLang RMSNorm mode fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Global mode selection and forced-residual FP32 weight multiplication now report as one numerical mode policy.", - "Global state restoration and the exact reference calculation remain checked." - ] - }, - { - "id": "TA-695", - "scope": "index-share caller cleanup fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Offline trainer and server runner failure cleanup now execute in one cross-caller lifecycle policy.", - "Both failure messages, mode handoffs, and retained-context release counts remain asserted." - ] - }, - { - "id": "TA-696", - "scope": "importance-sampling metric reduction fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default ratio aggregation and custom TIS extrema now form one no-DP metric policy.", - "Weighted means, extrema, valid-token aggregation, and Python-scalar output remain checked." - ] - }, - { - "id": "TA-697", - "scope": "GLM LoRA target resolution fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Raw-HF default inference and explicit-target precedence now share one resolution policy and one model config fixture.", - "All five default attention targets and the exact explicit override remain asserted." - ] - }, - { - "id": "TA-698", - "scope": "remote backend RPC wrapper fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Weight-sync timeout forwarding and optimizer sparse-delta capture now execute in one operation-payload policy.", - "Operation names, request ids, timeout, pause and cache modes, endpoints, and sparse-delta fields remain checked." - ] - }, - { - "id": "TA-699", - "scope": "DSv4 RoPE cache-length fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Config-derived length and environment override precedence now execute in one cache-sizing policy.", - "The independent context-parallel short-cache rejection remains separate." - ] - }, - { - "id": "TA-700", - "scope": "SGLang JIT and kernel RMSNorm CPU fallback fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "JIT and kernel mode residual numerics plus packed-shape handling now form one CPU fallback matrix.", - "Both exact residual comparisons, the packed output shape, and global mode restoration remain checked." - ] - }, - { - "id": "TA-701", - "scope": "nonresident LoRA kill-session checkpoint fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Missing-checkpoint preservation and later evicted-checkpoint promotion now form one session lifecycle on the same runner.", - "Path traversal rejection remains an independent security boundary." - ] - }, - { - "id": "TA-702", - "scope": "QARL calibration batch loading fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Truncation, calibration-size limiting, attention-mask alignment, and malformed token-shape rejection now form one input policy.", - "The model calibration and persistent-state roundtrip remains a separate lifecycle." - ] - }, - { - "id": "TA-703", - "scope": "RMSNorm SGLang CPU mode modules", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Global SGLang selection, forced-residual FP32 multiplication, JIT fallback, kernel fallback, and packed inputs now live in one mode policy.", - "The redundant one-test JIT module is removed and the original global mode is restored atomically." - ] - }, - { - "id": "TA-704", - "scope": "NCCL rendezvous port fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sticky ephemeral-port rotation and explicit-port pinning now execute in one training rendezvous lifecycle.", - "Store creation, process-group destruction, active-port state, and bind-failure admission remain checked." - ] - }, - { - "id": "TA-705", - "scope": "FutureStore test-local response helper assertions", - "decision": "remove", - "status": "applied", - "evidence": [ - "The removed assertions exercised response-builder functions defined inside the test file rather than production code.", - "Production FutureEntry defaults, expiry, terminal states, and queue-state transitions remain checked." - ] - }, - { - "id": "TA-706", - "scope": "FQN matcher utility fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Single, all, and any matching now form one pattern-policy report instead of three one-function reports.", - "Exact, wildcard, grouped-number, indexed, prefixed, empty, first-match, and invalid-input cases all remain." - ] - }, - { - "id": "TA-707", - "scope": "block-FP8 quantization imports and input fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Contiguity and divisibility rejection now execute inside the quantization shape, dtype, scale, and block-size policy.", - "Imports of XORL's own block-FP8 module now fail visibly instead of being converted into an unavailable-feature skip; two unused imports were removed." - ] - }, - { - "id": "TA-708", - "scope": "dataset source loading fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local file, saved directory, hub, URL, missing source, and string or list data_files now form one source-resolution policy.", - "Download counts and every original source-selection assertion remain checked." - ] - }, - { - "id": "TA-709", - "scope": "exact DCP skip-mode fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact-model deferral and non-exact-model rejection now form one load-mode admission policy.", - "FSDP deregistration and the prohibition on unintended HF shard reads remain asserted." - ] - }, - { - "id": "TA-710", - "scope": "inert sparse-delta SGLang compatibility module", - "decision": "remove", - "status": "applied", - "evidence": [ - "A module-level guard skipped every test because its zstd case targeted disk_compression arguments absent from the production writer.", - "The retained sparse-delta file and backend suites pass, and the stronger trainer-to-request-processor-to-SGLang E2E owns receiver application, checksum, validate-only, and final parameter parity." - ] - }, - { - "id": "TA-711", - "scope": "launcher rank-zero address precedence fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Remote discovery and explicit engine-host precedence are two outcomes of the same address-selection policy.", - "The retained report checks the discovery call and the explicit-host short circuit." - ] - }, - { - "id": "TA-712", - "scope": "launcher server-override parsing fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Schema-agnostic parsing and removed-field validation are consecutive boundaries of one override policy.", - "Arbitrary parsed values and the ZORL migration diagnostic remain asserted." - ] - }, - { - "id": "TA-713", - "scope": "DistSignSGD local hook ownership fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local-parameter registration and FSDP-managed exclusion now execute in one ownership report.", - "The retained policy checks hook installation, configuration state, and absence of a duplicate managed-parameter hook." - ] - }, - { - "id": "TA-714", - "scope": "NVFP4 QARL normalization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Alias defaults, activation override, and every invalid block size are outcomes of one normalization policy.", - "Weight-only defaults, fixed group size, explicit activation, and all original rejection literals remain checked." - ] - }, - { - "id": "TA-715", - "scope": "direct relu2 activation-registry probe", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained non-gated MoE reference test constructs relu2 experts and proves exact forward and gradient behavior against relu squared.", - "A separate assertion that relu2 appears in internal dictionaries provided no stronger production regression signal." - ] - }, - { - "id": "TA-716", - "scope": "server Adam configuration and initializer fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "ServerArguments conversion now feeds the ModelRunner initializer in the same test instead of ending at an intermediate dictionary.", - "Non-default betas and epsilon, defaults, optimizer parameter groups, and malformed-beta rejection remain checked." - ] - }, - { - "id": "TA-717", - "scope": "Pydantic API field-echo and automatic roundtrip reports", - "decision": "remove", - "status": "applied", - "evidence": [ - "Two broad reports primarily asserted that Pydantic returned constructor fields and rejected omitted required fields.", - "Real training, checkpoint, sampler, and session endpoint tests construct these models; the unique forward session-id alias was retained in the compatibility policy." - ] - }, - { - "id": "TA-718", - "scope": "runner protocol constructor and default-factory fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct payload-field echoes, UUID uniqueness, timestamp range, and optional-None assertions are replaced by one full typed-wire equality contract.", - "All operation payloads, success and error responses, tensors, JSON, ACK correlation, and pickle rejection remain checked." - ] - }, - { - "id": "TA-719", - "scope": "duplicate API-orchestrator request-response flow", - "decision": "remove", - "status": "applied", - "evidence": [ - "The removed flow repeated subsets of the adjacent request and output roundtrips, builders, validators, and streaming-error checks.", - "The retained orchestrator integration suite separately exercises real queue processing and request identity across the live protocol seam." - ] - }, - { - "id": "TA-720", - "scope": "direct EP checkpoint-mesh selection fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy and PP-parent selection are now exercised through the production restore consumer instead of separate direct helper reports.", - "Both selected shapes, named dimensions, placements, ModelState delegation, and malformed-dimension rejection remain checked." - ] - }, - { - "id": "TA-721", - "scope": "DeepEP buffer-size validation fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "RDMA overflow, no-RDMA allowance, 128-byte alignment, and the default two-GB case are one buffer admission policy.", - "Every original size, RDMA allocation, expected byte count, and overflow diagnostic remains checked." - ] - }, - { - "id": "TA-722", - "scope": "weight-sync quantization normalization fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "BF16 no-ops and valid FP8 forms now share one supported normalization policy; unsupported methods and malformed FP8 forms share one rejection policy.", - "All aliases, defaults, exclusions, invalid formats, activation schemes, scale storage, and contextual diagnostics remain checked." - ] - }, - { - "id": "TA-723", - "scope": "direct OPD teacher-sort helper probe", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Nested teacher_id, nested teacher_ids, and top-level precedence now execute through the real OPD model-pass sorting branch.", - "The retained transaction additionally proves that teacher sorting composes with packer datum order and Mooncake routing-payload order." - ] - }, - { - "id": "TA-724", - "scope": "SignSGD sparse-step and base-optimizer state fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense updates, decoupled decay, missing gradients, and sparse rejection now form one SignSGD step policy.", - "A separate report of PyTorch Optimizer state-dict behavior was removed; builder admission and parameter-group ownership remain independent." - ] - }, - { - "id": "TA-725", - "scope": "direct FP32 routing-scale helper probe", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained local fused-MoE backward oracle uses the same discriminating BF16 gradient and small FP32 routing value.", - "It compares every input, routing, gate-up, and down gradient exactly, so a pre-multiply BF16 routing cast still fails through production autograd." - ] - }, - { - "id": "TA-726", - "scope": "sparse-MLA machine-specific speed assertion", - "decision": "relocate", - "status": "applied", - "evidence": [ - "The production-shape H100 timing ratio is hardware certification rather than repository correctness and no longer runs under pytest.", - "An explicit certification script retains combined-versus-split warmup, median timing, environment restoration, and a configurable speedup gate; all three numerical kernel reports remain." - ] - }, - { - "id": "TA-727", - "scope": "unasserted vocab-parallel CE benchmark inside distributed correctness", - "decision": "relocate", - "status": "applied", - "evidence": [ - "The pytest worker previously ran 100 production-scale warmup and timed forward or backward iterations only to print tables after correctness had passed.", - "The retained two-rank test now ends after eager and compiled value and gradient parity; an explicit torchrun certification script owns timing and peak-memory reporting." - ] - }, - { - "id": "TA-728", - "scope": "single-GPU dense FP8 CLI training smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained checkpoint-and-resume test begins with the same eager one-GPU two-step FP8 training configuration and the same finite-loss, gradient, and module-usage assertions.", - "The stronger retained test passed both training phases and additionally verifies DCP checkpoint creation and resume." - ] - }, - { - "id": "TA-729", - "scope": "basic two-GPU DistSignSGD training smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained four-GPU test exercises the same FSDP2, gradient-accumulation, five-step optimizer path while adding Ulysses exact-sum and replicated data-parallel composition.", - "After selecting eager attention to avoid an unrelated tiny-shape FlashAttention compiler failure, the retained composition test passed; unit tests separately pin both no-SP and SP sign-reduction behavior." - ] - }, - { - "id": "TA-730", - "scope": "Nemotron-H packed all-block-types smoke", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed sequence boundaries now flow through the retained all-mixer loss and backward contract instead of a second finite-output smoke.", - "The retained test checks loss plus finite nonzero gradients for Mamba, attention, routed experts, latent projections, shared experts, and embeddings." - ] - }, - { - "id": "TA-731", - "scope": "Qwen3 unfused forward-shape smoke", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two independently initialized MLPs only proved that fused and unfused implementations returned their declared hidden shape; they did not compare values.", - "Direct projection replacement and model-wide unfuse ownership now form one CPU policy covering attention and MLP modules in every layer." - ] - }, - { - "id": "TA-732", - "scope": "fragmented PP NCCL sender and receiver protocol tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One composed mocked roundtrip now captures the sender's real metadata and flattened payload and replays both directly through the receiver.", - "Names, matrix, vector, and scalar shapes, BF16 storage, and exact reconstructed values remain checked; the empty-payload protocol remains separate." - ] - }, - { - "id": "TA-733", - "scope": "duplicate families-v2 fused-versus-split norm comparison", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained norm contract explicitly forces both realizations, proves the split implementation was reached, and compares more hidden sizes plus residual, plain, and zero-centered variants.", - "The retained dispatch module still independently proves that shipped sizes select fused execution and deep shapes reach split execution." - ] - }, - { - "id": "TA-734", - "scope": "three one-rank exact-GLM FSDP2 component wrappers", - "decision": "remove", - "status": "applied", - "evidence": [ - "Each worker already supports world sizes one and two; its retained two-rank wrapper executes every shared lifecycle, byte-parity, ownership, and gradient assertion.", - "The two-rank branches additionally prove that LoRA factors are genuinely sharded. Both wrapper variants share the same optional SGLang import gate, which is unavailable in the repository venv." - ] - }, - { - "id": "TA-735", - "scope": "standalone families-v2 dispatch helper suite", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The retained norm policy now spies on the production split implementation while exercising shipped hidden sizes and a deep split shape, so it proves actual dispatch rather than only the private predicate.", - "All direct boundary cases remain in that covering policy, and the separate fused-versus-split numerical contract remains independent." - ] - }, - { - "id": "TA-736", - "scope": "direct optimizer-step learning-rate resolver report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit request, registered session, server training config, create-model registration, missing-value rejection, and both legacy fallbacks now drive the actual optimizer-step payload path.", - "The standalone report only called the private resolver with synthetic namespaces and is no longer needed." - ] - }, - { - "id": "TA-737", - "scope": "direct default Qwen3.5 rotary helper comparison", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained dense-and-MoE attention integration reaches the production QKV projection and compares its rotated Q/K values with the same Hugging Face half-rotation reference.", - "It also rejects the pairwise alternative, so the direct helper probe was a strict subset of actual attention behavior." - ] - }, - { - "id": "TA-738", - "scope": "seven fragmented dense Qwen3.5 RMSNorm reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two retained policy reports cover exact-family selection, fused serving families, explicit v2 admission and rejection, every v2 norm site, GDN separation, layer-input modes, and final-norm modes.", - "No norm case was discarded; the old reports were narrow fragments of the same construction and site-assignment contracts." - ] - }, - { - "id": "TA-739", - "scope": "direct dense and MoE Qwen3.5 config-conversion reports", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "One retained family policy writes realistic local config files and loads both variants through the production local auto-config entry point.", - "The covering path preserves registry admission, derived layer schedules, head geometry, linear-attention fields, mRoPE extraction, and MoE geometry assertions." - ] - }, - { - "id": "TA-740", - "scope": "direct Qwen3.5 checkpoint skip-regex report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained grouped checkpoint loader now uses the real Qwen3.5 skip-pattern constant while routing dense and expert shards.", - "Both top-level MTP forms and the non-MTP negative case are asserted inside the production prefetch filter transaction." - ] - }, - { - "id": "TA-741", - "scope": "duplicate families-v2 environment kill-switch report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained family-selection policy already iterates every trainer and sampler environment variable and verifies default-on plus per-variable rollback.", - "Runtime dispatcher rollback remains separately covered with a kernel-reachability spy." - ] - }, - { - "id": "TA-742", - "scope": "one-line no-shared-prefix repack report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The no-op fallback is now the opening case of the complete repack and remap policy.", - "Shared grouping, layout, position ids, loss fields, cumulative lengths, and output remapping remain covered in the same report; the P-equals-one edge remains independent." - ] - }, - { - "id": "TA-743", - "scope": "fragmented active-LoRA admission truth table and topology reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One composite-admission policy now covers the complete family, every missing component, and the independent scoring-only marker.", - "The production model build now proves both complete flag derivation and non-TP16 rejection in one topology transaction." - ] - }, - { - "id": "TA-744", - "scope": "direct BI router batch-invariance and default-path reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Batch composition invariance now runs through the production MoEBlock route method and checks logits, selections, and normalized weights.", - "Exact-contract selection and the ordinary BF16 route remain together as the two branches of one production dispatch policy." - ] - }, - { - "id": "TA-745", - "scope": "separate standard and temperature BI fused-LM-head oracle reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One forward-and-backward policy compares fused and eager execution at unit and non-unit temperatures.", - "Per-token log probabilities, ignored-token loss, aggregate loss, hidden gradients, and weight gradients remain asserted." - ] - }, - { - "id": "TA-746", - "scope": "three local step phase and memory summary helper fragments", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical ordering, custom ordering, empty input, numeric coercion, phase aggregates, and memory aggregates now form one local-finalization policy.", - "All prior edge cases and exact aggregate values remain covered." - ] - }, - { - "id": "TA-747", - "scope": "optimizer and weights Pydantic field-echo report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The removed assertions primarily repeated automatic constructor assignment, defaults, schema properties, and serialization.", - "The unique legacy optimizer session-id alias now drives the actual optim-step endpoint and payload; retained create-model, create-session, and weights endpoints cover the request and response models." - ] - }, - { - "id": "TA-748", - "scope": "separate multi-part optimizer selection and update reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One optimizer lifecycle now proves multi-part selection, parameter coverage, updates for every part, gradient clearing, scheduler propagation, and single-part fallback.", - "Custom parameter-group rejection remains an independent admission boundary." - ] - }, - { - "id": "TA-749", - "scope": "flat pseudo-packed tensor-collator report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The old report described equal-length flat examples as packed even though the collator intentionally preserves them as separate samples.", - "The retained policy now exercises the real already-batched dict and nested packed-dataset branches while preserving general conversion and dtype coverage." - ] - }, - { - "id": "TA-750", - "scope": "direct OPD list-chunking helper report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The helper is a guarded slicing comprehension, and the production pipeline only enables chunking for positive sizes.", - "The removed report merely repeated Python list slicing; payload alignment, endpoint registration, version verification, and pipeline lifecycle behaviors remain." - ] - }, - { - "id": "TA-751", - "scope": "sequence-parallel no-group identity report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The no-group branch directly returns the input tuple and the test asserted only object values through that one-line identity path.", - "Padding roundtrips and real two-rank gather backward reduction remain as the behavioral sequence-parallel reports." - ] - }, - { - "id": "TA-752", - "scope": "direct SGLang FP32 grouped-GEMM accumulator probe", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained local and EP custom-autograd reports execute the accumulator through the production backward functions.", - "Their exact independent routing-gradient oracles use values that distinguish FP32 accumulation from BF16 rounding, so the direct helper probe was a strict subset." - ] - }, - { - "id": "TA-753", - "scope": "train-router argument and configuration field echoes", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The server default now feeds MoEBlock.from_config and proves by backward behavior that the gate remains detached.", - "The same policy retains the enabled all-to-all gradient path and the unsupported DeepEP rejection." - ] - }, - { - "id": "TA-754", - "scope": "standalone packing-cache string-format report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The old report only repeated the interpolation fields of generate_packing_hash and missed its document-alignment branch.", - "The retained PackingDataset lifecycle now obtains real cache paths and proves ring-attention alignment creates a distinct cache identity." - ] - }, - { - "id": "TA-755", - "scope": "thin MD5 and SHA256 wrapper report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report retested standard-library digest constants and string encoding through one-line wrappers.", - "Dataset split and configuration fingerprint policies retain higher-level determinism, sensitivity, format, and order-independence coverage for the production MD5 consumer; the SHA256 wrapper has no production caller." - ] - }, - { - "id": "TA-756", - "scope": "BI GEMM post-import environment and pinned-constant tautologies", - "decision": "remove", - "status": "applied", - "evidence": [ - "Setting legacy environment variables after importing the module could not prove import-time independence, and the named variables have no production readers.", - "The block-K report asserted that lookup returned the same constant it directly injects; retained CUDA reports instead prove table bit neutrality, cross-bucket row invariance, and DeepGEMM parity." - ] - }, - { - "id": "TA-757", - "scope": "four fragmented EP adapter backend argument-boundary reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Native, Triton, Triton MoE-act, and Quack variants now form one capability-aware backend matrix.", - "Common-argument consumption, explicit unsupported-FP8 rejection, and Quack activation-native forwarding all remain; registry signatures and numerical expert-score forwarding stay independent." - ] - }, - { - "id": "TA-758", - "scope": "separate routing-replay wire decode and weight tensor reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One wire-to-tensor policy now decodes SGLang base64, inferred-shape, Python, NumPy, and tensor forms before exercising float weights, padding, and sequence-parallel slicing.", - "Ring-attention and general sequence-parallel layout contracts remain independent." - ] - }, - { - "id": "TA-759", - "scope": "separate OPD output-edge and hidden-only reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All-ignored stability, per-token output, and zero-KL hidden-only mode now form one output-edge policy.", - "Numerical backend parity, gradient reduction, and OPRD hidden-distance contracts remain separate." - ] - }, - { - "id": "TA-760", - "scope": "separate DCP synchronization and metadata process-group selector reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One process-group policy now covers cached Gloo creation under NCCL, no-op behavior under Gloo, PP-disabled metadata, caller precedence, and global fallback.", - "All prior selection branches and call-count assertions remain." - ] - }, - { - "id": "TA-761", - "scope": "separate ParallelState default and custom-construction reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Defaults, uninitialized properties, invalid modes, distributed rank discovery, topology validation, and enabled flags now form one construction policy.", - "Mesh construction and singleton initialization lifecycles remain independent." - ] - }, - { - "id": "TA-762", - "scope": "separate EP LoRA initialization and plan-slicing reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One lifecycle now follows global GKN factors from initialization and zero-B state through plan registration into exact rank-zero and rank-one expert slices.", - "The retained CUDA report independently exercises EP and non-EP forward and gradient behavior." - ] - }, - { - "id": "TA-763", - "scope": "separate accepted and rejected external FP8 configuration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One compatibility policy now admits the supported NeMo blockwise form and rejects every unsupported recipe or external runtime configuration.", - "Blackwell hardware admission and BF16 layer-island injection remain independent policies." - ] - }, - { - "id": "TA-764", - "scope": "separate QARL activation-override nesting report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One context-manager lifecycle now covers mixed prior values, both override directions, plain-module exclusion, exception restoration, and nested inner-first restoration.", - "No activation-override branch was discarded." - ] - }, - { - "id": "TA-765", - "scope": "separate QARL calibration loading and persistent-state reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One calibration lifecycle now follows truncated JSON input and malformed-shape rejection through model calibration, metadata population, state serialization, and restoration.", - "The retained assertions cover both input admission and durable quantization state." - ] - }, - { - "id": "TA-766", - "scope": "standalone NVFP4 normalization report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "NVFP4 aliases, activation selection, and invalid group sizes now feed the QARLLinear forward and STE policy instead of ending at normalized dictionary fields.", - "Dense-model injection remains an independent production composition report." - ] - }, - { - "id": "TA-767", - "scope": "separate NVFP4 activation forward and STE reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One autograd contract now proves exact dequantized NVFP4 forward values, actual lossiness, and identity straight-through gradients.", - "The forward and backward halves exercise the same activation fake-quant primitive." - ] - }, - { - "id": "TA-768", - "scope": "separate QARL MoE backend-shadow admission and exception reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One shadow-context policy now covers armed Triton W4A4 selection, exception restoration, activation-disabled Triton, and activation-enabled eager no-ops.", - "Every backend and restoration branch remains asserted." - ] - }, - { - "id": "TA-769", - "scope": "separate QARL sync-configuration and handler-derivation reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One training-to-sync lifecycle derives folded modules, quantizes only the QARL weights, preserves BF16 islands, and passes the same configuration through the production WeightSyncHandler.", - "Malformed caller-supplied QARL sync configuration remains a separate rejection boundary." - ] - }, - { - "id": "TA-770", - "scope": "standalone stochastic-rounding seeded-generator report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Seeded-generator reproducibility now belongs to the same call contract as shape, device, output dtype, and invalid-input admission.", - "Unbiased expectation and adjacent-BF16-neighbor properties remain independent statistical and numerical policies." - ] - }, - { - "id": "TA-771", - "scope": "separate Triton QARL MoE weight-quantization disabled report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One real Triton policy now compares the quantized lossy forward and STE gradients with an exact unquantized passthrough from the same seeded weights and routing.", - "This removes a second backend launch that repeated setup without a distinct failure domain." - ] - }, - { - "id": "TA-772", - "scope": "fragmented API training response reports and duplicate legacy optimizer payload", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forward auto-load information and forward-backward executor timings now form one API response-projection policy.", - "The focused optimizer report retains response telemetry and current-control mapping; the broader optimizer-step policy already owns legacy aliases, Adam fields, defaults, precedence, and missing-value rejection." - ] - }, - { - "id": "TA-773", - "scope": "standalone RequestProcessor lifecycle and statistics reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One live processor now starts ready, carries three real forward and forward-backward operations, exposes their counters, and then stops.", - "The removed reports repeated processor construction or a model pass solely to inspect readiness and counter fields." - ] - }, - { - "id": "TA-774", - "scope": "scheduler default-field, repr, and raw FIFO helper report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report asserted constructor constants, class names, and low-level deque methods without carrying a schedulable request through execution.", - "Retained Scheduler transactions prove FIFO dispatch, capacity, pending removal, and clear behavior through the production boundary." - ] - }, - { - "id": "TA-775", - "scope": "separate scheduler terminal-state and statistics-history reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One lifecycle now completes, fails, aborts pending and actually running requests, checks terminal statistics, clears state, and verifies bounded history.", - "The old running-abort block dispatched an older FIFO item and therefore silently aborted its named request while it was still pending." - ] - }, - { - "id": "TA-776", - "scope": "GPU-marked MicroBatchCollator order and uneven-size report in the distributed data-loader suite", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report created no process group, rank, sampler, or sequence shard and duplicated the CPU micro-batch splitting and error contract.", - "Distributed partitioning, sequence sharding, and packed-pipeline reports remain independent." - ] - }, - { - "id": "TA-777", - "scope": "standalone packed position-id and missing-label report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Generated ignored labels, document-position resets, and caller-position replacement now execute after the retained pack-to-unpack roundtrip.", - "The behavior is metadata for the same packed transaction, not an independent failure domain." - ] - }, - { - "id": "TA-778", - "scope": "standalone sequential-packing legacy-layout report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact legacy greedy grouping now begins the all-strategy correctness and utilization policy.", - "Best-fit and balanced-DP document preservation remain checked against that sequential baseline in the same report." - ] - }, - { - "id": "TA-779", - "scope": "duplicate mixed-valid and oversized skip assertion in generic packing edge cases", - "decision": "remove", - "status": "applied", - "evidence": [ - "The packing-strategy admission contract already owns the exact mixed-input legacy-drop behavior.", - "The generic edge policy retains the distinct all-skipped rejection, empty, single, missing-input, and NumPy cases." - ] - }, - { - "id": "TA-780", - "scope": "standalone orchestrator statistics and health-classifier report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Scheduler and processor statistics are now asserted after the retained forward, optimizer, and end-to-end health transactions.", - "Callable-existence checks, repeated read-only getter calls, and private classifier truth tables added no behavior beyond that real health response." - ] - }, - { - "id": "TA-781", - "scope": "orchestrator integration abort block that asserted only the original request produced output", - "decision": "remove", - "status": "applied", - "evidence": [ - "The DummyBackend commonly completed before the abort arrived, and the assertion neither checked abort acknowledgement nor an aborted terminal state.", - "The scheduler lifecycle now deterministically proves pending and active abort semantics." - ] - }, - { - "id": "TA-782", - "scope": "standalone MoE weight-sync bucket-size precedence report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Transport defaults and environment precedence now select the cap used by the retained byte-based bucket split policy.", - "Oversized-first-item behavior remains asserted in the same production helper lifecycle." - ] - }, - { - "id": "TA-783", - "scope": "standalone endpoint cache-metadata normalization report", - "decision": "remove", - "status": "applied", - "evidence": [ - "Current cache_epoch metadata now returns through a complete streaming FP8 sync, including post-process configuration and transfer flags.", - "Legacy cache_version normalization remains covered by the complete sparse-delta sync response." - ] - }, - { - "id": "TA-784", - "scope": "standalone compile-wrapper weight-name normalization report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Compiled expert, broadcast, and Qwen linear-attention names now normalize inside the full inference-layout unfusion policy.", - "The retained assertions verify emitted receiver names and tensor values, not only the string helper." - ] - }, - { - "id": "TA-785", - "scope": "direct P2P warm-mode selector report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Warm-only mode now performs real buckets after cached and cold prepare states.", - "The fake Mooncake engine observes async submission only for cached prepare and synchronous transfer for cold prepare." - ] - }, - { - "id": "TA-786", - "scope": "direct P2P small-entry transfer helper report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Three GPU-direct entries now pass through transfer_bucket and flush twice with a two-entry Mooncake chunk limit.", - "Observed engine calls prove the 2-1 chunk sequence and receiver addresses through the production worker." - ] - }, - { - "id": "TA-787", - "scope": "direct persistent-source interval registration report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Repeated GPU-direct buckets backed by one stable allocation now register it exactly once and reuse it on the second bucket.", - "Backend destruction proves the persistent range is deregistered during cleanup." - ] - }, - { - "id": "TA-788", - "scope": "P2P direct-EP capability and sender-rank field echoes", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report stopped at constructor properties for implicit and explicit sender ranks.", - "Retained direct-EP policies use those configurations for filtered scatter, dense partitioning, process-group collectives, failure propagation, prewarming, and rank-owned transfer." - ] - }, - { - "id": "TA-789", - "scope": "fragmented EP checkpoint mesh selector, ModelState caller, and restore reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One restore policy now rejects ambiguous meshes, routes ModelState through named dimensions, and constructs the final DTensor across legacy and PP parent layouts.", - "Dropping the EP dimension remains an independent reverse-conversion report." - ] - }, - { - "id": "TA-790", - "scope": "separate checkpoint materialization success and zero-meta failure reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One load-state policy now proves successful materialization, missing-optimizer omission, restored counters, and both pre-restore and post-restore meta-storage rejection.", - "Each failure still asserts whether the checkpointer was reached." - ] - }, - { - "id": "TA-791", - "scope": "separate ModelRunner initial-load and restore-completion wrapper reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The production restore wrapper now calls the real initial-load method for optimizer-enabled and optimizer-disabled policies and synchronizes both counters.", - "A failing checkpoint manager proves the completion flag stays false after the actual load call raises." - ] - }, - { - "id": "TA-792", - "scope": "three copies of canonical-LoRA sampler-weight export across API suites", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained checkpoint-path contract starts from a normalized session spec and verifies save_lora_only, model identity, destination path, and returned xorl URI.", - "Two smaller copies asserted subsets of the same APIServer method with equivalent fake responses." - ] - }, - { - "id": "TA-793", - "scope": "separate create-model conflicting-recreate and worker-registration-failure reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One admission and rollback policy now covers an existing incompatible session, a mismatched base repository, and a cross-rank registration failure.", - "Failed registration is still proven not to mutate model configuration or registered IDs." - ] - }, - { - "id": "TA-794", - "scope": "direct base-model canonicalizer truth-table report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Create-model registration now succeeds when the server uses an HF snapshot path and the client uses the equivalent repository ID.", - "The same endpoint policy rejects a snapshot path for a distinct repository." - ] - }, - { - "id": "TA-795", - "scope": "standalone last-inference-endpoint adapter-tracking cleanup report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sampler listing, path resolution, recency tracking, and last-receiver removal now form one adapter-tracking lifecycle.", - "Removal proves both endpoint state and receiver-derived adapter state are cleared." - ] - }, - { - "id": "TA-796", - "scope": "separate receiver quantization detection, enrichment, and set-policy reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One receiver policy now reads config.json, normalizes receiver names, enriches a caller FP8 config, and admits that exact result as the sync default.", - "MTP, static activation, UE8M0, compressed-tensors, and BF16 boundaries remain asserted in the same flow." - ] - }, - { - "id": "TA-797", - "scope": "FutureEntry constructor, manually assigned terminal states, and standalone queue-state report", - "decision": "remove", - "status": "applied", - "evidence": [ - "Real FutureStore jobs already prove pending, processing, completed, failed, and expired states plus results and classified errors.", - "Queue pause state now follows actual concurrent processing and statistics rather than a freshly constructed empty store." - ] - }, - { - "id": "TA-798", - "scope": "separate DeepSeek training-builder router rejection and successful-freeze reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One builder policy now rejects unfrozen, QLoRA, and unmerged-QKV modes before exercising the admitted frozen-router construction.", - "The admitted path proves router parameters are frozen while ordinary attention parameters remain trainable." - ] - }, - { - "id": "TA-799", - "scope": "direct DeepSeek tensor-parallel validator report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained build_parallelize_model policy invokes the same validator through its production consumer.", - "It still proves DeepSeek tensor parallelism fails before unsupported parallelization can proceed." - ] - }, - { - "id": "TA-800", - "scope": "duplicate full Trainer bootstrap setup for causal-loss lm-head mode", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One bootstrap policy now covers parallel-state initialization and both eager and Quack-linear causal-loss parameterization.", - "The Quack branch still proves lm_head_fp32 is omitted rather than passed to the unsupported loss implementation." - ] - }, - { - "id": "TA-801", - "scope": "separate disabled and enabled manual CUDA-timing reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One state-machine lifecycle now starts disabled, rejects an invalid mode, records forward and recompute phases, and drains the accumulator.", - "The fake-event path still proves unrecorded CUDA event pairs are omitted." - ] - }, - { - "id": "TA-802", - "scope": "direct LoRA dtype and generic-upcast helper rows already exercised by the model builder", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The production builder policy observes BF16 base weights, FP32 adapters, and skip_param_upcast for mixed-precision LoRA.", - "Only the distinct QLoRA, explicit-skip, and dense-default helper branches remain alongside that lifecycle." - ] - }, - { - "id": "TA-803", - "scope": "P2P expert FP8 production geometry and all-global-index certification sweeps", - "decision": "remove", - "status": "applied", - "evidence": [ - "Global expert names are produced by the size-independent ep_rank times local expert count plus local index formula, so eight full EP-rank transactions selected no new branch.", - "Two retained transactions cover partial blocks plus block-128 quantization, single and multiple receivers, and a nonzero EP offset while checking exact bytes, scales, and dequantized values." - ] - }, - { - "id": "TA-804", - "scope": "production-sized Qwen3.6 shared-expert P2P parity report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained shared-expert transaction exercises singular and plural receiver namespaces, fused gate/up placement, down projection placement, scales, and the passthrough gate.", - "The removed 2048-by-512 allocation repeated those same locator and transfer branches with larger tensors only." - ] - }, - { - "id": "TA-805", - "scope": "separate P2P compiled-name, language-model-prefix, and tied-lm-head lookup reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One receiver-name compatibility policy now performs real transfers for _orig_mod stripping and language_model prefix fallback.", - "The same policy proves a missing tied lm_head locator is skipped without duplicating the embedding transfer." - ] - }, - { - "id": "TA-806", - "scope": "separate direct-EP dense manifest and outgoing-buffer ownership reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One ownership policy assigns every dense family to exactly one sender, keeps fused and split projection aliases together, and applies that assignment to both receiver manifests and outgoing buffers.", - "Expert entries remain rank-owned in manifests and excluded from the dense buffer filter." - ] - }, - { - "id": "TA-807", - "scope": "duplicate nonzero-sender initialization setup for default and prewarmed P2P engines", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One initialization policy runs both environment modes against the same scattered tensor-map contract.", - "It proves default construction follows broadcast while prewarming constructs the engine before broadcast." - ] - }, - { - "id": "TA-808", - "scope": "separate direct-EP rank-filter routed and all-filtered transfer reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One transfer policy routes each slice only to its owning receiver for two ranks.", - "The same policy covers a non-owning rank whose valid bucket is fully filtered and therefore emits no engine transfer." - ] - }, - { - "id": "TA-809", - "scope": "separate Qwen3-MoE layer and final RMSNorm family-declaration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One construction policy now checks layer-zero, later-layer, and final-model norm families.", - "It still executes the decoder call sites and proves the final norm relies on its module declaration rather than a per-call override." - ] - }, - { - "id": "TA-810", - "scope": "direct delayed-TP-shard residual-association helper report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The real Qwen3-MoE decoder-layer path now receives BF16 shard values for which sum-then-residual differs from sequential residual addition.", - "The retained test observes the materialized input, norm call, residual output, and diagnostics through the production consumer." - ] - }, - { - "id": "TA-811", - "scope": "direct O-projection partial-residual mode helper report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Both sum-then-residual and residual-then-partials modes now run through Qwen3MoeDecoderLayer._pre_mlp_forward with discriminating BF16 values.", - "The layer test verifies the norm input, returned residual, output, diagnostic captures, and the expected bitwise association difference." - ] - }, - { - "id": "TA-812", - "scope": "standalone cross-engine RMSNorm funnel and test-discriminator reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "SGLang single-tensor and fused funnel outputs now sit inside the corresponding qk, pre-summed residual-tree, and post-attention module policies across every adversarial shape.", - "The rare family-difference discriminator now guards the qk site policy directly rather than reporting a test of the test." - ] - }, - { - "id": "TA-813", - "scope": "masked cross-engine RMSNorm trunk-flag report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The fixture explicitly declared the no-residual family, which selects the batch-invariant wrapper before the trunk-flag branch and made the flag unable to change dispatch.", - "The retained qk module policy reaches that wrapper and compares both the serving kernel and serving family funnel across all adversarial shapes." - ] - }, - { - "id": "TA-814", - "scope": "mocked distributed data-loader partitioning, micro-batch, and sequence-parallel report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The partitioning outcome was computed from rank-index lists constructed by the test itself; the injected sampler was never observed through the dataloader.", - "The retained builder policy verifies sampler rank and replica ownership plus SP pipeline insertion, while the MicroBatchCollator policy verifies exact split values and failures." - ] - }, - { - "id": "TA-815", - "scope": "shallow mocked sequence-sharding shape report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report asserted only equal shard lengths and a loose padding range after iterating mocked ranks.", - "The retained TextSequenceShardCollator policies check exact rank slices, non-divisible padding, attention metadata, labels, and token-aligned side channels through the production collator." - ] - }, - { - "id": "TA-816", - "scope": "mock-rank packed data-loader shape examples", - "decision": "remove", - "status": "applied", - "evidence": [ - "Changing mocked DP ranks while asserting the same output shape could not detect incorrect partition ownership or data overlap.", - "The retained real data-loader lifecycle covers packed and variable-length samples, while PackingConcatCollator owns exact concatenated values, position resets, extra fields, padding, and flash-attention metadata." - ] - }, - { - "id": "TA-817", - "scope": "direct fresh-manager checkpoint optimizer-selection report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report loaded a SignSGD checkpoint into an AdamW-default manager and asserted only the selected optimizer type and normalized session field.", - "The retained real AdapterCoordinator lifecycle performs that same transition through both explicit checkpoint loading and eviction auto-loading, while the multi-adapter lifecycle also reloads mixed optimizer types." - ] - }, - { - "id": "TA-818", - "scope": "one-layer row of the GLM semantic MoE stack parameterization", - "decision": "remove", - "status": "applied", - "evidence": [ - "Changing the synthetic stack from four MoE layers to one changed only repetition count and selected no distinct model or canonicalization branch.", - "The retained four-layer transaction checks every boundary, final logprob parity, batch permutation, per-row composition, and a deliberately omitted first-layer canonicalization that changes the final result." - ] - }, - { - "id": "TA-819", - "scope": "shape-only Qwen Triton expert forward report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report initialized a Triton expert, ran one random forward, and asserted only that the output shape matched the input shape.", - "The retained eager-versus-Triton MoE transaction exercises the same backend with numerical output comparison and gradients for every LoRA factor." - ] - }, - { - "id": "TA-820", - "scope": "synthetic DeepSeek-like MLA LoRA target stub report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Its default-target assertions duplicated the retained real DeepSeek model policy over the same MLA projection names.", - "The unique explicit-target partition case now runs through inject_lora_into_model_with_moe on the real DeepSeek model and proves the untargeted output projection remains unchanged." - ] - }, - { - "id": "TA-821", - "scope": "base MoE expert constructor field, shape, mapping, and registry echoes", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report instantiated wrappers and backend variants but asserted only stored strings, parameter shapes, mapping identity, and registry membership.", - "Retained eager, native, Triton, non-gated, injection, and model-construction policies execute those registrations and layouts; the separate LoRA initialization policy still protects frozen bases and trainable factor state." - ] - }, - { - "id": "TA-822", - "scope": "DSv4 KV-QAT helper boolean and private-field report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The removed report stopped after calling dsv4_kv_qat_enabled and reading DeepSeekV4Attention._kv_qat_enabled.", - "The retained C0 attention forward/backward transaction now supplies FP8 quantization configuration and observes the QAT call on the exact no-RoPE KV slice with block size 64." - ] - }, - { - "id": "TA-823", - "scope": "standalone DSv4 RoPE cache-length precedence shape report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The report called the cache builder directly and stopped after checking two tensor shapes.", - "The retained C0/C128 attention forward-backward policy now constructs and consumes the environment-sized cache, while the C128 context-parallel compressor forward requires the config-sized fallback to cover its rank-one slice." - ] - }, - { - "id": "TA-824", - "scope": "packed DeepSeek checkpoint handler type and private quantization-field report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The former report ended after checking the selected handler type and its private group-size and bit-width fields.", - "The retained packed-loader transaction now obtains the handler through the model, parses the official nested config, and proves 8-bit group-64 packed expert tensors dequantize to the expected gate, up, and down values." - ] - }, - { - "id": "TA-825", - "scope": "DSv4 compressor and indexer constructor-presence report", - "decision": "remove", - "status": "applied", - "evidence": [ - "C0 and C128 component selection is already exercised by the retained attention forward-backward transaction, including the C128 compressor path.", - "The retained C4 synthetic checkpoint load constructs both compressor and indexer and validates their separately translated APE tensors, so an incorrect topology cannot pass that lifecycle." - ] - }, - { - "id": "TA-826", - "scope": "external FLA Hopper autotuner regression report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report imported fla.ops.gated_delta_rule directly and never called any XoRL module, wrapper, or integration seam.", - "Its pass/fail result was controlled entirely by an optional dependency's production-shape backward kernel, so an upstream autotuner change was not an XoRL repository regression." - ] - }, - { - "id": "TA-827", - "scope": "standalone batch-invariant full-mean dtype spelling report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The dtype keyword reaches the same full-reduction override already owned by the mean-versus-sum regression contract.", - "That retained contract now checks the BF16 input to FP32 output spelling alongside one- and two-dimensional full reductions in both supported dtypes." - ] - }, - { - "id": "TA-828", - "scope": "production-size head-v2 projection and batch-invariance repetitions", - "decision": "remove", - "status": "applied", - "evidence": [ - "The 1024-hidden and 20480-vocabulary reports selected no additional head-v2 launch or merge branch beyond the retained focused geometry.", - "The retained head-v2 policies prove exact v1 projection bits, shared decode/scoring statistics, selected-logprob composition, arbitrary-slice batch invariance, fused-loss gradients, and rollback behavior." - ] - }, - { - "id": "TA-829", - "scope": "orphan eager OPRD layer-cache gather report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The private _get_opd_teacher_layer_hidden_states wrapper has no production caller; its only call site was the removed test.", - "The live OPRD path uses _get_opd_teacher_layer_fetcher, whose retained policy verifies selected cache rows, multiple layer slices, layer count, and returned shapes." - ] - }, - { - "id": "TA-830", - "scope": "legacy IS metric accumulator CPU and distributed reports", - "decision": "remove", - "status": "applied", - "evidence": [ - "ModelRunner._accumulate_is_metrics and _finalize_is_metrics have no production callers; runtime forward-backward now uses _accumulate_loss_metrics and _finalize_loss_metrics.", - "The retained OPD runner policy covers current mean, extrema, empty-rank, and loss-specific accumulation, while the retained two-rank report exercises the still-live _sp_allreduce_kl_metrics collective." - ] - }, - { - "id": "TA-831", - "scope": "EP kernel forward-only routing-score repetition", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report loaded the Triton and Quack EP kernel classes through test stubs and compared only their forward outputs with a local torch reference.", - "The retained EP routing-score report exercises the same two kernel classes and reference computation while checking both forward values and routing-score gradients." - ] - }, - { - "id": "TA-832", - "scope": "direct QARLLinear gradient and state-dict smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "The direct wrapper smoke repeated gradient presence, dynamic scale updates, and state restoration without exercising model injection or an optimizer step.", - "The retained QARL training lifecycle performs injection, a real AdamW update, changed-logprob checks, model checkpoint restoration, and exact restored logprobs; the calibration lifecycle independently checks persistent activation and block-scale state." - ] - }, - { - "id": "TA-833", - "scope": "private DSv4 FWHT helper reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The former reports separately tested a private helper's algebra, an impossible non-power-of-two DSv4 width, and fallback dispatch with only shape and finiteness assertions.", - "One retained public rotate_activation transaction now disables the optional kernel and proves the known transform, self-inverse behavior, and norm preservation across supported widths." - ] - }, - { - "id": "TA-834", - "scope": "launcher worker command with contradictory missing server arguments", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report bypassed Launcher construction, paired a valid flat config path with server_args=None, and asserted only that model_path was absent from a mocked subprocess command.", - "Normal construction either resolves ServerArguments from that config or raises before worker launch; retained launcher policies exercise live address selection, readiness failure, and override admission." - ] - }, - { - "id": "TA-835", - "scope": "DSv4 routing replay corrupted-global-state report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report directly overwrote the private module global with a fabricated future stage value that no production caller emits.", - "Retained DSv4 and generic routing-replay lifecycles exercise every supported stage through record, forward replay, backward replay, checkpoint recomputation, and R3 preload paths." - ] - }, - { - "id": "TA-836", - "scope": "activation-offload None-limit standalone backward smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "The standalone report passed None directly to a helper even though both trainer and server argument surfaces define activation_gpu_limit as a float, then asserted only that an unrelated square operation produced a gradient.", - "It did not observe offload placement, byte accounting, prefetch, or a production training path, so it could not discriminate activation-offload behavior." - ] - }, - { - "id": "TA-837", - "scope": "unreferenced FileLockLoader support suite", - "decision": "remove", - "status": "applied", - "evidence": [ - "FileLockLoader is not exported from xorl.data.prepare and has no caller or import anywhere else under src; its class name appeared only in its three-test module.", - "The retained data-preparation tests cover the live packing cache, dataset loading, hashing, retries, and preprocessing paths without preserving an unused counter-file abstraction." - ] - }, - { - "id": "TA-838", - "scope": "orphan runner-protocol JSON compatibility assertions", - "decision": "remove", - "status": "applied", - "evidence": [ - "Production runner communication exclusively uses serialize_message and deserialize_message; BaseMessage.to_json and BaseMessage.from_json have no runtime caller.", - "The retained protocol transaction still round-trips every live message payload through the actual transport codec, rejects pickle bytes, preserves tensors, and creates request acknowledgements." - ] - }, - { - "id": "TA-839", - "scope": "FutureStore test-only convenience accessor assertions", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "get_status, get_result, get_error, list_by_model, get_queue_state, and set_queue_state have no production caller; only the test used those convenience surfaces.", - "The retained lifecycle now observes live FutureEntry state through FutureStore.get while continuing to prove scheduling, concurrency, result and failure storage, deletion, model cleanup, and TTL expiry." - ] - }, - { - "id": "TA-840", - "scope": "standalone FSDP reduce-op canonicalizer truth table", - "decision": "remove", - "status": "applied", - "evidence": [ - "The four-line CPU report called the private canonicalizer directly with hand-built wrappers and raw enum values.", - "The retained two-rank FSDP2 lifecycle installs BF16StochasticAllToAllReduceScatter through set_custom_reduce_scatter, so PyTorch supplies the real wrapped reduce operation before finite gradient and numerical-error checks." - ] - }, - { - "id": "TA-841", - "scope": "fake full-precision expert FSDP kwargs report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The standalone report defined a fake object with one boolean class field and asserted that one dictionary key was removed.", - "The retained topology policy now makes the same assertion on an actual Glm52ExactTP16SharedExpertBlockFP8QLoRA after proving it is selected as one topmost FSDP unit." - ] - }, - { - "id": "TA-842", - "scope": "unintegrated GDN decode-prep P5 certification suite", - "decision": "remove", - "status": "applied", - "evidence": [ - "The four SM90 reports certified an opt-in gdn_decode_prep candidate that is not exported, documented, or called by any production module.", - "Its only in-repository source dependency is another unused decode-solve candidate, so graph capture and bitwise campaign gates did not protect a reachable XoRL execution path." - ] - }, - { - "id": "TA-843", - "scope": "test-only analytic pipeline bubble formula", - "decision": "remove", - "status": "applied", - "evidence": [ - "analytic_bubble_fraction was exported only by pp_profiling itself and had no production, documentation, example, or script caller; its only consumer was a hand-written formula truth table.", - "The retained PPBubbleProfiler path is constructed by Trainer and measures actual schedule busy intervals, memory, and P2P estimates instead of predicting an idealized formula." - ] - }, - { - "id": "TA-844", - "scope": "orphan first-fit-decreasing feasibility checker assertions", - "decision": "remove", - "status": "applied", - "evidence": [ - "ffd_check had no source caller and was imported only by the packing test, where six literals exercised an algorithm that dataset preparation never invokes.", - "The retained report exercises pack_group, allocate_sequentially, and PackingDataset, which are the actual packing and rank-allocation paths." - ] - }, - { - "id": "TA-845", - "scope": "unused grouped and any-FQN matcher truth tables", - "decision": "remove", - "status": "applied", - "evidence": [ - "check_all_fqn_match and check_any_fqn_match had no runtime caller; only the distributed-utils test preserved their grouped-number, prefix, and return-index behavior.", - "The retained check_fqn_match report protects the matcher actually used throughout ParallelPlan and sharded adapter state resolution." - ] - }, - { - "id": "TA-846", - "scope": "standalone per-parameter EP gradient hook compatibility path", - "decision": "remove", - "status": "applied", - "evidence": [ - "The hook installer and single-gradient reducer were documented for standalone tests and diagnostics and had no production caller; one branch of the real multi-rank report was their only consumer.", - "The retained two-rank and three-rank lifecycles exercise the production coalesced optimizer-boundary reducer, participation masks, bucket accounting, clipping, and non-finite rejection; the test-only parameter_count alias was removed with the hook." - ] - }, - { - "id": "TA-847", - "scope": "unreachable DeepSeek-V4 indexer autograd and backward campaign", - "decision": "remove", - "status": "applied", - "evidence": [ - "The V4IndexerFunction wrapper and batched_indexer_bwd kernel were imported only by two parameterized test reports; no production, documentation, example, or script path referenced either module.", - "The live V4Indexer invokes batched_indexer_fwd directly and returns discrete top-k indices, so the retained forward-score, causal-mask, numerical-range, and zero-input gates cover the reachable kernel while eight dormant wrapper/backward cases are gone." - ] - }, - { - "id": "TA-848", - "scope": "unintegrated manual CUDA timing module and lifecycle report", - "decision": "remove", - "status": "applied", - "evidence": [ - "No production module, package export, documentation, example, or script imported xorl.utils.manual_cuda_timing; its setter, scope, and drain functions were consumed only by its own mocked-event test.", - "The report therefore certified an instrumentation lifecycle that no XoRL execution path could enable or observe." - ] - }, - { - "id": "TA-849", - "scope": "standalone repeat_kv shape, value, and device smoke", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The standalone helper report separately checked identity, repeated shapes, a tiny value pattern, and CUDA device preservation without entering attention.", - "The retained eager-attention transaction now compares GQA weights and outputs against an independent torch.repeat_interleave reference, proving the repeated KV values are consumed correctly through the live backend while retaining invalid-head-layout rejection." - ] - }, - { - "id": "TA-850", - "scope": "identity-expert synthetic full MoE pipeline report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report composed histogram, index, scatter, and gather around an identity expert and ended at the same hidden_states times topk invariant already exercised by the retained scatter-gather round trip.", - "Retained kernel reports independently compare histogram/index ordering and gather, scatter, and add-gather values, while real MoE model and backend suites cover non-identity expert computation and gradients." - ] - }, - { - "id": "TA-851", - "scope": "test-only pipeline single-stage schedule convenience predicate", - "decision": "remove", - "status": "applied", - "evidence": [ - "is_single_stage_schedule had no runtime caller and its only assertions repeated whether schedule_stage_style returned single for GPipe and 1F1B.", - "The retained schedule policy still validates every supported style, split-backward mode, virtual-stage constraint, and the real schedule-class decision inside build_pipeline_schedule." - ] - }, - { - "id": "TA-852", - "scope": "unconsumed GLM sparse-cache mapping and gather helpers", - "decision": "remove", - "status": "applied", - "evidence": [ - "physical_cache_to_logical_indices and gather_selected_logical_values had no source, documentation, example, or script caller; their only consumer was an appended assertion block in the sparse-selector test.", - "The retained selector contract exercises the live canonical logical-index producer, including ties, short rows, dead rows, sorted unique indices, valid counts, and the production boundary tail." - ] - }, - { - "id": "TA-853", - "scope": "test-only GLM inventory and layer-plan convenience projections", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Glm52AdapterInventory.role_counts and four Glm52LayerPlan filtered-layer properties had no runtime caller and merely precomputed views of their public target and layer tuples for tests.", - "The retained contracts derive those counts and schedules from the authoritative tuples while preserving exact official roles, full/shared indexer allocation, dense/sparse placement, and producer boundaries." - ] - }, - { - "id": "TA-854", - "scope": "test-only NVFP4 format-dispatch wrapper", - "decision": "remove", - "status": "applied", - "evidence": [ - "fake_quantize was neither package-exported nor called by source; tests alone checked that its only supported string forwarded to fake_quantize_nvfp4 and that an unrelated string raised.", - "The retained public fake_quantize_nvfp4 contract compares both supported dtypes with an independent reference, proves straight-through gradients, and rejects invalid tensor geometry." - ] - }, - { - "id": "TA-855", - "scope": "standalone native-FP8 metadata-dictionary preflight", - "decision": "remove", - "status": "applied", - "evidence": [ - "validate_native_fp8_state_metadata had no loader caller; its test constructed an artificial metadata dictionary solely to call the helper directly.", - "DistributedCheckpointer uses the retained validate_native_fp8_dcp_checkpoint path, whose real-DCP tests reject castable payload metadata before load and validate EP-restored expected shapes." - ] - }, - { - "id": "TA-856", - "scope": "unintegrated lightweight ModelState reference export path", - "decision": "remove", - "status": "applied", - "evidence": [ - "ModelState.reference_state_dict described a future direct-safetensors path but had no production, export-script, documentation, or example caller; only test fixtures invoked it.", - "The retained checkpoint policies exercise actual DCP state collection, persistent QARL buffer metadata, compatibility rejection, pipeline key unions, optimizer filtering, and save/load process groups." - ] - }, - { - "id": "TA-857", - "scope": "routing-replay cursor reset API used only by tests", - "decision": "remove", - "status": "applied", - "evidence": [ - "reset_forward, reset_backward, reset_all_forward, and reset_all_backward had no runtime caller; tests alone rewound synthetic cursor values and then expected a checkpoint replay cursor to return to zero.", - "The real trainer and R3 handler clear replay instances at transaction teardown, so retained tests now assert that backward replay advances the cursor and that clear_all performs the production cleanup." - ] - }, - { - "id": "TA-858", - "scope": "manual LoRA merged-weight cache invalidators with no caller", - "decision": "remove", - "status": "applied", - "evidence": [ - "The dense, fused-delta, and MoE invalidate_merged_weight_cache methods had no production caller; one test invoked the dense method only to isolate two executions.", - "Merged caches are keyed by tensor versions, storage pointers, active rank, and alpha, and retained tests prove automatic optimizer-step and runtime-configuration invalidation plus bounded generation release." - ] - }, - { - "id": "TA-859", - "scope": "undocumented router top-k diagnostic policy matrix", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_MOE_ROUTER_TOPK_POLICY and its stable-low-id, tie-bias, and raw-logit branches had no documentation, script, example, or production configuration consumer; only a synthetic test truth table selected them.", - "The retained router contracts use the live torch.topk selection and still cover softmax weighting, normalization, DSv4 correction bias, hash routing, balanced synthetic profiling, and exact batch-invariant routing." - ] - }, - { - "id": "TA-860", - "scope": "layer-list router-FP32 environment diagnostic and standalone report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "XORL_MOE_ROUTER_FP32_LAYERS was absent from runtime configuration, documentation, scripts, and examples; its parser and layer-index test were their own consumers.", - "The retained configuration contract now exercises the real _router_fp32 model setting through MoEBlock and proves that hidden states and gate weights enter the FP32 projection without preserving the parallel environment override." - ] - }, - { - "id": "TA-861", - "scope": "unqualified FP64 MoE parity detour and dispatch assertion", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_MOE_FP64_ACCUM had no configuration, documentation, launcher, example, or serving-side implementation in the repository; its only test replaced the FP64 method with a mock and asserted dispatch precedence rather than numerical parity.", - "The retained MoE reports exercise the shipped eager, Triton, fused-SGLang, EP, and TP-simulation paths with real forward, gradient, layout, determinism, and admission checks." - ] - }, - { - "id": "TA-862", - "scope": "Qwen3-MoE delayed-residual and partial-residual experiment family", - "decision": "remove", - "status": "applied", - "evidence": [ - "The delayed residual pair, TP-shard carry, post-attention partial residual modes, alternate RMSNorm force flags, and candidate-capture matrix were controlled only by undocumented process-wide environment variables; no repository configuration, launcher, example, or end-to-end model test enabled them.", - "Four reports constructed private tuple inputs or attached tensor attributes in test doubles. The retained Qwen3-MoE report executes the normal decoder and final-norm consumers and verifies the explicit no-residual versus residual-tree family contract." - ] - }, - { - "id": "TA-863", - "scope": "unused stacked-LoRA initialization and merge utility surface", - "decision": "remove", - "status": "applied", - "evidence": [ - "The stacked initialization, delta, merge, and unmerge helpers had no production, documentation, example, or script caller; their package exports and one arithmetic truth-table report were their only consumers.", - "The live compute_lora_scaling function now remains at the group-GEMM package boundary used by dense, MoE, and quantized adapters, whose retained construction, loading, gradient, and optimizer reports all exercise it." - ] - }, - { - "id": "TA-864", - "scope": "tautological eager SwiGLU parity wrapper and report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The private _native_silu_and_mul helper had no source caller and was exactly one torch.nn.functional.silu expression; its sole report compared that wrapper with the identical expression labeled as the SGLang reference.", - "Real fused SwiGLU forward and backward behavior remains covered through the Triton operator, exact GLM MLP composition, and model-level LoRA and QLoRA paths; the independent dense RoPE cross-engine report remains in the file." - ] - }, - { - "id": "TA-865", - "scope": "unwired families-v2 specialized QK-norm kernel and golden", - "decision": "remove", - "status": "applied", - "evidence": [ - "qk_norm_v2 and its Triton kernel had no model, trainer, dispatcher, configuration, documentation, example, or script caller; only a standalone strided-view report and one frozen golden invocation reached them.", - "The Qwen3.5 and Qwen3.5-MoE attention paths use their declared RMSNorm modules. Retained families-v2 gates continue to cover the live hidden-state RMSNorm fused and split realizations, dispatch boundary, exact-model selection, and frozen numerical trees." - ] - }, - { - "id": "TA-866", - "scope": "undocumented SGLang MoE TP simulation diagnostic matrix", - "decision": "remove", - "status": "applied", - "evidence": [ - "The XORL_SGLANG_MOE_TP_SIM environment family and its direct, cache, Triton, alternate-reduce, DeepGEMM, fused-kernel, and runner modes had no configuration, launcher, example, documentation, or non-diagnostic consumer; history identifies the lane as a K3 parity diagnostic and later experiment.", - "Nine reports constructed tiny full-local tensors and replaced each optional backend with a Python fake. The retained SGLang fused-expert suites exercise the supported local and EP dispatch, weight layouts, loader admission, autograd, cache invalidation, and failure boundaries without preserving the parallel simulation product." - ] - }, - { - "id": "TA-867", - "scope": "CPU-only SGLang JIT and kernel RMSNorm diagnostic report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report selected sglang_jit and sglang_kernel but ran only CPU tensors, so neither optional CUDA implementation, loader, ABI boundary, nor serving arithmetic was exercised; both modes reduced to an ordinary eager fallback formula.", - "The modes remain available for explicit diagnostics. Retained RMSNorm suites cover CPU fallback arithmetic, fused forward and backward, exact Qwen site integration, family admission, global configuration forwarding, and real CUDA kernels when available." - ] - }, - { - "id": "TA-868", - "scope": "unintegrated sparse-delta receiver sharding and translation-future helpers", - "decision": "remove", - "status": "applied", - "evidence": [ - "The contiguous sharder, per-rank raw and encoded writers, translation-future collector, and terminal future writer formed a closed source/test-only subgraph with no runtime, configuration, documentation, example, or script caller.", - "Three reports fabricated receiver shards and the complete optional delta_encoding future API. The retained source-capture lifecycle exercises the live ModelRunner and dispatcher boundary, manifest aggregation, validated single-file packing, and sparse-delta backend consumption." - ] - }, - { - "id": "TA-869", - "scope": "deprecated R3 payload configuration and RequestProcessor aliases", - "decision": "remove", - "status": "applied", - "evidence": [ - "externalize_r3_payloads, keep_r3_payloads, routing_payload_dir, and keep_routing_payloads appeared only in compatibility branches and two test inputs; all launchers and runtime constructors already use r3_payload_transport, r3_payload_dir, and r3_payload_keep.", - "The canonical Mooncake and filesystem tests remain and still cover payload creation, slicing, cleanup, retention, namespace validation, and server configuration serialization without preserving duplicate names." - ] - }, - { - "id": "TA-870", - "scope": "fully mocked Trainer bootstrap forwarding report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report instantiated Trainer through __new__, supplied two large SimpleNamespace configurations, and replaced every bootstrap dependency to assert one direct ep_intranode keyword assignment and the presence or absence of one loss dictionary key.", - "Retained distributed tests execute both EP mesh layouts, loss tests exercise quack_linear numerics and admission, argument tests validate the settings, and the trainer model-boundary report covers meaningful configuration forwarding without reproducing bootstrap internals." - ] - }, - { - "id": "TA-871", - "scope": "undocumented Muon Quack tuning environment override", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_MUON_QUACK_TUNED had no configuration, documentation, example, or script consumer; its only report replaced Quack GEMMs with lambdas and observed the tuned keyword.", - "Muon retains the qualified tuned=False default used by trainer Quack paths. Retained reports cover backend import failure, architecture and dtype selection, real optimizer updates, grouped Gram Newton-Schulz execution, and CUDA compute dtype." - ] - }, - { - "id": "TA-872", - "scope": "fake-arithmetic CollatePipeline suite and test-only constructor forms", - "decision": "remove", - "status": "applied", - "evidence": [ - "Two reports composed fake collators that added and multiplied token IDs, then asserted those fake operations plus single-callable, tuple, and empty-list constructor forms absent from runtime callers.", - "The retained DataLoaderBuilder integration exercises the live non-empty list pipeline through tensor conversion, flattening, shifting, packing, micro-batch splitting, and optional sequence sharding." - ] - }, - { - "id": "TA-873", - "scope": "orphan data-preparation support implementations after test removal", - "decision": "remove", - "status": "applied", - "evidence": [ - "FileLockLoader remained unexported and unreferenced after its test-only suite was removed, while the SHA256 string wrapper had no caller after its standard-library restatement report was removed.", - "Live dataset preparation retains packing-cache persistence, source loading, split and configuration fingerprints, retries, preprocessing, and dataloader lifecycles." - ] - }, - { - "id": "TA-874", - "scope": "test-only retry strategies and wall-clock timing assertions", - "decision": "remove", - "status": "applied", - "evidence": [ - "Linear and constant retry strategies were selected only by the unit test; the sole production decorator use relies on the exponential default.", - "The retained retry contract covers success, retryable request and Hub failures, exhaustion, unrelated exceptions, and exponential delays by observing requested sleeps without real-time thresholds." - ] - }, - { - "id": "TA-875", - "scope": "fully mocked Trainer numerical-flag forwarding report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report constructed Trainer via __new__, supplied a large fake argument tree, replaced foundation-model construction, and asserted direct keyword copies for numerical flags and LoRA scalars.", - "Retained argument and model-policy suites cover parsing, resolution, admission, construction, and numerical behavior without reproducing the Trainer call signature." - ] - }, - { - "id": "TA-876", - "scope": "test-only orchestrator response builders and validation conveniences", - "decision": "remove", - "status": "applied", - "evidence": [ - "Twelve response builders and four validation or introspection helpers had no orchestrator, API server, scheduler, dispatcher, documentation example, or runtime caller; only their protocol test and package re-exports referenced them.", - "The retained protocol contract round-trips the live OrchestratorRequest and OrchestratorOutputs dataclasses through msgpack and reconstructs the typed operation payload consumed by ZMQ communication." - ] - }, - { - "id": "TA-877", - "scope": "redundant dense NVFP4 wrapper-injection smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report wrapped two Linear modules in a Sequential and asserted the wrapper count, format string, and output shape without checking a distinct numerical or lifecycle boundary.", - "Retained QARL reports cover targeted dense injection and exclusions, counters and summaries, NVFP4 arithmetic and straight-through gradients, a real optimizer update, changed log-probabilities, and checkpoint restoration." - ] - }, - { - "id": "TA-878", - "scope": "mock-only weight-version forwarding report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report bound one handler method to a MagicMock and replaced the NCCL synchronizer, so it observed only direct keyword forwarding through mocked collaborators.", - "The retained handler policy contract verifies flush_cache and weight_version at transfer_bucket, and the P2P protocol contract verifies the requested version in the real completion request body." - ] - }, - { - "id": "TA-879", - "scope": "mocked sharded-LM-head builder keyword-copy report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report replaced foundation-model construction and parallelization, then asserted only that fsdp_sharded_lm_head_loss=True crossed the adjacent function call.", - "Retained distributed and training-utility suites execute the sharded LM-head loss under FSDP and cover its admission, chunking, normalization, and gradients." - ] - }, - { - "id": "TA-880", - "scope": "mocked ModelRunner FP8, QARL, and sharded-loss builder keyword copies", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report replaced build_training_model and compared unchanged train-config values with the fake builder's kwargs; it constructed no FP8, QARL, calibrated, or sharded-loss model.", - "Retained builder and QARL lifecycle suites perform real dense and MoE injection, validate FP8 policy, calibrate before parallelization, update parameters, and restore checkpoints." - ] - }, - { - "id": "TA-881", - "scope": "Dr.GRPO outer-loop dispatch into a fake forward loop", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report replaced ModelRunner._forward_loop and asserted that the drgrpo string, input objects, and generic runner fields reached the fake, without executing Dr.GRPO computation.", - "The retained runner report executes the actual Dr.GRPO branch and its clipping, KL, temperature, legacy-field, output, and K3 policies; independent lifecycle suites cover completion, failure, routing, identity, and step accounting." - ] - }, - { - "id": "TA-882", - "scope": "raw routing-replay object forwarding into mocked backends", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report passed routed-expert lists through forward and forward-backward requests, replaced both backend methods with AsyncMock, and asserted object equality at the immediate kwargs seam.", - "Retained routing lifecycles exercise Mooncake and filesystem encoding, datum ordering, slice loading, cleanup, wire decoding, rank-zero selection, model identity, and failures." - ] - }, - { - "id": "TA-883", - "scope": "fake ModelRunner block-FP8 QLoRA builder report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report replaced build_training_model, asserted direct configuration copies, and constructed no quantized or adapter model; its remaining target-set assertion duplicated the dedicated GLM target-resolution contract.", - "Retained real builder suites cover foundation and injection settings, enabling preconditions, adapter inventory ownership, targets, quantized construction, and QLoRA execution." - ] - }, - { - "id": "TA-884", - "scope": "standalone module-path and FQN-matcher utility truth tables", - "decision": "remove", - "status": "applied", - "evidence": [ - "The two reports exercised recursive getattr and setattr plus regex wildcard examples on a toy Sequential and ModuleDict without reaching a sharding or ownership outcome.", - "Retained ParallelPlan and sharded adapter-state suites invoke the same helpers through exact and wildcard FQNs while slicing parameters, assigning placements and gradient domains, and materializing real adapter layouts." - ] - }, - { - "id": "TA-885", - "scope": "unused singular launcher free-port helper", - "decision": "remove", - "status": "applied", - "evidence": [ - "find_free_port had no source, test, documentation, example, or script caller; similarly named test-local helpers were independent definitions.", - "The live launcher retains find_free_ports, which allocates the three- and four-port rendezvous layouts consumed by worker startup." - ] - }, - { - "id": "TA-886", - "scope": "mock-only RemoteBackend payload wrapper suite", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report replaced RemoteBackend._execute and asserted operation names, request ids, timeouts, and fields copied into two payload constructors; it never serialized or transported a request.", - "Retained protocol, request-processor, dispatcher, sparse-delta, and weight-sync suites exercise typed payload reconstruction and the downstream behavior of the same fields." - ] - }, - { - "id": "TA-887", - "scope": "public xorl.rl primitive package and direct contract suite", - "decision": "remove", - "status": "applied", - "evidence": [ - "Repository and organization-wide code searches find no consumer of the six exports beyond their direct self-test. XoRL Client PR 14 imports xorl_client.rl, not xorl.rl, and sends importance_sampling, cispo, or policy_loss names over the API.", - "The trainer dispatches those names directly to the integrated xorl.ops.loss implementations. PR 67 removed the isolated package and self-test from main; this pull request inherits that merge and carries no source or test diff for the removal." - ] - }, - { - "id": "TA-888", - "scope": "test-only runner protocol acknowledgement and response factories", - "decision": "remove", - "status": "applied", - "evidence": [ - "create_ack_for_request and create_response_for_request were exported but had no runtime, documentation, example, or script caller; only the acknowledgement helper had a direct field-copy assertion.", - "Live transport code constructs RunnerAck and RunnerResponse directly, while the retained protocol report round-trips those actual dataclasses through the bounded MessagePack wire format." - ] - }, - { - "id": "TA-889", - "scope": "orphan sparse-delta translation-input loader and fake dependency report fragment", - "decision": "remove", - "status": "applied", - "evidence": [ - "load_sparse_source_delta_inputs had no translation engine, receiver, CLI, example, documentation, or source caller; its only consumer installed a synthetic delta_encoding package and checked fabricated empty shards.", - "Retained sparse-delta tests cover live source capture, manifests, packed-file validation, backend upload, receiver application, and end-to-end trainer-to-SGLang behavior." - ] - }, - { - "id": "TA-890", - "scope": "production-embedded canonical MoE test oracle and oracle self-test", - "decision": "remove", - "status": "applied", - "evidence": [ - "canonical_moe_reduce_reference had no runtime caller and shared _adjacent_pairwise_bf16 with the implementation it was used to validate; its standalone report then compared that shared helper with a hand-written tree.", - "The distributed and GLM model contracts now compute their expected adjacent BF16 tree independently in test code, and the real multi-process transport, permutation, chunking, output-distribution, and backward gate passes." - ] - }, - { - "id": "TA-891", - "scope": "unconsumed SequencePartial reducer and synthetic layout matrix", - "decision": "remove", - "status": "applied", - "evidence": [ - "SequencePartial had no loss, trainer, runner, CLI, example, documentation, or other source caller; only its dense, packed, and hand-sliced context-parallel unit report instantiated it.", - "TokenPartial remains the sole production reducer and retains integrated causal-LM, policy, importance-sampling, OPD, Dr.GRPO, TP, and FSDP coverage." - ] - }, - { - "id": "TA-892", - "scope": "duplicate Tinker weights-info and create-model compatibility report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The older report repeated session-spec persistence, normalized LoRA and optimizer metadata, and weights-info loading already covered by the focused session-endpoint lifecycle.", - "The retained report now also asserts Tinker's flat lora_rank response field while preserving disk-over-memory metadata, path confinement, and full-weight cases." - ] - }, - { - "id": "TA-893", - "scope": "fake-model sampler prefill-length forwarding report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report supplied one literal list to _compute_micro_batch_loss, captured kwargs on a fake model, and asserted only tensor dtype, device, and unchanged value.", - "Retained GLM indexer and sparse-attention contracts exercise prefill-boundary validation and behavior; the Dr.GRPO runner report remains focused on executing the actual loss branch." - ] - }, - { - "id": "TA-894", - "scope": "private trainer telemetry formatting reports", - "decision": "remove", - "status": "applied", - "evidence": [ - "The two reports instantiated no trainer lifecycle; they asserted byte-to-GB field naming, private key ordering, duplicated local summary floats, call counts, and empty dictionaries on synthetic namespaces.", - "Retained component-timer coverage runs real GLM- and Qwen-shaped forward and backward hooks on CUDA and preserves unrecorded-event recovery, while trainer suites cover optimization and synchronization behavior." - ] - }, - { - "id": "TA-895", - "scope": "split fake seams for runner forward session and R3 propagation", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "One report replaced _execute_and_gather and the other called _execute_compute directly, so neither covered the handler-to-trainer chain.", - "The replacement executes the real rank-zero handler, gather wrapper, and compute dispatch through the trainer while stubbing only distributed side effects." - ] - }, - { - "id": "TA-896", - "scope": "families-v2 source-text import lint", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report opened bi_families_v2.__file__ and banned four import substrings without executing a numerical or cross-engine behavior.", - "Retained suites cover model-program selection, rollback, dispatch, fused-versus-split equality, cross-engine bytes, and CUDA bit gates; the audit now identifies module-source reads as source inspection." - ] - }, - { - "id": "TA-897", - "scope": "repository-wide private-reference pytest scan", - "decision": "relocate", - "status": "applied", - "evidence": [ - "The report exercised no XoRL behavior; it listed every tracked file with Git, decoded source and documentation, and matched repository-policy regexes.", - "The same zero-dependency scan now runs as scripts/check_public_tree.py from the pre-commit lint workflow, preserving pull-request enforcement outside pytest." - ] - }, - { - "id": "TA-898", - "scope": "duplicate fake register-session dispatcher report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report called _handle_register_session on a fake coordinator and repeated payload and response assertions from the dedicated runner session-ops suite.", - "The retained session-ops report also covers cross-rank rejection, while the retained request-processor lifecycle separately reaches DummyBackend registration." - ] - }, - { - "id": "TA-899", - "scope": "self-fulfilling orchestrator-client forward and optimizer socket report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report's mock engine computed num_samples from its own received list and echoed the request learning rate, so both asserted results were authored by the test double rather than XoRL behavior.", - "Retained real-ZMQ reports verify interleaved forward and optimizer requests plus exact serialization, while the orchestrator end-to-end error report rejects empty model-pass batches." - ] - }, - { - "id": "TA-900", - "scope": "duplicate request-processor empty-batch rejection fragment", - "decision": "remove", - "status": "applied", - "evidence": [ - "The processor report repeated empty-list rejection already exercised through the real orchestrator request lifecycle.", - "Its distinct nonempty batch without valid targets remains as a focused processor validation report." - ] - }, - { - "id": "TA-901", - "scope": "duplicate API registration and explicit optimizer forwarding paths", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The legacy create-model report repeated focused full-weight and normalized LoRA worker registration; the retained full-weight report now also supplies empty optional configs to preserve that compatibility boundary.", - "The optimizer bundle's explicit learning-rate forwarding repeated the focused training-ops report, while its legacy Adam payload and learning-rate fallback priority remain covered." - ] - }, - { - "id": "TA-902", - "scope": "microscopic exact GLM and native FP8 contract wrappers", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Twelve collected items existed only to call one to three adjacent assertion helpers for the same CPU component, fragmenting topology, operand, byte, gradient, construction, admission, and checkpoint facets into narrow reports.", - "Component-level contracts now execute every original helper, with monkeypatch state explicitly reset between independent seams; separate Hopper and behaviorally distinct router or expert reports remain separate." - ] - }, - { - "id": "TA-903", - "scope": "DeepSeek-V4 private checkpoint-name and APE helper report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The report asserted private string rewrites and called the APE inverse on a synthetic forward transform even though the retained synthetic checkpoint transaction loads window, C4, hash, shared-expert, and routed-expert families through the real handler.", - "The retained transaction now checks values at every major destination, including norms, attention, HC, router bias, shared and fused experts, C4 APE tensors, and renamed indexer projections." - ] - }, - { - "id": "TA-904", - "scope": "fake GLM-5.2 block-FP8 QLoRA builder plumbing report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report replaced both foundation-model construction and QLoRA injection, then asserted that flags, rank, alpha, quant format, group size, and a fabricated inventory crossed those fakes.", - "Retained GLM-5.2 suites construct the full 700-target, 1700-factor model and exercise exact component admission; the builder's fail-closed requirement for LoRA plus QLoRA remains in the quantized-mode admission report." - ] - }, - { - "id": "TA-905", - "scope": "static OLMo-2 tensor-parallel plan dictionary report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report inspected TP_PLAN and MODEL_TP_PLAN entry types and missing keys without constructing a device mesh or applying tensor parallelism.", - "Retained two-rank CPU reports apply the production plan to OLMo-2, execute forward and backward through local-axis QK RMSNorm, rowwise and colwise projections, post-norm residual flow, and the vocab-sharded LM head, and compare the custom QK norm numerically." - ] - }, - { - "id": "TA-906", - "scope": "direct GLM sparse-MLA auto-dispatch self-comparison", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report compared auto dispatch on CPU with a direct call to the same torch reference implementation, duplicating the retained full-model sparse-versus-dense numerical path.", - "The full GLM attention integration report reaches auto dispatch through Glm5Model and checks dense parity; the distinct unknown-backend rejection was moved into that report rather than removed." - ] - }, - { - "id": "TA-907", - "scope": "standalone LM-head TP CP-DP-HSDP topology matrix", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report launched the same four-rank CP, DP, and HSDP layouts as the retained LM-head FSDP end-to-end suite, then asserted group membership, sizes, and mesh labels without executing a model or loss.", - "The retained end-to-end cases build a real FSDP-sharded LM head on those meshes and compare parameter synchronization, vocab ranges, global loss, full weight gradients, and local hidden gradients with eager references; the distinct EP-overlay topology report remains." - ] - }, - { - "id": "TA-908", - "scope": "private checkpoint URI construction and parsing snapshot", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The report called _to_xorl_uri and _from_xorl_uri directly for the same five documented path spellings already exercised by the adjacent save-and-load lifecycle, plus an undocumented arbitrary raw path.", - "The retained lifecycle creates real checkpoint directories and loads xorl URI, explicit weights path, model/checkpoint, checkpoint-only, and legacy weights/checkpoint inputs; its save assertion now pins the exact public xorl URI." - ] - }, - { - "id": "TA-909", - "scope": "direct FlashAttention metadata helper report", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The report called the metadata helper on synthetic dictionaries even though production reaches it through the packing and sequence-shard collators.", - "The retained packing-collator report now checks exact multi-sequence and single-sequence cumulative lengths and maximum lengths after real concatenation." - ] - }, - { - "id": "TA-910", - "scope": "direct sequence-shard slicing and padding primitive reports", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The reports called sp_slice and sp_padding directly and included a zero-length padding no-op.", - "The retained full-collator report now checks exact CP rank-zero and rank-one slices plus constant, sequential, label, and position padding on the real last-rank path." - ] - }, - { - "id": "TA-911", - "scope": "standalone API-orchestrator message roundtrip", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report serialized a request and response directly while the retained ZMQ client-engine tests send and receive those same message types over the production socket protocol.", - "The real communication lifecycle now pins the request payload, sequence id, timestamp, response id, type, payload, and terminal flag after crossing both sockets." - ] - }, - { - "id": "TA-912", - "scope": "fabricated DSv4 RoPE context-parallel slice guard", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The report invoked get_freqs_cis_for_cp with a fake rank and an arbitrary tensor rather than a model consumer.", - "The retained DSv4 compressor report now constructs a real short-cache CP compressor and reaches the same fail-loud capacity guard through forward_raw after also proving the supported C128 path." - ] - }, - { - "id": "TA-913", - "scope": "fragmented deferred QLoRA key-plan and cache-lifetime reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense merged, EP16 expert, and missing-pair cases are branches of the same prequantized module-key planner and now report as one policy contract.", - "Retained-cache clear and release behavior now runs inside the per-module deferred-loader residency lifecycle; every original assertion remains." - ] - }, - { - "id": "TA-914", - "scope": "standalone DeepSeek router rejection wrapper", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Foundation-model and training-builder router rejection are two entry points to the same DeepSeek router admission rule and now share one report with the admitted freeze behavior.", - "The tensor-parallel guard remains separate because it protects a different topology boundary." - ] - }, - { - "id": "TA-915", - "scope": "duplicate TeacherActivationCache selection report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ordinary rank-2 and rank-3 selection is already exercised through the Mooncake teacher-store consumer and the runner producer-consumer lifecycle.", - "Unique async completion, bounds, device residency, dtype reload, and rank-3 layer-slice checks remain in the cache lifecycle report." - ] - }, - { - "id": "TA-916", - "scope": "optimizer snapshot and immediate-load helper reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Recursive CPU snapshot semantics now run with the optimizer transaction failure policy that consumes those snapshots.", - "Immediate moment and step restoration now runs with the stronger uninterrupted-versus-resumed trajectory contract; all bytewise assertions remain." - ] - }, - { - "id": "TA-917", - "scope": "single-branch request-processor reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "A nonempty batch with no valid targets is the failure branch of the retained forward-backward processor lifecycle.", - "Register-session forwarding now runs with the processor control-operation lifecycle instead of presenting one backend roundtrip as a separate product behavior." - ] - }, - { - "id": "TA-918", - "scope": "standalone create-model normalized registration report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Normalized LoRA and optimizer registration is the successful branch of the same create-model lifecycle that handles recreation and registration rollback.", - "The create-session route remains separate because it has distinct refresh and override behavior." - ] - }, - { - "id": "TA-919", - "scope": "standalone folded-LoRA gradient-dtype helper report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FSDP gradient metadata controls the same folded-weight autograd boundary as the straight-through gradient comparisons.", - "The metadata, dtype, nonzero-gradient, shared-factor, fused gate-up, and linear assertions now execute in one autograd contract." - ] - }, - { - "id": "TA-920", - "scope": "separate QARL checkpoint buffer-key report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "QARL persistent-buffer mismatch detection and pipeline parameter or buffer key unions are facets of the same checkpoint model-key compatibility contract.", - "Strict and non-strict mismatch assertions remain alongside non-pipeline and pipeline metadata behavior." - ] - }, - { - "id": "TA-921", - "scope": "direct checkpoint expert-key classifier report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Expert-key classification exists to route dense and expert tensors in grouped checkpoint loading and is now checked inside that consumer lifecycle.", - "All supported expert, shared-expert, dense-MLP, and attention name cases remain asserted." - ] - }, - { - "id": "TA-922", - "scope": "standalone packing allocation primitive report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Bin capacity, safe-mode, offset, rank coverage, and non-overlap assertions describe the allocation used by PackingDataset.", - "They now run in the dataset lifecycle that also constructs sequential and multipack bins, loads cached bins, and checks cache identity." - ] - }, - { - "id": "TA-923", - "scope": "standalone empty TokenPartial reducer report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Zero scale with an empty mask is the boundary branch of the same denominator and additive-composition policy.", - "The exact zero assertion remains in the TokenPartial component report." - ] - }, - { - "id": "TA-924", - "scope": "standalone legacy TopK router wrapper", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy softmax selection, V4-input isolation, routed scaling, configuration selection, and FP32 gate routing are one router policy matrix.", - "Synthetic balanced, sqrt-softplus, and hash-routing behaviors remain separate because they execute distinct routing modes." - ] - }, - { - "id": "TA-925", - "scope": "scattered server runtime configuration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The one-helper R3 transport report is now part of the server runtime roundtrip contract.", - "Adapter-gradient bucket serialization and zero-value rejection moved from the gradient-math suite into that same user-facing configuration lifecycle." - ] - }, - { - "id": "TA-926", - "scope": "fragmented exact GLM attention and MoE construction rejection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dependency flags, EP16, lm-head TP16, rank-one alpha-one, sparse-MLA, and all-to-all requirements are branches of exact-component construction admission.", - "Every fail-before-mutation assertion remains in one attention and one complete-MoE admission report; successful inventory and post-EP ownership reports stay separate." - ] - }, - { - "id": "TA-927", - "scope": "standalone kill-session path validation report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Path-like model ID rejection is a security branch of the same kill-session checkpoint lifecycle.", - "The lifecycle still proves failure preservation, evicted-checkpoint promotion, metadata cleanup, and path rejection." - ] - }, - { - "id": "TA-928", - "scope": "standalone multi-part optimizer custom-group rejection report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Custom parameter-group rejection is the admission branch of multi-part optimizer construction.", - "It now runs with real multi-part updates, zeroing, scheduler propagation, model mapping, and single-part fallback." - ] - }, - { - "id": "TA-929", - "scope": "separate adapter ownership declaration rejection report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Overlapping and foreign pending authority are declaration-level branches of the compiler's fail-closed ownership policy.", - "They now run with missing-universe, unsupported-TP, managed-FSDP, and false-EP ownership rejection." - ] - }, - { - "id": "TA-930", - "scope": "split EP checkpoint mesh restore and drop reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Restoring the EP dimension and dropping it back to the stage-local expert-FSDP mesh are opposite directions of one checkpoint mesh policy.", - "Legacy and PP-parent mesh validation, placement, target shape, and local tensor assertions all remain." - ] - }, - { - "id": "TA-931", - "scope": "weight-sync protocol edge reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Empty PP-NCCL payload handling now runs with the named-tensor roundtrip instead of reporting a two-call mock branch separately.", - "Sparse-delta baseline priming and receiver-failure retry now run in the transfer state-machine lifecycle; post-packed transfer and initialization remain separate boundaries." - ] - }, - { - "id": "TA-932", - "scope": "fragmented packing and sequence-shard collator reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Generic extras, token-aligned fields, pre-shifted labels, and packed boundaries are field policies of their respective production collators.", - "All concatenation, padding, label, position, and FlashAttention metadata assertions remain; the CP16 token-side-channel contract stays separate." - ] - }, - { - "id": "TA-933", - "scope": "standalone R3 side-payload validation report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Malformed reference, bounds, and count failures are branches of the same R3 put, sliced-load, and cleanup lifecycle.", - "The low-level Mooncake tensor codec remains separate because it exercises serialization rather than R3 reference ownership." - ] - }, - { - "id": "TA-934", - "scope": "split sync-quantization normalization and rejection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Supported defaults and normalization plus unsupported methods, formats, schemes, list shapes, and explicit reasons form one configuration policy matrix.", - "Every accepted output and rejection message remains asserted through normalize_sync_quantization_config." - ] - }, - { - "id": "TA-935", - "scope": "NCCL rendezvous bind-failure branch report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Store-bind failure before inference initialization is the fail-closed branch of NCCL rendezvous initialization.", - "It now runs with ephemeral-port rotation and explicit-port pinning in one training rendezvous lifecycle." - ] - }, - { - "id": "TA-936", - "scope": "P2P async status-timeout branch report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Status polling timeout is the failure branch of asynchronous size and cutoff dispatch.", - "The prepare-request timeout remains separate because it governs a different HTTP phase." - ] - }, - { - "id": "TA-937", - "scope": "standalone SGL cross-attention cu-seqlens rejection report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Cross-attention cu-seqlens rejection belongs to the page-size-one SGL KV-cache adapter policy.", - "Generic fixed-length, variable-length, eager, backend-registry, and alternate FlashAttention paths remain separate contracts." - ] - }, - { - "id": "TA-938", - "scope": "split Rank0 ready-handshake branch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Normal acknowledgement, request-before-acknowledgement, client identity, unexpected message, and receive failure are outcomes of one Rank0Protocol ready handshake.", - "All wire-message, queue, acknowledgement, identity, and request-count assertions remain in the handshake lifecycle." - ] - }, - { - "id": "TA-939", - "scope": "split runner load-state preparation and routing reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Path confinement and preparation failure are admission phases of the same load-state operation that routes multi-adapter and single-tenant restores.", - "Error preservation, artifact-root enforcement, adapter conversion, trainer routing, and step reset remain asserted." - ] - }, - { - "id": "TA-940", - "scope": "split scheduler dispatch and terminal-state reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FIFO admission, capacity, dispatch, completion, failure, abort, statistics, clearing, and bounded history form one scheduler lifecycle.", - "A fresh scheduler isolates the terminal-transition matrix from the capacity scenario without creating a second product report." - ] - }, - { - "id": "TA-941", - "scope": "standalone multi-adapter Adam override report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Multi-adapter Adam beta and epsilon propagation is a branch of ModelRunner.optim_step's full, partial, omitted, and non-Adam override policy.", - "The adapter-manager call and adapter optimizer parameter-group assertions remain in the optimizer-step report." - ] - }, - { - "id": "TA-942", - "scope": "split ParallelState construction and singleton initialization reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Defaults, validation, enabled flags, initialization, automatic DP sharding, access, and reinitialization protection are one ParallelState lifecycle.", - "EP mesh construction and requires-mesh behavior remain separate because they exercise a different helper boundary." - ] - }, - { - "id": "TA-943", - "scope": "fragmented DeepEP internode-preflight outcome reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Skip, uninitialized, intranode, transport failure, identity roundtrip, and corruption are branches of preflight_internode_transport.", - "Topology detection and buffer-size admission remain separate contracts; every node diagnostic and corruption assertion is retained." - ] - }, - { - "id": "TA-944", - "scope": "standalone ParallelPlan rejection branch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Indivisible generic meta slicing now runs with the successful metadata-preserving slice policy.", - "Malformed exact-GLM already-local and force-shard singletons now run with the exact meta EP disposition contract; the materialized real-tensor shard remains separate." - ] - }, - { - "id": "TA-945", - "scope": "fragmented exact absorbed-KV-B checkpoint pair reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Arrival order, duplicate members, incomplete pairs, dtype mismatch, and shape mismatch are outcomes of one NativeBlockFP8PairBuffer transaction.", - "The exact-attention source inventory remains separate because it validates model construction rather than checkpoint pair state." - ] - }, - { - "id": "TA-946", - "scope": "split block-FP8 quantization, dequantization, and edge reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shapes, scales, admission, roundtrip error, determinism, storage, magnitude edges, signs, and dimensional consistency are one block-FP8 codec contract.", - "All CUDA assertions execute in the same component report under the existing GPU gate." - ] - }, - { - "id": "TA-947", - "scope": "split GKN block-FP8 quantize and dequantize reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "GKN output geometry, roundtrip accuracy, tail blocks, zero blocks, scale range, output dtype, contiguity, and rank admission form one two-dimensional codec policy.", - "The large-matrix path and dequantization failure assertions remain under the same CUDA report." - ] - }, - { - "id": "TA-948", - "scope": "split shared-prefix attention edge reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "A singleton member and a one-token prompt are degeneracies of the same shared-prefix attention layout.", - "The general dtype, head-size, GQA, forward, and backward matrix remains separate from this edge-layout report." - ] - }, - { - "id": "TA-949", - "scope": "standalone one-token shared-prefix repack report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The empty shared block for a one-token prompt is an edge branch of shared-prefix detection, repacking, and remapping.", - "Its exact decoded blocks and empty cross-attention indices remain asserted in the full repack lifecycle." - ] - }, - { - "id": "TA-950", - "scope": "split exact dense gate-up pair-buffer reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Out-of-order gate and up emission plus missing, duplicate, invalid-dtype, and non-finite members are outcomes of one exact dense checkpoint transaction.", - "Fused byte layout, scale order, model installation, base-loaded state, and every failure remain asserted." - ] - }, - { - "id": "TA-951", - "scope": "fragmented NVFP4 fake-quant reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two-dimensional reference parity, STE behavior, rank and block admission form one base fake-quant contract.", - "Three-dimensional projection STE, expert-independent scaling, and fused gate-up per-half scaling form one expert fake-quant contract." - ] - }, - { - "id": "TA-952", - "scope": "standalone DSV4 attention TP rejection report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unsupported tensor-parallel construction is an admission branch of the DSV4 attention storage and backend-call contract.", - "Window-only and C128 forward-backward variants remain separately parameterized because they execute different attention structures." - ] - }, - { - "id": "TA-953", - "scope": "separate eager-versus-native MoE determinism report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Backend determinism and all-tokens-to-one-expert behavior are edge branches of the eager-versus-native forward and backward parity matrix.", - "All expert-count, hidden-size, top-k, batch, sequence, gradient, and edge assertions remain under the same CUDA gate." - ] - }, - { - "id": "TA-954", - "scope": "standalone non-gated MoE constructor rejection report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unsupported backend and gated-activation rejection are admission branches of the CPU eager non-gated expert contract.", - "GPU Triton and native parity remains separate because it crosses optional backend and device boundaries." - ] - }, - { - "id": "TA-955", - "scope": "split gradient-checkpoint configuration and runtime-gate reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default and nondefault method propagation configure the same GradientCheckpointingLayer gate exercised by training, enabled, and method combinations.", - "Both ordinary and MoE layer configuration plus the exact checkpoint-call truth table remain asserted." - ] - }, - { - "id": "TA-956", - "scope": "split Kimi wrapper conversion and local registry-loading reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Kimi wrapper mapping, official auxiliary defaults, registry resolution, and local text-config unwrapping are one DeepSeek-V3 configuration-loading lifecycle.", - "Every MLA, MoE, routing, RoPE, registry, and local-load assertion remains." - ] - }, - { - "id": "TA-957", - "scope": "split local Kimi tokenizer and fallback-loader policy reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The dedicated local TikToken path and generic tokenizer or processor fallback are branches of the public auto-loader policy.", - "Token IDs, text roundtrip, right padding, and absence of implicit remote-code trust remain asserted." - ] - }, - { - "id": "TA-958", - "scope": "split sqrt-softplus and softmax routing-regather reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sqrt-softplus scaling and dtype plus unchanged softmax gather and renormalization are modes of MoEBlock._regather_routing.", - "Both eager-router comparisons and the independent softmax formula remain in one mode matrix." - ] - }, - { - "id": "TA-959", - "scope": "standalone all-invalid FlashMLA autograd report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All-invalid rows are the zero-row branch of the same FlashMLA compaction, TileLang backward, and scatter transaction.", - "Compacted valid rows, zero-scattered invalid gradients, all-zero output, and backward bypass remain asserted; dispatch-envelope admission stays separate." - ] - }, - { - "id": "TA-960", - "scope": "fragmented causal-LM Z-loss CPU reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Positive coefficient, zero coefficient, and tensor-parallel rejection are outcomes of one causal-LM Z-loss policy.", - "Reference CE and Z-loss values, gradients, absent metrics at zero, and fail-before-collective behavior remain; compiled CUDA parity stays separate." - ] - }, - { - "id": "TA-961", - "scope": "fragmented streaming forward-KL reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense-reference gradients, chunk invariance, ignore masking, and low-memory parity are one streaming-kernel contract.", - "OPD backend parity and unsupported logprob clamping form one dispatch contract; the independent FP64 gradcheck remains separate." - ] - }, - { - "id": "TA-962", - "scope": "fragmented LM-head module selection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local and TP module selection, FP32 bypass, and logprob temperature are one per-token CE policy matrix.", - "Importance-sampling module use plus causal-LM FP32 bypass and TP hidden-gradient reduction form one outer loss-dispatch contract." - ] - }, - { - "id": "TA-963", - "scope": "standalone Dr.GRPO objective branch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Zero and empty boundaries, positive-advantage direction, and KL reference admission are branches of the forward, backward, and metric contract.", - "Logprob-temperature behavior and microbatch composition remain separate numerical contracts." - ] - }, - { - "id": "TA-964", - "scope": "fragmented fused selected-logprob reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Frozen-output input gradients and irregular tails are shape and autograd branches of fused selected-logprob forward-backward parity.", - "Per-token, causal-LM, quack-linear, and importance-sampling dispatch now form one integration report; production-vocabulary finiteness and no-full-logits memory remain separate regressions." - ] - }, - { - "id": "TA-965", - "scope": "split families-v2 RMSNorm property reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP64 proximity, residual and zero-centered behavior, batch composition invariance, and run determinism are properties of one RMSNorm-v2 tree.", - "Forced fused-versus-split equivalence and production dispatch remain separate because they certify realization and selection boundaries." - ] - }, - { - "id": "TA-966", - "scope": "fragmented BI fused LM-head reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Eager parity, gradients, deterministic batch composition, and unsupported-mode guards form one BI fused loss contract.", - "Unit temperature identity and near-probability-one clamping form one kernel edge policy." - ] - }, - { - "id": "TA-967", - "scope": "split TileLang V4 indexer edge reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Causal masking, large-value stability, and zero-input output are edge branches of the same batched indexer kernel.", - "The parameterized forward matrix remains separate because it covers four execution geometries." - ] - }, - { - "id": "TA-968", - "scope": "standalone NF4 codebook roundtrip report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Codebook ordering and exact-value roundtrip are foundational branches of the flat NF4 codec contract.", - "The GKN codec remains separate because it uses a different packed layout and scale geometry." - ] - }, - { - "id": "TA-969", - "scope": "fragmented OPD full-vocabulary policy reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Diagnostics, loss clamping, reward weighting, stable metric keys, and unsupported top-k dispatch are branches of one full-vocabulary OPD policy.", - "Policy-gradient behavior, KL estimators, and compiled sampled-token logprobs remain separate because they exercise different objectives or callables." - ] - }, - { - "id": "TA-970", - "scope": "split GDN convolution forward and backward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forward parity, backward parity, variable-length and batch behavior, and repeat determinism form one causal-convolution numerical contract.", - "End-to-end GDN integration, optional SGLang parity, input admission, and contract-state lifecycle remain separate boundaries." - ] - }, - { - "id": "TA-971", - "scope": "fragmented exact TP1 QLoRA wrapper reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Static configuration and fail-before-import runtime admission are one wrapper policy; rounded forward values and surrogate gradients are one numerical transaction.", - "Factor-only VJP behavior and saved-master mutation rejection form one backward safety policy; packed-state dtype movement and literal CUDA parity remain separate." - ] - }, - { - "id": "TA-972", - "scope": "split FP8 linear numerical-mode reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Padding recipes, full residual correction, and activation-only correction are modes of the same FP8 matmul numerical contract.", - "Backend selection, profiling, CPU fallback, injection, and an optimizer train step remain separate implementation boundaries." - ] - }, - { - "id": "TA-973", - "scope": "fragmented TileLang sparse-MLA feature reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Attention-sink reference parity and observable effect form one sink policy; partial-invalid forward and backward behavior form one masking transaction.", - "Parameterized geometries and deterministic-versus-atomic backward remain independent kernel reports." - ] - }, - { - "id": "TA-974", - "scope": "fragmented exact routed-expert factor-buffer and empty-route reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Global and post-EP local factor banks are two inputs to the same physical sampler-buffer contract.", - "Zero-token bank gradients, all-sentinel zero gradients, and stride admission are edge branches of one routed-gradient policy." - ] - }, - { - "id": "TA-975", - "scope": "split canonical GLM52 MoE configuration and selection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Routing-replay rejection, internal transport resolution, exact canonical selection, and ordinary noncanonical selection form one canonical-MoE mode policy.", - "Sparse-selector arithmetic, codecs, loaders, and runtime dispatch remain separate production boundaries." - ] - }, - { - "id": "TA-976", - "scope": "fragmented exact shared-expert admission and state reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Construction restrictions and fail-before-kernel runtime checks form one component admission policy.", - "FP32 logical masters, dtype movement, canonical checkpoint sources, and immutable binding form one persistent-state policy; optional SGLang views remain separate from native base views." - ] - }, - { - "id": "TA-977", - "scope": "split topmost mixed-precision FSDP selection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Generic nested-module selection and the exact shared-expert specialization exercise the same topmost protected-unit selector.", - "Expert mixed precision, reduce dtype, sequence-parallel folding, and prefetch direction remain separate configuration policies." - ] - }, - { - "id": "TA-978", - "scope": "fragmented model-runner token-diagnostic reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Selection, empty and top-k boundaries, loss-logprob cross-checks, raw-weight references, and hidden summaries are output branches of one token-diagnostic callable.", - "KL token diagnostics, tensor dumps, component hooks, and trusted diagnostic inputs remain separate callables or lifecycle boundaries." - ] - }, - { - "id": "TA-979", - "scope": "split stochastic-rounding property reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shape and dtype, input admission, seeded repeatability, expectation, and neighboring-value bounds are properties of one BF16 stochastic-rounding operation.", - "Every original deterministic and statistical assertion remains in one CPU numerical contract." - ] - }, - { - "id": "TA-980", - "scope": "one-report-per-mode learning-rate scheduler tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Constant, linear, and cosine schedules are branches of the same scheduler builder and now form one mode matrix.", - "Invalid configuration remains separate because it is the builder admission boundary." - ] - }, - { - "id": "TA-981", - "scope": "split DistSignSGD hook and topology-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local hook registration, FSDP-managed exclusion, and unsupported HSDP, CP, EP, and non-FSDP DTensor rejection are outcomes of configure_distsignsgd.", - "Reduce-scatter arithmetic, optimizer construction, and parameter updates remain separate production boundaries." - ] - }, - { - "id": "TA-982", - "scope": "standalone default-session kill and unload protection report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ordinary LoRA teardown, checkpoint URI return, re-registration, and default-session protection form one API session-termination lifecycle.", - "Model creation, weight metadata, and legacy session-spec loading remain separate endpoints or persistence boundaries." - ] - }, - { - "id": "TA-983", - "scope": "fragmented inference-endpoint registration and sync reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Port routing, explicit worker health, adapter routing, auto-sync, server topology, and FP8 KV-cache admission are branches of endpoint registration.", - "Weight-sync quantization admission and FP8 KV-cache invalidation now form one sync request policy; listing and receiver enrichment remain separate." - ] - }, - { - "id": "TA-984", - "scope": "one-report-per-layout P2P FP8 receiver tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fused and unfused attention, partial blocks, Qwen3.6 QKVZ and full attention, nonexpert namespaces, mixed passthrough entries, shared experts, and routed experts are receiver-layout branches of one transfer protocol.", - "All byte, scale, slice, dequantization, endpoint, and expert-coverage assertions remain in one FP8 receiver-layout matrix." - ] - }, - { - "id": "TA-985", - "scope": "fragmented P2P multi-sender initialization reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-zero scatter, nonzero-rank adoption, explicit sender groups, peer-failure propagation, and engine prewarm ordering are branches of one multi-sender initialization state machine.", - "Locator copy modes, dense sharding, and rank-filtered transfer remain separate data-partitioning policies." - ] - }, - { - "id": "TA-986", - "scope": "split P2P transfer manifest rejection and name compatibility reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Missing locators, incompatible shapes, invalid source ranks, canonical name aliases, orig_mod stripping, and language-model prefix fallback form one receiver-manifest resolution policy.", - "Replicated locator staging and transfer metadata remain separate transport behaviors." - ] - }, - { - "id": "TA-987", - "scope": "split P2P source-staging mode reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Small CPU pooling, GPU-direct persistent registration and chunking, and aligned mixed-dtype scratch views are staging modes of transfer_bucket.", - "Receiver-handle coalescing and failure diagnostics remain separate scheduling and observability boundaries." - ] - }, - { - "id": "TA-988", - "scope": "split P2P pending-transfer and destroy reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pending failure propagation, skip-completion cleanup, failed-transfer draining, receiver completion, endpoint results, deregistration, and completion errors form one teardown lifecycle.", - "Preparation, transfer, and explicit complete_sync metadata remain separate protocol stages." - ] - }, - { - "id": "TA-989", - "scope": "one-wrapper-per-topology lm-head TP FSDP tests", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CP-replica, DP, no-CP DP, no-CP HSDP, and OPD variants invoke the same four-process launcher and embedded loss-gradient oracle.", - "All six topology and loss-mode programs remain executed with per-case failure identifiers in one distributed matrix." - ] - }, - { - "id": "TA-990", - "scope": "split empty and aborted adapter gradient-epoch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Empty-step rejection, idempotent abort, scratch reset, publication state, and poisoned or pending rejection form one pre-mutation epoch lifecycle.", - "Gradient capture and successful optimizer commit remain separate transactions." - ] - }, - { - "id": "TA-991", - "scope": "split authoritative adapter optimizer success reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Analytical clipping, AdamW parameters and moments, scratch reuse, global-step commit, and the single logical-norm collective are properties of one successful optim_step transaction.", - "Exact LM-head replicated optimizer coherence remains separate because it validates a different topology-specific helper." - ] - }, - { - "id": "TA-992", - "scope": "split authoritative adapter optimizer outcome reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pre-mutation semantic rejection, partial optimizer failure, and collective failure are outcome branches of optim_step.", - "Recoverability, poison state, publication gates, parameter mutation or preservation, and restart guidance remain asserted for every branch." - ] - }, - { - "id": "TA-993", - "scope": "split adapter checkpoint path and target-manifest reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Trusted-root confinement, strict target-manifest persistence, validation, and mismatch rejection form one checkpoint save and validation policy.", - "Restore compatibility and checkpoint structure remain separate load boundaries." - ] - }, - { - "id": "TA-994", - "scope": "split authoritative adapter checkpoint restore-plan reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Lifecycle reset, plan fingerprint restoration, compatible replacement, direct topology mismatch rejection, and atomic nonmutation form one authoritative restore policy.", - "Coordinator materialization and general session compatibility remain separate orchestration and user-policy boundaries." - ] - }, - { - "id": "TA-995", - "scope": "fragmented adapter optimizer save, resume, and admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shard emission and manifest identity rejection form one save contract; uninterrupted, evicted public, LR override, and weights-only control paths form one resume lifecycle.", - "Legacy and incomplete artifact rejection now includes resident-state atomicity; parameter identity, transactional commit, and logical resharding remain separate boundaries." - ] - }, - { - "id": "TA-996", - "scope": "split FP8 grouped forward and weight-gradient reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Same-NK forward and same-MN weight-gradient kernels are the two arithmetic halves of one grouped FP8 training contract.", - "Block-loop and Triton references, empty groups, tail shapes, block sizes, precomputed sequence offsets, and scalar-Quack dispatch remain asserted." - ] - }, - { - "id": "TA-997", - "scope": "fragmented SGLang fused-expert EP activation reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "DeepEP and FP8 exclusion, missing runtime, flag-off stock dispatch, score dtype, empty-rank behavior, and compute guards are branches of one EP dispatch admission boundary.", - "Happy-path compute, slot combine, trainable autograd, and weight presentation remain separate numerical or ownership contracts." - ] - }, - { - "id": "TA-998", - "scope": "split model-runner expert-factor compiler reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Certified unquantized backends, registered-rank specialization, block-FP8 DeepEP, NF4, NVFP4, and generic quantized admission all compile the same expert-factor ownership plan.", - "Module versus fused producer families, quantization guards, metadata mismatch, shape drift, and uncertified parallelism remain asserted in one compiler matrix." - ] - }, - { - "id": "TA-999", - "scope": "fragmented Qwen-235B simulator calibration and built-in pack reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Markdown ingestion, leave-one-out evaluation, calibrated and extrapolated scenarios, OOM boundaries, topology what-if and auto sweeps form one Qwen-235B calibration workflow.", - "Built-in pack replay now includes consolidated cross-pack validation; portable ledgers, security admission, and kernel ranking remain separate simulator boundaries." - ] - }, - { - "id": "TA-1000", - "scope": "split grouped checkpoint load routing and fallback reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense and expert routing, fused and FFN source formats, local dense fallback, absent EP-group fallback, strict rejection, and persistent-buffer filtering are branches of grouped_load_weights.", - "State-dict resolution, object broadcast, DTensor copying, and strict post-processing remain separate callables." - ] - }, - { - "id": "TA-1001", - "scope": "fragmented FP8 weight-sync selection and CPU expert reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Projection inclusion and exclusion, receiver skip lists, broad selection, stacked quantization, and already-FP8 passthrough form one input and layout policy.", - "CPU expert projection, zero padding, deferred formatting, exclusion, reusable workspace staging, and workspace quantization form one expert CPU pipeline." - ] - }, - { - "id": "TA-1002", - "scope": "split weight-sync quantization admission and enrichment reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unsupported quantization rejection and FP8 BF16-island enrichment are branches of handle_sync_inference_weights.", - "Backend post-processing, adapter materialization, parameter extraction, and sparse-delta sync remain separate transactions." - ] - }, - { - "id": "TA-1003", - "scope": "fragmented request-processor R3 side-payload reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Mooncake externalization, datum-order preservation, and normal, exceptional, and default cleanup form one model-pass side-payload lifecycle.", - "NCCL sync, dispatcher forwarding, optimizer and checkpoint operations, and token unpacking remain separate request boundaries." - ] - }, - { - "id": "TA-1004", - "scope": "split GLM5 indexer construction and selection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Parameter geometry, FP32 projection, sentinel masking, padding detection, and sorted and blocked selection form one indexer component contract.", - "The TileLang fast path and sparse-attention model integration remain separate implementation boundaries." - ] - }, - { - "id": "TA-1005", - "scope": "fragmented GLM52 sparse-selector pipeline reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Logical selection, Hadamard transport, fused projection, sampler key preparation, portable codecs, runtime dispatch, and dependency loading form one sparse-selector pipeline.", - "Production-shape SGLang CUDA codec parity is now a separate optional report so its runtime skip cannot mask the portable pipeline assertions." - ] - }, - { - "id": "TA-1006", - "scope": "split Muon builder configuration and SGD fallback reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Threading the fallback choice through build_optimizer and proving its state-free SGD update are one builder configuration contract.", - "Direct Gram-Newton-Schulz arithmetic, backend dispatch, and grouped update geometry remain separate numerical boundaries." - ] - }, - { - "id": "TA-1007", - "scope": "split Muon fused-expert classification and Nemotron integration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fused gate-up detection, FSDP attribute-loss recovery, gated and non-gated family classification, and a real Nemotron-H optimizer step form one parameter-ownership policy.", - "Matrix grouping and Newton-Schulz implementation details remain separate optimizer arithmetic reports." - ] - }, - { - "id": "TA-1008", - "scope": "fragmented distributed-checkpointer metadata, load, and save reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Metadata admission and synchronous, no-dist, custom-group, and asynchronous load and save routing form one DistributedCheckpointer I/O policy.", - "Optimizer metadata-key filtering now runs with the optimizer-state contract; model-key and LoRA compatibility remain separate schema boundaries." - ] - }, - { - "id": "TA-1009", - "scope": "fragmented exact GLM52 and Qwen3.5 training-program admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical GLM52 numerical resolution and official geometry form one model program; exact Qwen3.5 numerical, MoE, topology, and model-scope admission form another.", - "The family-independent RoPE selector remains separate because it also covers ordinary non-GLM behavior." - ] - }, - { - "id": "TA-1010", - "scope": "split quantized-export CLI parsing and base-directory reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "YAML and CLI precedence, size parsing, module invocation, BF16 islands, configuration output, and shard indexing form one end-to-end export command contract.", - "Projection layouts, MoE layouts, QARL logprob preservation, and source admission remain separate export boundaries." - ] - }, - { - "id": "TA-1011", - "scope": "fragmented create-model endpoint lifecycle reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Session-spec normalization, recreation admission, registration rollback, reserved-checkpoint initialization, and full-weight admission are outcomes of create_model_endpoint.", - "The lower-level create-session endpoint, termination, and checkpoint metadata lookup remain separate API lifecycles." - ] - }, - { - "id": "TA-1012", - "scope": "split adapter load admission, failure, and rank-zero broadcast reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct load success, trusted-path admission, synchronized failure, pipeline rejection, and auto-registration rollback form one handle_load_adapter_state outcome policy.", - "Rank-zero routing, sharded restoration, session-spec mismatch, and transactional optimizer rejection form one broadcast-load mode; eviction and registration remain separate lifecycles." - ] - }, - { - "id": "TA-1013", - "scope": "split packing edge, validation, and unpacking reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Empty, missing, oversized, NumPy, valid, and malformed inputs form one packing admission and output-validation policy.", - "Per-token unpacking modes now run with the end-to-end pack, metadata, forward-output, and sample-boundary round trip." - ] - }, - { - "id": "TA-1014", - "scope": "split teacher-cache contributor and distributed assembly reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CP and EP contributor selection, legacy duplicate mode, SP trimming, and cross-rank writer gathering form one distributed teacher-cache producer policy.", - "Mooncake storage round trips, OPD loss execution, and debug artifacts remain separate transport, numerical, and observability boundaries." - ] - }, - { - "id": "TA-1015", - "scope": "split OPD pipeline shift and teacher-cache transport reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Causal teacher shifting, cache-index alignment and rejection, and Mooncake metadata admission form one OPD pipeline payload contract.", - "Endpoint reuse, student-version verification, and preparation-worker queueing remain separate orchestration boundaries." - ] - }, - { - "id": "TA-1016", - "scope": "fragmented unquantized LoRA checkpoint export and round-trip reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Runtime-rank export, PEFT hybrid-shared layout, SGLang shared-outer layout, and adapter-manager loading for both ownership modes form one unquantized checkpoint workflow.", - "Low-level EP slicing and quantized projection-subset round trips remain separate conversion and representation boundaries." - ] - }, - { - "id": "TA-1017", - "scope": "split optimizer publication and fatal dispatcher failure reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Successful publication, commit and handler-tail poisoning, and rank-zero fatal termination are outcomes of one post-mutation optimizer publication lifecycle.", - "Forward-backward completion and explicit epoch abort remain separate gradient-epoch transactions." - ] - }, - { - "id": "TA-1018", - "scope": "split empty-layout and deterministic adapter initialization reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Empty logical packing, discovery, replica classification, and ownership compilation form one empty-shard layout policy.", - "Coordinate, replica, LoRA-B, session, and FQN-order invariance form one deterministic initialization contract; real Gloo and explicit EP composition remain separate topology reports." - ] - }, - { - "id": "TA-1019", - "scope": "split server filesystem and compile-worker security reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Authoritative artifact roots, symlink rejection, and private diagnostic input admission form one server filesystem trust boundary.", - "Compile-target allowlisting, safe protocol type round trips, and oversized-frame rejection form one worker IPC trust boundary; outbound endpoint validation remains separate." - ] - }, - { - "id": "TA-1020", - "scope": "fragmented adapter-manager optimizer and checkpoint-load reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Optimizer construction, hyperparameter persistence, learning-rate updates, and mixed-rank multi-optimizer reload form one manager configuration and persistence lifecycle.", - "Session compatibility, weights-only behavior, structure admission, PEFT suffixes, sharded indices, and rank capacity form one load compatibility policy." - ] - }, - { - "id": "TA-1021", - "scope": "split FSDP mixed-precision selection and reduce-dtype reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Topmost protected-module selection, expert FSDP policy stripping, mesh-dependent reduce dtype, explicit overrides, and dtype admission form one mixed-precision policy.", - "Sequence-parallel folding, optional-boolean parsing, and manual prefetch direction remain separate configuration boundaries." - ] - }, - { - "id": "TA-1022", - "scope": "golden-bit reports split below BI contract-tree version", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All family, normalization, mean, softmax, matrix, and LM-head frozen hashes now form one v1 golden-tree gate; normalization and head hashes form one v2 gate.", - "Every failure retains its case and output label, and both versioned gates share the same H100 capability contract." - ] - }, - { - "id": "TA-1023", - "scope": "split FlashQLA Gate 2 shape-invariance reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed-versus-individual rows, total-token invariance, and block-DV tile invariance are the three shape branches of one Gate 2 bitwise contract.", - "Auto-CP admission and Gate 4 chunk-state handoff remain separate because they exercise different production decisions." - ] - }, - { - "id": "TA-1024", - "scope": "split MiniMax M3 activation, router, and text-runtime reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Clamped SwiGLU, biased sigmoid routing, text forward and backward, multimodal-token rejection, and parallel-mode admission form one MiniMax M3 runtime program.", - "Configuration, checkpoint mapping, and sparse-attention paging remain separate serialization and kernel boundaries." - ] - }, - { - "id": "TA-1025", - "scope": "cross-engine RMSNorm reports split by serving site class", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "QK, pre-summed residual-tree, post-attention residual, zero-centered family-1, and families-v2 cases share one cross-engine bitwise oracle and shape matrix.", - "The module-level SGLang dependency gate remains visible; native family admission, module dispatch, and fused backward stay in separate suites." - ] - }, - { - "id": "TA-1026", - "scope": "split Qwen3.5 Class-B and attention rotary reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Class-B dispatch precedence and fail-closed CPU or dtype admission form one rotary admission policy.", - "Dense and MoE half-rotate behavior plus exact post-RoPE BF16 casting form one attention projection policy; the low-level interleaved reference remains separate." - ] - }, - { - "id": "TA-1027", - "scope": "split Mooncake tensor codec and hidden-transport reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Tensor byte round trips, canonical dtype strings, metadata emission, suffixed object keys, and rank-2 and rank-3 fetches form one hidden transport contract.", - "Teacher cache consumption, malformed metadata, removal, and configuration precedence remain separate consumer, admission, and lifecycle boundaries." - ] - }, - { - "id": "TA-1028", - "scope": "split fused-GDN merged-forward and cache reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical LoRA folding, slice-local gradients, exact output projection, cache reuse, bounded generations, and release after adapter publication form one merged-weight lifecycle.", - "Delta arithmetic, manifest filtering, and sharded checkpoint loading remain separate representation and serialization boundaries." - ] - }, - { - "id": "TA-1029", - "scope": "split exact GDN module-dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fused gated RMSNorm routing, unsupported residual rejection, and full GatedDeltaNet exact-program routing form one model-program dispatch contract.", - "Gating and normalization forward and backward numerics plus the solve-tril warp pin remain separate kernel contracts." - ] - }, - { - "id": "TA-1030", - "scope": "split BI router GEMM forward and backward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP32 reference agreement, empty and dtype admission, and analytical hidden and weight gradients form one router GEMM numerical contract.", - "Leading-dimension linear behavior, top-k weight processing, and MoE block integration remain separate API and model boundaries." - ] - }, - { - "id": "TA-1031", - "scope": "split RMSNorm family structure and numerical-routing reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Family-name admission, residual-shape rejection, and Qwen site declarations form one explicit family API structure policy.", - "Funnel equivalence, family vitality, zero-centered folding, and module dispatch form one GPU numerical-routing contract; undeclared-family enforcement remains a separate tripwire." - ] - }, - { - "id": "TA-1032", - "scope": "split SGLang-fused RMSNorm forward and backward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Residual and no-residual forward bit exactness and their analytical backward comparisons share one fused numerical implementation and CUDA gate.", - "CPU fallback, model integration, and trunk-contract dispatch remain separate portability and integration boundaries." - ] - }, - { - "id": "TA-1033", - "scope": "split DeepSeek V4 shared and routed SwiGLU limit reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The shared MLP and eager routed-expert checks jointly prove that the same configured SwiGLU limit reaches both expert implementations.", - "MoE routing, hash-table admission, and routing replay remain separate structural and lifecycle contracts." - ] - }, - { - "id": "TA-1034", - "scope": "split Nemotron H published-layout load, parity, and save reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Strict load accounting, HF numerical parity, ignored MTP admission, and byte-exact save reconstruction form one published-checkpoint codec transaction.", - "The supported published key set is explicit so ignored MTP input is not incorrectly required in saved output; stacked HF input and EP ownership remain separate representations." - ] - }, - { - "id": "TA-1035", - "scope": "split Qwen MoE fused-expert checkpoint-handler reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Qwen3 per-expert and Qwen3.5 stacked layouts now report through one fused-expert load/save policy, including deferred QLoRA expert loading.", - "Expert parameter registration and model-weight QARL filtering remain separate module and filesystem boundaries." - ] - }, - { - "id": "TA-1036", - "scope": "split QLoRA quantized execution and NVFP4 scale-merge reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quantized storage, forward and backward, dequantization, prequantized loading, EMA scale convention, and LoRA merge-requantization form one quantized-weight lifecycle.", - "Injection, block-FP8 checkpoint representation, and optimizer-state reset remain separate integration and lifecycle contracts." - ] - }, - { - "id": "TA-1037", - "scope": "split DeepSeek V4 decoder runtime and gradient-checkpoint reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "C128 execution, causal-LM forward and backward, hash-layer threading, and decoder checkpoint wrapping form one model runtime contract.", - "Construction and topology admission plus dtype preservation remain separate structural policies." - ] - }, - { - "id": "TA-1038", - "scope": "split DTensor checkpoint and rank-zero broadcast loading reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Copying full checkpoint tensors into existing replicated or sharded DTensors and materializing multi-axis DTensors for one writer are the load and save directions of one tensor checkpoint codec.", - "Object payload transport, NCCL device selection, weight-load group routing, and handler-filtered rank-zero loading form one broadcast loading transaction; state-dict resolution and grouped expert routing remain separate policies." - ] - }, - { - "id": "TA-1039", - "scope": "split Muon Gram-Newton-Schulz construction and execution reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Builder keyword propagation, byte-limit admission, SGD fallback, an actual Gram-Newton-Schulz parameter update, and restart autotuning form one configured optimizer lifecycle.", - "Quack backend selection, grouped matrix scheduling, standard Newton-Schulz, CUDA compute dtype, and model parameter classification remain separate branches." - ] - }, - { - "id": "TA-1040", - "scope": "split BI trunk-linear forward and backward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Persistent-GEMM forward bit exactness, batch invariance, dtype admission, and cuBLAS input, weight, and bias gradients form one wrapped-linear numerical contract under the same CUDA gate.", - "Wrapper selection and global-interpose training admission remain separate structural and global-mode boundaries." - ] - }, - { - "id": "TA-1041", - "scope": "split DeepSeek V3 expert checkpoint layout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "External per-expert and internal fused layouts, dense and packed EP slicing, requested device and dtype, and quantization-config discovery form one expert checkpoint codec policy.", - "Every prior dense, packed, internal, EP-local, and configured quantization assertion remains in the surviving report." - ] - }, - { - "id": "TA-1042", - "scope": "split Nemotron H model runtime and gradient-checkpoint reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Router output, loss backward through every mixer type, and full-layer gradient checkpointing form one model training runtime contract.", - "Packed variable-length equivalence remains separate because it changes document-boundary state propagation." - ] - }, - { - "id": "TA-1043", - "scope": "split RoPE registry precision and exact-lane bit reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Registry-wide FP32 frequency construction, BF16 consumption, and unchanged zero-K3 contract-lane cosine and sine bits form one frequency-precision policy.", - "Lazy cache growth and architecture-specific CUDA construction remain separate cache-lifecycle and device-placement boundaries." - ] - }, - { - "id": "TA-1044", - "scope": "split DistSignSGD builder and direct-step reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FSDP2 builder admission, weight-decay grouping, hook configuration, and a preaggregated update with decoupled decay form one configured optimizer contract.", - "Reduce-scatter sign timing and local-versus-FSDP hook ownership remain separate communication and gradient-ownership boundaries." - ] - }, - { - "id": "TA-1045", - "scope": "split sequence-parallel and LM-head explicit synchronization reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sequence-parallel gradient sums, DTensor skipping, adapter-finalization exclusions, LM-head replica gradient handling, and marked-parameter broadcast form one explicit synchronization policy.", - "Token counting, gradient clipping, pipeline loss chunking, and their independent collectives remain separate trainer contracts." - ] - }, - { - "id": "TA-1046", - "scope": "split dense and MoE hidden-component hook reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Residual, attention, normalization, dense MLP, routed-MoE callback, and shared-expert components are alternate producers for one hidden-component capture pipeline.", - "Summary formatting, ranked tensor dumps, and trusted override loading remain separate output and trust boundaries." - ] - }, - { - "id": "TA-1047", - "scope": "split base, pipeline, and LoRA checkpoint compatibility reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local and pipeline key discovery, metadata unions, QARL buffer mismatch, base-to-LoRA loading, and LoRA-only loading form one checkpoint model-compatibility policy.", - "Distributed checkpoint transport and optimizer-state filtering remain separate IO and payload boundaries." - ] - }, - { - "id": "TA-1048", - "scope": "split gradient-epoch completion, abort, and failure reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forward-backward rendezvous and commit, explicit abort routing, uniform rejection, and rank-asymmetric failure conversion form one dispatcher gradient-epoch lifecycle.", - "Session registration, save operations, and post-optimizer publication poisoning remain separate RPC and mutation boundaries." - ] - }, - { - "id": "TA-1049", - "scope": "split DeepSeek V3 router-output and replay reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Aux-loss-driven router-logit emission across sparse layers and recording selected indices and weights form one router observability and replay contract.", - "Model backward and router freezing plus LoRA target injection remain separate training and adapter policies." - ] - }, - { - "id": "TA-1050", - "scope": "split MoE token permutation and unpermutation reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Expert-sorted routing weights and scatter-add reconstruction are the encode and decode directions of one memory-efficient token permutation codec.", - "All-to-all score ordering and hidden-dimension chunking remain separate transport boundaries." - ] - }, - { - "id": "TA-1051", - "scope": "split SignSGD builder and direct-step reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Builder construction, decay and no-decay grouping, dense sign updates, decoupled weight decay, and sparse-gradient rejection form one configured SignSGD contract.", - "DistSignSGD distributed communication and cautious-decay behavior remain in their dedicated suites." - ] - }, - { - "id": "TA-1052", - "scope": "split Qwen3.5 families-v2 zero-centered backward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Effective folded-weight gradients and the residual twin's output and residual gradient paths are two call shapes of one zero-centered families-v2 backward contract.", - "Both CPU reference comparisons and every input, residual, and weight gradient assertion remain in the surviving report." - ] - }, - { - "id": "TA-1053", - "scope": "split FP8 config translation and Blackwell admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Legacy FP8 config normalization, incompatible external runtime rejection, and explicit Blackwell artifact admission form one FP8 configuration boundary.", - "BF16 layer-island resolution and injection remain a separate model transformation policy." - ] - }, - { - "id": "TA-1054", - "scope": "split Quack PTX compilation output and entry-discovery reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unique temporary outputs, bounded ptxas execution, cleanup, and exact kernel-entry discovery form one PTX compilation process-safety contract.", - "Worker framing and cache-key hashing remain separate IPC and cache-trust boundaries." - ] - }, - { - "id": "TA-1055", - "scope": "split merged-LoRA fold, cache, and fused-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical linear and expert folding plus straight-through factor gradients now form one low-level fold numerical contract.", - "LoraLinear merged selection, gradient parity, optimizer and active-rank cache invalidation form one linear lifecycle; MoE canonical merged views, versioned caches, and fused-expert admission form one expert lifecycle.", - "Native EP execution and trunk-linear wrapping remain separate integration boundaries." - ] - }, - { - "id": "TA-1056", - "scope": "split MoE-LoRA construction and GPU numerical reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Backend-specific initialization, frozen and trainable ownership, runtime rank slicing, from-module conversion, model injection, and block injection form one construction policy.", - "Zero-delta base equivalence and eager-versus-native or Triton output and gradient agreement form one GPU numerical policy under the same capability gate.", - "CPU eager execution, zero-token gradients, and EP router-score application remain separate runtime boundaries." - ] - }, - { - "id": "TA-1057", - "scope": "split NVFP4 2D and expert fake-quant reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Independent 2D reference agreement, shape admission, linear STE, 3D expert STE, expert isolation, and fused gate-up scale ownership form one NVFP4 fake-quant contract.", - "Every dense, expert, and fused projection representation remains covered in the surviving report." - ] - }, - { - "id": "TA-1058", - "scope": "split P2P prepare, completion, and cleanup reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Cold and cached prepare behavior are states of one P2P initialization handshake under the same backend contract.", - "Successful completion, pending-transfer failure, receiver notification, and destroy cleanup form one terminal synchronization lifecycle.", - "Fanout, slicing, coalescing, diagnostics, and multi-sender routing remain separate transport boundaries." - ] - }, - { - "id": "TA-1059", - "scope": "split NCCL endpoint and flattened transfer reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Initialization and transfer endpoint ports plus the optional two-phase receiver protocol form one NCCL endpoint transaction policy.", - "Flat, chunked-flat, and receiver-fenced hybrid buckets are load-format branches of one flattened transfer contract.", - "Endpoint health and invalid multi-rank direct format remain independent admission boundaries." - ] - }, - { - "id": "TA-1060", - "scope": "split GDN contract packing, routing, admission, and state reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed convolution weights, armed routing, fail-closed admission, call-scoped state, and checkpoint recomputation form one exact GDN contract lifecycle.", - "Low-level CUDA parity, full-block integration, and optional SGLang tree-kernel parity remain distinct numerical boundaries." - ] - }, - { - "id": "TA-1061", - "scope": "split FlashAttention API and page-size-one cache path reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fixed-length and variable-length calls are public shapes of one FlashAttention API behavior contract.", - "SGL, paged FlashAttention, flags-off, and FA4 selection are branches of one page-size-one KV-cache routing policy.", - "Backend registry resolution and eager head-layout numerics remain separate implementation boundaries." - ] - }, - { - "id": "TA-1062", - "scope": "split authoritative adapter optimizer outcome reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Empty and aborted epochs, a successful clipped update, nonfinite input, optimizer failure, and collective failure are states of one authoritative optimizer lifecycle.", - "Capture ownership, publication admission, exact LM-head coherence, and checkpoint restore remain separate component boundaries." - ] - }, - { - "id": "TA-1063", - "scope": "split optimizer checkpoint save and resume reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sharded manifest emission, identity rejection, moment restoration, and bitwise continuation are the write and read sides of one optimizer checkpoint codec.", - "Artifact admission and logical cross-layout resharding remain separate compatibility boundaries." - ] - }, - { - "id": "TA-1064", - "scope": "split sampling adapter reconciliation and scoped tracking reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Stale-state reconciliation, query failure preservation, model-scoped tracking, and failed-load atomicity form one sampling-session adapter lifecycle.", - "Sampler checkpoint storage and adapter-only export remain separate filesystem and orchestrator operations." - ] - }, - { - "id": "TA-1065", - "scope": "split inference weight-sync forwarding and admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Endpoint forwarding, pool selection, quantization admission, and cache invalidation are request branches of one inference weight-sync API transaction.", - "Endpoint registration, health refresh, and receiver capability detection remain separate lifecycle boundaries." - ] - }, - { - "id": "TA-1066", - "scope": "split model-session creation and kill reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Normalized registration, duplicate and topology admission, reserved checkpoint handling, kill, optional final save, and default-session protection form one model-session lifecycle.", - "The lightweight create-session alias remains a separate endpoint registration contract." - ] - }, - { - "id": "TA-1067", - "scope": "split checkpoint weights-info and legacy session metadata reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Disk-backed weights information, path admission, full-weight metadata, and legacy SignSGD metadata upgrade form one checkpoint session-spec decoding policy.", - "No metadata assertion or legacy compatibility case was removed." - ] - }, - { - "id": "TA-1068", - "scope": "split quantized exporter projection and MoE layout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fused QKV, MLA, linear-attention, GKN expert, and fused gate-up layouts are architecture branches of one exported-model tensor layout contract.", - "CLI and directory behavior, source admission, QARL folding, and low-level FP8 quantization remain separate boundaries." - ] - }, - { - "id": "TA-1069", - "scope": "split expert-adapter backend capability and factor ownership reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Backend capability and plan identity plus factor ownership, reduction domains, and checkpoint persistence form one expert-adapter structural contract.", - "Construction and semantic preservation remain independently reported." - ] - }, - { - "id": "TA-1070", - "scope": "split generic and model-family expert-adapter injection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact target-subset forwarding and GLM, Qwen3, and Qwen3.5 wrapper construction form one expert-adapter injection policy.", - "Each family, backend, target set, quantization format, and checkpoint-buffer assertion remains covered." - ] - }, - { - "id": "TA-1071", - "scope": "split expert-adapter semantic preservation and fail-closed reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Supported SiLU preservation and rejection of incompatible activations, biases, quantization groups, target sets, and model-family semantics form one fail-closed semantic contract.", - "Runtime numerical parity remains outside this structural suite." - ] - }, - { - "id": "TA-1072", - "scope": "split canonical MoE plan and group-alias topology reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Trainer and sampler plan identity, topology admission, logical ordinals, and world-32 CP, EP, and expert-FSDP group aliases form one canonical MoE topology contract.", - "Graph metadata and transport selection remain separate boundaries." - ] - }, - { - "id": "TA-1073", - "scope": "split canonical MoE distributed transport numerical reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two- and eight-contributor dense transport plus sixteen-contributor packed and CP-sharded parity are topology cases of one distributed transport numerical contract.", - "All 2-, 8-, and 16-process subprocess checks still execute in the surviving report." - ] - }, - { - "id": "TA-1074", - "scope": "split adapter auto-load, explicit load, and rank-zero broadcast reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Evicted auto-load, fresh materialization, explicit path load, rollback, and all-rank or rank-zero-broadcast restoration are branches of one adapter load lifecycle.", - "Adapter registration and save admission remain separate mutation boundaries." - ] - }, - { - "id": "TA-1075", - "scope": "split token selection and CP-sharded KL position reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Target selection, boundary behavior, raw-weight cross-checking, hidden summaries, and CP-sharded KL position mapping form one token-diagnostics policy.", - "Hidden-component capture, tensor-dump output, and trusted override input remain separate boundaries." - ] - }, - { - "id": "TA-1076", - "scope": "split OPD backend numerics and gradient-reduction reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Reference, streaming, low-memory, and sharded-store forward agreement plus backward, partial reduction, and output dtype form one OPD numerical backend contract.", - "Output-shape edge behavior and hidden-only distance remain separately reported." - ] - }, - { - "id": "TA-1077", - "scope": "split fused selected-logprob small and production-vocabulary numerical reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dtype, bias, temperature, frozen-head, irregular-tail, Qwen, and GPT-OSS vocabulary cases form one fused selected-logprob forward and backward numerical contract.", - "Loss-dispatch agreement and the no-full-logits memory gate remain separate integration and resource boundaries." - ] - }, - { - "id": "TA-1078", - "scope": "split DRGRPO temperature behavior from forward numerical report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forward value, gradients, metrics, zero boundaries, advantage direction, KL penalty, and logprob-temperature behavior form one DRGRPO objective contract.", - "Microbatch reducer composition remains a separate aggregation boundary." - ] - }, - { - "id": "TA-1079", - "scope": "split packing capacity and edge-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Concatenation, capacity splitting, mixed lengths, empty and single input, oversize admission, missing fields, NumPy conversion, and microbatch validation form one core packing policy.", - "Packed metadata, disabled mode, and full pack-to-unpack roundtrip remain separate contracts." - ] - }, - { - "id": "TA-1080", - "scope": "split receiver postprocess and sync quantization configuration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Receiver postprocess selection, FP8 KV-cache requirements, unsupported-format admission, and generated BF16 islands form one quantized weight-sync configuration contract.", - "Tensor quantization numerics remain in the dedicated FP8 sync suite." - ] - }, - { - "id": "TA-1081", - "scope": "split LoRA preparation, parameter extraction, and inference-layout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Adapter materialization, parameter filtering and tied aliases, compile-name normalization, and architecture-specific unfusing form one sync-source tensor preparation pipeline.", - "Bucket sizing, transport routing, and sparse-delta selection remain separate boundaries." - ] - }, - { - "id": "TA-1082", - "scope": "split EP collection admission and expert-data reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "P2P sender selection, direct-EP collection admission, and gated or nongated local expert projection collection form one EP synchronization-source policy.", - "Actual transport remains independently tested." - ] - }, - { - "id": "TA-1083", - "scope": "split checkpoint save failure and exact-active-LoRA admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Factor-only admission, rank-zero artifact failures, LoRA-only failures, and pre-barrier error surfacing form one fail-closed checkpoint save policy.", - "No downstream conversion or collective is allowed after an admission or write failure." - ] - }, - { - "id": "TA-1084", - "scope": "split dense and MoE LoRA save-format reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Live dense target resolution and collective stacked-MoE factor slicing are format branches of one LoRA checkpoint export contract.", - "Optimizer-state checkpointing remains in the adapter optimizer resume suite." - ] - }, - { - "id": "TA-1085", - "scope": "split removed and unsupported server configuration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Removed fields, incompatible quantized modes, vLLM-only runtime knobs, broadcast loading, and unsupported multi-adapter modes form one fail-closed server configuration boundary.", - "Valid shipped and feature-specific configurations remain independently reported." - ] - }, - { - "id": "TA-1086", - "scope": "split general, optimizer, runner, and model-specific server runtime reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical defaults, nested runtime controls, receiver cache dtype, R3 transport, gradient buckets, Muon options, runner compatibility, sparse MLA, and MoE routing controls form one runtime configuration roundtrip.", - "Quantized-training and parallel-topology configuration remain separate specialized boundaries." - ] - }, - { - "id": "TA-1087", - "scope": "split teacher-cache distributed assembly and Mooncake integration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Contributor selection, CP and DP assembly, valid-label trimming, Mooncake metadata emission, byte roundtrip, and activation-cache consumption form one teacher hidden-cache lifecycle.", - "OPD loss execution and debug artifacts remain separate consumers and outputs." - ] - }, - { - "id": "TA-1088", - "scope": "split Muon Gram-Newton-Schulz configuration and grouping reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Builder options, fallback behavior, restart autotuning, grouped shapes, transpose equivalence, fused halves, and byte-limit chunking form one configured Gram-Newton-Schulz optimizer contract.", - "Quack backend selection, standard Newton-Schulz, and CUDA compute dtype remain separate algorithm and platform boundaries." - ] - }, - { - "id": "TA-1089", - "scope": "split Triton-grouped and scalar-Quack FP8 numerical reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Block-loop and Triton-grouped forward and weight gradients plus scalar-Quack per-expert scaling are backend branches of one grouped FP8 GEMM numerical contract.", - "DeepGEMM subprocess isolation and model train-step integration remain separate gates." - ] - }, - { - "id": "TA-1090", - "scope": "split dispatcher selection, routing slice, and row-provenance reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "DP, EP, CP, and legacy rank selection, routing-payload slicing, rank-local row grouping, and source provenance form one dispatcher input-distribution policy.", - "Packing strategy and R3 payload storage remain separate upstream boundaries." - ] - }, - { - "id": "TA-1091", - "scope": "split completion rendezvous and per-token result merging reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local payload trimming, rank rendezvous, CP replica deduplication, disagreement rejection, and rank-zero per-token merging form one dispatcher completion transaction.", - "Diagnostic dumping remains a separate output boundary." - ] - }, - { - "id": "TA-1092", - "scope": "split request-processor and runner-dispatcher forward lifecycle reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Processor readiness, forward and backward execution, timing propagation, invalid-target rejection, shutdown, model identity, auto-load, and R3 forwarding form one request-to-runner compute lifecycle.", - "NCCL sync, optimizer and checkpoint RPCs, and payload storage remain separate operations." - ] - }, - { - "id": "TA-1093", - "scope": "split raw-numerator accumulation and staged gradient-capture reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Raw numerator accumulation, model-gradient clearing, FP32 scratch reuse, staged commit, DTensor preservation, and atomic prevalidation form one gradient-capture transaction.", - "Ownership-plan compilation and optimizer mutation remain separate lifecycle boundaries." - ] - }, - { - "id": "TA-1094", - "scope": "standalone exact LM-head optimizer coherence report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Scalar-tensor optimizer-state coherence is a distributed validation branch of the authoritative optimizer lifecycle, not an independently meaningful user behavior.", - "Its collective mocks and assertions now run inside the optimizer lifecycle contract." - ] - }, - { - "id": "TA-1095", - "scope": "split adapter checkpoint trust, restore-lifecycle, and compatibility reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Trusted paths, strict target manifests, lifecycle reset, ownership-plan admission, optimizer compatibility, learning-rate rules, checkpoint structure, and PEFT sharding form one checkpoint restore and admission policy.", - "Coordinator-driven checkpoint materialization remains a separate server integration boundary." - ] - }, - { - "id": "TA-1096", - "scope": "split adapter eviction and mixed-adapter manager reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Mixed ranks and optimizers, adapter switching, training, checkpoint reload, capacity eviction, dirty-state protection, multi-rank rejection, and save-failure rollback form one multi-adapter lifecycle.", - "No eviction or failure assertion was removed." - ] - }, - { - "id": "TA-1097", - "scope": "split state-dict resolution and rank-zero loading reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local-versus-broadcast shard resolution is the input phase of rank-zero checkpoint transport and loading.", - "Object transport, group selection, handler-filtered prefetch, and state-dict resolution now form one rank-zero loading policy." - ] - }, - { - "id": "TA-1098", - "scope": "split grouped-load routing and strict postprocess coverage reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense and expert routing, format conversion, group fallback, and strict parameter and persistent-buffer coverage are phases of one grouped checkpoint-loading transaction.", - "Distributed DTensor checkpoint materialization remains a separate save-side correctness gate." - ] - }, - { - "id": "TA-1099", - "scope": "split canonical GLM-5.2 trainer topology and layer-plan reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Certified world, EP, and CP topology plus the official producer schedule, pipeline split, malformed-plan rejection, and indexer allocation form one canonical layer-plan contract.", - "The topology admission assertions still execute before layer-plan validation." - ] - }, - { - "id": "TA-1100", - "scope": "split index-share lifecycle and FSDP identity reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Publication, reuse, concurrency rejection, exception cleanup, and identity preservation across FSDP input casting form one index-share lifecycle.", - "All lifecycle and model-integration assertions remain intact." - ] - }, - { - "id": "TA-1101", - "scope": "split correction-bias checkpoint and canonical MoE configuration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Correction-bias FP32 preservation and checkpoint admission are router branches of the canonical MoE configuration and selection contract.", - "Native sampler codec parity and end-to-end semantic logprob composition remain independent gates." - ] - }, - { - "id": "TA-1102", - "scope": "split SGLang fused-MoE resolution and block-dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Automatic and explicit enablement, eligibility, logging, flag-off preservation, and block entrypoint selection form one fused-MoE resolution and dispatch policy.", - "Real-kernel parity remains an independent GPU gate." - ] - }, - { - "id": "TA-1103", - "scope": "split fused-expert admission and trainable-dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Semantic admission failures and the gradient-sensitive choice between the autograd function and plain kernel are branches of one fused-expert dispatch contract.", - "Gradient numerics remain separately capability-gated." - ] - }, - { - "id": "TA-1104", - "scope": "split fused-MoE weight-mode and strided-adapter layout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Transient, cached, and zero-copy strided modes, cache invalidation, serving tensor layout, and adapter gate-up layout form one kernel weight-layout contract.", - "No cache, storage-alias, or layout assertion was removed." - ] - }, - { - "id": "TA-1105", - "scope": "split simulator topology, shape, and analytical-ledger reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Topology resolution, local-token shapes, FLOPs, activation storage, and communication bytes form one analytical accounting contract.", - "Observed benchmark ingestion and kernel ranking remain separate empirical policies." - ] - }, - { - "id": "TA-1106", - "scope": "split simulator config and metadata resolution from path-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Config fingerprinting, cached and known-model metadata resolution, calibration-pack containment, symlink rejection, and restricted local reads form one input resolution and admission policy.", - "All traversal and trust-boundary assertions remain intact." - ] - }, - { - "id": "TA-1107", - "scope": "split ad-hoc Qwen and built-in calibration-pack reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Markdown ingestion, calibration evaluation, scenario planning, built-in pack replay, feasibility boundaries, and consolidated validation form one calibration lifecycle.", - "Generic observed-run ingestion remains an independent input format and planning contract." - ] - }, - { - "id": "TA-1108", - "scope": "split direct-output and expert eFSDP real-autograd reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct-output sharding plus shared-owner and all-owner expert layouts are branches of one four-GPU FSDP adapter-gradient ownership contract.", - "Each distributed subprocess and certification marker remains required." - ] - }, - { - "id": "TA-1109", - "scope": "split unquantized and quantized expert AllToAll reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Eager, Triton, native, Quack, NF4, NVFP4, block-FP8, projection-subset, and all-owner cases are backend and representation branches of one AllToAll ownership policy.", - "Every two-GPU subprocess still executes." - ] - }, - { - "id": "TA-1110", - "scope": "split unquantized and quantized DeepEP expert reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Hybrid-shared, all-owner, and quantized Quack cases form one optional DeepEP adapter-gradient ownership contract.", - "The optional dependency and two-GPU capability gate remain on the report." - ] - }, - { - "id": "TA-1111", - "scope": "split positive SGLang EP presentation and dispatch-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Top-k-one pair presentation, FP32 routing weights, local weight layout, flag-off preservation, empty-rank handling, and semantic rejection form one EP dispatch and admission policy.", - "Slot combine, trainable dispatch, weight modes, and real gradient parity remain separate contracts." - ] - }, - { - "id": "TA-1112", - "scope": "split FP8 sync quantization and projection-selection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "BF16 islands, block scale and zero-padding semantics, projection selection, skip lists, stacked tensors, and existing FP8 values form one CPU sync-quantization contract.", - "Expert workspace and GPU execution remain independently reported." - ] - }, - { - "id": "TA-1113", - "scope": "split dense, quantized, MoE, and fused-GDN LoRA sync-extraction reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense LoRA, QLoRA, quantized MoE factors, and fused GDN factors are architecture branches of one adapter-folding sync-source policy.", - "Every merged-weight and raw-factor exclusion assertion remains intact." - ] - }, - { - "id": "TA-1114", - "scope": "split generic, exact-LM-head, and replica-topology runner compiler reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Generic module and direct-output ownership, exact TP16 LM-head VJP masks, replica divisors, group coverage, and fail-closed topology admission form one runner gradient-ownership compiler policy.", - "Expert-factor compilation remains a separate specialized contract." - ] - }, - { - "id": "TA-1115", - "scope": "split effective LM-head selection and authoritative analytical-step reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Merged and legacy effective-head selection are input branches of the direct-output analytical gradient capture and optimizer step.", - "The surviving contract checks selection bytes, gradients, capture, norm, and parameter mutation end to end." - ] - }, - { - "id": "TA-1116", - "scope": "split AnyPrecision AdamW state-strategy and DTensor-offload reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Denominator chunking, Kahan compensation, gradient reuse, CPU state offload, DTensor local-shard wrapping, and device restoration form one optimizer state-strategy policy.", - "Cautious decay math remains independently reported." - ] - }, - { - "id": "TA-1117", - "scope": "split routed-expert topology, owner remap, and physical-buffer reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP16 and MoE-TP1 admission, all owner-slot remaps, global and owner-local factor banks, sampler buffer shapes, dtypes, and zero padding form one routed-bank layout policy.", - "Literal sampler numerics remain a separate GPU gate." - ] - }, - { - "id": "TA-1118", - "scope": "split zero-token, sentinel, and mixed-owner routed-gradient reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Owned zero-base gradients, all-sentinel structural zeros, input-layout rejection, and top-k-eight mixed-owner VJPs are branches of one routed-gradient edge policy.", - "The Hopper and SGLang capability gate remains on the combined report." - ] - }, - { - "id": "TA-1119", - "scope": "split ownership compiler topology, fingerprint, declaration, and replica reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Topology declarations, authority masks, rank-local fingerprint invariance, fail-closed structural admission, and orthogonal replica coverage are phases of one ownership-plan compilation transaction.", - "Compiled producer execution and residual gradient transport remain separate runtime contracts." - ] - }, - { - "id": "TA-1120", - "scope": "split shared-expert construction and logical checkpoint-state reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Construction, runtime admission, logical state, and checkpoint-state behavior form one shared-expert contract.", - "Physical SGLang views remain separate so optional dependency admission cannot hide the CPU structural report." - ] - }, - { - "id": "TA-1121", - "scope": "split exact TP1 configuration and dtype-state identity reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Runtime admission and dtype moves are lifecycle phases of the exact TP1 configuration contract.", - "Packed-state master dtype and object identity assertions remain intact." - ] - }, - { - "id": "TA-1122", - "scope": "split exact TP1 forward VJP and backward-safety reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Reference forward values, surrogate gradients, and rejected unsafe backward paths form one numerical and autograd policy.", - "All value, gradient, and error assertions remain intact." - ] - }, - { - "id": "TA-1123", - "scope": "split request-scoped NCCL synchronization from orchestrator control operations", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Request-scoped NCCL group naming is a synchronization branch of the optimizer, checkpoint, registration, and lifecycle control report.", - "The exact group-name and operation assertions remain intact." - ] - }, - { - "id": "TA-1124", - "scope": "split GLM5 indexer construction and DSA mask reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Indexer construction and its DSA mask behavior form one indexer-selection policy.", - "The mask branch runs in an isolated monkeypatch context within the surviving report." - ] - }, - { - "id": "TA-1125", - "scope": "split sparse MLA reference wrapper and attention-integration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Reference behavior, wrapper semantics, and attention integration are layers of one sparse-MLA policy.", - "Full forward and recompute behavior remain separate end-to-end gates." - ] - }, - { - "id": "TA-1126", - "scope": "split GLM5 sparse-KV adapter weight and dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sparse-KV adapter weights, adapter dispatch, and MoE dispatch form one adapter-and-routing policy.", - "The sparse-KV branch runs in an isolated monkeypatch context." - ] - }, - { - "id": "TA-1127", - "scope": "split native routed-partial module entry and actual-operand capture reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FSDP pre-forward hook entry is part of the native-combine execution boundary whose actual operands are captured.", - "Collective padding and fused-gate gradient parity remain independent contracts." - ] - }, - { - "id": "TA-1128", - "scope": "split optimizer parameter identity and transaction snapshot failure reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical identity, binding validation, recursive snapshots, and failed-collective commit behavior form one optimizer transaction policy.", - "Logical resharding remains independently reported." - ] - }, - { - "id": "TA-1129", - "scope": "split optimizer checkpoint resume and artifact-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Successful sharded save and bitwise resume and failed artifact admission are complementary branches of one checkpoint recovery policy.", - "Legacy pickle, missing artifact, and resident-state preservation assertions remain intact." - ] - }, - { - "id": "TA-1130", - "scope": "split P2P receiver placement and replicated staged-source reuse reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-receiver slices and replicated locators are branches of one receiver-placement policy.", - "The session, pointer, length, and staged-source identity assertions remain intact." - ] - }, - { - "id": "TA-1131", - "scope": "split P2P staging and receiver-handle coalescing reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Typed staging, registration lifetime, alignment, and coalescing are phases of one transfer staging policy.", - "Coalescing with and without receiver handles runs in an isolated environment context." - ] - }, - { - "id": "TA-1132", - "scope": "split P2P flush and weight-version payload from sync completion", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Flush-cache and weight-version propagation are completion-payload branches of the sync lifecycle.", - "Cache retention, tied aliases, failure completion, and cleanup remain in the same lifecycle report." - ] - }, - { - "id": "TA-1133", - "scope": "split sampler adapter reconciliation from adapter tracking and invalidation", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Reconciliation, model-scoped atomic tracking, receiver invalidation, listing, deletion, and resolution form one sampler adapter-state policy.", - "Sampler-weight export remains a separate orchestration contract." - ] - }, - { - "id": "TA-1134", - "scope": "split zero-token LoRA structural gradients from eager expert forward and backward", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Zero-token output is the empty-input branch of eager expert forward and backward behavior.", - "Every local-factor structural-gradient assertion remains intact." - ] - }, - { - "id": "TA-1135", - "scope": "split prequantized exclude metadata parsing from checkpoint-handler exclusion behavior", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Metadata formats, precedence, malformed inputs, dense and MoE skip behavior, and auxiliary-key passthrough form one checkpoint exclusion policy.", - "General prequantized detection and non-excluded loading remain independently reported." - ] - }, - { - "id": "TA-1136", - "scope": "split exact GLM52 attention inventory and construction-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The canonical attention-factor inventory and fail-closed input branches define one exact construction policy.", - "All rank, alpha, component, dispatch, sparse-MLA, source, dtype, and identity assertions remain intact." - ] - }, - { - "id": "TA-1137", - "scope": "split exact GLM52 MoE inventory and construction-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The shared and routed inventory and invalid dependency and topology branches define one exact MoE construction policy.", - "Post-EP layout and selected-logprob LM-head specialization remain independent reports." - ] - }, - { - "id": "TA-1138", - "scope": "split DeepSeek-V4 construction topology and precision-preservation reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pipeline admission, parallel-group wiring, FP32 markers, dtype casts, and complex RoPE buffers are one construction-state policy.", - "Full model forward, backward, hash routing, and checkpoint recomputation remain a separate runtime report." - ] - }, - { - "id": "TA-1139", - "scope": "split routing-weight position numerics and configuration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Before-down and after-down numerical behavior and the setting that selects them define one routing-position contract.", - "Reference gradients, error classes, lazy configuration, environment, auto, explicit, and invalid branches remain intact." - ] - }, - { - "id": "TA-1140", - "scope": "split dense and MoE FP32 cast-once LoRA merge reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense linear and MoE expert layouts exercise the same zero-preserving FP32-add-then-cast invariant.", - "All dtype, fused gate-up, down-projection, zero, and nonzero assertions remain intact." - ] - }, - { - "id": "TA-1141", - "scope": "split FlashQLA forward and backward parity reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forward output, final state, and input gradients form one numerical parity report per head shape.", - "Both parameterized head shapes and the Hopper and TileLang capability gate remain unchanged." - ] - }, - { - "id": "TA-1142", - "scope": "split NVFP4 and block-FP8 expert-load reference reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both formats are branches of one prequantized expert-load policy.", - "Packed bytes, scales, global factors, amax, projection, shape, and dequantization assertions remain intact." - ] - }, - { - "id": "TA-1143", - "scope": "split flat and GKN NF4 codec reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Flat and GKN layouts implement one NF4 codebook and quantization-error contract.", - "Packing, scale, dtype, zero, shape, and error assertions remain intact." - ] - }, - { - "id": "TA-1144", - "scope": "split block-FP8 and NVFP4 GNK-to-GKN conversion reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both prequantized formats exercise one checkpoint-to-runtime layout-conversion invariant.", - "Direct parity, roundtrip, non-square, stacking, scale, and error assertions remain intact." - ] - }, - { - "id": "TA-1145", - "scope": "split same-NK and same-MN grouped GEMM reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Same-NK and same-MN geometries are branches of one grouped GEMM kernel-family contract.", - "Transpose, uneven, empty-group, contiguity, device, shape, and numerical assertions remain intact." - ] - }, - { - "id": "TA-1146", - "scope": "split dense and expert-layout Muon full-gradient oracle reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense 2-D and expert 3-D sharding are layout branches of one full-gradient oracle-parity claim.", - "Both two-GPU subprocesses remain independently executed." - ] - }, - { - "id": "TA-1147", - "scope": "split dense and expert-layout Muon shard-local negative controls", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense 2-D and expert 3-D sharding are layout branches of one shard-local negative-control claim.", - "Both two-GPU subprocesses remain independently executed." - ] - }, - { - "id": "TA-1148", - "scope": "split dense Qwen3.5 RMSNorm dispatch and site-assignment reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Structural selection, family dispatch, and assignment to every model site form one CPU RMSNorm resolution contract.", - "Invalid mode, v1, v2, coexistence, GDN exclusion, layer, and final-norm assertions remain intact." - ] - }, - { - "id": "TA-1149", - "scope": "split Qwen3.5-MoE RMSNorm dispatch and site-assignment reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Family dispatch and assignment to every MoE model site form one CPU RMSNorm resolution contract.", - "Ordinary-mode, exact-mode, v2, layer, and final-norm assertions remain intact." - ] - }, - { - "id": "TA-1150", - "scope": "split Qwen3.5-MoE family-1 integration and family-2 residual GPU reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Family-1 interpose and full-layer parity and family-2 residual composition form one GPU bit-exact integration contract.", - "The CUDA capability gate and every bitwise assertion remain intact." - ] - }, - { - "id": "TA-1151", - "scope": "split DeepSeek-V4 converter meta-dtype and roundtrip reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Meta-model dtype preparation is the input phase of the same HF-to-DCP roundtrip conversion policy.", - "FP32 destinations, BF16 tensors, DCP output, exact loads, and legacy sidecar-free LoRA assertions remain intact." - ] - }, - { - "id": "TA-1152", - "scope": "split ordinary and cross-shard DeepSeek-V4 conversion reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Single-shard roundtrip and deferred cross-shard weight and scale pairing are branches of one conversion policy.", - "Process-group state is explicitly reset between the two converter invocations." - ] - }, - { - "id": "TA-1153", - "scope": "split MoE-block and decoder-layer torch compile reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Block and decoder composition are lower-level phases of one compiler compatibility policy.", - "Every available expert backend, AOT eager, Inductor, forward, backward, and numerical assertion remains intact." - ] - }, - { - "id": "TA-1154", - "scope": "split local and EP fused-expert FP32 routing-gradient reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unfiltered local, filtered local, and EP backward paths exercise one FP32 routing-gradient oracle.", - "The local cases run as an internal loop and the EP case runs once, preserving all gradient assertions without duplicate work." - ] - }, - { - "id": "TA-1155", - "scope": "split EP adapter registry and backend argument FP8-boundary reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Optional registration, live signatures, shared arguments, activation forwarding, and FP8 admission define one adapter boundary contract.", - "The isolated optional-import context and every available backend branch remain intact." - ] - }, - { - "id": "TA-1156", - "scope": "split exact Qwen hook preparation and hybrid trunk selection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Merged-LoRA preparation and selected trunk modules are phases of one exact wrapping policy.", - "The stale fixture now satisfies the resolved-family contract, restores process-wide family state, and checks that wrapping intentionally arms the contract lane." - ] - }, - { - "id": "TA-1157", - "scope": "split GLM52 native-FP8 configuration model and buffer checkpoint reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Configuration admission, module replacement, pair-buffer materialization, and checkpoint ownership form one native-FP8 state policy.", - "Canonical router dispatch and frozen expert scoring remain separate behavioral contracts." - ] - }, - { - "id": "TA-1158", - "scope": "split direct and MoEExperts Quack FP8 train-step reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct TP-FP8 execution and the module wrapper are layers of one Quack expert train-step policy.", - "TP reduction, grouped backends, bias, activation, gradient, finiteness, and master-update assertions remain intact." - ] - }, - { - "id": "TA-1159", - "scope": "split scoring and trainable SGLang EP dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Scoring and trainable execution are branches of one SGLang EP dispatch and admission policy.", - "Slot combine, weight presentation, and live stock-Triton gradient parity remain separate contracts." - ] - }, - { - "id": "TA-1160", - "scope": "split fused RMSNorm kernel and model-integration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Residual and no-residual kernel numerics and their dense-model call sites form one GPU integration contract.", - "CPU fallback and trunk-specific dispatch remain independently reported." - ] - }, - { - "id": "TA-1161", - "scope": "split Qwen3.5 pairwise rotary numerics from attention rotary behavior", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pairwise-interleaved reference numerics and the attention modules that select half-rotate semantics form one rotary policy.", - "Dense and MoE attention and mRoPE assertions remain intact." - ] - }, - { - "id": "TA-1162", - "scope": "split Qwen3.5 Class-B rotary admission from attention rotary behavior", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Class-B fused admission and fail-closed CPU behavior are configuration branches of the same rotary policy.", - "Fused-call and CUDA-rejection assertions remain intact." - ] - }, - { - "id": "TA-1163", - "scope": "split DeepSeek-V4 SwiGLU clamp from non-hash MoE behavior", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shared and routed SwiGLU clamping are activation branches of the non-hash MoE runtime policy.", - "Hash routing and record/replay remain separate contracts." - ] - }, - { - "id": "TA-1164", - "scope": "split GDN gating and gated RMSNorm numerical reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Gating and gated normalization are the two numerical primitives of one GDN forward and backward contract.", - "Exact-model module dispatch and triangular-solve geometry remain separate." - ] - }, - { - "id": "TA-1165", - "scope": "split native block-FP8 encoding checkpoint and execution admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Byte encoding, protected state, checkpoint validation, execution entry, and fail-closed admission form one native block-FP8 state machine.", - "Every byte, dtype, identity, rollback, hook, range, and error assertion remains intact." - ] - }, - { - "id": "TA-1166", - "scope": "split sparse-MLA backward reference and deterministic-atomic parity reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Reference correctness and deterministic-versus-atomic parity are complementary branches of one backward policy.", - "The H100 and TileLang gates and all gradient and finiteness assertions remain intact." - ] - }, - { - "id": "TA-1167", - "scope": "split RMSNorm v2 realization parity and dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forced fused-versus-split bit identity and the heuristic that selects those realizations form one realization and dispatch policy.", - "Reference accuracy and batch and run invariance remain a separate numerical-tree contract." - ] - }, - { - "id": "TA-1168", - "scope": "split Class-B RoPE dtype admission and shape-backward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fail-closed table dtype admission and partial-rotary forward and backward behavior are branches of one Class-B primitive contract.", - "All shape cases, untouched-tail bytes, gradients, and error assertions remain intact." - ] - }, - { - "id": "TA-1169", - "scope": "split Class-B RoPE table-layout report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Unique-frequency-half table construction is the input-layout edge of the same Class-B primitive contract.", - "The exact output shape and value assertions remain intact." - ] - }, - { - "id": "TA-1170", - "scope": "split EP gradient backend and metadata-domain admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Backend declarations and parameter-metadata declarations select the same gradient-reduction domain contract and must both fail closed.", - "The real two-rank reduction report remains separate." - ] - }, - { - "id": "TA-1171", - "scope": "split exact dense and LM-head legacy weight-sync rejection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense, ordinary projection, and LM-head components share one prohibition on legacy merged-weight publication.", - "Separate-factor checkpoint byte preservation remains a distinct successful-publication report." - ] - }, - { - "id": "TA-1172", - "scope": "split Qwen3.5 context-parallel positive and negative subprocess reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ulysses execution and ring-plus-FLA rejection are positive and negative branches of one context-parallel admission contract.", - "Both two-GPU subprocesses and their independent success checks still run." - ] - }, - { - "id": "TA-1173", - "scope": "split QLoRA NVFP4 and prequantized block-FP8 lifecycle reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "NVFP4 and block-FP8 are format branches of the same quantized QLoRA execution and merge lifecycle.", - "All live CUDA loading, forward, backward, scale, merge, and requantization assertions remain intact." - ] - }, - { - "id": "TA-1174", - "scope": "split training-model FP8 construction and quantized-mode admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Successful FP8 construction and rejected incompatible modes are the positive and negative branches of one quantized model-builder policy.", - "Former monkeypatch isolation is preserved with scoped contexts." - ] - }, - { - "id": "TA-1175", - "scope": "split training-model QARL lifecycle report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "QARL construction, calibration order, and admission belong to the same model-builder quantization policy as FP8.", - "Dense-only restrictions, calibration state, and parallelization-order assertions remain intact." - ] - }, - { - "id": "TA-1176", - "scope": "split routing-replay sequence-parallel and ring-attention layout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sequence slicing and ring zigzag placement are topology branches of one context-parallel routing-layout contract.", - "Wire decoding and weight tensor construction remain separate." - ] - }, - { - "id": "TA-1177", - "scope": "split sparse-delta single-file and source-capture reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Single-file encoding is the artifact boundary consumed by source capture and manifest publication.", - "Scoped monkeypatch contexts preserve the former module-stub isolation." - ] - }, - { - "id": "TA-1178", - "scope": "split DeepSeek-V4 checkpoint codec and handler-ownership reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quantized decoding and EP-aware expert fusion form one checkpoint conversion and ownership policy.", - "End-to-end synthetic model loading remains separate." - ] - }, - { - "id": "TA-1179", - "scope": "split active-LoRA composite admission and atomic flag reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Atomic setting and clearing establishes the composite flag state consumed by exact-family admission.", - "All missing-component and scoring-only branches remain intact." - ] - }, - { - "id": "TA-1180", - "scope": "split active-LoRA server derivation and cached-surface activation reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-one server derivation and indexer and MoE activation jointly describe complete-composite propagation.", - "Topology rejection and partial-composite negative controls remain intact." - ] - }, - { - "id": "TA-1181", - "scope": "split Nemotron-H published and stacked expert checkpoint layout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-expert published weights and stacked in-memory weights are two accepted input layouts for one bidirectional checkpoint handler.", - "EP ownership remains separate." - ] - }, - { - "id": "TA-1182", - "scope": "split LoRA target-manifest success and failure reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Runtime target selection and schema or coverage rejection are positive and negative branches of one manifest contract.", - "Every schema, rank, count, target, and unlisted-module assertion remains intact." - ] - }, - { - "id": "TA-1183", - "scope": "split Qwen2 model construction and checkpoint-handler reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HF configuration conversion, unfusing, and bidirectional checkpoint translation form one architecture-support contract.", - "All model-layout and HF parity assertions remain intact." - ] - }, - { - "id": "TA-1184", - "scope": "split OLMo2 model construction and checkpoint-handler reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "HF configuration conversion, post-norm construction, unfusing, and checkpoint translation form one architecture-support contract.", - "All model-layout and HF parity assertions remain intact." - ] - }, - { - "id": "TA-1185", - "scope": "split Quack worker-protocol and PTXAS process-safety reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Worker response framing and PTXAS timeout and temporary-output handling are process-boundary safety checks for one compilation path.", - "All timeout, truncation, uniqueness, cleanup, and entry-selection assertions remain intact." - ] - }, - { - "id": "TA-1186", - "scope": "split Quack cache-key hashing report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Deterministic safe cache identity is the persistence edge of the same Quack compilation process contract.", - "Collision boundaries and rejection of unsafe pickle hooks remain intact." - ] - }, - { - "id": "TA-1187", - "scope": "split shared-prefix matrix and singleton attention reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Multi-member and singleton groups are shape branches of one shared-prefix forward and backward equivalence contract.", - "The optional FA3 capability gate remains unchanged." - ] - }, - { - "id": "TA-1188", - "scope": "split OPD output-edge and hidden-distance reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ignored tokens, per-token output, hidden-only loss, and chunked hidden distance describe one OPD edge-behavior contract.", - "Backend numerics and gradient reduction remain separate." - ] - }, - { - "id": "TA-1189", - "scope": "split packing-strategy admission and correctness reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Strategy validation, oversized handling, token preservation, capacity, and utilization form one generic packing-policy contract.", - "Balanced-DP-specific scheduling remains separate." - ] - }, - { - "id": "TA-1190", - "scope": "split packing-strategy determinism and datum-order report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Determinism and datum order are output invariants of the same generic packing-policy contract.", - "All strategies, repeated builds, and reordered-index assertions remain intact." - ] - }, - { - "id": "TA-1191", - "scope": "split Tinker session OpenAPI and activity-lifecycle reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Schema publication, creation, follow-up use, heartbeat, and canonical configuration form one public session-endpoint lifecycle.", - "All HTTP-boundary and server-state assertions remain intact." - ] - }, - { - "id": "TA-1192", - "scope": "split Mooncake side-payload store and R3 slice reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Tensor roundtrip is the storage primitive used by R3 reference publication, selective loading, and cleanup.", - "Integer, float, missing-key, slice, validation, and cleanup assertions remain intact." - ] - }, - { - "id": "TA-1193", - "scope": "split checkpoint-save failure and live-adapter-state reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fail-closed admission, write failures, and successful dense and MoE factor publication are branches of one adapter-save policy.", - "Scoped monkeypatch contexts preserve former isolation." - ] - }, - { - "id": "TA-1194", - "scope": "split launcher worker discovery and readiness reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Address resolution and readiness or early-exit handling are successive phases of one launcher worker-control lifecycle.", - "Server override parsing remains separate." - ] - }, - { - "id": "TA-1195", - "scope": "split P2P async dispatch and prepare-timeout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Transfer cutoff, status timeout, and prepare-request timeout are environment-controlled branches of one P2P async API policy.", - "All sync-versus-async, timeout, request payload, and transport assertions remain intact." - ] - }, - { - "id": "TA-1196", - "scope": "split K3 debug metrics and logprob-temperature reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Tail metrics and temperature-matched zero K3 jointly define behavior-logprob observability for both loss implementations.", - "TokenPartial reducer identity remains separate." - ] - }, - { - "id": "TA-1197", - "scope": "split Mooncake hidden transport and teacher-consumer reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Tensor codec, metadata publication, retrieval, and indexed teacher consumption form one hidden-transport lifecycle.", - "Rank-two, rank-three, and multi-teacher cases remain intact." - ] - }, - { - "id": "TA-1198", - "scope": "split Mooncake metadata admission and store-lifecycle reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Malformed or legacy metadata rejection, object removal, and configuration precedence form one store-admission lifecycle.", - "All missing-key, size, schema, cleanup, and environment assertions remain intact." - ] - }, - { - "id": "TA-1199", - "scope": "split dense QARL weight fake-quant and injection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Export-equivalent fake quantization and model injection are numerical and structural phases of one dense QARL policy.", - "Target selection, summary, forward counts, and model admission remain intact." - ] - }, - { - "id": "TA-1200", - "scope": "split dense QARL configuration-normalization report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Recipe normalization and rejection establish the configuration admitted by the same dense fake-quant lifecycle.", - "Static, unsupported-format, and invalid-block failures remain intact." - ] - }, - { - "id": "TA-1201", - "scope": "split NVFP4 QARL MoE conversion and eager-execution reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Identity-preserving class conversion and eager forward and backward behavior are consecutive phases of one expert lifecycle.", - "Parameter identity, quantization effect, gradients, passthrough, and admission remain intact." - ] - }, - { - "id": "TA-1202", - "scope": "split NVFP4 QARL MoE injection report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Model injection selects and invokes the same expert conversion lifecycle.", - "NVFP4 admission, FP8 rejection, and independent target selection remain intact." - ] - }, - { - "id": "TA-1203", - "scope": "split QARL weight-sync success and bad-configuration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Derived FP8 sync configuration and explicit incompatible configuration are positive and negative branches of one QARL sync policy.", - "Folded and excluded module behavior remains intact." - ] - }, - { - "id": "TA-1204", - "scope": "split expert-adapter capability ownership and semantics reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Backend capability, factor ownership, and preserved activation semantics jointly define the generic expert-adapter contract.", - "Injection and model-family construction remain separate." - ] - }, - { - "id": "TA-1205", - "scope": "split teacher-head storage and manager-residency reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Weight discovery, sharded storage, cross-shard views, residency, dtype reload, and prefetch form one teacher-head lifecycle.", - "Teacher activation caching remains separate." - ] - }, - { - "id": "TA-1206", - "scope": "split exact-server trunk and numerical-family selection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Trunk wrapping and v2 family selection are coordinated pre-parallelization model-program choices.", - "Scoped monkeypatch contexts and the autouse global-state reset preserve isolation." - ] - }, - { - "id": "TA-1207", - "scope": "split NVFP4 quantization roundtrip and directory-export reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed codec correctness is the primitive exercised by end-to-end NVFP4 directory export.", - "Shared scales, BF16 islands, W4A4 inputs, requantization rejection, and reconstruction error remain intact." - ] - }, - { - "id": "TA-1208", - "scope": "split FP8 export CLI and model-layout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CLI configuration, output directory construction, and architecture-specific layout transforms form one export command contract.", - "Every transform runs in its own named temporary case." - ] - }, - { - "id": "TA-1209", - "scope": "split FP8 export admission report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Source preflight and QARL fold rejection are fail-closed branches of the same export command contract.", - "Primitive quantization and trained-logprob preservation remain separate." - ] - }, - { - "id": "TA-1210", - "scope": "split OPD endpoint-registration and student-version verifier reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Existing endpoint identity and expected student weight version jointly establish OPD endpoint admission.", - "Matching, mismatch, and endpoint-error branches remain intact." - ] - }, - { - "id": "TA-1211", - "scope": "split OPD prepare-worker and payload-transport reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Chunk preparation, queue completion, causal shifting, cache-index alignment, and Mooncake metadata form one pipeline-preparation contract.", - "All queue and transport assertions remain intact." - ] - }, - { - "id": "TA-1212", - "scope": "split DeepSeek-V3 trainer and parallelizer admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Router-freeze construction and downstream tensor-parallel rejection are successive admission layers of one DeepSeek training policy.", - "The successful router-freeze path and all incompatible configurations remain intact." - ] - }, - { - "id": "TA-1213", - "scope": "split Class-B RoPE selection and canonical GLM numerical-program reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Class-B selection is one required component of the canonical GLM-5.2 numerical program and its fail-closed overrides.", - "Exact Qwen3.5 program admission remains separate." - ] - }, - { - "id": "TA-1214", - "scope": "split simulator topology-ledger and config-metadata admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Topology resolution and analytical ledgers are derived from the same trusted training configuration and model metadata boundary.", - "Observed benchmarking, calibration, and kernel ranking remain separate." - ] - }, - { - "id": "TA-1215", - "scope": "split DeepEP internode topology and preflight reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Node-spanning topology detection directly controls whether the internode transport preflight runs.", - "Scoped patches preserve every skip, intranode, failure-diagnostic, identity, and corruption branch." - ] - }, - { - "id": "TA-1216", - "scope": "split DeepEP buffer-size and RDMA admission report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "NVL alignment and RDMA byte validation are resource-admission edges of the same internode transport policy.", - "The int32 limit and all byte-layout cases remain intact." - ] - }, - { - "id": "TA-1217", - "scope": "split generic parallel-plan meta slicing and gradient-domain reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Parameter slicing, placement, and explicit replicated-gradient metadata are outputs of one generic EP plan application.", - "Shape, dtype, requires-grad, divisibility, and reduction assertions remain intact." - ] - }, - { - "id": "TA-1218", - "scope": "split exact GLM meta and materialized already-local EP plan reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Meta and materialized tensors are allocation branches of the same exact routed-expert disposition policy.", - "Already-local bases, force-sharded banks, replicated factors, and malformed singleton guards remain intact." - ] - }, - { - "id": "TA-1219", - "scope": "split pipeline FQN partition and stage-placement reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Module partitioning and rank ownership are consecutive outputs of one pipeline layout plan.", - "Single, loop, virtual, v-style, pinned, weighted, and infeasible cases remain intact." - ] - }, - { - "id": "TA-1220", - "scope": "split pipeline schedule metadata and admission report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Schedule style and microbatch admission consume the same stage layout plan.", - "Every supported schedule and invalid virtual-stage or microbatch case remains intact." - ] - }, - { - "id": "TA-1221", - "scope": "split PP profiler interval-union and P2P-byte accounting reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Busy interval union and P2P bytes are pure accounting primitives feeding one bubble-profile report.", - "Profiler patch lifecycle and live CUDA schedule execution remain separate." - ] - }, - { - "id": "TA-1222", - "scope": "split Muon full-gradient oracle and shard-local negative-control reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full-gradient parity and shard-local divergence are positive and negative controls of one distributed Muon policy.", - "All four two-GPU subprocesses across dense and MoE layouts still run." - ] - }, - { - "id": "TA-1223", - "scope": "split DeepSeek-V4 C128 and C4 compression admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "C128 ratio divisibility and C4 overlap divisibility are compression-regime branches of one context-parallel policy.", - "Output shape, cache capacity, and overlap failure assertions remain intact." - ] - }, - { - "id": "TA-1224", - "scope": "split BI fused LM-head integration and kernel edge reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Loss forward and backward parity and unit-temperature or near-one probability edges exercise one selected-logprob kernel contract.", - "Determinism, batch invariance, guards, identity bits, and nonpositive logprobs remain intact." - ] - }, - { - "id": "TA-1225", - "scope": "split BI full and dimension mean reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Full reduction and explicit dimension reductions are dispatch branches of one batch-invariant mean policy.", - "FP32 and BF16 accuracy and bitwise dimension behavior remain intact." - ] - }, - { - "id": "TA-1226", - "scope": "split BI head-v2 projection and trainability reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Projection and statistics bits, batch invariance, fused CE, gradients, and rollback form one head-v2 lifecycle.", - "The live CUDA capability gate remains unchanged." - ] - }, - { - "id": "TA-1227", - "scope": "split GDN primitive numerics and exact-model module-dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Gating and gated RMSNorm numerics are the kernel behavior selected by the exact-model GDN dispatch program.", - "Triangular-solve geometry remains separate." - ] - }, - { - "id": "TA-1228", - "scope": "split Quack EP Triton parity and half-concat reference reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Backend parity and the independent half-concatenated activation reference jointly establish one Quack EP numerical contract.", - "CPU gradient-arity behavior remains separate." - ] - }, - { - "id": "TA-1229", - "scope": "split runner batch conversion and sequence-shard reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Float side-channel conversion, ragged padding, and sequence sharding form one runner batch-materialization policy.", - "Teacher hidden-state dtype, padding, shape, and shard assertions remain intact." - ] - }, - { - "id": "TA-1230", - "scope": "split exact dense and routed adapter-gradient ownership reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense managed-factor compilation and routed fail-closed ownership are topology branches of one exact adapter-gradient policy.", - "Scoped patches preserve manager and parallel-state isolation." - ] - }, - { - "id": "TA-1231", - "scope": "split DeepSeek-V4 checkpoint codec-handler and synthetic-load reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Codec and EP-handler behavior culminate in the same synthetic end-to-end checkpoint load contract.", - "FP8, MXFP4, window, C4, hash, strict-buffer, and ownership cases remain intact." - ] - }, - { - "id": "TA-1232", - "scope": "split DeepSeek-V4 construction and runtime reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Construction, topology, precision preservation, forward, backward, hash routing, and recomputation form one model contract.", - "A scoped patch restores the former construction-test isolation before runtime execution." - ] - }, - { - "id": "TA-1233", - "scope": "split exact attention checkpoint inventory and pair-state reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Source inventory selects the native absorbed-kv pair state machine exercised by order, byte, completion, dtype, and shape checks.", - "All exact factor inventory and handler assertions remain intact." - ] - }, - { - "id": "TA-1234", - "scope": "split canonical MoE routed and shared boundary reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Routed and shared partials are the two contributor boundaries of one canonical MoE block policy.", - "Global and local IDs, scale, root invocation, and contributor ordinal remain intact." - ] - }, - { - "id": "TA-1235", - "scope": "split canonical native-FP8 router and expert-state reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Serving-compatible routing and frozen scoring-only expert execution form one canonical native-FP8 runtime policy.", - "Configuration, buffer, and checkpoint ownership remain separate." - ] - }, - { - "id": "TA-1236", - "scope": "split Nemotron-H EP ownership and checkpoint-layout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP admission and slicing govern the same published and stacked layouts exercised by the bidirectional handler.", - "HF parity, exact saved bytes, skips, and plan targeting remain intact." - ] - }, - { - "id": "TA-1237", - "scope": "split fused RMSNorm model integration and trunk reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ordinary model integration and trunk-specific no-residual dispatch are branches of one GPU fused RMSNorm policy.", - "CPU fallback remains a separate capability domain." - ] - }, - { - "id": "TA-1238", - "scope": "split RoPE registry precision and native-cache reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Registry FP32 recipes and lazy device-local cache materialization form one CPU RoPE precision and cache policy.", - "Exact architecture serving-device execution remains a separate GPU report." - ] - }, - { - "id": "TA-1239", - "scope": "split adapter-gradient pre-rendezvous and ModelRunner tail-failure reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both subprocesses exercise bounded fail-closed behavior before a capture can commit.", - "The independent two-rank subprocesses and their original assertions remain intact." - ] - }, - { - "id": "TA-1240", - "scope": "split adapter-gradient publication-commit failure report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Publication commit is the final CPU failure phase of the same bounded fatal-gradient lifecycle.", - "The asymmetric post-mutation GPU boundary remains a separate capability report." - ] - }, - { - "id": "TA-1241", - "scope": "split live two-rank clip and three-rank participation reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Norm clipping, nonfinite admission, and all-rank participation form one live distributed clipping policy.", - "Both independent subprocess topologies still execute." - ] - }, - { - "id": "TA-1242", - "scope": "split FutureStore creation-processing and model-expiration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Creation, processing, concurrency, model operations, expiration, and cleanup form one async store lifecycle.", - "A fresh identical store instance preserves the former fixture isolation between phases." - ] - }, - { - "id": "TA-1243", - "scope": "split orchestrator communication roundtrip and edge-lifecycle reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Roundtrip, interleaving, exceptions, and shutdown are phases of one client communication lifecycle.", - "All async assertions and engine interactions remain intact." - ] - }, - { - "id": "TA-1244", - "scope": "split sparse-delta initialization and runtime-helper reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Post-only initialization, prepacked admission, and runtime delta loading govern one backend policy.", - "Encoding and load assertions remain intact." - ] - }, - { - "id": "TA-1245", - "scope": "split FP8 LM-head CE selection and loss-dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-token LM-head selection and dispatcher routing are consecutive layers of one FP8 loss policy.", - "Numerical selection and dispatch assertions remain intact." - ] - }, - { - "id": "TA-1246", - "scope": "split sequence-shard core and token-side-channel reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sharding, padding, flash-attention metadata, and per-token side channels form one collator materialization policy.", - "The original patched parallel-state contexts remain intact." - ] - }, - { - "id": "TA-1247", - "scope": "split server-CLI sequence-boundary and shard-preservation reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Server-CLI boundary parity culminates in preservation and regeneration by the sequence-shard collator.", - "Dtype, padding, stale-metadata, and original-position assertions remain intact." - ] - }, - { - "id": "TA-1248", - "scope": "split sequence LCM-padding and collator-divisibility report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "LCM padding is the topology branch of the same sequence-metadata alignment policy.", - "RequestProcessor and post-shard divisibility assertions remain intact." - ] - }, - { - "id": "TA-1249", - "scope": "split padded-unpacking boundary report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Padded output unpacking is the consumer boundary of the same packed sequence policy.", - "Padded and unpadded sample-count assertions remain intact." - ] - }, - { - "id": "TA-1250", - "scope": "split AnyPrecision AdamW cautious execution and state-strategy reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Cautious decay, chunked denominators, state reuse, and DTensor offload form one optimizer lifecycle.", - "The temporary-path fixture and all numerical and state assertions remain intact." - ] - }, - { - "id": "TA-1251", - "scope": "split inference-endpoint registration and list-health reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Registration and subsequent health-aware listing form one public endpoint lifecycle.", - "Explicit worker, auto-sync, FP8 KV-cache, and v1-model fallback assertions remain intact." - ] - }, - { - "id": "TA-1252", - "scope": "split P2P prepare and initialize-fanout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Prepare payloads, cached maps, endpoint fanout, and fanout cleanup are phases of one initialization handshake.", - "Scoped monkeypatch contexts preserve environment isolation between initialization modes." - ] - }, - { - "id": "TA-1253", - "scope": "split P2P complete-sync report from initialization handshake", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Receiver completion, tied aliases, cache flushing, version forwarding, and cleanup close the same P2P lifecycle.", - "All completion and failure assertions remain intact." - ] - }, - { - "id": "TA-1254", - "scope": "split P2P receiver placement and source-staging reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Receiver placement, replicated-source reuse, scratch alignment, coalescing, and registration form one staging policy.", - "FP8 receiver layouts remain a separate format contract." - ] - }, - { - "id": "TA-1255", - "scope": "split P2P invalid-manifest and transfer-failure reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Source, receiver, shape, rank, and name admission failures belong with runtime transfer diagnostics.", - "Diagnostic detail, sample caps, and disabled-sampling assertions remain intact." - ] - }, - { - "id": "TA-1256", - "scope": "split microbatch splitting and DataLoaderBuilder configuration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Builder batch sizing and collator selection culminate in microbatch splitting and epoch delegation.", - "Sampler, sequence-parallel, custom-collator, edge, and set-epoch assertions remain intact." - ] - }, - { - "id": "TA-1257", - "scope": "split dataset expansion-type and split-merge reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dataset expansion and type selection feed the same composition policy as train-validation splitting and merging.", - "Raw loading and preprocessed persistence remain separate I/O boundaries." - ] - }, - { - "id": "TA-1258", - "scope": "split MiniMax-M3 configuration-registration and text-runtime reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Configuration conversion and registry admission select the text-only model runtime exercised by forward, backward, and rejection checks.", - "Checkpoint ownership and MSA paging remain separate mechanisms." - ] - }, - { - "id": "TA-1259", - "scope": "split attention backend resolution and eager-layout reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CPU backend selection now executes the eager numerical head-layout contract it resolves.", - "FlashAttention and SGL page-cache paths retain their independent optional capability gates." - ] - }, - { - "id": "TA-1260", - "scope": "split FP8 training and other low-precision argument reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP8 training, block-FP8 QLoRA, QARL, aliases, defaults, and incompatible combinations form one low-precision parsing policy.", - "Fresh temporary subdirectories preserve the former configuration-file isolation." - ] - }, - { - "id": "TA-1261", - "scope": "split direct-EP multi-sender initialization and scatter-copy reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Scatter-copy ownership is an initialization branch of the same direct-EP multi-sender lifecycle.", - "List, deep, and locator-reuse modes remain intact under a scoped environment." - ] - }, - { - "id": "TA-1262", - "scope": "split direct-EP dense-sharding manifest report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense and expert manifest partitioning determines the buffers published by the same multi-sender lifecycle.", - "QKV and gate-up ownership affinity assertions remain intact." - ] - }, - { - "id": "TA-1263", - "scope": "split direct-EP rank-filter transfer report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-filtered receiver transfers are the execution boundary of the direct-EP manifest policy.", - "Owning-rank and empty-rank transfer assertions remain intact." - ] - }, - { - "id": "TA-1264", - "scope": "split FP8Linear CUDA matmul and train-step reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Padded recipes and residual correction now culminate in live FP8 forward, backward, and master-weight mutation.", - "Both phases retain the same CUDA capability gate and every numerical assertion." - ] - }, - { - "id": "TA-1265", - "scope": "split dense and packed SSD recurrence reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense, multichunk, boundary-crossing packed, convolution, and mixer behavior form one CPU recurrence contract.", - "Unavailable-kernel admission and live GPU kernel parity remain separate capability reports." - ] - }, - { - "id": "TA-1266", - "scope": "split adapter checkpoint materialization and restore-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Coordinator materialization and auto-load are entry points into the same checkpoint restore and admission lifecycle.", - "A fresh temporary subtree preserves manager and optimizer isolation." - ] - }, - { - "id": "TA-1267", - "scope": "split OPD loss execution and metric aggregation reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-microbatch OPD execution now culminates in the aggregation, extrema, reduction, and empty-rank policy consuming its metrics.", - "Scoped patches preserve device and parallel-state isolation." - ] - }, - { - "id": "TA-1268", - "scope": "split request-processor forward and packed-row batching reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed-row batching is an OPD branch of the same live forward-backward processor lifecycle.", - "The helper uses a fresh internal processor while retaining the shared-processor rejection case." - ] - }, - { - "id": "TA-1269", - "scope": "split model-scoped and sampler-scoped checkpoint listing reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Listing, deletion, resolution, and isolation now cover both storage namespaces in one public checkpoint policy.", - "Explicit setup and teardown give the sampler namespace a fresh APIServer and temporary root." - ] - }, - { - "id": "TA-1270", - "scope": "split sampler adapter tracking and normalized export reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Adapter reconciliation, model-scoped tracking, and adapter-only export form one sampler-weight lifecycle.", - "The normalized session-spec request and output URI assertions remain intact." - ] - }, - { - "id": "TA-1271", - "scope": "split Muon Gram-Newton-Schulz configuration and Quack backend reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quack import, dispatch, tuned-mode, and dtype admission are backend branches of the configured Gram-Newton-Schulz optimizer.", - "Nested monkeypatch contexts preserve cache and import isolation." - ] - }, - { - "id": "TA-1272", - "scope": "split FP8 MoE expert and injected full-model train-step reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct expert variants now culminate in injected dense-expert-dense forward, backward, and master-weight mutation.", - "Both phases retain the same CUDA capability gate." - ] - }, - { - "id": "TA-1273", - "scope": "split canonical LoRA fold and LoraLinear merged-forward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical folding and straight-through gradients now execute through LoraLinear selection, ordinary isolation, and cache invalidation.", - "All exact and legacy forward assertions remain intact." - ] - }, - { - "id": "TA-1274", - "scope": "split canonical LoRA fold and MoE merged-weight cache reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Expert gate-up and down folding, weight-sync views, versioned caches, and fused admission are the MoE branch of the same fold policy.", - "Native EP execution and trunk wrapping remain separate integrations." - ] - }, - { - "id": "TA-1275", - "scope": "split GDN delta-linear product and merged-forward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The explicit low-rank product now feeds sliced canonical folding, gradient ownership, GDN projection, and bounded cache behavior.", - "Geometry-manifest and checkpoint roundtrip remain separate boundaries." - ] - }, - { - "id": "TA-1276", - "scope": "split exact LM-head per-token and causal-loss routing reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Per-token exact-head selection and causal-loss TP-group admission are consecutive layers of one loss-routing contract.", - "A scoped patch preserves each dispatcher replacement." - ] - }, - { - "id": "TA-1277", - "scope": "split exact LM-head weight and server selector report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Weight and module selection prove that the exact loss route never materializes a merged delta.", - "Identity assertions remain intact in the complete exact-head policy." - ] - }, - { - "id": "TA-1278", - "scope": "split exact LM-head FSDP replicated-factor report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FSDP replicated-factor admission is the parallel ownership boundary of the same exact-head loss policy.", - "The lora-A-only fail-closed assertion remains intact." - ] - }, - { - "id": "TA-1279", - "scope": "split absorbed-KV contract and dtype-move state reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Native base bytes, FP32 logical masters, identities, and dtype moves form one CPU state contract.", - "The official CUDA Q/V program remains a separate capability report." - ] - }, - { - "id": "TA-1280", - "scope": "split absorbed-KV state and direct-projection admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct projection, branch, materialization, and factor-dtype rejection define the fail-closed edge of the same CPU component contract.", - "Every negative assertion remains intact." - ] - }, - { - "id": "TA-1281", - "scope": "split canonical MoE graph metadata and transport-admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Capacity metadata and transport resolution jointly define the canonical MoE planning boundary.", - "Dense, packed, CP-sharded, graph, and output-distribution admission cases remain intact." - ] - }, - { - "id": "TA-1282", - "scope": "split canonical MoE metadata and exact parallel-plan reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Trainer and sampler plan hashes, group layouts, topology rejection, and metadata now form one CPU planning policy.", - "The distributed reduction and backward subprocess remains separate." - ] - }, - { - "id": "TA-1283", - "scope": "split BI router GEMM and leading-dimension linear reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The leading-dimension wrapper is a shape branch of the same BF16-input FP32-output router GEMM contract.", - "Forward and backward comparisons remain intact." - ] - }, - { - "id": "TA-1284", - "scope": "split BI router GEMM and top-k weight reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP32 logits and top-k normalization/casting are consecutive stages of exact router selection.", - "Renormalized, cast-only, and dtype-rejection cases remain intact." - ] - }, - { - "id": "TA-1285", - "scope": "split BI router primitives and MoEBlock dispatch report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "MoEBlock exact and ordinary routing now exercise the complete GEMM and top-k primitive policy under the same CUDA gate.", - "Batch-composition and stock-path assertions remain intact." - ] - }, - { - "id": "TA-1286", - "scope": "split NCCL endpoint transfer and flattened or hybrid bucket reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Endpoint initialization, two-phase completion, and flattened, chunked, and hybrid broadcasts now form one NCCL transfer policy.", - "Scoped environment contexts retain independent load-format setup and receiver-fence assertions." - ] - }, - { - "id": "TA-1287", - "scope": "split NCCL transfer and multi-rank direct-format admission reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Multi-rank direct-load rejection is the admission edge of the same endpoint transfer policy.", - "The health-check lifecycle remains a separate report." - ] - }, - { - "id": "TA-1288", - "scope": "split weight-sync source and bucket-sizing reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Adapter selection, parameter extraction, inference layout, and byte-capped chunking jointly define the sync-source preparation policy.", - "Default, shared override, MoE override, split, and oversize-item assertions remain intact." - ] - }, - { - "id": "TA-1289", - "scope": "split weight-sync source and direct-EP sender-selection reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Direct-EP sender mapping and collection gating determine which prepared source tensors enter the transfer.", - "Scoped environment contexts isolate default and round-robin replica strategies." - ] - }, - { - "id": "TA-1290", - "scope": "split shipped adapter examples and quantized server-configuration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shipped MoE LoRA and QLoRA parsing now culminates the low-precision server-configuration contract.", - "Clean-process parsing and every certified Quack target assertion remain intact." - ] - }, - { - "id": "TA-1291", - "scope": "split sharded-adapter layout and deterministic-initialization reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local packing, empty-shard ownership, and topology-invariant deterministic initialization now form one CPU adapter-state policy.", - "A fresh temporary subtree preserves manager and checkpoint isolation." - ] - }, - { - "id": "TA-1292", - "scope": "split sharded-adapter state and explicit-EP layout-discovery reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit expert sharding and generic replication are layout-discovery branches of the same CPU adapter-state policy.", - "The real two-rank Gloo DTensor report remains separate." - ] - }, - { - "id": "TA-1293", - "scope": "split dispatcher save and session-registration reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Registration, cross-rank failure, nonresident auto-load, and state or adapter saves now form one session checkpoint lifecycle.", - "A scoped patch preserves the rank-zero failure boundary." - ] - }, - { - "id": "TA-1294", - "scope": "split optimizer publication and gradient-epoch completion reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forward-backward completion, abort, uniform rejection, optimizer publication, and fatal tail failures now form one mutation lifecycle.", - "Commit ordering, poisoning, and process-termination assertions remain intact." - ] - }, - { - "id": "TA-1295", - "scope": "split token diagnostics and hidden-component hook reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense and MoE component hooks now feed the same diagnostic computation and summary policy they support.", - "Equation-term tensors, hook cleanup, selection, loss cross-checks, and CP mapping remain asserted." - ] - }, - { - "id": "TA-1296", - "scope": "split diagnostic tensor-dump and trusted-override reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Ranked tensor persistence and trusted diagnostic replay now form one artifact-boundary policy.", - "A fresh subtree preserves file isolation and the missing-root rejection remains intact." - ] - }, - { - "id": "TA-1297", - "scope": "split P2P engine construction and initialization reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Hostname precedence and Mooncake fallback construction now precede the prepare, fanout, cache, completion, and cleanup lifecycle.", - "A scoped environment context preserves every resolution branch." - ] - }, - { - "id": "TA-1298", - "scope": "split adapter weight-publication and authoritative optimizer reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Clean mid-epoch weight admission and strict checkpoint rejection are publication branches of the authoritative optimizer lifecycle.", - "A fresh manager subtree preserves state isolation." - ] - }, - { - "id": "TA-1299", - "scope": "split packed side-metadata and full pipeline reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Teacher, cache, weight, RL, and hidden-state metadata now flow through the same full pack, forward, and unpack pipeline.", - "Capacity and packing-disabled policies remain separate." - ] - }, - { - "id": "TA-1300", - "scope": "split model-pass R3 payload and token-diagnostic unpacking reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed token diagnostics are now unpacked in the model-pass side-payload lifecycle they support.", - "Position rebasing and every aligned diagnostic field remain asserted." - ] - }, - { - "id": "TA-1301", - "scope": "split API optimizer and forward response-metric reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Optimizer, forward, and forward-backward response shaping now form one public training-operation response policy.", - "Each phase constructs a fresh API server." - ] - }, - { - "id": "TA-1302", - "scope": "split runner gradient compiler and staged-capture abort reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forward-backward abort now closes the failure edge of the runner gradient-ownership compilation policy.", - "The abort phase uses a fresh adapter-manager path." - ] - }, - { - "id": "TA-1303", - "scope": "split OPD cache shaping and loss-execution reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed cache rows, teacher boundaries, last-k weights, loss, gradients, and profiling now form one OPD execution policy.", - "Distributed cache assembly and debug artifacts remain separate capabilities." - ] - }, - { - "id": "TA-1304", - "scope": "split dispatcher batch distribution and packing-dummy reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Balanced packing and sequential dummy behavior are input-construction branches of dispatcher sharding and provenance.", - "A scoped parallel-state patch preserves rank isolation." - ] - }, - { - "id": "TA-1305", - "scope": "split routing payload and microbatch diagnostic artifact reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Routing references and diagnostic dumps now form one dispatcher side-payload artifact and security policy.", - "Raw manifests, Mooncake slices, legacy-pickle rejection, symlink rejection, and R3 dump contents remain intact." - ] - }, - { - "id": "TA-1306", - "scope": "split orchestrator success and error or concurrency reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Initialization, operations, errors, end-to-end completion, concurrent requests, statistics, and shutdown now form one lifecycle.", - "The same live fixture carries the successful and failure phases without restarting the capability." - ] - }, - { - "id": "TA-1307", - "scope": "split cautious primitive and SignSGD report from optimizer construction", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Primitive masking and SignSGD execution now begin the cautious weight-decay optimizer policy.", - "Zero decay, ordinary decay, aligned, misaligned, and zero-direction assertions remain intact." - ] - }, - { - "id": "TA-1308", - "scope": "split AnyPrecisionAdamW cautious report from optimizer construction", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "AnyPrecisionAdamW numerics, chunked state, gradient reuse, and DTensor offload now run inside the complete cautious optimizer policy.", - "The temporary path remains isolated within the parent report." - ] - }, - { - "id": "TA-1309", - "scope": "split Muon cautious report from optimizer construction", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Muon and its AdamW fallback now exercise the same cautious feature before builder routing and kwarg admission.", - "Post-Newton-Schulz masking and ordinary-decay equivalence remain asserted." - ] - }, - { - "id": "TA-1310", - "scope": "split synthetic balanced TopK routing report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Synthetic balanced routing is now an environment-selected branch of the complete TopK router contract.", - "A scoped environment context prevents the synthetic mode from leaking into ordinary routing cases." - ] - }, - { - "id": "TA-1311", - "scope": "split sqrtsoftplus noaux TopK routing report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Biased selection, unbiased weights, normalization, and missing-bias rejection now run with the router configuration policy.", - "All DSv4-specific assertions remain intact." - ] - }, - { - "id": "TA-1312", - "scope": "split hash-table TopK routing report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Token-to-expert lookup, bias independence, input admission, softmax, scaling, and from-config behavior now form one router contract.", - "Each routing algorithm still uses independent tensors and router instances." - ] - }, - { - "id": "TA-1313", - "scope": "split optional boolean coercion from parallel policy configuration", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Boolean admission now accompanies mixed-precision and reduce-dtype configuration in one CPU parallel-policy report.", - "Sequence-parallel folding and manual prefetch remain separate topology policies." - ] - }, - { - "id": "TA-1314", - "scope": "split FP8 module injection and CPU fallback reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Injected module identity, recipes, exclusions, CPU execution, output dtype, and fail-fast fallback now form one CPU policy.", - "The CPU profiler remains a separate observability capability." - ] - }, - { - "id": "TA-1315", - "scope": "split CUDA FP8 profiler and live train-step reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CUDA operand-error breakdown now culminates the live FP8 matmul, correction, backward, and master-weight mutation policy.", - "Both phases retain the same CUDA capability gate and scoped profiler environment." - ] - }, - { - "id": "TA-1316", - "scope": "split GLM52 canonical MoE configuration and sparse selector reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical routing selection, transport, configuration rejection, codecs, and sparse selector pipeline now form one selection policy.", - "Layer-plan allocation and semantic logprob parity remain separate contracts." - ] - }, - { - "id": "TA-1317", - "scope": "split exact dense MLP root-state and forward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical factor ownership and immutable checkpoint-source binding now precede exact fused gate-up, activation, and down execution.", - "All unique-path and state-dictionary assertions remain intact." - ] - }, - { - "id": "TA-1318", - "scope": "split exact dense MLP runtime admission and forward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Rank-alpha construction, atomic runtime updates, and pre-forward consistency rejection now bound the exact MLP execution policy.", - "Every fail-closed assertion remains intact." - ] - }, - { - "id": "TA-1319", - "scope": "split exact dense MLP checkpoint roundtrip and forward reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "XoRL load and PEFT export now close the same six-factor exact MLP component lifecycle.", - "The roundtrip uses a fresh pytest temporary directory." - ] - }, - { - "id": "TA-1320", - "scope": "split sample metadata from packing-dataset lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Position metadata and trainable-token filtering now begin the PackingDataset construction policy.", - "Single, batched, missing-field, and rejection assertions remain intact." - ] - }, - { - "id": "TA-1321", - "scope": "split dataset preprocessing from packing-dataset lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Record filtering and optional evaluation preprocessing now feed the dataset packing, allocation, cache, and missing-column behavior they support.", - "All preprocessing assertions execute before PackingDataset construction." - ] - }, - { - "id": "TA-1322", - "scope": "split pipeline-profiler patching from interval and P2P accounting", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Instance patching, restoration, double-patch admission, interval merging, and P2P byte estimation now form one CPU profiler policy.", - "The live CUDA GPipe step remains a separate capability report." - ] - }, - { - "id": "TA-1323", - "scope": "split selected QLoRA shard-cache loading from deferred loader", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Requested-key-only shard reads now begin the bounded deferred prequantized-loader lifecycle.", - "A scoped monkeypatch context preserves fake shard and cache isolation." - ] - }, - { - "id": "TA-1324", - "scope": "split prequantized QLoRA key planning from deferred loader", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Merged projections, EP16 expert slices, missing-pair rejection, per-module loads, and cache release now form one loader policy.", - "Every exact-key and peak-residency assertion remains intact." - ] - }, - { - "id": "TA-1325", - "scope": "split prequantized checkpoint detection from handler policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "NVFP4 and block-FP8 detection now precede the Qwen checkpoint-handler behavior selected by those formats.", - "A dedicated temporary subtree preserves all file-format cases." - ] - }, - { - "id": "TA-1326", - "scope": "split prequantized skip and load behavior from exclusion policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Quantized-key skipping, QKV and bias merges, exclusion parsing, dense handling, and MoE handling now form one checkpoint policy.", - "Normal and prequantized paths retain independent handler instances." - ] - }, - { - "id": "TA-1327", - "scope": "split fused selected-logprob primitive and loss dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Primitive forward and backward parity now flow through per-token CE, causal LM, Quack, and importance-sampling dispatch.", - "The full-logits memory-bound regression remains a separate performance contract." - ] - }, - { - "id": "TA-1328", - "scope": "split streaming forward-KL primitive and OPD dispatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense parity, chunking, masking, low-memory execution, compiled OPD dispatch, and clamp rejection now form one forward-KL execution policy.", - "The fp64 autograd gradcheck remains an independent numerical guard." - ] - }, - { - "id": "TA-1329", - "scope": "split DistSign reduce-scatter from optimizer lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "SP summation, post-sum sign, forced SUM reduction, builder selection, parameter grouping, and update numerics now form one optimizer policy.", - "A scoped distributed patch preserves communication isolation." - ] - }, - { - "id": "TA-1330", - "scope": "split DistSign hook configuration from optimizer lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Local and FSDP-managed hook ownership plus unsupported-topology admission now run before optimizer construction and stepping.", - "Each fake parallel topology remains independently asserted." - ] - }, - { - "id": "TA-1331", - "scope": "split batch-invariant global interpose and trunk-linear reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Wrapped trunk forward and backward parity now culminate in global-interpose training rejection and no-grad admission under the same CUDA gate.", - "The trunk contract is explicitly reset before the global-interpose phase." - ] - }, - { - "id": "TA-1332", - "scope": "split EP shared-replica classification from clipping policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Shared-replica ownership classification now begins the complete CPU EP clipping policy.", - "The live multi-rank reduction remains separate." - ] - }, - { - "id": "TA-1333", - "scope": "split EP norm modes and empty-gradient report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Inf norm, empty groups, missing gradients, and mixed-mesh clipping now form one local clipping policy.", - "Every norm and scaling assertion remains intact." - ] - }, - { - "id": "TA-1334", - "scope": "split skip-FSDP classification and local clipping report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Skip-FSDP expert classification, raw local norms, and uniform clipping now execute in the complete EP policy.", - "The no-reduction and no-division assertions remain intact." - ] - }, - { - "id": "TA-1335", - "scope": "split clip-grad dispatch report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP-aware and ordinary FSDP dispatch now precede mixed-mesh foreach behavior in one CPU report.", - "The world-one Gloo fixture still backs real DTensor mesh handling." - ] - }, - { - "id": "TA-1336", - "scope": "split MoE histogram and index kernels from gather-scatter report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Histogram and expert-slot index construction now feed scatter, gather, add-gather, and roundtrip execution.", - "All kernels retain the same CUDA capability gate." - ] - }, - { - "id": "TA-1337", - "scope": "split deterministic MoE scatter report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Stable ordering, stock coverage, dtype handling, and escape-hatch routing now run inside the complete MoE kernel policy.", - "A scoped environment context prevents the injected failure path from leaking." - ] - }, - { - "id": "TA-1338", - "scope": "split BI GEMM row invariance from table neutrality report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Table-entry bit neutrality now culminates in row invariance across M sizes and bucket boundaries.", - "Optional DeepGEMM parity remains an independent capability report." - ] - }, - { - "id": "TA-1339", - "scope": "split QLoRA injection from quantized execution lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Module injection and checkpoint-format selection now precede NVFP4 and block-FP8 execution, loading, and merging.", - "All phases retain the same CUDA gate." - ] - }, - { - "id": "TA-1340", - "scope": "split QLoRA optimizer reset from quantized execution lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "LoRA optimizer-state clearing and interval-triggered merge integration now close the QLoRA lifecycle.", - "Non-LoRA state preservation remains asserted." - ] - }, - { - "id": "TA-1341", - "scope": "split exact base DCP key contract and official-state load reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Projected dense and scale aliases now feed an official base-DCP load into exact runtime state.", - "A fresh checkpoint subtree and scoped load configuration preserve isolation." - ] - }, - { - "id": "TA-1342", - "scope": "split exact shared-expert native base views from construction report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Construction, runtime admission, logical factors, checkpoint sources, and native TP16 base slices now form one CPU component policy.", - "Optional SGLang factor slicing and Hopper execution remain separate." - ] - }, - { - "id": "TA-1343", - "scope": "split MoE LoRA initialization from eager execution report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Backend construction, frozen and trainable ownership, active-rank views, injection, eager forward, backward, and MoEBlock integration now form one CPU lifecycle.", - "Cross-backend CUDA numerics remain separate." - ] - }, - { - "id": "TA-1344", - "scope": "split EP LoRA router-score report from eager component lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All-to-all and DeepEP score application, identity behavior, and gradient flow now close the mocked CPU MoE LoRA policy.", - "Dispatch, compute, and combine mocks remain isolated in local contexts." - ] - }, - { - "id": "TA-1345", - "scope": "split asynchronous routing-replay record report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "CUDA stream ordering now begins the MoEBlock routing-replay integration under the same hardware gate.", - "The CPU registry and stage-management unit report remains independently runnable." - ] - }, - { - "id": "TA-1346", - "scope": "split multi-layer and pipeline routing-replay report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Single-layer record and replay now proceeds through multi-layer, multi-microbatch, and 1F1B scheduling in one CUDA lifecycle.", - "Global replay state is explicitly reset between phases." - ] - }, - { - "id": "TA-1347", - "scope": "split base-model and R3 routing-replay integration report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Checkpoint enabling, non-MoE admission, full backward replay, and R3 forward preload now close the routing-replay lifecycle.", - "Global replay state is explicitly reset before this phase." - ] - }, - { - "id": "TA-1348", - "scope": "split DTensor checkpoint materialization from grouped loading", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Replicated and sharded DTensor copy and four-rank save materialization now begin the grouped checkpoint-load policy.", - "The real CPU process workers remain intact." - ] - }, - { - "id": "TA-1349", - "scope": "split rank-zero checkpoint resolution and transport from grouped loading", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Object transport, rank-zero filtered prefetch, local resolution, grouped expert routing, and strict coverage now form one load lifecycle.", - "Scoped monkeypatch contexts preserve transport isolation." - ] - }, - { - "id": "TA-1350", - "scope": "split checkpoint model-key compatibility from distributed IO", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Pipeline key unions, QARL buffer admission, base-to-LoRA compatibility, metadata, load groups, and save groups now form one state lifecycle.", - "A fresh compatibility subtree preserves file isolation." - ] - }, - { - "id": "TA-1351", - "scope": "split optimizer-state filtering from distributed checkpoint IO", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Metadata-selected optimizer keys and multi-optimizer child filtering now close the distributed checkpoint policy.", - "A fresh optimizer subtree and scoped patches preserve isolation." - ] - }, - { - "id": "TA-1352", - "scope": "split exact MoE post-EP layout from construction inventory", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The 1700-factor inventory now proceeds through EP placement and logical owner-shape discovery in one construction policy.", - "A scoped EP16 rank patch isolates the layout phase." - ] - }, - { - "id": "TA-1353", - "scope": "split exact selected-logprob head attachment from MoE construction", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Selected-logprob LM-head ownership and fail-closed execution now close the complete exact construction policy.", - "The head phase uses an independent world16 patch context." - ] - }, - { - "id": "TA-1354", - "scope": "split fused GDN delta and merged-forward report from geometry", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Manifest selection and geometry now feed low-rank products, canonical folding, gradients, output projection, and bounded caches.", - "Every gradient-slice and cache-release assertion remains intact." - ] - }, - { - "id": "TA-1355", - "scope": "split fused GDN sharded checkpoint load from component lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Two-shard PEFT load now closes the same fused GDN injection, execution, and serialization lifecycle.", - "A fresh sharded checkpoint subtree preserves the original export." - ] - }, - { - "id": "TA-1356", - "scope": "split FlashMLA input flattening from backward compaction", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Batch offsets and invalid-index normalization now feed valid-row backward compaction and zero-gradient scattering.", - "All behavior remains in the hermetic CPU/mock policy." - ] - }, - { - "id": "TA-1357", - "scope": "split FlashMLA production-envelope admission from backward policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Device, shape, and Hopper admission now close the FlashMLA mock execution policy.", - "A scoped CUDA-capability patch isolates the admission phase." - ] - }, - { - "id": "TA-1358", - "scope": "split exact TP1 configuration from CPU forward policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Construction, rank-alpha admission, dtype identity, input rejection, exact forward, surrogate backward, and safety now form one CPU component policy.", - "The literal CUDA direct-program report remains separate." - ] - }, - { - "id": "TA-1359", - "scope": "split RMSNorm family declaration tripwire from funnel execution", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Undeclared warnings and required-family rejection now precede bitwise family funnel and module dispatch under one CUDA gate.", - "CPU structural guards remain independently runnable." - ] - }, - { - "id": "TA-1360", - "scope": "split fused gate-up registration from MoE checkpoint export", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Base and LoRA fused parameter ownership now begins the MoE checkpoint and HF export policy.", - "All registration shapes and aliases remain asserted." - ] - }, - { - "id": "TA-1361", - "scope": "split fused expert checkpoint-handler roundtrip from model export", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Qwen3 and Qwen3.5 fused expert load, save, deferred-skip, QKV unfusing, and QARL-buffer filtering now form one checkpoint policy.", - "Every handler uses an independent instance." - ] - }, - { - "id": "TA-1362", - "scope": "split OPD KL-estimator report from full-vocab policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "VERL estimator values, k3-plus straight-through gradients, and estimator dispatch now run with full-vocab modes and diagnostics.", - "All estimator formulas remain asserted." - ] - }, - { - "id": "TA-1363", - "scope": "split OPD policy-gradient mode report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Input admission, clipping metrics, PPO KL, and finite policy-gradient loss now form a branch of the complete OPD policy.", - "Backward-compatible metrics remain asserted." - ] - }, - { - "id": "TA-1364", - "scope": "split compiled OPD sampled-logprob ignored-label report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Compiled sampled-token logprobs and ignored-label zeroing now close the OPD dispatch policy.", - "Student and teacher output shapes and masks remain asserted." - ] - }, - { - "id": "TA-1365", - "scope": "split GDN convolution primitive and end-to-end block reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Primitive forward, backward, varlen, batching, and determinism now culminate in full GatedDeltaNet output and gradient parity.", - "Optional SGLang parity and CPU admission remain separate reports." - ] - }, - { - "id": "TA-1366", - "scope": "split SGLang fused-expert admission from CPU resolution and dispatch", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Feature resolution and block dispatch now proceed through unsupported activation, bias, clamp, and trainable-dispatch admission.", - "Each phase uses a scoped monkeypatch context while preserving every guard assertion." - ] - }, - { - "id": "TA-1367", - "scope": "split SGLang fused-expert weight mode and layout report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Transient, cached, and strided weight ownership now closes the CPU fused-expert policy.", - "Cache invalidation, zero-copy views, kernel layout, and split gate-up assertions remain intact." - ] - }, - { - "id": "TA-1368", - "scope": "split SGLang runtime-context report from fused-expert policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Runtime creation, compatible reuse, and incompatible-context rejection now form the final CPU configuration phase.", - "The real optional SGLang execution gate remains separate." - ] - }, - { - "id": "TA-1369", - "scope": "split SGLang trainable-gradient numerics from real parity", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Stock-Triton gradients and masked routing gradients now precede strided and auto-mode bitwise parity under one CUDA and SGLang capability gate.", - "CPU policy coverage is not hidden by the optional dependency gate." - ] - }, - { - "id": "TA-1370", - "scope": "split sparse-MLA attention-sink report from forward parity sweep", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Zero and signed sink parity plus the non-ignored effect check now run in the representative first forward specialization.", - "All four compiled top-k forward specializations remain collected." - ] - }, - { - "id": "TA-1371", - "scope": "split DeepSeek-V3 default LoRA target report from tiny model lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default and explicit MLA and MoE adapter targets now follow the tiny forward, backward, and router-freeze transaction.", - "Every projection type assertion remains in a named scenario helper." - ] - }, - { - "id": "TA-1372", - "scope": "split DeepSeek-V3 router observability and replay report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sparse-layer router output counts and recorded routing weights now close the DeepSeek-V3 model lifecycle.", - "Dense-prefix and all-MoE schedules remain covered." - ] - }, - { - "id": "TA-1373", - "scope": "split DeepSeek-V4 hash-layer structure report from non-hash MoE policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Non-hash routing, shared experts, SwiGLU clamps, hash admission, table selection, and gradients now form one CPU MoE policy.", - "Hash and non-hash branches retain independent fixtures inside named helpers." - ] - }, - { - "id": "TA-1374", - "scope": "split DeepSeek-V4 hash routing replay report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Hash-table routing now culminates in record and replay-backward behavior in the same architecture MoE report.", - "Recorded indices and replay gradients remain asserted." - ] - }, - { - "id": "TA-1375", - "scope": "split MiniMax-M3 checkpoint mapping from architecture support", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Configuration, registry, text runtime, checkpoint fusion, EP ownership, and expert aliases now form one architecture-support report.", - "Raw-key skip accounting and all projection mappings remain asserted." - ] - }, - { - "id": "TA-1376", - "scope": "split MiniMax-M3 paging and CPU MSA admission report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Stable paged-KV layout and loud CPU rejection now close the MiniMax-M3 architecture-support report.", - "The helper remains independent of optional CUDA execution." - ] - }, - { - "id": "TA-1377", - "scope": "split GLM52 routed-bank EP checkpoint slice report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The exact full-block QLoRA inventory now includes scoped EP16 local-bank ownership and offset selection.", - "All 75 routed banks retain local-expert and global-offset assertions." - ] - }, - { - "id": "TA-1378", - "scope": "split GLM52 exact component and admission report from QLoRA inventory", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Logical inventory now proceeds through exact dense roots, rank-alpha rejection, target validation, and supported training-mode admission.", - "The meta-device construction keeps the combined policy hermetic." - ] - }, - { - "id": "TA-1379", - "scope": "split GLM52 routed-expert gradient edges from literal sampler coverage", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "All 256 owner-slot remaps now culminate in sentinel, mixed-owner, and top-k logical VJP behavior under one hardware gate.", - "CPU topology and physical-buffer policy remains separately runnable." - ] - }, - { - "id": "TA-1380", - "scope": "sparse-MLA deterministic-dKV and invalid-index backward reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Both reports pass independently and protect distinct deterministic accumulation and invalid-slot gradient behavior.", - "The larger base backward sweep terminates the current H100 pytest process, so folding these passing boundaries into that sweep would destroy failure isolation." - ] - }, - { - "id": "TA-1381", - "scope": "FP8 grouped-kernel numerics and full MoE optimizer-step reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The grouped forward and weight-gradient report passes independently on H100.", - "The full Quack optimizer-step report terminates the current pytest process, which is a real lifecycle failure that must not erase the passing kernel-level result." - ] - }, - { - "id": "TA-1382", - "scope": "split GLM5 indexer construction from architecture support", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "GLM5 configuration and registry loading now proceed through indexer geometry, FP32 head projection, selection, and masking.", - "Each mutable indexer scenario runs in a scoped monkeypatch context." - ] - }, - { - "id": "TA-1383", - "scope": "split GLM5 sparse-MLA wrapper and attention integration report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Reference arithmetic, dispatch shaping, Ulysses integration, dense parity, and backend rejection now form the sparse-attention phase of one GLM5 support policy.", - "The real TileLang fast path remains an independent CUDA report." - ] - }, - { - "id": "TA-1384", - "scope": "split GLM5 checkpoint filtering from architecture support", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Partial-layer and EP expert-key filtering now precede adapterization and model execution in the architecture lifecycle.", - "Layer, non-layer, local-expert, and out-of-range key assertions remain intact." - ] - }, - { - "id": "TA-1385", - "scope": "split GLM5 adapter sparse-KV and MoE dispatch report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Default MLA, shared-expert, and routed-expert targets now feed EP dispatch and sparse absorbed-KV LoRA execution.", - "Indexer exclusions and live delta contribution remain asserted." - ] - }, - { - "id": "TA-1386", - "scope": "split GLM5 forward and recompute report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense and MoE layer construction now culminates in end-to-end hidden-state output and recompute-before-dispatch checkpoint routing.", - "Optional HF-reference logits remain separately runnable." - ] - }, - { - "id": "TA-1387", - "scope": "split native EP-combine variable-row collective report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "EP8 admission now proceeds through token and ID padding, reduce-scatter gradients, and shared maximum-row selection.", - "Mocked collectives run in an isolated monkeypatch context." - ] - }, - { - "id": "TA-1388", - "scope": "split native EP-combine serving fused-gate gradient report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Serving fused-gate output and trainer gradients now form the differentiable gate phase of the native-combine policy.", - "Hidden, gate, shared, and routed gradients still match the eager reference." - ] - }, - { - "id": "TA-1389", - "scope": "split native EP-combine dispatch and actual-operand report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FSDP module entry, local routed partials, shared projection, chain sum, diagnostics, and final output now close one native EP transaction.", - "Every captured exact-combine boundary remains asserted." - ] - }, - { - "id": "TA-1390", - "scope": "split SGLang EP slot-combine report from dispatch and admission", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Flag, backend, empty-rank, and trainable admission now proceed through slot-ordered reduction and pair-count guards.", - "The CPU EP fixture and mock kernel boundary are identical." - ] - }, - { - "id": "TA-1391", - "scope": "split SGLang EP weight-presentation report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Strided zero-copy views plus transient and cached presentation now close the CPU SGLang EP policy.", - "The real optional stock-Triton gradient comparison remains separate." - ] - }, - { - "id": "TA-1392", - "scope": "FlashQLA auto-CP pin, shape invariance, and chunk-chaining reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The reports protect distinct control-flow pinning, packed and batch-composition invariance, and recurrent state-handoff exactness.", - "Gate 2 and Gate 4 are production numerical boundaries rather than repeated shape smoke cases." - ] - }, - { - "id": "TA-1393", - "scope": "training utility clipping, metadata, pipeline loss, and synchronization reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The four reports exercise unrelated gradient clipping, distributed counting, chunked cross-entropy, and explicit synchronization APIs.", - "Sharing a utility module is not evidence that these failure boundaries are equivalent." - ] - }, - { - "id": "TA-1394", - "scope": "GLM52 index sharing, codec parity, semantic MoE, layer plan, and selector reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The five reports separate FSDP identity, optional sampler bytes, end-to-end logprob composition, static topology, and sparse-selector implementation behavior.", - "Only the sampler codec requires the paired CUDA serving stack; folding would hide the four hermetic CPU contracts." - ] - }, - { - "id": "TA-1395", - "scope": "split MoE all-to-all pre-dispatch score-order report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Token permutation and routing-weight order now feed the mocked all-to-all pre-dispatch transaction.", - "Received ordering, expert cumsums, and routing-weight gradients remain asserted." - ] - }, - { - "id": "TA-1396", - "scope": "split MoE post-all-to-all hidden-chunking report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The same permutation lifecycle now closes with chunked and unchunked post-dispatch output and gradient parity.", - "Pre- and post-dispatch collectives use isolated monkeypatch contexts." - ] - }, - { - "id": "TA-1397", - "scope": "split packing full-pipeline roundtrip from packing-on policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Capacity, batching, token metadata, generated labels, simulated forward output, and sample-boundary unpacking now form one packing-on lifecycle.", - "Packing-disabled behavior remains a separate supported mode." - ] - }, - { - "id": "TA-1398", - "scope": "split trained-QARL logprob preservation from quantized export policy", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "A trained QARL directory now proves exact dequantized logprobs before the CLI layout and admission matrix runs.", - "The scenario uses a dedicated temporary export directory." - ] - }, - { - "id": "TA-1399", - "scope": "split optimizer-step Adam override report from server initialization", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Server defaults and validation now feed full, partial, omitted, adapter, and non-Adam optimizer-step overrides.", - "All mutable runner patches are scoped to the step phase." - ] - }, - { - "id": "TA-1400", - "scope": "split optimizer dispatcher payload-forwarding report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Explicit and omitted Adam fields now traverse configuration, runner mutation, and dispatcher forwarding as one server lifecycle.", - "Backward-compatible None values remain asserted." - ] - }, - { - "id": "TA-1401", - "scope": "split sparse-delta prepacked-path posting report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Sparse encoding, baseline updates, receiver failures, per-rank packed paths, cache metadata, and endpoint accounting now form one backend lifecycle.", - "The prepacked phase uses its own temporary directory." - ] - }, - { - "id": "TA-1402", - "scope": "split sparse-delta initialization admission report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Post-only and prepacked-only admission plus runtime helper loading now close the sparse-delta backend policy.", - "Initialization artifacts use a dedicated temporary directory." - ] - }, - { - "id": "TA-1403", - "scope": "packing-on and packing-disabled reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Packing-disabled is a supported control-flow mode with different batching, warning, shifting, and loss-mask behavior.", - "It is not a narrow input variation of the packing-on roundtrip." - ] - }, - { - "id": "TA-1404", - "scope": "FP8 weight quantization primitive and directory export reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The primitive report protects block-scale numerical and layout behavior without filesystem or CLI dependencies.", - "The export report protects model-state transformation, artifact layout, configuration, and admission." - ] - }, - { - "id": "TA-1405", - "scope": "split checkpoint compatibility argument parsing report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Optimizer, packing, and numeric configuration parsing now proceeds through EP checkpoint compatibility, automatic checkpoint resolution, and optimizer-state loading.", - "The checkpoint phase uses a dedicated configuration directory and monkeypatch context." - ] - }, - { - "id": "TA-1406", - "scope": "split low-precision argument parsing report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The same parser lifecycle now closes with FP8 aliases, fail-fast defaults, vLLM knob rejection, and supported low-precision modes.", - "All environment and argv mutations remain isolated." - ] - }, - { - "id": "TA-1407", - "scope": "split dataset local, hub, URL, and data-files loading report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dataset name and shard expansion plus type inference now feed local-file, saved-directory, hub, URL, and data-files loading.", - "The loading phase uses an isolated cache and temporary directory." - ] - }, - { - "id": "TA-1408", - "scope": "split preprocessed dataset save-load report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Expanded and loaded datasets now proceed through split, merge, save, reload, and missing-cache behavior as one preparation lifecycle.", - "The persisted artifact uses a dedicated temporary directory." - ] - }, - { - "id": "TA-1409", - "scope": "split P2P trainer abort-marker report from device selection", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "IB-device mapping precedence now proceeds through sync-abort publication, peer observation, and cleanup.", - "Abort state is scoped to a temporary transfer directory." - ] - }, - { - "id": "TA-1410", - "scope": "split P2P trainer peer-status gather report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The trainer transfer policy now closes by gathering local success and remote failure status across ranks.", - "Distributed mocks run in an isolated monkeypatch context." - ] - }, - { - "id": "TA-1411", - "scope": "split FP8 adapter-merge weight-sync report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "BF16-island selection and block layout now feed dense QLoRA and MoE adapter folding before FP8 quantization.", - "Projection selection, skip lists, existing FP8 state, and stack behavior remain asserted." - ] - }, - { - "id": "TA-1412", - "scope": "split FP8 CPU expert projection and workspace report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The CPU weight-sync policy now proceeds through expert transposition, zero padding, deferred quantization, streaming workspaces, flush reset, and completion metadata.", - "The independent live GPU parity and device-transfer report remains separate." - ] - }, - { - "id": "TA-1413", - "scope": "checkpoint save-load, list-delete, and model-ID validation API reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The reports protect distinct mutating checkpoint I/O, listing and deletion isolation, and request-boundary path validation.", - "Collapsing public endpoint and security failures into one large transaction would reduce actionable failure isolation." - ] - }, - { - "id": "TA-1414", - "scope": "weight-sync receiver, source preparation, and sparse-delta handler reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Receiver post-processing, trainer-side parameter extraction, and sparse prepacked transport are different endpoints and data paths.", - "Each report already consolidates its internal configuration, layout, and admission variants." - ] - }, - { - "id": "TA-1415", - "scope": "inference endpoint registration, weight sync, and quantization normalization reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Registration and health discovery, explicit synchronization, and quantization schema validation expose distinct public API boundaries.", - "Each report contains its own success, rejection, and compatibility lifecycle rather than shape-only variants." - ] - }, - { - "id": "TA-1416", - "scope": "split adapter registration report from coordinated load lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Fresh and evicted adapter loading now proceeds through adapter and session registration, broadcasts, cross-rank failure rollback, and worker exceptions.", - "Registration uses a dedicated temporary checkpoint root." - ] - }, - { - "id": "TA-1417", - "scope": "split adapter save admission report from coordinated load lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The coordinator policy now closes with save admission that refuses to recreate missing evicted state.", - "Load, registration, and save paths retain independent fixtures inside named helpers." - ] - }, - { - "id": "TA-1418", - "scope": "split optimizer checkpoint save-resume report from identity transaction", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Canonical parameter identity and transactional collective failure now feed sharded manifest creation, artifact admission, and bitwise resumed training.", - "The save-resume phase uses an isolated monkeypatch context and artifact root." - ] - }, - { - "id": "TA-1419", - "scope": "split optimizer logical reshard report from checkpoint resume", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Saved optimizer moments now proceed through one-dimensional, multidimensional, replicated, same-world, and invalid-source logical resharding.", - "Reshard fixtures use a dedicated checkpoint directory and patch context." - ] - }, - { - "id": "TA-1420", - "scope": "split model-runner initial checkpoint restore report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Checkpoint materialization and zero-meta admission now feed runner step restoration, optimizer selection, and failure-state publication.", - "Manager-level patches are undone before the runner phase." - ] - }, - { - "id": "TA-1421", - "scope": "split default-adapter initialization report from checkpoint loading", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The base restore lifecycle now culminates in guarded fresh default-adapter initialization and ownership compilation.", - "Uninitialized to-empty storage remains explicitly rejected before registration." - ] - }, - { - "id": "TA-1422", - "scope": "split dispatcher routing-payload transport and security report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Batch sharding and provenance now feed filesystem and Mooncake routing payload slicing, legacy and symlink rejection, and diagnostic artifacts.", - "Routing transport uses an isolated temporary root and monkeypatch context." - ] - }, - { - "id": "TA-1423", - "scope": "split dispatcher completion rendezvous and per-token merge report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The dispatcher lifecycle now closes with CP replica deduplication, disagreement rejection, bounded completion payloads, and rank-zero merge.", - "All completion cases remain in a named scenario helper." - ] - }, - { - "id": "TA-1424", - "scope": "split endpoint health preflight from NCCL transfer report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Endpoint model discovery and all-failure diagnostics now precede NCCL initialization, direct bucket routing, and two-phase receiver completion.", - "The combined report covers one endpoint health-to-transfer lifecycle." - ] - }, - { - "id": "TA-1425", - "scope": "split adapter ownership compiler topology report from producer execution", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "A declared module-managed producer now feeds all four topology families, stable fingerprints, fail-closed admission, and replica-domain coverage.", - "Fullgraph producer gradients and compiler structure remain asserted." - ] - }, - { - "id": "TA-1426", - "scope": "split adapter residual transport report from ownership compilation", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The compiled ownership plan now culminates in bucketed residual reduction, immutable raw accumulators, and logical norm accounting.", - "Distributed finalizer mocks use an isolated monkeypatch context." - ] - }, - { - "id": "TA-1427", - "scope": "split runner expert-factor compilation and admission report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense, exact LM-head, and replica topology compilation now proceeds through unquantized, block-FP8, NVFP4, NF4, and session-rank expert-factor contracts.", - "Certified and rejected backend combinations retain dedicated helper fixtures." - ] - }, - { - "id": "TA-1428", - "scope": "split direct-output analytical-step report from runner ownership compiler", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The runner compiler policy now closes with authoritative analytical gradients, effective LM-head folding, capture finalization, and parameter mutation.", - "The direct-output phase uses a dedicated manager root and monkeypatch context." - ] - }, - { - "id": "TA-1429", - "scope": "split OPD distributed teacher-cache assembly report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Teacher contributor selection, CP gathering, Mooncake metadata, and cache row consumption now feed the OPD loss execution lifecycle.", - "Patch-decorated helpers are invoked with isolated fixture contexts." - ] - }, - { - "id": "TA-1430", - "scope": "split OPD debug-artifact report from loss execution", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Packed cache and weight shaping plus teacher assembly now culminate in loss metrics and ranked vocab-parallel debug artifacts.", - "Loss and debug files use separate temporary directories." - ] - }, - { - "id": "TA-1431", - "scope": "split fused gate-up and Nemotron-H Muon classification report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Muon construction, grouping, fallback, Gram-Newton-Schulz stepping, and Quack backend admission now include fused gate-up discovery and a tiny Nemotron-H update.", - "Standard Newton-Schulz arithmetic and CUDA FP32-compute preservation remain separate." - ] - }, - { - "id": "TA-1432", - "scope": "attention backend registry, FlashAttention API, and SGL page-cache reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Backend resolution, variable-length FlashAttention behavior, and paged KV-cache semantics exercise different public APIs and storage models.", - "Their shared module location does not make their regressions equivalent." - ] - }, - { - "id": "TA-1433", - "scope": "mixed precision, folded sequence parallelism, and manual FSDP prefetch reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The reports protect independent dtype selection, model transformation, and execution-order configuration boundaries.", - "None is an input variation of another policy." - ] - }, - { - "id": "TA-1434", - "scope": "Muon standard Newton-Schulz and CUDA compute-dtype reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Batched leading-dimension preservation tests a different orthogonalization algorithm from the main Gram-Newton-Schulz policy.", - "Live CUDA FP32 backend dtype preservation cannot be replaced by CPU construction or model-classification assertions." - ] - }, - { - "id": "TA-1435", - "scope": "split dense primitive and model batch-composition reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Batch-invariant matmul, RMSNorm, log-softmax, and mean now establish the primitive contract before the padded dense-model consumer runs under the same CUDA gate.", - "The surviving report retains every primitive equality and full-model hidden-state assertion." - ] - }, - { - "id": "TA-1436", - "scope": "split DeepSeek-V4 attention runtime and sink-dtype reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Window and compressed attention forward-backward coverage now includes sink storage, TileLang call-dtype conversion, RoPE state, and TP admission in one attention lifecycle.", - "The TileLang boundary remains mocked and hermetic inside the representative window-attention case." - ] - }, - { - "id": "TA-1437", - "scope": "split DeepSeek-V4 wo_a and full attention-LoRA training reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The grouped wo_a delta and gradient regression now precedes all-target injection, base freezing, autograd reachability, and first-step adapter gradients.", - "Both scenarios retain independent model construction and inputs within one supported attention-LoRA policy." - ] - }, - { - "id": "TA-1438", - "scope": "standalone TileLang indexer causal-mask report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Every forward geometry now verifies both valid numerical scores and all invalid future positions from the same kernel output.", - "A vectorized negative-infinity assertion replaces the redundant kernel invocation and Python element walk; large-value and zero-input cases still run once." - ] - }, - { - "id": "TA-1439", - "scope": "dataset split-fingerprint and configuration-hash reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Split derivation and complete dataset-configuration identity are different production algorithms with different inputs and consumers.", - "Their shared digest representation does not make either report an example of the other." - ] - }, - { - "id": "TA-1440", - "scope": "dataloader builder configuration and packed pipeline integration reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The first report isolates sampler, microbatch, parallel-state, and collator construction; the second executes real loader batches and packed sequence layouts.", - "Keeping the unit boundary separately runnable preserves useful failure localization without one report per input example." - ] - }, - { - "id": "TA-1441", - "scope": "Mooncake hidden transport and metadata-admission reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Successful tensor transport and TeacherActivationCache consumption are independent from malformed metadata, missing objects, removal, and environment configuration failures.", - "The positive data path and fail-closed admission path already consolidate their internal variants." - ] - }, - { - "id": "TA-1442", - "scope": "DeepSeek-V4 HF-to-DCP conversion and AutoModel loading reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Distributed-checkpoint conversion and Transformers AutoModel dispatch are distinct public ingestion paths even though both consume a synthetic HF snapshot.", - "A failure in one does not imply or diagnose a failure in the other." - ] - }, - { - "id": "TA-1443", - "scope": "GLM52 native-FP8 routing-runtime and checkpoint-buffer reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Canonical routing and scoring-only expert execution are runtime behavior, while configuration validation, module replacement, byte packing, and checkpoint ownership are construction and persistence boundaries.", - "Each report already joins the narrow examples inside its own production boundary." - ] - }, - { - "id": "TA-1444", - "scope": "split families-v2 RMSNorm reference and realization reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "FP64 wrongness bounds, residual behavior, zero-centered mode, batch invariance, and repeatability now precede forced fused-versus-split bit equality and dispatch selection.", - "All checks exercise one frozen RMSNorm numerical tree under the same CUDA gate." - ] - }, - { - "id": "TA-1445", - "scope": "split exact factor-only weight-sync and checkpoint-publication reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Exact dense, projection, LM-head, streaming, and sparse-delta preflight rejection now culminate in separate A and B factor persistence and byte verification.", - "The filesystem publication uses an isolated temporary checkpoint inside the same factor-only contract." - ] - }, - { - "id": "TA-1446", - "scope": "parameterized index-share checkpoint modes and separate cleanup report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Reentrant and non-reentrant checkpoint modes are now an internal two-case input matrix rather than two product report IDs.", - "Producer reuse, detached payloads, gradients, forward-only completion, forward failure, backward failure, and idempotent cleanup form one index-share lifecycle." - ] - }, - { - "id": "TA-1447", - "scope": "standalone solve_tril two-warp configuration pin", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The warp count selects the association tree of a Triton reduction and is therefore part of the exact arithmetic contract, not a cosmetic tuning constant.", - "The tolerant GDN runtime parity report cannot detect a bit-level drift in that source configuration." - ] - }, - { - "id": "TA-1448", - "scope": "standalone two-GPU PP 1F1B convergence job", - "decision": "remove", - "status": "applied", - "evidence": [ - "The schedule-parity report already launches the same PP2 FSDP1 1F1B baseline before comparing all virtual-stage schedules.", - "Baseline convergence is now asserted in that surviving transaction, eliminating a separate full training subprocess." - ] - }, - { - "id": "TA-1449", - "scope": "separate LoRA FSDP2 convergence and checkpoint-resume jobs", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The two-phase FSDP2 checkpoint transaction now resumes to the former 20-step convergence horizon and asserts the explicit checkpoint-load marker.", - "The surviving report passed with checkpoint restoration, global step 20, and the original loss-convergence threshold." - ] - }, - { - "id": "TA-1450", - "scope": "test-authored OPD packing-strategy loss invariance report", - "decision": "remove", - "status": "applied", - "evidence": [ - "Its fake backend computed KL and global normalization entirely with test helpers rather than invoking XoRL's OPD loss implementation.", - "Production OPD numerics and reducer composition plus sequential, best-fit, and balanced-DP packing remain covered directly." - ] - }, - { - "id": "TA-1451", - "scope": "fake teacher-cache metadata echo report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The backend returned fixed metadata and timing literals that the assertions merely read back from the request-processor response.", - "Real ModelRunner Mooncake publication, cache-index construction, activation-cache consumption, and OPD pipeline transport remain covered." - ] - }, - { - "id": "TA-1452", - "scope": "remaining full-weight FP8 end-to-end topology matrix", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The seven survivors select distinct checkpoint-resume, tensor-parallel, Ulysses, Ring, hybrid long-tail packing, local MoE, and DeepEP EP/eFSDP mechanisms.", - "Earlier shape-ladder and checkpoint-by-topology cross-products are already removed; no survivor is only a larger input example of another." - ] - }, - { - "id": "TA-1453", - "scope": "P2P staging, FP8 layout, and transfer-failure reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Source staging and receiver placement, fused FP8 byte layouts, and fail-closed manifest and runtime diagnostics exercise different transport boundaries.", - "Each report already consolidates the narrow locator, shape, retry, and layout variants within its boundary." - ] - }, - { - "id": "TA-1454", - "scope": "adapter manager ownership, capture, optimizer, restore, and multi-adapter reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Plan compilation, gradient capture, authoritative mutation, checkpoint admission, and multi-adapter eviction are independent state-machine transitions with different rollback obligations.", - "The reports already aggregate their per-input and per-optimizer branches rather than multiplying product IDs." - ] - }, - { - "id": "TA-1455", - "scope": "duplicate eight-GPU PP2 FSDP4 AdamW convergence job", - "decision": "remove", - "status": "applied", - "evidence": [ - "The retained Muon job exercises the same PP2, FSDP4, two-microbatch, packed training topology while also covering optimizer partitioning.", - "AdamW construction and stepping are covered independently; changing the optimizer does not select a different pipeline transport or loss-normalization mechanism." - ] - }, - { - "id": "TA-1456", - "scope": "dead Nemotron-H end-to-end training module", - "decision": "remove", - "status": "applied", - "evidence": [ - "All three static reports required tiny_nemotron_h_model_dir, a fixture absent from every repository conftest and absent since the module was introduced.", - "Every report therefore errored during pytest setup without constructing a model; real Nemotron runtime, packed-varlen, checkpoint, gradient, and Muon-step policies remain covered." - ] - }, - { - "id": "TA-1457", - "scope": "native-FP8 plain-conversion family parametrization", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Linear and expert modules are two inputs to the same frozen-FP32 ordinary conversion contract and now execute in one internal family matrix.", - "Both families still prove non-DTensor materialization, byte preservation, device placement, dtype, and requires-grad state." - ] - }, - { - "id": "TA-1458", - "scope": "FlashQLA small-head and production-head report IDs", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Four-head and 32-head inputs now run sequentially inside one FlashQLA-versus-FLA numerical policy instead of producing separate pytest reports.", - "Both shapes retain forward, final-state, and every input-gradient cosine and finiteness check on Hopper." - ] - }, - { - "id": "TA-1459", - "scope": "PEFT MoE EP-slice orientation parametrization", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Down-projection A and gate-projection B are transpose orientations of the same global-to-local expert slice conversion and now form one report.", - "Both orientations still construct all eight published experts and verify the exact rank-two local shard." - ] - }, - { - "id": "TA-1460", - "scope": "non-gated MoE backend report parametrization", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Installed Triton and native implementations now feed one eager-reference policy instead of emitting one report per backend.", - "Each available backend still verifies forward values plus input, gate-up, and down-projection gradients." - ] - }, - { - "id": "TA-1461", - "scope": "surviving PP2 FSDP4 Muon topology gate", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The report is the only direct-trainer eight-GPU composition of pipeline parallelism and FSDP4 after duplicate removal.", - "A live run reached the 1F1B schedule and exposed the current pipeline backward product failure, so it is an effective behavioral gate rather than a configuration smoke." - ] - }, - { - "id": "TA-1462", - "scope": "unreachable FP8 hybrid Ulysses Ring long-tail E2E report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report requested tiny_agent_context_dense_model_dir_with_weights, a fixture that has never existed in repository history, so it failed during setup before constructing a trainer.", - "The surviving FP8 E2E matrix separately exercises Ulysses, Ring, checkpoint-resume, local MoE, and DeepEP EP/eFSDP, while packing behavior has functioning lower-level coverage." - ] - }, - { - "id": "TA-1463", - "scope": "DeepSeek-V4 window and compressed attention report parametrization", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Window attention and C128 compression are internal modes of one forward-backward shape policy and now execute as an isolated in-report matrix.", - "Both modes retain output, parameter-gradient, RoPE, QAT, sink-dtype, TileLang call-dtype, and topology-admission assertions." - ] - }, - { - "id": "TA-1464", - "scope": "exact GLM sparse attention CP report parametrization", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Non-CP and Ulysses-CP select two branches of the same exact q and v projection routing contract and now run as an isolated Boolean mode matrix.", - "Both branches still verify call order, factor-only execution, query offsets, tensor shapes, and the absence of frozen-weight materialization." - ] - }, - { - "id": "TA-1465", - "scope": "stale optional GLM5 FLOPs-counter module gate", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The test skipped the entire module on a missing GLM5 configuration that is now a shipped source-tree dependency, not an optional package.", - "Direct imports make a broken GLM5 package or configuration import fail the report instead of silently removing coverage." - ] - }, - { - "id": "TA-1466", - "scope": "stale Dr.GRPO model-runner implementation gate", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "ModelRunner now has both the Dr.GRPO dispatch branch and its loss-exclusion inventory, making the upstream-WIP module skip obsolete.", - "The report now unconditionally covers dispatch, legacy field names, temperature, output suppression, and K3 output requests." - ] - }, - { - "id": "TA-1467", - "scope": "eager versus native MoE broad import skip", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The helper caught every exception while importing shipped internal MoE modules and converted production import regressions into skips.", - "Lazy imports remain collection-safe, but a failure now fails the live forward, backward, determinism, and edge-case policy." - ] - }, - { - "id": "TA-1468", - "scope": "Triton and Quack EP routing-score report parametrization and import skip", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both backends implement the same routing-score forward and gradient boundary and now run inside one isolated backend matrix rather than emitting separate product reports.", - "The Quack case had always skipped because its test-owned module stub omitted a required grouped-GEMM symbol; the repaired stub now executes Quack, and internal import failures are no longer suppressed." - ] - }, - { - "id": "TA-1469", - "scope": "opt-in DeepGEMM grouped-FP8 subprocess report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report skipped unless XORL_TEST_DEEP_GEMM_FP8=1, and no repository workflow, script, or configuration ever sets that flag.", - "A permanently dormant manual diagnostic is not suite coverage; the remaining FP8-MoE reports execute injection, grouped forward and weight gradients, Quack and Triton backends, and full optimizer steps." - ] - }, - { - "id": "TA-1470", - "scope": "Quack EP parity broad internal-import skips", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Both CPU gradient-arity and GPU Quack-versus-Triton reports caught every exception from shipped internal modules and converted implementation regressions into skips.", - "Quack is a pinned project dependency; explicit CUDA gates remain, while internal imports now fail closed and both reports pass." - ] - }, - { - "id": "TA-1471", - "scope": "GKN checkpoint and eager-backend internal-import skips", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The reports described transformers as optional even though it is a core dependency and the imported checkpoint buffer and eager backend are shipped source files.", - "Direct imports preserve the GKN checkpoint, eager, native, Triton, and cross-backend assertions while exposing packaging or import regressions." - ] - }, - { - "id": "TA-1472", - "scope": "GLM5 TileLang indexer broad import skip", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The supported-CUDA report caught every exception importing the shipped TileLang indexer and could hide a broken kernel module behind an availability skip.", - "TileLang is pinned by the project; the explicit CUDA gate remains and the direct-import fast-path parity report passes." - ] - }, - { - "id": "TA-1473", - "scope": "FlashQLA contract and numerical parity broad import skips", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The reports already gate on CUDA architecture and the required TileLang feature, so an additional catch-all around shipped FlashQLA imports only concealed backend regressions.", - "SM90 and prefer_instruction skips remain explicit; all four exact-contract reports and the FlashQLA-versus-FLA numerical report pass with direct imports." - ] - }, - { - "id": "TA-1474", - "scope": "standalone forced-SSM-kernel-unavailable report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Forcing use_kernel=True with the kernel removed is the admission branch of the same CPU SSD fallback and packed-recurrence policy.", - "The RuntimeError assertion now runs after dense, packed, chunking, gradient, and full-mixer recurrence checks without a separate pytest report." - ] - }, - { - "id": "TA-1475", - "scope": "standalone learning-rate scheduler invalid-configuration report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Invalid learning rate, warmup ratio, and mode inputs are builder admission branches adjacent to constant, linear, and cosine scheduling.", - "All five rejection cases now close the scheduler mode policy instead of producing an independent product report." - ] - }, - { - "id": "TA-1476", - "scope": "direct pytest.main launch blocks in test modules", - "decision": "remove", - "status": "applied", - "evidence": [ - "Thirteen test modules carried __main__ blocks that are never reached by repository pytest invocation, collection, or CI.", - "Removing the duplicate direct-launch surface changes no test behavior; individual modules remain runnable through pytest paths and node IDs." - ] - }, - { - "id": "TA-1477", - "scope": "pytest GPU and skip markers attached to private FP8 assertion helpers", - "decision": "remove", - "status": "applied", - "evidence": [ - "Pytest markers on directly called _assert helpers do not enforce selection or skipping and falsely implied that each helper had an independent hardware gate.", - "The seven collected FP8 policies retain their real GPU markers, execute every helper, and pass after the inert helper decorations are removed." - ] - }, - { - "id": "TA-1478", - "scope": "pytest markers attached to private assertion helpers repository-wide", - "decision": "remove", - "status": "applied", - "evidence": [ - "One hundred twenty private helpers across 30 files carried MarkDecorator metadata even though their names prevent collection and callers invoke them directly.", - "All public test reports retain their CPU, GPU, async, distributed, architecture, and optional-dependency markers; representative model, operator, distributed, server, and weight-sync reports pass." - ] - }, - { - "id": "TA-1479", - "scope": "unused abstract-method lookalikes in the FP8 weight-sync QLoRA fake", - "decision": "remove", - "status": "applied", - "evidence": [ - "The fake subclass overrode four base methods only to raise NotImplementedError, but the exercised merge-and-sync path never calls them and the base methods already fail identically.", - "The retained dequantize_expert implementation supplies the only fake behavior consumed by the production weight-sync path, whose focused policy passes." - ] - }, - { - "id": "TA-1480", - "scope": "test-tree Ruff defects in fixture and path-bootstrap scaffolding", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The packed-dataset fixture assigned an unused current_pos variable, while three standalone distributed or E2E modules intentionally imported after path bootstrapping without local E402 annotations.", - "The dead assignment is removed, the intentional imports are explicit, and the complete tests tree now passes Ruff." - ] - }, - { - "id": "TA-1481", - "scope": "duplicate balanced synthetic TopK routing report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The dedicated TopK router policy already proves balanced expert selection, uniform weights, count balance, and override precedence for synthetic routing.", - "Only MoEBlock replay regather behavior was unique in the second report; that assertion now closes the train-router dispatch policy and both surviving router policies pass." - ] - }, - { - "id": "TA-1482", - "scope": "unused repository-wide and E2E fixtures plus their dead helper types", - "decision": "remove", - "status": "applied", - "evidence": [ - "An AST fixture dependency map covering function parameters, fixture-to-fixture dependencies, usefixtures, indirect parametrization, and getfixturevalue found no consumer for fake_packed_dataset or small_dense_model_dir_with_weights.", - "FakePackedDataset and the root SimpleCollator were reachable only from that dead fixture or nowhere; full pytest setup planning succeeds after all four dead setup surfaces are removed." - ] - }, - { - "id": "TA-1483", - "scope": "standalone vocab-parallel reverse-KL pseudo-test", - "decision": "remove", - "status": "applied", - "evidence": [ - "The file instructed users to invoke pytest but defined only main and worker functions, so collection produced zero reports and no repository workflow called it directly.", - "The retained four-rank lm-head TP FSDP policy reaches the gathered vocab-parallel OPD implementation through production code and passes loss plus hidden and weight gradient comparison across six CP, DP, and HSDP topologies." - ] - }, - { - "id": "TA-1484", - "scope": "manual DeepEP, uneven vocab-parallel OPD, and real-model QLoRA campaign workloads under tests", - "decision": "relocate", - "status": "applied", - "evidence": [ - "These modules define no pytest report and are not called by a repository workflow; they require direct torchrun or Python execution, or two 100-step eight-H100 real-checkpoint training runs.", - "Their diagnostic value is preserved under certification/deepep, certification/opd, and certification/qwen3_30b while the tests tree now contains only repository regression coverage and shared test support." - ] - }, - { - "id": "TA-1485", - "scope": "fabricated Quack DeepEP None-gradient backward report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report called custom-autograd backward methods directly with a SimpleNamespace context and grad_output=None, then asserted a tuple length derived from that same fake context; PyTorch autograd and production dispatch were never entered.", - "The retained Quack grouped-GEMM and DeepEP no-permute reports execute real forward and backward graphs, compare outputs and every trainable gradient against trusted implementations, and pass." - ] - }, - { - "id": "TA-1486", - "scope": "context-parallel FLOPs report using an unsupported GLM model type", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "XorlFlopsCounter has no xorl_glm5 estimator, so both sides of the former cp_size comparison were always zero and the report could not detect double-counted sequence lengths.", - "The rewritten report uses the supported qwen3_moe estimator, proves the baseline is nonzero, and then checks that changing cp_size does not alter global-sequence FLOPs." - ] - }, - { - "id": "TA-1487", - "scope": "standalone QARL activation NVFP4 value and STE report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report used the production internal quantizer as its numerical reference and repeated the two-dimensional STE already owned by the independent pure-PyTorch NVFP4 operator policy.", - "Its unique leading-dimension reshape contract now executes inside that independent policy, preserving value, shape, and gradient assertions while removing one pytest report." - ] - }, - { - "id": "TA-1488", - "scope": "standalone W4A4 MoE temporary backend-name report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report exercised only the temporary triton-to-triton_w4a4 name switch and restoration; it did not run a quantized down projection or grouped GEMM.", - "That admission and exception-restoration lifecycle now closes the existing CPU NVFP4 MoE conversion, eager execution, and injection policy, which passes with all former assertions." - ] - }, - { - "id": "TA-1489", - "scope": "duplicate Mooncake byte-store fake in distillation and utility suites", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The distillation file redefined the same in-memory put, get, existence, and removal object API already provided by tests._helpers.opd solely to record call keys.", - "The shared fake now records those calls for both consumers; transport and teacher-cache report boundaries remain unchanged and all four reports pass." - ] - }, - { - "id": "TA-1490", - "scope": "remaining checkpoint, experiment-planning, exporter, and OPD script reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Checkpoint process-group selection is independent of expert-mesh restore; experiment ingestion, path admission, calibration, and correctness-gated ranking enter different production branches.", - "Quantization primitives are independent of on-disk CLI layout, while OPD endpoint-version verification is independent of payload preparation and transport, so merging these reports would hide distinct failure boundaries." - ] - }, - { - "id": "TA-1491", - "scope": "standalone launcher CLI override parsing and removed-field report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report covered one schema-agnostic parse and one removed ZORL override rejection, both part of the existing server removed-configuration admission boundary.", - "Those exact assertions now run with YAML and direct override rejection in test_removed_configuration_boundary; launcher worker discovery and readiness remain a separate report." - ] - }, - { - "id": "TA-1492", - "scope": "standalone two-token ModelRunner causal-LM loss summation report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report invoked _compute_micro_batch_loss on a zero-logit toy model solely to prove that two equal token losses are summed rather than averaged.", - "The same raw-sum and per-token assertions now precede the retained ModelRunner loss-dispatch policy, which exercises DR-GRPO fields, options, temperature, and output controls through the same production method." - ] - }, - { - "id": "TA-1493", - "scope": "standalone QLoRA clean-interpreter import smoke report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report only checked that importing QLoRA utilities and expert modules returned zero, so it did not prove its stated model-package decoupling contract.", - "A strengthened clean-interpreter check now closes the expert capability and ownership policy by asserting that neither xorl.models nor any child module is loaded; the standalone report and file are removed." - ] - }, - { - "id": "TA-1494", - "scope": "optional Torch import inside the server protocol serialization policy", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The protocol report had already imported and exercised XoRL server modules that require Torch, yet used pytest.importorskip before its tensor round trip.", - "Torch is a core dependency and is now imported normally, so a broken installation fails closed instead of silently dropping the tensor serialization branch." - ] - }, - { - "id": "TA-1495", - "scope": "remaining server API, path security, worker protocol, and lifecycle report boundaries", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Outbound endpoint validation, artifact and diagnostic path confinement, and compile-worker function and message admission protect different trust boundaries.", - "API configuration validation, TensorData re-nesting, session publication, optimizer fallback, training response metrics, and ready-handshake queuing reach separate consumers and failure paths rather than input variants of one branch." - ] - }, - { - "id": "TA-1496", - "scope": "standalone DeepSeek-V4 AutoConfig and meta-builder registration report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The standalone report fabricated the same tiny standard HF snapshot shape used by the retained AutoModel from_pretrained loader report, then stopped after AutoConfig and meta construction.", - "AutoConfig class resolution, HF mapping, XoRL meta construction, actual AutoModel weight loading, and tensor equality now form one standard-snapshot loader policy; DCP conversion remains a separate report." - ] - }, - { - "id": "TA-1497", - "scope": "standalone Nemotron-H registry and local-config construction report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report proved registry lookup and Ultra-style config normalization but never constructed or executed the resulting model family.", - "Those loader assertions now open the retained Nemotron-H runtime, backward, router, and gradient-checkpointing policy; packed variable-length behavior and checkpoint parity remain independent reports." - ] - }, - { - "id": "TA-1498", - "scope": "remaining model-family config, registry, checkpoint, and runtime report boundaries", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Kimi-wrapped DeepSeek-V3 conversion covers nested text-config aliases absent from the base runtime model, while Qwen3.5 dense and MoE config normalization has no general model-construction owner to absorb it.", - "MiniMax M3 already owns config, registry, runtime, admission, checkpoint, and paging in one policy; Qwen2 and OLMo2 each combine HF construction, fused and unfused layouts, checkpoint round trips, and numerical HF parity." - ] - }, - { - "id": "TA-1499", - "scope": "standalone request-retry decorator report in data preparation", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "retry_on_request_exceptions has one production consumer: the high-level prepare_datasets operation, so its immediate success, transient HTTP failures, exhaustion, backoff, and unrelated-exception behavior are part of dataset preparation resilience.", - "Those assertions now close the existing dataset expansion, split, loader, merge, save, and reload lifecycle; the standalone utility report and file are removed." - ] - }, - { - "id": "TA-1500", - "scope": "stochastic-rounding unbiasedness stress loop", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The former CPU check accumulated 4,000 separately rounded 64 by 64 tensors, performing more than 16 million element updates to estimate a generic relative-error bound.", - "One seeded 65,536-element sample at exactly one quarter of a BF16 interval now checks the expected 25 percent round-up probability, sample mean, and legal neighbors directly and passes." - ] - }, - { - "id": "TA-1501", - "scope": "two-hundred-trial distributed stochastic-rounding bias tail", - "decision": "remove", - "status": "applied", - "evidence": [ - "After its four-rank reduce-scatter comparison, the report launched 200 additional all-to-all collectives solely to repeat the unbiased-expectation property owned by the CPU primitive policy.", - "The distributed report retains the distinct native-FP32 comparison and per-element BF16 transit error bound and passes on four GPUs in 18 seconds; FSDP2 integration remains separate." - ] - }, - { - "id": "TA-1502", - "scope": "remaining optimizer, trainer-utility, and data-preparation report boundaries", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Optimizer reports already aggregate construction, grouping, numerical updates, state strategy, cautious decay, and backend admission by optimizer family rather than by input size.", - "Gradient clipping, token and microbatch metadata, pipeline chunked CE, explicit gradient synchronization, timer fail-soft handling, live CUDA hooks, collator layouts, fingerprints, and packing each reach separate production consumers." - ] - }, - { - "id": "TA-1503", - "scope": "standalone single-GPU Qwen3-8B LoRA convergence job", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The job repeated the retained two-GPU checkpoint transaction's real model, LoRA rank and alpha, learning rate, 20-step horizon, and exact convergence threshold without selecting another production branch.", - "The surviving FSDP2 job adds checkpoint save and load, an explicit load marker, and final-step validation; model and server LoRA policies own non-FSDP construction, forward, backward, optimizer, and checkpoint behavior." - ] - }, - { - "id": "TA-1504", - "scope": "pseudo-E2E CUDA OPD free-tensor convergence loop", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report bypassed ModelRunner initialization, called a private loss helper directly, and optimized hidden-state and lm-head tensors as free parameters for eight steps, so decreasing loss did not represent a trainer or server lifecycle.", - "The runner policy already owns two-teacher cache loading, metrics, loss, and backward through that helper, while the real GPU OPD server report owns ModelRunner startup, forward_backward, and optim_step; the pseudo-E2E file is removed." - ] - }, - { - "id": "TA-1505", - "scope": "remaining end-to-end topology and integration report boundaries", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The FP8 jobs select dense checkpoint-resume, tensor parallel, Ulysses, Ring, plain MoE, and DeepEP expert-sharded paths; pipeline jobs separately reach direct trainer, schedule parity, server ModelRunner, FSDP, and folded PP-EP-CP topologies.", - "The retained OPD reports own request packing and Mooncake grouping, a real SGLang teacher, and the complete sampler-teacher-Mooncake-trainer-weight-sync loop; DistSignSGD, hybrid shared-LoRA MoE telemetry, and LoRA checkpoint resume exercise distinct production mechanisms." - ] - }, - { - "id": "TA-1506", - "scope": "Qwen3.5 trunk-wrap finite-output GPU report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report wrapped a tiny Qwen3.5 model and asserted only BF16 dtype and finite hidden states, after the same file had already proved the exact full-attention, linear-attention, dense-MLP, shared-expert, and exclusion inventory.", - "Generic trunk policies already prove bitwise forward and backward, batch invariance, BF16 admission, serving-lane equality, and two-rank FSDP2 composition, so the model-specific finite-output report selected no uncovered runtime behavior." - ] - }, - { - "id": "TA-1507", - "scope": "scale-only Quack DeepEP parity and checkpoint-training cases", - "decision": "remove", - "status": "applied", - "evidence": [ - "The 16K-token parity case repeated the same production geometry, balanced routing, unchunked no-permute path, reference comparison, and gradients already exercised at 4K tokens.", - "The 32K-token checkpoint case repeated the same checkpointed three-step optimizer path already exercised at 8K tokens; random routing, empty experts, explicit chunking, and checkpoint versus non-checkpoint training remain." - ] - }, - { - "id": "TA-1508", - "scope": "tautological and scale-only block-FP8 workload tails", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The generic codec allocated a 1024 by 2048 tensor only to recompute storage bytes from dtypes and element counts that the same report had already asserted; the result could not detect a codec defect.", - "The GKN codec's 4096-square roundtrip repeated the same multi-program two-dimensional kernel and error threshold covered by divisible and tail-tile shapes; removing both tails preserves geometry, accuracy, admission, edge, and determinism coverage." - ] - }, - { - "id": "TA-1509", - "scope": "remaining smoke-named and shape-named model, operator, and distributed reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "DeepSeek-V4 attention checks both window and compressed-KV forward and backward, every trainable gradient, FP8-QAT dispatch, sink dtype transfer, and TP rejection rather than shapes alone.", - "FP8 DeepEP uniquely composes no-permute transport with clamped-SwiGLU, native activation, expert biases, grouped FP8 backward, and all gradients; BF16 stochastic reduction separately proves custom-hook installation and FSDP2 gradient agreement." - ] - }, - { - "id": "TA-1510", - "scope": "standalone GatedDeltaNet FlashQLA environment-dispatch report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report monkeypatched the FlashQLA chunk function and asserted one call plus input and output shapes, making it a backend-selection branch rather than a numerical GDN policy.", - "Its assertions now close the CPU FlashQLA selection and exact-contract precedence policy; real CUDA numerical, state-chaining, and batch-invariance gates remain separate." - ] - }, - { - "id": "TA-1511", - "scope": "standalone runtime-rank MoE LoRA scaling report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both reports exercised WeightSyncHandler's inference-buffer construction, while the standalone file stopped at a hard-coded active-rank delta and a three-name buffer.", - "The active-rank scaling, emitted names, values, dtypes, shapes, and source cleanup now close the broader QLoRA merge and FP8 sync lifecycle; the standalone file is removed." - ] - }, - { - "id": "TA-1512", - "scope": "standalone families-v2 selector report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "GLM exact selection and nonexact environment rollback are properties of model-builder structure, while Qwen's v1 pin is applied by its exact-model hook; the selector-only report called private setters without either production consumer.", - "Both legacy environment aliases now close the trainer model-selection policy, and the Qwen hook proves its v1 pin overrides the legacy v2 request; CUDA norm reachability and numerical-tree reports remain distinct." - ] - }, - { - "id": "TA-1513", - "scope": "remaining small BI, DeepEP, DSV4, checkpoint, loss, and server reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The remaining small files protect distinct behavior: DeepEP async-combine safety and internode preflight are different failure modes, DSV4 rotation fallback owns optional-kernel-free orthonormality, and MTP checkpoint remapping has no broader GLM4 loader owner.", - "Token-loss composition, gradient-accumulation group routing, batch-slice topology mapping, OPD layer-cache slicing, KKT contract geometry, and families-v2 norm reachability each exercise a separate production consumer or dispatch boundary rather than construction metadata alone." - ] - }, - { - "id": "TA-1514", - "scope": "family-specific ModelRunner LoRA target-resolution reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The GLM and Kimi files rebuilt large model configurations even though ModelRunner reads only the top-level model_type, and both repeated the same explicit-target precedence branch.", - "One compact cross-family policy now retains GLM defaults, Kimi defaults including lm_head, explicit targets, and manifest targets; one duplicate branch and more than one hundred lines of irrelevant fixture data are removed." - ] - }, - { - "id": "TA-1515", - "scope": "standalone distributed-checkpointer process-group selector report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The selector report called private helpers with synthetic distributed namespaces, while the retained distributed-checkpointer I/O policy already owns load and save process-group routing.", - "NCCL-to-Gloo selection, one-time caching, native-Gloo reuse, non-pipeline omission, and pipeline custom/default metadata groups now close that I/O policy; the standalone file is removed." - ] - }, - { - "id": "TA-1516", - "scope": "standalone server batch-slice rank helper report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The dispatcher policy already exercises distinct EP slices, CP sharing, EP-FSDP coordinates, padding, routing side payloads, and the rollback switch through real batch selection.", - "The remaining replicated-DP and direct helper mappings now close that dispatcher policy, with environment restoration verified by the subsequent cases; the standalone file is removed." - ] - }, - { - "id": "TA-1517", - "scope": "remaining mock-heavy checkpoint, rendezvous, protocol, and model reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "NCCL rendezvous fault injection protects port rotation and bind-before-inference ordering; checkpoint restore protects zero-meta admission and base-before-adapter initialization; protocol round trips reject pickle and preserve tensors.", - "Gradient-checkpoint gating, native-EP combine, LoRA target manifests, sparse-delta artifacts, and exact factor publication each retain numerical, persistence, security, or production-lifecycle outcomes beyond captured argument plumbing." - ] - }, - { - "id": "TA-1518", - "scope": "test-only kernel-variant comparison API and assertion tail", - "decision": "remove", - "status": "applied", - "evidence": [ - "compare_kernel_variants had no runtime, CLI, documentation, or example consumer; its only caller computed a speedup from two literals after the production ranker had already ordered the same rows.", - "The retained rank policy still proves that a faster unvalidated candidate cannot displace the validated winner, which is the simulator's actual promotion contract." - ] - }, - { - "id": "TA-1519", - "scope": "test-only NVFP4 export dequantizer and duplicate numerical assertions", - "decision": "remove", - "status": "applied", - "evidence": [ - "dequantize_nvfp4_export was imported only by the exporter test, so production shipped an inverse implementation solely to grade its own quantizer twice.", - "The independent fake-quant policy retains exact numerical-reference coverage; the exporter policy retains packed bytes, scale shapes and dtypes, fused shared scales, BF16 islands, activation scales, directory metadata, and requantization rejection." - ] - }, - { - "id": "TA-1520", - "scope": "orphaned simulator reference_counter_total_flops helper", - "decision": "remove", - "status": "applied", - "evidence": [ - "The helper described itself as test support but had no callers anywhere in source, tests, documentation, examples, or scripts after simulator-policy consolidation.", - "Removing the 42-line adapter also removes its sole SimpleNamespace dependency without changing the analytical ledger or trainer FLOPs counter." - ] - }, - { - "id": "TA-1521", - "scope": "remaining low-reference cache, profiling, export, and weight-sync APIs", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Sparse-delta baseline reset is an explicit operational hook, the FP8 error recorder is called by FP8Linear, and dense-buffer rank filtering is invoked dynamically by WeightSyncHandler.", - "Teacher-store preparation and QARL export are public package or CLI surfaces; fused-expert cache invalidation and Mooncake hidden-store methods own runtime state rather than serving only as assertion oracles." - ] - }, - { - "id": "TA-1522", - "scope": "test-only GLM exact LM-head group binding and factor-view APIs", - "decision": "remove", - "status": "applied", - "evidence": [ - "Production supplies the TP group to the constructor and converts FP32 factor masters inside both real autograd functions; neither public-looking helper had a runtime caller.", - "The test now constructs the real object with its group, while the retained custom-boundary policy directly proves the actual autograd function saves the BF16 factor bytes." - ] - }, - { - "id": "TA-1523", - "scope": "GLM routed-expert trace-only hook path and public factor-buffer wrapper", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "sampler_value_trace installed clone-heavy wrapper hooks and returned a test-specific dataclass through a branch production always disabled; physical_factor_buffers only converted masters before calling the real internal builder.", - "The trace machinery and wrapper are removed; the GPU policy now compares real production forwards with zero and live LoRA factors and proves routing-scale linearity through the actual module path." - ] - }, - { - "id": "TA-1524", - "scope": "test-only GLM shared-expert physical-factor convenience view", - "decision": "remove", - "status": "applied", - "evidence": [ - "The public wrapper had no production consumer and only duplicated the BF16 conversion performed before the runtime-owned _physical_factor_views_from_effective builder.", - "CPU SGLang slice parity and the official CUDA operand policy now exercise that same builder directly, preserving physical layout and byte assertions." - ] - }, - { - "id": "TA-1525", - "scope": "test-only IndexShare context-manager convenience path", - "decision": "remove", - "status": "applied", - "evidence": [ - "The model uses begin plus finish_forward in its own try/finally and never calls the context-manager wrapper; only the unit helper exercised that alternate lifecycle.", - "The retained policy now drives begin and finish_forward exactly as production does for failed and successful forward-only invocations." - ] - }, - { - "id": "TA-1526", - "scope": "remaining source symbols with test-dominant lexical references", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Runner and RequestProcessor handlers are reached through command dispatch, API endpoint functions are framework-registered, and dense rank filtering is selected dynamically by WeightSyncHandler.", - "DataLoader collator mutation is a documented extension surface, while adapter transactions, exact-factor ownership, and cache invalidation participate in runtime lifecycle contracts despite sparse direct-name references." - ] - }, - { - "id": "TA-1527", - "scope": "generic, dense-Qwen3.5, and MoE-Qwen3.5 RMSNorm policies", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The generic fused policy owns kernel forward, backward, bitwise, packed-shape, and trunk behavior, while the family policy owns explicit dispatch admission and fail-closed structure.", - "Dense and MoE Qwen3.5 use separate module implementations and call sites; each policy executes its own construction and dispatch rather than repeating inputs against one owner." - ] - }, - { - "id": "TA-1528", - "scope": "remaining shape, dtype, finiteness, registry, and configuration reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The apparent weak-output reports also contain numerical reference, gradient, cache, checkpoint, or cross-backend assertions that simple assert-shape classification missed.", - "Compressor capacity, GKN format, ragged batching, RoPE bytes, and model-family construction each protect a distinct runtime consumer." - ] - }, - { - "id": "TA-1529", - "scope": "split OLMo-2 QK-norm and full tensor-parallel subprocess reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both reports initialized the same two-rank Gloo CPU tensor-parallel mesh and exercised LocalAxisRMSNormShard with Olmo2QKRMSNorm.", - "The surviving full-model subprocess now begins with plain and sharded numerical RMSNorm oracles, then applies the production TP plan and proves forward, lm-head, and all-parameter backward execution." - ] - }, - { - "id": "TA-1530", - "scope": "standalone FlashAttention diagnostic decode causal-flag report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The two boolean branches exist only to support diagnostic decode-cache behavior and do not describe an independent product policy.", - "Both causal-flag assertions now open the retained Qwen3-MoE natural and routing-replay cached-forward parity lifecycle." - ] - }, - { - "id": "TA-1531", - "scope": "native-FP8 materialization exact Torch-wheel and ambient-future assertions", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The dependency lock already pins Torch 2.12.1+cu132, so repeating the wheel string inside a behavior test made compatible environment changes fail before materialization ran.", - "The worker now explicitly disables swap_module_params_on_conversion to select the replacement path that exposed the regression, then proves plain and two-rank FSDP2 frozen-state behavior." - ] - }, - { - "id": "TA-1532", - "scope": "split MoE block, decoder, and full-model torch.compile reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both reports shared the same CUDA capability gate and compiler-compatibility owner; the full-model report only continued the lower-level block and decoder sequence.", - "The surviving policy still runs every available MoE backend through AOT eager and Inductor at lower levels, then runs native and eager compiled layers through full-model forward and backward." - ] - }, - { - "id": "TA-1533", - "scope": "remaining multi-report numerical and lifecycle files", - "decision": "keep", - "status": "accepted", - "evidence": [ - "CPU and CUDA RMSNorm, Z-loss, non-gated MoE, and native-FP8 reports require separate capability outcomes so CPU coverage is not hidden behind a GPU skip.", - "FP64 gradcheck, forward and backward sparse-MLA kernels, routing wire decode versus context layout, and enabled versus disabled packing own distinct numerical or runtime failure meanings." - ] - }, - { - "id": "TA-1534", - "scope": "unscheduled sparse-delta trainer-to-SGLang external integration report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report required both an unpinned delta-encoding source tree and an SGLang sparse-delta receiver absent from the pinned submodule; no workflow, script, or configuration supplies either test path.", - "A catch-all converted every missing or broken external implementation into a skip, so the 522-line fake trainer and orchestrator never established repository coverage; retained policies own XORL artifact encoding, source capture, hashing, posting, and transport lifecycles." - ] - }, - { - "id": "TA-1535", - "scope": "standalone TokenPartial component report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Caller-scaled denominator, microbatch additivity, raw-sum, sequence-mean-token-sum, and empty-mask behavior are branches of the shared loss reducer contract rather than an independent product policy.", - "Every direct reducer assertion now closes the retained policy and importance-sampling loss identity report, eliminating one file and pytest identity without losing an oracle." - ] - }, - { - "id": "TA-1536", - "scope": "remaining pinned MoE and TileLang availability exception shields", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The MoE compiler helper broadly suppressed failures importing a shipped availability utility and wrapped an infallible Quack list append; the sparse-MLA reports treated hard-pinned TileLang as optional.", - "Explicit CUDA and Hopper admission remains, while internal dependency failures now fail closed; the real MoE compile matrix and both TileLang sparse-MLA policies pass." - ] - }, - { - "id": "TA-1537", - "scope": "NVSHMEM library-path catch-all helpers", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "All three callers already use pytest.importorskip for nvidia.nvshmem before launching their DeepEP workers, so swallowing every subsequent path-resolution error only hid malformed installations.", - "The helpers now resolve the admitted package directly; all affected reports collect and the installed package path is importable." - ] - }, - { - "id": "TA-1538", - "scope": "standalone runtime FLOPs context-parallel denominator report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The runtime counter and simulator analytical ledger share the same global-sequence-length numerator contract; context parallelism must change placement rather than multiply global work.", - "The exact cp1-versus-cp64 assertion now closes the retained topology, shape, and analytical-ledger policy instead of occupying a one-test utility file." - ] - }, - { - "id": "TA-1539", - "scope": "standalone gradient-accumulation loss and HSDP deferral reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Loss-group forwarding, local/global token normalization, and backward scaling are the consumer half of the trainer metadata-counting policy.", - "HSDP microbatch deferral and restoration are branches of the explicit SP and LM-head gradient-synchronization policy; all assertions remain in those two owner reports." - ] - }, - { - "id": "TA-1540", - "scope": "standalone private OPD layer-cache fetcher report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Index filtering, streamed layer slices, and layer-count reporting exist only to feed the retained OPD microbatch loss lifecycle.", - "The exact requested indices, slice ranges, and output shapes now execute before the real streaming loss, gradient, cache, metric, and debug-artifact assertions." - ] - }, - { - "id": "TA-1541", - "scope": "standalone ModelRunner LoRA target-resolution and kill-session reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Implicit family targets, explicit lists, and manifest precedence now close the runner's adapter ownership compiler policy that consumes the selected targets.", - "Nonresident checkpoint promotion, failed-kill preservation, registry cleanup, and path rejection now close the existing optimizer, checkpoint-load, and session-registry lifecycle." - ] - }, - { - "id": "TA-1542", - "scope": "standalone sync-quantization dictionary normalization report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "BF16 no-ops, valid FP8 normalization, module exclusions, and malformed or unsupported forms are admission branches of receiver quantization detection and enrichment.", - "All direct normalization assertions now run in the API policy that detects receiver configuration, propagates unsupported markers, enriches user input, and persists the default." - ] - }, - { - "id": "TA-1543", - "scope": "standalone IndexShare trainer and server caller-cleanup report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Offline and server failure cleanup are the public-caller completion branches of the same IndexShare lifecycle that owns forward-only, backward-retained, checkpoint-recompute, and idempotent close behavior.", - "Both real caller wrappers and their exact release counts now close the retained checkpointed lifecycle policy." - ] - }, - { - "id": "TA-1544", - "scope": "standalone rank-zero ready-handshake report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "ACK handling, request-before-ACK queueing, client identity, malformed frames, and message preservation are runtime branches of the orchestrator-runner wire protocol.", - "The async handshake now follows serialization, tensor roundtrip, command creation, and pickle rejection in one protocol owner." - ] - }, - { - "id": "TA-1545", - "scope": "standalone cross-rank packed-padding synchronization report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Tensor padding, ignored labels, attention masks, cumulative sequence boundaries, and max-length updates are the distributed completion of server-versus-CLI packed sequence metadata alignment.", - "The exact 176-to-512 cross-rank case now runs alongside local padding, SP sharding, stale-metadata replacement, LCM admission, and unpacking behavior." - ] - }, - { - "id": "TA-1546", - "scope": "standalone lm-head TP plus EP mesh-membership subprocess", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The mesh-only report stopped after constructing four-rank TP, replica, and EP groups, while the retained lm-head FSDP policy already used the same DP2 by CP2 topology for real loss and gradients.", - "That full transaction now enables EP2, asserts exact group membership, and continues through parameter sync, vocab-sharded loss, global loss parity, weight gradients, and hidden gradients." - ] - }, - { - "id": "TA-1547", - "scope": "server runner reports that execute source-loaded production module copies", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Seven reports loaded ModelRunner, RunnerDispatcher, AdapterCoordinator, CheckpointManager, or LoRAAdapterManager from file paths under synthetic module names even though the canonical package modules import successfully.", - "The reports now patch and exercise the actual runtime module objects, preserving all 11 adapter, checkpoint, optimizer, session, and dispatcher assertions while removing duplicate module state." - ] - }, - { - "id": "TA-1548", - "scope": "source-loaded launcher copy and fake dependency graph in server-argument policy", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The report imported the canonical launcher for override parsing but re-executed the same source file with fake API-server, orchestrator, session, QARL, and packing modules solely to obtain load_server_arguments.", - "All four server-argument reports now call the canonical launcher and real dependency graph, including shipped-config subprocess parsing and sparse-MLA propagation." - ] - }, - { - "id": "TA-1549", - "scope": "smallest remaining operator, model, and distributed report boundaries", - "decision": "keep", - "status": "accepted", - "evidence": [ - "DSV4 fallback rotation, GLM4 MTP checkpoint remapping, DeepEP async-combine admission, KKT launch geometry, families-v2 dispatch, BI mean, and Class-B RoPE each reach a different production boundary rather than repeating size-only examples.", - "These reports retain numerical or fail-closed behavior not owned by their adjacent kernel, model, or topology policies; line count alone is not a deletion signal." - ] - }, - { - "id": "TA-1550", - "scope": "private ring-attention zigzag helper report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report mixed assertions on the private _get_zigzag_step_section implementation with the public packed-sequence reorder consumed by TextSequenceShardCollator.", - "The public single-document, packed-document, multi-rank, identity, and invalid-length behavior now closes the collator policy, while the private section-helper assertions are removed." - ] - }, - { - "id": "TA-1551", - "scope": "optional FA3 and SGLang exact-runtime reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "FA3 ring-attention merging and shared-prefix attention still exercise production numerical kernels, while the GLM exact SGLang joins cross real checkpoint export, adapter parsing, and memory-pool boundaries.", - "Their environment contract is intentionally isolated: the default XoRL profile remains Torch 2.12.1 without sglang-kernel, and pinned SGLang declares Torch 2.11.0 with sglang-kernel 0.4.5. Lazy wrapper imports are not ABI validation." - ] - }, - { - "id": "TA-1552", - "scope": "source-loaded EP routing-score report and synthetic backend registration graph", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The routing-score report rebuilt Triton and Quack from source under test-only names with fake kernel packages, while the EP adapter policy already owned expert_scores forwarding and backend registration.", - "Its complete forward and routing-score gradient oracle now closes the canonical EP adapter policy; the routing-position policy shares only explicit CPU kernel doubles, and the standalone report is removed." - ] - }, - { - "id": "TA-1553", - "scope": "source-loaded Quack process, compiler, and cache module copies", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The Quack safety policy executed worker protocol, ptxas, and cache files under synthetic module names with fake cutlass and tvm_ffi modules even though their canonical package imports succeed.", - "Timeout, truncation, temporary-output, entry-point, and safe cache-key assertions now run against the actual runtime module objects." - ] - }, - { - "id": "TA-1554", - "scope": "synthetic Torch FSDP module in HSDP gradient-sync policy", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The policy replaced torch.distributed._composable.fsdp.fully_shard in sys.modules solely to make a fake object pass the FSDPModule type gate.", - "It now uses a lightweight instance of Torch 2.12's real FSDPModule API type and preserves the complete deferral, last-microbatch, restoration, and replicate-size assertions." - ] - }, - { - "id": "TA-1555", - "scope": "fake FA4 package graph and runtime module reloads in attention registry policy", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The default profile already installs FA4 without flash_attn_interface, but the policy replaced flash_attn modules and reloaded both production registry modules to recreate that state.", - "Registry admission and resolution now run against the installed canonical FA3 or FA4 availability and retain the eager fallback and unavailable-flash failure boundaries." - ] - }, - { - "id": "TA-1556", - "scope": "split OPD driver version-verification and payload-transport reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Student endpoint matching and weight-version success, mismatch, and failure branches are setup and publication checks of the same OPD pipeline that owns worker preparation, causal payloads, and Mooncake metadata transport.", - "All assertions now execute in one driver lifecycle, and the standalone executable script is loaded once instead of re-executed for every helper." - ] - }, - { - "id": "TA-1557", - "scope": "remaining synthetic optional-dependency modules", - "decision": "keep", - "status": "accepted", - "evidence": [ - "delta_encoding is absent from the default environment, Mooncake cannot load without its CUDA runtime, and the default Torch 2.12 lane intentionally lacks the pinned SGLang and sgl_kernel runtime.", - "The retained doubles exercise explicit serialization, fallback, runtime-context, and slot-combine contracts only; they are not used as evidence that the compiled dependency ABI or real kernel operation works." - ] - }, - { - "id": "TA-1558", - "scope": "conditional skips for shipped EP and non-gated MoE backends", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The default profile ships Triton, native Torch, and Quack support, but the EP adapter report could skip midway and the non-gated report tested whichever subset happened to register.", - "Both policies now fail closed if a shipped backend disappears and execute every expected forwarding, rejection, forward, and gradient branch." - ] - }, - { - "id": "TA-1559", - "scope": "duplicate three-GPU skip inside OPD full-pipeline report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The module-level marker already skips unless three CUDA devices are visible before fixtures or the test body run.", - "The second in-body device-count branch repeated the same admission rule after model artifacts had already been created." - ] - }, - { - "id": "TA-1560", - "scope": "remaining conditional runtime reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The remaining reports require Hopper-only FlashQLA or exact GLM kernels, two-rank FSDP, CUDA profiler events, a real SGLang and sgl_kernel lane, or optional DeepGEMM execution.", - "Each gate protects a numerical, distributed, or compiled-runtime transaction that has no ordinary CPU branch hidden behind the skip." - ] - }, - { - "id": "TA-1561", - "scope": "tiny dense QARL AdamW training smoke", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report only proved that one ordinary AdamW step changes a tiny model's weights and logprobs, behavior owned by PyTorch rather than QARL.", - "QARL injection and summaries remain in the fake-quant policy, while forward state and exact state-dict restoration remain in the calibration lifecycle." - ] - }, - { - "id": "TA-1562", - "scope": "standalone QARL activation-quant override report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The override changes the same QARLLinear and QARLMoEExperts fake-quant state already constructed by the dense fake-quant owner.", - "Enable, disable, per-module restoration, exception safety, non-QARL exclusion, and nested restoration now close that owner lifecycle without a separate report." - ] - }, - { - "id": "TA-1563", - "scope": "standalone Qwen3.5 families-v2 backward candidate report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report replaced both numerical kernels with CPU references and tested only the zero-centered autograd wrapper wiring.", - "That wiring now closes the Qwen3.5 norm dispatch and site-assignment owner, preserving effective-weight and dual residual-gradient parity." - ] - }, - { - "id": "TA-1564", - "scope": "standalone LoRA mixed-precision model-builder report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report patched the same build_training_model construction boundary as the existing FP8 and QARL model-builder lifecycle.", - "Base BF16 retention, FP32 adapter factors, trainability, dtype resolution, and generic-upcast admission now execute in that owner policy." - ] - }, - { - "id": "TA-1565", - "scope": "test-only canonical-MoE sampler plan and serialization surface", - "decision": "remove", - "status": "applied", - "evidence": [ - "ParallelRole.SAMPLER, glm52_sampler, launcher_tp_size, logical_ordinal, as_dict, and digest had no production, documentation, example, or launcher consumer; only the canonical-MoE test called them.", - "The retained trainer plan still validates exact group membership, logical ordinals, topology rejection, and the real distributed collective contract." - ] - }, - { - "id": "TA-1566", - "scope": "test-only one-call adapter gradient capture convenience", - "decision": "remove", - "status": "applied", - "evidence": [ - "Production ModelRunner owns the two-phase stage_gradient_numerators and commit_gradient_capture transaction; capture_gradient_numerators had only test callers.", - "Tests now drive the production boundary directly, including explicit abort behavior and real multi-rank fatal paths, with setup-only repetition isolated in test helpers." - ] - }, - { - "id": "TA-1567", - "scope": "standalone DeepEP async-combine opt-in report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The unsafe async-combine environment gate is part of the same DeepEP transport admission surface as internode topology, preflight, and buffer sizing.", - "Both default-safe and explicit-opt-in branches now close the existing DeepEP admission policy without a separate report." - ] - }, - { - "id": "TA-1568", - "scope": "standalone GLM52 exact-MoE modeling forwarding report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report constructed exact routed and shared experts solely to inspect arguments forwarded by Glm5MoEBlock.", - "Routed IDs, scaling, and shared contributor-ordinal forwarding now close the exact-MoE construction and inventory owner that already validates those concrete modules." - ] - }, - { - "id": "TA-1569", - "scope": "standalone Kimi tokenizer loader report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Kimi's local tokenizer selection is part of the same model-family loading boundary as its wrapper config and registry resolution.", - "Local TikToken decoding and generic tokenizer/processor fallback behavior now close the Kimi/DeepSeek registry owner without a separate report." - ] - }, - { - "id": "TA-1570", - "scope": "standalone SignSGD builder and step report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Its no-decay parameter grouping repeated the generic builder contract already exercised by DistSignSGD, while the cautious optimizer policy already constructed and stepped SignSGD.", - "Dense sign updates, decoupled decay, missing-gradient behavior, and sparse-gradient rejection now close the retained optimizer policy." - ] - }, - { - "id": "TA-1571", - "scope": "isolated stochastic-rounding report and four-GPU BF16 all-to-all gate", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "stochastic_round_to_bf16 has one production consumer: BF16StochasticAllToAllReduceScatter, so two independent reports overstated confidence in one transaction.", - "One default-runtime policy now proves deterministic admission, neighbor distribution, unbiased expectation, and the real two-rank Gloo all-to-all with FP32 accumulation against reduce-scatter." - ] - }, - { - "id": "TA-1572", - "scope": "standalone Qwen3 projection-unfusing structure report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report carried a distributed marker but initialized no process group and asserted only module replacement shapes.", - "The production model-level unfuse path, TP plan, checkpoint-handler transition, and every layer's projection inventory now close the torch_parallelize policy owner." - ] - }, - { - "id": "TA-1573", - "scope": "duplicate balanced synthetic-routing fragment in train-router report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The fragment checked only that cached expert IDs survived regather and weights became uniform under the balanced environment switch.", - "The retained TopKRouter owner already proves the exact cyclic expert sequence, balance bound, uniform weights, and override of softmax, hash-table, and bias inputs." - ] - }, - { - "id": "TA-1574", - "scope": "standalone NVFP4 QARL linear report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report constructed the same QARLLinear and normalized the same recipe surface as the dense fake-quant owner, differing only by format.", - "NVFP4 group admission, weight-only forward, straight-through gradient, and disabled-weight behavior now close the format-spanning QARL policy." - ] - }, - { - "id": "TA-1575", - "scope": "standalone MoE sqrtsoftplus regather report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Routing replay regather is a continuation of TopKRouter's softmax, sqrtsoftplus, selected-expert, dtype, and scaling contract.", - "Both regather branches now close the router owner that already proves selection bias, hash routing, normalization, and configured FP32 execution." - ] - }, - { - "id": "TA-1576", - "scope": "standalone families-v2 norm reachability report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report invoked the same v2 RMSNorm tree and payload as the numerical, realization, and dispatch policy, then inspected only trainer entry-point reachability.", - "Trainer dispatch and the v1 kill switch now close the v2 norm owner under isolated monkeypatch contexts." - ] - }, - { - "id": "TA-1577", - "scope": "standalone families-v2 LM-head report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Projection, selected-logprob, batch-invariance, backward, and kill-switch assertions target the same final-token probability transaction as the bi_fused LM-head policy.", - "The v2 realization now closes that owner alongside eager parity, temperature, determinism, guards, and probability-boundary behavior." - ] - }, - { - "id": "TA-1578", - "scope": "standalone batch-invariant mean interpose regression", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The regression exercises set_batch_invariant_mode's global Torch interpose, already owned by the trunk-linear forward and gradient-admission policy.", - "Full, typed, one-dimensional, keepdim, and multi-dimensional reductions now close that interpose owner while preserving the former sum-versus-mean bug oracle." - ] - }, - { - "id": "TA-1579", - "scope": "standalone DSV4 rotate-activation fallback report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "rotate_activation is consumed by the DSV4 compressor and indexer, and the standalone report only disabled the optional fast transform to inspect its fallback.", - "Fallback basis mapping, involution, and norm preservation now close the compressor's context-parallel shape and admission policy." - ] - }, - { - "id": "TA-1580", - "scope": "standalone Qwen3-MoE RMSNorm family declaration report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Dense Qwen and shared-attention family declarations already belong to the generic RMSNorm family contract.", - "Qwen3-MoE layer-zero, residual-tree, explicit call-site, and bare final-norm behavior now close that same structural owner." - ] - }, - { - "id": "TA-1581", - "scope": "standalone LoRA permanent cast-once merge report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report rebuilt the same LoraLinear, MoEExpertsLoRA, factor deltas, and canonical FP32 fold owned by the merged-forward contract.", - "Zero-adapter preservation and permanent BF16/FP16 cast-once merge behavior now close the canonical fold owner on CPU instead of requiring CUDA unconditionally." - ] - }, - { - "id": "TA-1582", - "scope": "standalone dataset hash and split-fingerprint reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both helpers exist to key the dataset split, loading, saving, and preparation lifecycle already exercised by the shared-data owner.", - "Determinism, split separation, input sensitivity, fractional sizes, multi-dataset order independence, and tokenizer/column sensitivity now close that lifecycle." - ] - }, - { - "id": "TA-1583", - "scope": "standalone virtual-stage MultiOptimizer report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report's end-to-end assertion was that build_lr_scheduler wraps every child optimizer and decays all parameter groups after delegated steps.", - "Multi-part construction, delegation, model mapping, single-part fallback, invalid explicit groups, and scheduler fanout now close the scheduler owner; checkpoint state filtering remains separately covered." - ] - }, - { - "id": "TA-1584", - "scope": "standalone QARL-to-FP8 weight-sync report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report derived an FP8 sync configuration from QARL, invoked WeightSyncHandler's quantizer, and exercised its request handler.", - "Folded-module metadata, skip-list behavior, derived request configuration, quantized buffers, and incompatible overrides now close the FP8 weight-sync owner." - ] - }, - { - "id": "TA-1585", - "scope": "synthetic identity-layer gradient-checkpoint truth table", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report attached MagicMock checkpoint functions to an identity layer and exhaustively restated the outer Python gate's training, flag, and method condition.", - "Real Nemotron-H training already proves default full-layer checkpoint execution and gradients, while the GLM-5 lifecycle proves recompute-before-dispatch bypasses the outer checkpoint and invokes the layer checkpoint." - ] - }, - { - "id": "TA-1586", - "scope": "standalone generic train-router dispatch report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Trainable and frozen router-gradient behavior is already exercised by routing-replay and real model lifecycles; the standalone tiny MoE repeated those backward assertions.", - "The unique DeepEP rejection now closes TopKRouter's MoEBlock configuration policy, and the frozen server default now closes server configuration serialization." - ] - }, - { - "id": "TA-1587", - "scope": "standalone mocked GDN KKT launch-geometry report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report replaced both KKT kernels with launch recorders solely to inspect the same gdn_contract switch owned by the GDN exact-contract policy.", - "Pinned BK, warp, stage, safety, and autotuned-off-lane behavior now close the existing GDN serving-geometry owner beside its solve-tril geometry." - ] - }, - { - "id": "TA-1588", - "scope": "standalone DSV4 attention shape smoke report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report directly constructed and manually initialized isolated C0 and C128 attention layers, then repeated forward shape, finiteness, and backward reachability already covered by the fully initialized DSV4 model lifecycle.", - "The unique FP8-QAT dispatch and TP rejection now run through the full model owner; direct-component dtype behavior that production model casting deliberately overrides was removed." - ] - }, - { - "id": "TA-1589", - "scope": "standalone synthetic LoRA target-manifest report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report built a fake two-layer attention tree even though the fused GDN lifecycle already consumes strict target manifests for real fused projection paths.", - "Count, rank, configured-target, unlisted-module, Boolean, schema, and integer validation now fail closed against the real fused-GDN manifest owner." - ] - }, - { - "id": "TA-1590", - "scope": "standalone GLM4 MTP tail-remap fragment", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report instantiated only Glm4MoeCheckpointHandler and checked three MTP tail aliases plus three ignored tail fields.", - "Those aliases and exclusions now close the GLM4 model-family construction and checkpoint lifecycle that already creates both ordinary and prequantized handlers." - ] - }, - { - "id": "TA-1591", - "scope": "remaining orchestrator and API communication reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The small reports exercise different live boundaries: runner rank-zero readiness and tensor-safe serialization, API-engine ZMQ request handling, and APIServer response metrics.", - "Their mocks isolate external processes but do not duplicate the protocol transaction or its failure modes, so combining them would obscure ownership rather than remove repeated behavior." - ] - }, - { - "id": "TA-1592", - "scope": "standalone importance-sampling and policy-loss microbatch reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both one-test files rebuilt the same masked tensors, TokenPartial denominator, full-batch call, microbatch calls, and summable-metric loop already parameterized by the shared loss-contract owner.", - "The shared owner now checks legacy identity and microbatch composition together for basic, KL, TIS, and IcePop modes; both copied reports are removed." - ] - }, - { - "id": "TA-1593", - "scope": "standalone dense-Qwen eager RoPE parity report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The report's remaining assertion compared RotaryEmbedding table consumption and Q/K application with an inlined serving arithmetic reference.", - "That exact CUDA oracle now closes the existing RoPE frequency-table, lazy-cache, serving-device, and zero-K3 lifecycle; the standalone owner is removed without dropping its bitwise checks." - ] - }, - { - "id": "TA-1594", - "scope": "deprecated adapter optimizer broadcast no-op and negative spies", - "decision": "remove", - "status": "applied", - "evidence": [ - "AdapterCoordinator.broadcast_adapter_optimizer_state had no production caller and deliberately performed no optimizer transfer; topology-specific optimizer state is restored through the all-ranks checkpoint path.", - "The method is removed together with seven lifecycle spies whose only assertion was that the dead no-op stayed uncalled; the real transactional optimizer-restore rejection remains covered." - ] - }, - { - "id": "TA-1595", - "scope": "standalone trainer-P2P and PP-NCCL handler reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Both files directly bound or called private WeightSyncHandler helpers for HCA selection, abort markers, peer status gathering, and PP named-tensor flatten/reconstruction.", - "Those policies now close the existing handler configuration, sender-selection, bucketing, and inference-layout owner; both one-test modules are removed while their complete cases remain." - ] - }, - { - "id": "TA-1596", - "scope": "deprecated ep_outside and moe_checkpoint_method configuration aliases", - "decision": "remove", - "status": "applied", - "evidence": [ - "Neither alias appears in shipped configurations or current documentation; ep_intranode and gradient_checkpointing_method are the documented native fields.", - "The aliases, parser remapping branches, duplicated simulator dimension, and compatibility-only parser inputs are removed. The active gradient_checkpointing_method='moe_act' execution mode remains available." - ] - }, - { - "id": "TA-1597", - "scope": "standalone GLM52 exact-attention construction lifecycle", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Its rank-alpha, dense-component, sparse-MLA, and all-to-all admission failures duplicated the complete exact-MoE constructor's cases.", - "The unique 780 attention-factor names, projection classes, source FQNs, per-layer trainable sets, and three dense roots now close the complete 1,700-factor constructor; the standalone report is removed." - ] - }, - { - "id": "TA-1598", - "scope": "synthetic prequantized GNK-to-GKN transpose report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report never invoked prequantized loading; it quantized tensors in both orientations and re-derived transpose equivariance for block-FP8 and NVFP4.", - "The retained QLoRA expert-loader owner exercises the real GNK-to-GKN byte and scale transforms for both formats and multiple shapes, while the codec owners cover numerical roundtrips." - ] - }, - { - "id": "TA-1599", - "scope": "standalone synthetic GKN format and backend report", - "decision": "remove", - "status": "applied", - "evidence": [ - "Its checkpoint half rebuilt ExpertWeightBuffer behavior already exercised through the DeepSeek-V3 checkpoint handler's exact load/save layout contract.", - "Its eager/native comparisons duplicate the dedicated forward/backward backend owner, while grouped-GEMM primitives and combined Quack/SGLang parity own the Triton path; the report's optional imports could silently bypass that branch." - ] - }, - { - "id": "TA-1600", - "scope": "parallel dense and MoE Qwen3.5 RMSNorm reports", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The two files independently enumerated the same copied zero-centered dispatch matrix, v2 admission policy, norm-site propagation, and residual-family selection.", - "One family-wide owner now drives both production classes through the shared CPU contract, retains the dense-only GDN and v2 backward boundaries, and uses one GPU model lifecycle for their shared normalization kernels." - ] - }, - { - "id": "TA-1601", - "scope": "SGLang compiled-kernel dependency and smoke contract", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The default Torch-2.12 profile intentionally excludes sglang-kernel, while pinned SGLang requires Torch 2.11.0 and sglang-kernel 0.4.5 in an isolated environment.", - "The retained smoke fails when the exact lane lacks the package, eagerly loads the compiled extension and wrapper symbols, and executes a real RMSNorm operation; it passes in the new .venv-sglang and skips in the default profile." - ] - }, - { - "id": "TA-1602", - "scope": "legacy SGLang fused-expert cache environment alias", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_MOE_SGLANG_FUSED_EXPERTS_CACHE_WEIGHTS appeared only in implementation and tests; no shipped configuration or documentation used it.", - "XORL_MOE_SGLANG_FUSED_EXPERTS_WEIGHT_MODE=cached remains the sole cache policy, with reuse, invalidation, transient, and zero-copy strided coverage retained." - ] - }, - { - "id": "TA-1603", - "scope": "legacy EP-wide duplicate server batch rollback mode", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_SERVER_EP_DUPLICATE_BATCHES was an undocumented rollback switch for ep_size-times redundant server compute; no shipped configuration selected it.", - "The correct per-rank EP slice mapping remains covered across EP, CP, padding, routing payload, and OPD teacher-cache assembly, while compatibility-only duplicate-broadcast assertions are removed." - ] - }, - { - "id": "TA-1604", - "scope": "orphaned Qwen linear-attention P2P compatibility branches", - "decision": "remove", - "status": "applied", - "evidence": [ - "The disabled fused QKV slicer produced a combined layout for which pinned SGLang has no receiver locator; its env gate, dimension plumbing, fake bypass assertion, and dead implementation are removed in favor of canonical locator slices.", - "Cold-prepare invalidation now has one policy boundary: cache_invalidation_mode=none disables it. The undocumented duplicate env override is removed, and cold, cached, and disabled prepare cases remain covered." - ] - }, - { - "id": "TA-1605", - "scope": "duplicate optimizer empty-cache environment override", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_SKIP_EMPTY_CACHE_AFTER_OPTIM_STEP appeared only in ModelRunner and its test, duplicating the native skip_empty_cache_after_optim_step train configuration field.", - "The optimizer-step owner now selects both active cache policies through train_config while retaining gradient scaling, clipping, mutation, synchronization, and result-metric coverage." - ] - }, - { - "id": "TA-1606", - "scope": "legacy and debug-only weight-sync environment policies", - "decision": "remove", - "status": "applied", - "evidence": [ - "Both legacy receiver post-process overrides and XORL_WEIGHT_SYNC_BUCKET_BYTES were compatibility-only; pinned P2P writes already target receiver-native FP8 storage, while KV-cache finalization is selected from endpoint requirements.", - "Direct-EP scatter's shallow/deep copy modes and legacy boolean alias only tested optional copies of immutable prepared locators. The production default reuses locators, and scatter serialization plus the retained manifest owner preserve the recipient boundary." - ] - }, - { - "id": "TA-1607", - "scope": "NeMo fp8_cfg compatibility translation", - "decision": "remove", - "status": "applied", - "evidence": [ - "No shipped XoRL configuration or documentation used fp8_cfg, while native enable_fp8_training and fp8_training_* fields already own the complete supported policy.", - "The dataclass aliases, NeMo policy extraction, normalization API, launcher remapping, and acceptance assertions are removed; the shared configuration tombstone now fails fast with the native-field migration." - ] - }, - { - "id": "TA-1608", - "scope": "tests/fp8_training/test_config_compat.py", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "Its direct external-runtime rejection matrix duplicated the public train and server parser owners after fp8_cfg translation was retired.", - "BF16 island selection now runs inside the FP8 injection owner and Blackwell rejection runs through build_training_model, eliminating the standalone file and two collected tests without losing production execution." - ] - }, - { - "id": "TA-1609", - "scope": "legacy numerical-family environment rollback switches", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_FAMILIES_V2 and SGLANG_FAMILIES_V2 duplicated the model program's structural v1/v2 selection and allowed trainer and sampler processes to drift independently.", - "Ordinary models now use v2, exact Qwen3.5 selects its qualified v1 program, and canonical GLM-5.2 selects v2. The surviving norm and LM-head tests select those production programs structurally instead of testing kill switches." - ] - }, - { - "id": "TA-1610", - "scope": "routing-weight position environment override", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_MOE_ROUTING_WEIGHTS_BEFORE_DOWN duplicated the model and server moe_routing_weights_before_down configuration and could silently override the resolved model program per process.", - "The numerical forward/backward oracle and auto, explicit, router-training, dispatch, and SGLang-parity policies remain in the production configuration owner; only the redundant lazy environment assertion is removed." - ] - }, - { - "id": "TA-1611", - "scope": "tests/qarl/test_calibration.py", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The standalone report mostly enumerated private JSON and JSONL loader shapes around a synthetic model, while the training-model owner already executes the real calibration lifecycle before parallelization.", - "Persistent QARL calibration state is retained in the dense fake-quant owner; the production builder still proves real batch loading, observer population, ordering, and calibrated-module counts." - ] - }, - { - "id": "TA-1612", - "scope": "tests/server/test_side_payloads.py", - "decision": "remove", - "status": "applied", - "evidence": [ - "The standalone fake-store report repeated Mooncake metadata encoding and error permutations outside the request and dispatcher lifecycles that own the feature.", - "The retained request processor writes and cleans R3 references on success and failure, and the retained dispatcher loads only the rank-local packed slice in production order." - ] - }, - { - "id": "TA-1613", - "scope": "DSV4 RoPE cache environment override", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_DSV4_ROPE_MAX_SEQ_LEN was a test and profiling override for a value already owned by config.max_position_embeddings, creating two cache-capacity authorities.", - "DSV4 model, LoRA, compressor, and context-parallel capacity owners now select explicit model configuration; the production capacity guard still fails loudly when that configured cache is too short." - ] - }, - { - "id": "TA-1614", - "scope": "tests/data/collators/test_tensor_collator.py permutation matrix", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "After earlier count consolidation, the single test still hid a broad matrix of scalar, boolean, string, dimensionality, and batch-size permutations for a mechanical converter.", - "One production-shaped contract now covers the four supported pipeline forms: flat features, already-batched dictionaries, nested packed features, and empty input, including dtype and tensor-passthrough boundaries." - ] - }, - { - "id": "TA-1615", - "scope": "SGLang EP slot-combine experiment and fake-kernel assertions", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_MOE_SGLANG_FUSED_EXPERTS_SLOT_COMBINE was an undocumented, default-off, scoring-only branch with no shipped configuration or runtime owner.", - "Its assertions replaced both moe_sum_reduce and distributed transport with world-size-one fakes. The qualified SGLang EP compute, stock all-to-all combine, real extension smoke, and independent FP32-routing backward oracle remain." - ] - }, - { - "id": "TA-1616", - "scope": "tests/server/runner/test_batch_utils.py", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The standalone report called the non-packed sequence sharder with batch_size=1, although production routes batch_size=1 through TextSequenceShardCollator.", - "Its useful float-side-channel and teacher-hidden-state behavior now runs through RunnerDispatcher with a real two-row batch, ragged conversion, and CP sharding; the data-collator owners retain packed behavior." - ] - }, - { - "id": "TA-1617", - "scope": "duplicate and unowned dispatcher diagnostic/dummy environment branches", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_MICROBATCH_DIAGNOSTIC_TENSORS duplicated diagnostic_microbatch_dump_tensors, and XORL_MICROBATCH_DIAGNOSTIC_DIR duplicated diagnostic_microbatch_dump_dir.", - "XORL_SERVER_MINIMAL_DUMMY_BATCH_TOKENS selected a second undocumented dummy constructor with no tests or callers. Diagnostics now use request parameters and padding uses the single retained zero-loss dummy lifecycle." - ] - }, - { - "id": "TA-1618", - "scope": "malformed TensorData compatibility assertions", - "decision": "remove", - "status": "applied", - "evidence": [ - "The API-type report treated mismatched and zero-sized tensor metadata falling back to a flat list as a compatibility promise, despite those inputs being invalid and prone to downstream misclassification.", - "The retained production-shaped contract covers rank-1 token IDs plus rank-2 teacher states and rank-3 routing tensors, which are the shapes the server must preserve." - ] - }, - { - "id": "TA-1619", - "scope": "test_ep_trainable_grads_match_stock_triton", - "decision": "remove", - "status": "applied", - "evidence": [ - "The test was skipped in the default environment and failed when the isolated SGLang environment made it executable: stock Triton and the serving wrapper intentionally use different routing-rounding programs.", - "The custom backward preserves the serving FP32-routing boundary, so stock bitwise equality is the wrong oracle. The retained independent eager oracle proves local and EP input, routing, and weight gradients against that actual contract." - ] - }, - { - "id": "TA-1620", - "scope": "isolated SGLang exact-kernel environment resolution", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Pinned SGLang resolves Quack 0.6 and CUTLASS DSL 4.6, while this XoRL source imports the Quack 0.5 and CUTLASS 4.5 APIs; the prior smoke-only environment therefore could not collect XoRL model tests.", - "The isolated setup and uv profile now override those trainer-side packages after installing SGLang. Fresh uv resolution succeeds, the real sgl_kernel operation passes, and the FP32-routing plus DSV4 model, LoRA, and compressor owners all pass under Torch 2.11." - ] - }, - { - "id": "TA-1621", - "scope": "mocked XORL_DCP_LOAD_NO_DIST routing assertions", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The no-dist environment is a real shared-filesystem recovery mode and remains covered by the GLM exact-DCP round trip, but two ModelState helpers only captured process_group=None and no_dist=True from a fake dcp.load call.", - "Those mocks are removed. The retained ModelState owner now performs a real model-only DCP save/load while proving a requested optimizer remains untouched; pipeline custom-group ordering stays separately covered." - ] - }, - { - "id": "TA-1622", - "scope": "test-only OPD loss-family inference", - "decision": "remove", - "status": "applied", - "evidence": [ - "Every production _finalize_loss_metrics call supplies its resolved loss_fn, while one direct test omitted it and forced runtime code to guess OPD from metric-name prefixes.", - "The private finalizer now requires the production argument and the retained OPD aggregation owner uses that actual call contract." - ] - }, - { - "id": "TA-1623", - "scope": "shards plus preprocess_shards compatibility behavior", - "decision": "remove", - "status": "applied", - "evidence": [ - "Documentation declares shards and preprocess_shards mutually exclusive and no shipped configuration combines them, but the preparation test promised that shards silently won.", - "DatasetConfig now rejects the invalid combination and the generator no longer carries the compatibility-only precedence branch." - ] - }, - { - "id": "TA-1624", - "scope": "dataset merge length-only shuffle assertions", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Both shuffle modes previously asserted only that concatenation preserved row count, so they passed if production ignored the shuffle flags entirely.", - "The retained lifecycle now distinguishes ordered concatenation, whole-merge permutation, and within-dataset permutation while proving row preservation." - ] - }, - { - "id": "TA-1625", - "scope": "data_files JSON-only routing assertions", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The string/list cases only counted Hub downloads and always selected JSON, masking that production ignored the documented ds_type field for downloaded data_files.", - "The loader now selects the configured dataset type and the retained cases verify JSON string and Parquet list routing plus their resolved files." - ] - }, - { - "id": "TA-1626", - "scope": "immediate-success retry decorator assertion", - "decision": "remove", - "status": "applied", - "evidence": [ - "The one-call success row added no behavior beyond the retained transient-success case, which already proves return-value propagation after the retry boundary.", - "Request and Hub transient classification, exponential backoff, exhaustion, and unrelated-exception propagation remain covered." - ] - }, - { - "id": "TA-1627", - "scope": "required SGLang ABI smoke and remaining default-profile installation path", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The server-training guide still installed pinned SGLang into the active environment and advertised Torch 2.9.1, contradicting the Torch-2.12 default and Torch-2.11 SGLang split.", - "Required smoke mode now fails unless CUDA is available for a real compiled RMSNorm operation and reports both import-loader error forms; the isolated Torch 2.11.0 and sglang-kernel 0.4.5 lane passes while the default profile contains no kernel wheel." - ] - }, - { - "id": "TA-1628", - "scope": "standalone block-FP8 activation and tiled-weight edge matrices", - "decision": "remove", - "status": "applied", - "evidence": [ - "The two component reports enumerated random scale ranges, dimensionality, block sizes, determinism, magnitude and sign examples, plus internal assertion failures without entering a model operation.", - "The retained FP8-linear owner executes activation and tiled-weight quantize/dequantize at block sizes 64 and 128, consumes both scale layouts in a real GEMM, and now checks each codec against its original production-shaped operand; five unexported and unreferenced compatibility aliases are also removed." - ] - }, - { - "id": "TA-1629", - "scope": "test-owned sequence-parallel metric argument-order compatibility", - "decision": "remove", - "status": "applied", - "evidence": [ - "Both production callers pass metrics, process group, and metric operations in the declared order, while the retained distributed test alone used the obsolete metrics, operations, group order.", - "The type-based argument swap is removed and the real two-GPU NCCL reduction now passes through the production signature without weakening its partial-sum or extrema assertions." - ] - }, - { - "id": "TA-1630", - "scope": "private generic GLM absorbed-projection einsum decomposition", - "decision": "remove", - "status": "applied", - "evidence": [ - "The test replaced GLM projections and rotary execution, then asserted the exact private _project_qkv_absorb and _project_absorbed_value einsum decomposition.", - "The retained full-model GLM sparse-versus-dense policy executes both generic absorbed projections numerically, while the exact-kv_b owner separately proves factor-only branch routing and non-materialization." - ] - }, - { - "id": "TA-1631", - "scope": "standalone NF4 codec and bandwidth matrices", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report repeated flat and GKN shapes, dtypes, constants, three group sizes, and large random tensors without entering a shipped module; its bandwidth cases converted a missed performance target into a successful skip.", - "The retained QLoRA linear owner now executes flat NF4 quantization through forward, backward, storage, and reconstruction, while the real two-GPU Triton expert transaction reconstructs each production-shaped GKN base before two optimizer steps." - ] - }, - { - "id": "TA-1632", - "scope": "ToTensorCollator pipeline-shape report", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The report covers the four inputs admitted by the production collator: flat samples, an already-batched dictionary, packed nested samples, and empty input.", - "The larger dataloader owner starts from tensors and therefore does not prove list and NumPy conversion, field-specific integer dtypes, tensor identity, or structural preservation." - ] - }, - { - "id": "TA-1633", - "scope": "QARL Triton expert integration report", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The CPU fake-quant report independently owns rounding and STE math, but cannot prove that QARLMoEExperts exposes quantized Parameters to the production grouped-GEMM implementation.", - "The retained GPU transaction executes the real Triton expert kernel, distinguishes quantized and disabled paths numerically, and proves finite gradients reach both expert projections." - ] - }, - { - "id": "TA-1634", - "scope": "local model config and registry reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The Qwen3.5 and Kimi reports load real local config.json and tokenizer artifacts through the auto-loader rather than merely asserting registry dictionary entries.", - "Tiny model forward/backward reports construct configs directly, so they do not cover wrapper conversion, derived hybrid layer types, official auxiliary-field spelling, or local tokenizer dispatch." - ] - }, - { - "id": "TA-1635", - "scope": "small server API, scheduler, and rank-zero protocol reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The reports cross real boundaries: API response and metric mapping, typed wire serialization, rank-zero ready and acknowledgement handling, and FIFO pending, running, terminal, and bounded-history transitions.", - "Their doubles isolate sockets and worker processes but do not replace the serialization, payload conversion, queue mutation, or failure behavior being asserted." - ] - }, - { - "id": "TA-1636", - "scope": "server security, checkpoint-path, and weight-sync routing reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The security and checkpoint owners enforce traversal, symlink, SSRF and DNS-pinning, tenant isolation, reserved-checkpoint, and destructive-operation boundaries that ordinary success lifecycles cannot replace.", - "The mock-heavy endpoint report owns two-phase receiver fencing, flattened byte layout, chunking, cache metadata, and health fallback rather than merely checking that HTTP helpers were called." - ] - }, - { - "id": "TA-1637", - "scope": "remaining legacy-labeled server compatibility behavior", - "decision": "keep", - "status": "accepted", - "evidence": [ - "DRGRPO still consumes both old_logprobs and rollout logprobs at the live loss boundary, and the public API still maps Tinker session payloads into current model identifiers and optimizer controls.", - "Checkpoint and optimizer compatibility cases reject unsafe pickle state or resolve supported on-disk and URI forms; none exists solely to satisfy a test-only branch." - ] - }, - { - "id": "TA-1638", - "scope": "fake Gram Newton-Schulz CUDA dtype report", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The standalone report replaced every backend operation with logging functions and therefore proved only that its fakes received FP32 tensors, not that the shipped CUDA program preserved FP32 arithmetic.", - "The retained two-GPU full-gradient Muon transaction now runs a real CUDA Gram Newton-Schulz tree and matches an independent FP32 program exactly in addition to its single-rank optimizer oracle." - ] - }, - { - "id": "TA-1639", - "scope": "disabled timer and manually injected unrecorded-event report", - "decision": "remove", - "status": "applied", - "evidence": [ - "The report asserted disabled no-ops and wrote fake objects directly into private event-pair dictionaries; last_skipped_event_pair_count existed only to expose that synthetic path to the test.", - "The test-only counter is removed, invalid event pairs remain safely ignored, and the retained real CUDA lifecycle attaches hooks to GLM-style and Qwen-style decoder layers and records forward and backward phase timings." - ] - }, - { - "id": "TA-1640", - "scope": "remaining optimizer and training utility reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Learning-rate reports own exact warmup, decay, floor, and multi-optimizer behavior, while DistSignSGD owns sign-after-SP-sum ordering, forced SUM reduction, local versus FSDP-managed hooks, and unsupported-topology admission.", - "Cautious decay, standard batched Newton-Schulz, token counting, chunked CE, and explicit gradient synchronization retain independent numerical or collective oracles that a successful training smoke cannot replace." - ] - }, - { - "id": "TA-1641", - "scope": "scheduler-only stepping in learning-rate traces", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The schedule report advanced LambdaLR repeatedly without optimizer steps, unlike Trainer._clip_and_step, and emitted PyTorch warnings about skipping the first learning-rate value.", - "Single and multi-optimizer traces now use the production optimizer-step then scheduler-step order while preserving the exact warmup, decay, and floor assertions without warnings." - ] - }, - { - "id": "TA-1642", - "scope": "quantized-export size parsing and direct sharding helpers", - "decision": "consolidate", - "status": "applied", - "evidence": [ - "The standalone parser examples and direct-function sharding setup exercised the same implementation path separately without proving that a configured command-line export honored its string-valued shard size.", - "The retained subprocess CLI transaction now consumes a 24B YAML value, emits a multi-shard index, reconciles its total size and shard count, reloads every artifact, and checks the converted tensor dtypes and layouts." - ] - }, - { - "id": "TA-1643", - "scope": "model-support default and private-dispatch assertions", - "decision": "remove", - "status": "applied", - "evidence": [ - "The GLM bare-config field report repeated values already exercised through an official-shaped local config artifact, while its monkeypatched exact-indexer dispatch probe is a subset of the canonical GLM-5.2 contract report's router and indexer selection transaction.", - "MiniMax's three private expert-key aliases are a strict subset of the centralized checkpoint expert-key classifier matrix; its real forward, backward, checkpoint packing, and EP ownership checks remain." - ] - }, - { - "id": "TA-1644", - "scope": "remaining command-line and model-support transactions", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The NVFP4 and FP8 export owners write and reload real safetensors/config artifacts, enforce conversion and fail-closed admission, and preserve trained QARL logprobs after folding rather than merely asserting helper mappings.", - "Qwen2 and OLMo2 load Hugging Face checkpoints into fused models and compare hidden states and logits; OPD exercises causal payload alignment and Mooncake metadata transport; private FSDP policy reports uniquely own CP-folding admission and prefetch direction." - ] - }, - { - "id": "TA-1645", - "scope": "undocumented FSDP prefetch boolean coercion matrix", - "decision": "remove", - "status": "applied", - "evidence": [ - "The authoritative training and model-builder fields are bool or Optional[bool], and every production caller supplies those typed values; integer and yes/no/on/off spellings were accepted only by a direct private-helper test.", - "The compatibility parser and its ten-case matrix are removed. FSDP2 now fails fast on non-boolean values while preserving forward defaults and backward inheritance." - ] - }, - { - "id": "TA-1646", - "scope": "context-parallel folding and manual FSDP prefetch reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The CP folding truth table decides whether expert parameters need a separate FSDP root across Ulysses, Ring Attention, and mixed layouts.", - "The prefetch transaction uniquely checks forward and backward neighbor direction plus the no-op topology; these are live topology and performance policies rather than configuration aliases." - ] - }, - { - "id": "TA-1647", - "scope": "runner session, optimizer, and P2P async reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Optimizer hyperparameters cross ServerArguments, wire payloads, dispatcher forwarding, dense optimizers, and adapter optimizers; the session registry crosses update, checkpoint load, eviction promotion, and kill.", - "The P2P async report owns size-based transfer selection and timeout failure at the real transport entry point. Its doubles isolate HTTP and Mooncake engines without replacing the production decision." - ] - }, - { - "id": "TA-1648", - "scope": "permissive optimizer cache-policy truthy adapter", - "decision": "remove", - "status": "applied", - "evidence": [ - "The skip_empty_cache_after_optim_step field has one boolean runtime owner; accepting arbitrary truthy objects and undocumented yes/on string spellings widened the contract without a configuration source.", - "The optimizer lifecycle still executes both cache policies using actual booleans and reports whether cache release was skipped." - ] - }, - { - "id": "TA-1649", - "scope": "deterministic MoE scatter rollback aliases and relaxed-atomic kernel", - "decision": "remove", - "status": "applied", - "evidence": [ - "XORL_MOE_DETERMINISTIC_SCATTER was absent from shipped arguments, examples, and documentation; it existed only to restore the older run-variant relaxed-atomic slot assignment.", - "The environment parser, four false-value aliases, route mocks, and obsolete Triton kernel are removed. The retained CUDA transaction proves stable order, full slot coverage, per-expert cumsum regions, and int32/int64 routing inputs." - ] - }, - { - "id": "TA-1650", - "scope": "mock-heavy checkpoint broadcast and adapter optimizer reports", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The broadcast owner includes real four-process Gloo DTensor materialization and checks grouped dense/expert transfer, strict coverage, checkpoint-handler filtering, and fallback admission.", - "Adapter optimizer resume retains bitwise uninterrupted comparison, artifact identity, topology-changing reshard, corruption rejection, atomic nonmutation, and a separate real two-rank rollback transaction; private-helper calls support those lifecycle guarantees rather than form isolated examples." - ] - }, - { - "id": "TA-1651", - "scope": "remaining conditional-runtime backend gates", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The SGLang smoke executes a compiled operation under the pinned ABI, FlashQLA compares a real two-rank context-parallel program with its local reference, and the FSDP/GLM reports compose exact kernels with real distributed ownership.", - "The optional DeepGEMM and real SGLang MoE reports exercise documented production dispatches with independent numerical or gradient oracles; their conditional admission reflects hardware and dependency availability rather than an absent assertion." - ] - }, - { - "id": "TA-1652", - "scope": "numeric-string routing-weight position aliases", - "decision": "remove", - "status": "applied", - "evidence": [ - "The public field is a bool or one of auto, true, and false; configuration loaders can therefore supply actual booleans or those declared strings, but no shipped caller produces the undocumented strings 1 and 0.", - "The resolver and retained policy test preserve typed booleans, declared string spellings, automatic regime selection, explicit overrides, and invalid-value rejection without expanding the interface for a test-only matrix." - ] - }, - { - "id": "TA-1653", - "scope": "deprecated AdapterState.lora_params compatibility property", - "decision": "remove", - "status": "applied", - "evidence": [ - "Production, documentation, and examples use AdapterState.local_params; only tests accessed the deprecated lora_params alias or reproduced it on fake state objects.", - "The alias and fake properties are removed, while optimizer resume, coordination, checkpoint round-trip, gradient ownership, and rollback tests continue through the authoritative local parameter storage." - ] - }, - { - "id": "TA-1654", - "scope": "legacy EP replicated-gradient group fallback", - "decision": "remove", - "status": "applied", - "evidence": [ - "The production EP classifier always emits ep_replicated_gradient_sync when synchronization is enabled; falling back to the broader ep_replicated group masked incomplete model metadata and was sustained only by a test-created group dictionary.", - "Synchronization now fails fast when the authoritative group is missing, and the retained real multi-rank transaction uses the same four-group metadata built by production while preserving coalesced, missing-gradient, nonfinite, and clipping coverage." - ] - }, - { - "id": "TA-1655", - "scope": "all current-main production source", - "decision": "keep", - "status": "accepted", - "evidence": [ - "Replaying the historical audit onto current main showed that helpers previously classified as test-only are now consumed by the exact trainer-serving stack merged in PR 57.", - "After merging main through PR 67, the pull request has zero src diff against origin/main. The upstream removal of src/xorl/rl is recorded by TA-887; earlier production-surface removals remain historical audit evidence rather than changes in this PR." - ] - }, - { - "id": "TA-1656", - "scope": "all current-main conflict paths", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The conflicted files contain behavior and coverage added or modified after the original audit, including canonical combine, exact sampling, byte-contract, full-parameter, and adapter ownership programs.", - "The initial replay resolved 71 conflict paths to current main, and the final rebase resolved 13 additional conflict paths the same way; the historical consolidation applies only where it merges cleanly, avoiding silent deletion of newer regression owners." - ] - }, - { - "id": "TA-1657", - "scope": "five remaining exact duplicate-body groups", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The OLMo2 and vocabulary-parallel wrappers intentionally delegate to different distributed workers despite identical wrapper bodies.", - "The GLM-5.2 and DSV4 temperature-gradient bodies bind to distinct architecture implementations through same-named local imports, so identical ASTs provide cross-architecture numerical coverage rather than duplicated execution.", - "Two MoE duplicate-body groups live in current-main files that conflicted during the final rebase; they remain unchanged under TA-1656 rather than being removed by the historical audit." - ] - }, - { - "id": "TA-1658", - "scope": "rebased current-main audit candidates", - "decision": "keep", - "status": "accepted", - "evidence": [ - "The seventeen conditional-runtime reports cover real CUDA, distributed, SGLang, DeepGEMM, and exact-kernel programs; the twenty-one no-inline-outcome reports delegate to assertion helpers, process exit gates, or successful validation boundaries.", - "The two source-inspection reports were introduced by the current exact stack to guard autograd saved-state and frozen-trunk detach properties. They are retained with all other current-main conflicts rather than re-audited away in this compatibility rebase." - ] - }, - { - "id": "TA-1659", - "scope": "rebase-only dataset rejection and loader-call assertions", - "decision": "remove", - "status": "applied", - "evidence": [ - "Current main does not reject a DatasetConfig containing both shards and preprocess_shards; that rejection belonged to a historical production change excluded by TA-1655 and cannot remain as a test-only contract.", - "The consolidation also added exact loader-format and data_files call assertions that over-specified private dispatch and contradicted current-main behavior. The retained lifecycle still exercises name and shard expansion, file resolution, download counts, local and hub loading, splitting, merging, persistence, hashing, and retries." - ] - }, - { - "id": "TA-1660", - "scope": "GLM indexer padding-mask CPU coverage", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "Consolidation moved padding-mask detection into a CPU-marked aggregate but retained one unconditional CUDA allocation from the former CUDA-gated owner.", - "The detection policy is device-independent, so both the valid-prefix and interspersed-mask inputs now use the same CPU device; real TileLang execution remains separately CUDA-gated." - ] - }, - { - "id": "TA-1661", - "scope": "optional fused-MoE CPU kernel doubles", - "decision": "rewrite", - "status": "applied", - "evidence": [ - "The canonical Triton MoE module exports grouped-GEMM globals only when CUDA and Triton are available, while the consolidated CPU numerical owners intentionally replace those kernels with independent torch references.", - "The shared helper now installs missing optional globals on CPU-only Torch instead of requiring them to pre-exist; the full CPU misc shard passes with CUDA hidden." - ] - } - ] -} diff --git a/tests/distributed/test_bi_trunk_linear_fsdp.py b/tests/distributed/test_bi_trunk_linear_fsdp.py index 882ba11e..8f8d1df3 100644 --- a/tests/distributed/test_bi_trunk_linear_fsdp.py +++ b/tests/distributed/test_bi_trunk_linear_fsdp.py @@ -1,6 +1,6 @@ """FSDP2 composition smoke for the scoped batch-invariant trunk-linear contract. -Verifies XORL_BI_TRUNK_LINEAR's module wrap composes with fully_shard (2 GPUs): +Verifies the trunk-linear contract's module wrap composes with fully_shard (2 GPUs): the sharded forward must be bit-identical to the unsharded wrapped module (same persistent-GEMM kernel over the all-gathered bf16 params), gradients must be finite, and the grad norm must match an unwrapped cuBLAS FSDP2 reference within @@ -123,7 +123,7 @@ def _run() -> None: @skip_if_gpu_count_less_than(2) def test_bi_trunk_linear_composes_with_fsdp2(): result = run_distributed_script(__file__, num_gpus=2, timeout=180) - result.assert_success("XORL_BI_TRUNK_LINEAR wrap should compose with fully_shard") + result.assert_success("trunk-linear contract wrap should compose with fully_shard") if __name__ == "__main__": diff --git a/tests/models/test_rmsnorm_sglang_fused.py b/tests/models/test_rmsnorm_sglang_fused.py index 31ae9108..86d70374 100644 --- a/tests/models/test_rmsnorm_sglang_fused.py +++ b/tests/models/test_rmsnorm_sglang_fused.py @@ -305,7 +305,7 @@ def leaf(t): # --------------------------------------------------------------------------- # -# Trunk contract lane (XORL_BI_TRUNK_LINEAR): the no-residual dispatch (qk-norm) +# Trunk contract lane: the no-residual dispatch (qk-norm) # must bit-match serving's family-1 batch-invariant kernel — which is the # aten::rms_norm interpose kernel, NOT the fused sglang residual tree (the two # disagree at 1 ulp on rare bf16 boundary values). diff --git a/tests/ops/test_bi_trunk_linear.py b/tests/ops/test_bi_trunk_linear.py index e5c844f0..cc4c46e0 100644 --- a/tests/ops/test_bi_trunk_linear.py +++ b/tests/ops/test_bi_trunk_linear.py @@ -1,4 +1,4 @@ -"""Scoped batch-invariant trunk-linear contract (XORL_BI_TRUNK_LINEAR). +"""Scoped batch-invariant trunk-linear contract. The trunk contract routes ONLY transformer-trunk nn.Linear forwards through the batch-invariant persistent GEMM (bit-identical to the aten::mm interpose lane and @@ -243,15 +243,15 @@ def test_global_interpose_raises_on_grad_requiring_inputs(): wn = torch.randn(HIDDEN, device="cuda", dtype=torch.bfloat16, requires_grad=True) xb = torch.randn(2, 16, HIDDEN, device="cuda", dtype=torch.bfloat16, requires_grad=True) with set_batch_invariant_mode(True): - with pytest.raises(RuntimeError, match="XORL_BI_TRUNK_LINEAR"): + with pytest.raises(RuntimeError, match="inference/verification-only"): _ = x @ w - with pytest.raises(RuntimeError, match="XORL_BI_TRUNK_LINEAR"): + with pytest.raises(RuntimeError, match="inference/verification-only"): _ = torch.rms_norm(x, (HIDDEN,), wn, 1e-6) - with pytest.raises(RuntimeError, match="XORL_BI_TRUNK_LINEAR"): + with pytest.raises(RuntimeError, match="inference/verification-only"): _ = torch.bmm(xb, xb.transpose(1, 2)) - with pytest.raises(RuntimeError, match="XORL_BI_TRUNK_LINEAR"): + with pytest.raises(RuntimeError, match="inference/verification-only"): _ = torch.log_softmax(x.float(), dim=-1) - with pytest.raises(RuntimeError, match="XORL_BI_TRUNK_LINEAR"): + with pytest.raises(RuntimeError, match="inference/verification-only"): _ = x.float().mean(-1)