All pull requests

Pull requests

• feat(p3): T01 operator contracts, deterministic Top-6 and sealed start kit

  #498·maxiaosong1124 opened Oct 8, 2026Collaborator·

  ·
  1

• refactor: organize models, backends, contracts and train and rollout engines

  Ascendcomponent: alignmentcomponent: executorscomponent: kernelsenhancementMUSAplatform: cudaplatform: rocmplatform: tritonpriority: hightype: refactor
  #497·Flink-ddd opened Oct 8, 2026Collaborator·5 tasks·

  ·
  1

• [WS1][Qwen-Image][CUDA] Flow SDE step and log-probability

  #496·LG-0927 opened Oct 8, 2026·

  ·
  1

• perf(logp): fused zero-workspace backward for the generic CUDA fused logp (#174)

  component: kernelsplatform: cudatype: performance
  #495·fusheng-ji opened Oct 7, 2026·

  ·
  1

• [WS2][CUDA][MiniMax-H3] sp_norm_adaln: sequence-parallel norm/AdaLN/gated residual, byte-equal to WS1 (RFC #420 stack 9/9)

  #494·fusheng-ji opened Oct 7, 2026·

  ·
  1

• [WS2][CUDA][MiniMax-H3] tp_adaln_3mod: column-parallel AdaLN projection, byte-equal to WS1 (RFC #420 stack 8/9)

  #493·fusheng-ji opened Oct 7, 2026·

  ·
  1

• perf(cuda): vectorize the deterministic all-gather copy and right-size its single-block path

  platform: cudatype: performance
  #491·fusheng-ji opened Oct 7, 2026·

  ·
  9

• ci: enable CodeRabbit automatic reviews for test-h3

  type: ci-cd
  #490·fusheng-ji opened Oct 7, 2026·

  ·
  3

• [WS1][CUDA][MiniMax-H3] final_adaln_out: norm_out projection + final norm/modulation (RFC #420 stack 7/9)

  MiniMax-H3new-modelplatform: cudaWS1
  #489·fusheng-ji opened Oct 7, 2026·

  ·
  14

• [WS1][CUDA][MiniMax-H3] adaln_gate_residual: gated residual with in-kernel gate gather (RFC #420 stack 6/9)

  MiniMax-H3new-modelplatform: cudaWS1
  #488·fusheng-ji opened Oct 7, 2026·

  ·
  19

• [WS1][CUDA][MiniMax-H3] h3_rmsnorm: RMSNorm with fused AdaLN modulation (RFC #420 stack 5/9)

  MiniMax-H3new-modelplatform: cudaWS1
  #487·fusheng-ji opened Oct 7, 2026·

  ·
  22

• [WS1][CUDA][MiniMax-H3] adaln_row_gather: fused six-way AdaLN row gather (RFC #420 stack 4/9)

  MiniMax-H3new-modelplatform: cudaWS1
  #486·fusheng-ji opened Oct 7, 2026·

  ·
  18

• [WS1][CUDA][MiniMax-H3] adaln_projection_3mod: three-modality AdaLN projection (RFC #420 stack 3/9)

  MiniMax-H3new-modelplatform: cudaWS1
  #485·fusheng-ji opened Oct 7, 2026·

  ·
  37

• [WS1][CUDA][MiniMax-H3] timestep_mlp_fp32: deterministic FP32 timestep MLP (RFC #420 stack 2/9)

  MiniMax-H3new-modelplatform: cudaWS1
  #484·fusheng-ji opened Oct 7, 2026·

  ·
  10

• [WS1][CUDA][MiniMax-H3] timestep_sinusoid_h3: FP32 timestep features + shared H3 harness (RFC #420 stack 1/9)

  MiniMax-H3new-modelplatform: cudaWS1
  #483·fusheng-ji opened Oct 7, 2026·

  ·
  16

• [DSv4][P3-9] Add router validation infrastructure

  deepseek-P3DSv4
  #481·quan-yi-ai opened Oct 6, 2026·

  ·
  1

• [WS1][Qwen-Image] Add timestep MLP with sm80/sm90 validation and reference audit

  new-modelplatform: cudaqwenWS1
  #478·zzf2026-svg opened Oct 6, 2026·

  ·
  1

• feat: select paired BI execution plans with RL_KERNEL_BI=1

  #477·Flink-ddd opened Oct 6, 2026Collaborator·

  ·
  1

• [WS1][CUDA][MiniMax-H3] Add conditioning_noise_mix forward and backward

  #476·ThreeMonth03 opened Oct 5, 2026·

  ·
  1

• [DSv4][T06] MQA attention and grouped o-proj

  deepseek-P2DSv4
  #474·carsontung666 opened Oct 5, 2026·

  ·
  27

• MiniMax-H3 Euler step kernels

  MiniMax-H3new-model
  #471·Jerry536-coco opened Oct 4, 2026·

  ·
  9

• fix(ci): pin runpodctl v2.9.0 and drop the deprecated config step (RUNPOD_API_KEY auth)

  type: ci-cd
  #470·fusheng-ji opened Oct 4, 2026·

  ·
  5

• [WS1][CUDA][Qwen3-Next] C6: GDN decode-step goldens and recurrent/conv state contract (RFC #428 stack 4/4)

  new-modelplatform: cudaqwenWS1
  #469·fusheng-ji opened Oct 4, 2026·

  ·
  8

• [WS1][CUDA][Qwen3-Next] Gated RMSNorm CUDA kernel for the GDN block (RFC #428 stack 3/4)

  new-modelplatform: cudaqwenWS1
  #468·fusheng-ji opened Oct 4, 2026·

  ·
  4

• [WS1][CUDA][Qwen3-Next] C1: zero-centred RMSNorm references and CUDA kernel (RFC #428 stack 2/4)

  new-modelplatform: cudaqwenWS1
  #467·fusheng-ji opened Oct 4, 2026·

  ·
  7