[Common][PyTorch] EP dispatch with unfused MXFP8 quantization

[phu0ngng]

Description

This PR adds MXFP8 support to the dispatch op of the NCCL EP path. The dispatch op is used in two places, and MXFP8 applies to both:

• Dispatch forward bfloat16 tokens are quantized to MXFP8 internally and dispatched to the target experts; recv is returned as a per-expert GroupedTensor.
• Combine backward the result-grad is scattered back to expert positions through the same (reverse) dispatch op, quantized to MXFP8, returning the expert-output grad as a per-expert GroupedTensor.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

PyTorch frontend (transformer_engine/pytorch/ep.py, distributed.py, csrc/extensions/ep.cpp)**

• The dispatch op quantizes bfloat16 tokens to MXFP8 internally when the buffer's dispatch_quant_recipe is set (MXFP8BlockScaling only for now); dispatch-forward recv is returned as a per-expert GroupedTensor. A pre-quantized input is rejected.
• Combine backward reuses the dispatch op to scatter the result-grad: it quantizes the grad to MXFP8 and returns the expert-output grad as a per-expert GroupedTensor. Combine forward is unchanged (high-precision).
• Recv data and block scales share a single caller-supplied (optionally symm-mem-backed) buffer, sliced into data-then-scale regions; the same convention is used for the combine backward grad buffer.

Common backend (common/ep/ep_backend.cpp, include/.../ep.h, comm_window.h)**

• Backend and public headers extended to carry block-scale buffers/windows through the dispatch primitive.

NCCL EP submodule**

• Bumped 3rdparty/nccl-extensions to the revision providing block-scaled dispatch.

Tests (tests/cpp_distributed/test_ep.cu, tests/pytorch/distributed/run_ep.py, run_test_ep.sh)**

• Added C++ distributed coverage for the MXFP8 dispatch path.
• Added PyTorch MXFP8 test passes for dispatch forward (normal, zero-copy, eager IO modes) and combine backward, gated behind a dedicated NVTE_EP_MXFP8_PASS run since the grouped path pins the per-expert alignment process-wide.

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

Adds unfused MXFP8 quantization to NCCL expert-parallel dispatch and combine backward.

• Routes compact MXFP8 block scales alongside token payloads through the common and PyTorch backends.
• Returns per-expert quantized results as GroupedTensor objects and supports caller-provided or symmetric-memory-backed buffers.
• Extends symmetric-memory pool lifecycle handling and adds distributed MXFP8 coverage.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/ep.py	Adds recipe-controlled MXFP8 dispatch and combine-backward quantization, grouped outputs, and combined data/scale buffer handling.
transformer_engine/pytorch/csrc/extensions/ep.cpp	Extends EP bindings to validate and pass compact MXFP8 scale tensors and symmetric-memory window offsets.
transformer_engine/common/ep/ep_backend.cpp	Adds NCCL descriptors and dispatch configuration for routing MXFP8 scales with token data.
transformer_engine/pytorch/distributed.py	Adds explicit release handling for the process-wide symmetric-memory pool.
tests/pytorch/distributed/run_ep.py	Adds normal, eager, zero-copy, caller-buffer, and combine-backward MXFP8 coverage.

Sequence Diagram

sequenceDiagram
  participant Caller
  participant PyTorch as PyTorch EP
  participant Quantizer as MXFP8 Quantizer
  participant Backend as NCCL EP Backend
  participant Expert
  Caller->>PyTorch: ep_dispatch(BF16 tokens)
  PyTorch->>Quantizer: Quantize data and compact block scales
  Quantizer-->>PyTorch: E4M3 data + E8M0 scale_inv
  PyTorch->>Backend: Dispatch data and scales
  Backend-->>Expert: Grouped MXFP8 tokens
  Expert->>PyTorch: BF16 expert output
  PyTorch->>Backend: Combine forward
  Backend-->>Caller: BF16 result
  Caller->>PyTorch: Result gradient
  PyTorch->>Quantizer: Quantize gradient
  PyTorch->>Backend: Reverse dispatch data and scales
  Backend-->>Expert: Grouped MXFP8 expert-output gradient

Loading

_{Reviews (6): Last reviewed commit: "Merge branch 'main' into ep_mxfp8" | Re-trigger Greptile}

[phu0ngng]

/te-ci L1 pytorch

[YangFei1990]

This is great, shall we also call it in ep_finalize?

[phu0ngng]

release_symm_mem_pool is only required before destroy_process_group().
User may call ep_finalize without destroy_process_group().

[YangFei1990]

Hmm why is that? I did not have it in mcore but it seems working (could be luck)

[YangFei1990]

Also just see input device is never used, shall we remove?

[phu0ngng]

On why

release_symm_mem_pool is only required before destroy_process_group().

I think destroy_process_group() deletes the NCCL communicator that is used to create the symmem, so the cached symmem-s become limbo.

Are you calling destroy_process_group() in MCore? Why?

[YangFei1990]

Mcore shutdown will call destroy_process_group

[phu0ngng]

Good to know.
Anyway, I found out that release_symm_mem_pool is idempotent, so I added it to ep_finalize as you suggested.

[phu0ngng]

/te-ci L1

[phu0ngng]

/te-ci L1