All pull requests

Pull requests

• fix(swiglu): guard Triton launches by the input device

  bug
  #665·Jungle430 opened Nov 6, 2026Contributor·

  ·
  1

  6

• [WS1][CUDA][MiniMax-H3] Add deterministic packed index select

  #663·Jungle430 opened Nov 5, 2026Contributor·3 of 4 tasks·

  ·
  1

• [ROCm][DSV4][P1] Integrate batch-invariant mHC pre, post, head and a shared linear on aiter (tasks 1, 2, 3, 4, 8, 9)

  deepseek-P1DSv4platform: rocm
  #662·zitian-wang opened Nov 5, 2026·

  ·
  9

• [Nemotron][CUDA] Add deterministic moe_experts_combine

  Nemotronnew-modelplatform: cuda
  #661·luoyueyuguang opened Nov 4, 2026Contributor·

  ·
  11

• GDN shared state contract and CPU sequence reference (RFC #642 Task 20)

  #660·hongleng opened Nov 4, 2026Contributor·

  ·
  1

• feat(h3): add partial multimodal RoPE operators

  component: kernelsMiniMax-H3new-model
  #656·Jungle430 opened Nov 2, 2026Contributor·

  ·
  1

• [gemma4] add scaled_token_embedding for Gemma 4

  gemmanew-model
  #653·haotianxue31 opened Nov 2, 2026·1 task·

  ·
  11

• docs: add a practical guide to train–rollout consistency

  documentation
  #652·Flink-ddd opened Nov 2, 2026Collaborator·

  ·
  9

• feat(gemma): add tied_lm_head_projection for Gemma 4 vocab

  gemmanew-modelplatform: cuda
  #645·Dnoob opened Oct 31, 2026Contributor·2 tasks·

  ·
  7

• [DSv4][P3-T08] Placement, token ownership, P7 binding, and gate-gradient collective provenance

  deepseek-P3DSv4platform: cuda
  #639·AlanX-Yan opened Oct 30, 2026·

  ·
  2

• [DSv4][CUDA] DSv4 testing control threads.

  do-not-closeDSv4platform: cudapriority: highstatus: needs-reviewWIP
  #635·KJLdefeated opened Oct 30, 2026Collaborator·

  ·
  24

• [WS1][Qwen-Image][CUDA] Add stride-aware multi-axis RoPE with compile support

  new-modelplatform: cudaqwenWS1
  #623·dandan20031031 opened Oct 28, 2026·

  ·
  19

• [ROCm] Add provider lock and execution checks

  platform: rocmqwen
  #622·Anoymoguy opened Oct 28, 2026·

  ·
  8

• feat(adaln): add indexed CUDA/Triton modulation and explicit strict qualification

  new-modelplatform: cudaqwen
  #621·hongleng opened Oct 28, 2026Contributor·

  ·
  80

• [ws1][Muse-Glimmer-30B] SwiGLU: FP64 references, Muse shapes and backend parity

  Muse-GlimmerWS1
  #610·sevleete opened Oct 27, 2026Contributor·

  ·
  30

• feat(reduction): add shared fixed-order Triton and Gluon reduction helpers

  enhancementstatus: needs-review
  #595·Jungle430 opened Oct 25, 2026Contributor·

  ·
  25

• docs: fix RoPE reference broadcasting and device placement

  documentation
  #594·Flink-ddd opened Oct 25, 2026Collaborator·

  ·
  4

• feat(rope): add Muse Glimmer F2 CUDA/Triton norm-scale-RoPE

  Muse-Glimmernew-model
  #589·cr-zhao opened Oct 25, 2026·

  ·
  1

• feat(cuda): add deterministic Qwen-Image latent normalization

  new-modelplatform: cudaqwen
  #577·JLiu4Coding opened Oct 23, 2026Contributor·

  ·
  2

• [WS1][CUDA][MiniMax-H3] checkpoint mapper and the vLLM-Omni rollout adapter (RFC #420 stack 7/7)

  MiniMax-H3new-modelplatform: cudaWS1
  #575·fusheng-ji opened Oct 22, 2026Contributor·

  ·
  25

• [WS2][CUDA][MiniMax-H3] ws2 tensor / sequence / context parallel and the hierarchical cross-node collective (RFC #420 stack 3/7)

  MiniMax-H3new-modelplatform: cudaWS2
  #569·fusheng-ji opened Oct 22, 2026Contributor·

  ·
  15

• [WIP] [WS1][kernels] joint_attn_qk_gemm for Qwen-Image (python + cuda)

  #560·ryankert01 opened Oct 16, 2026Collaborator·

  ·
  3

  

• feat(adaln): add shared FP32 modulation backends and H100 validation

  new-modelplatform: cudaqwen
  #559·hongleng opened Oct 16, 2026Contributor·

  ·
  15

• [WS1][Muse-Glimmer-30B] Foundation: architecture fingerprint, strict loader and operator trace (RFC #501)

  Muse-Glimmernew-modelWS1
  #558·ryankert01 opened Oct 16, 2026Collaborator·

  ·
  14

  

• [WS1][CUDA][Qwen3-Next] EP4 distributed logprob parity: expert-parallel MoE, full-model L2 bitwise on TP4 x EP4 (RFC #428 stack 10/11)

  new-modelplatform: cudaqwenWS1
  #557·fusheng-ji opened Oct 15, 2026Contributor·

  ·
  52