All pull requests

Pull requests

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

  #595·Jungle430 opened Oct 25, 2026Contributor·

  ·
  1

• docs: fix RoPE reference broadcasting and device placement

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

  ·
  3

• [DSV4][ROCm] Add contract-bound BI Indexer projections

  DSv4platform: rocm
  #590·inaniloquentee opened Oct 25, 2026Collaborator·

  ·
  1

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

  #589·cr-zhao opened Oct 25, 2026·

  ·
  1

• [WS1][CUDA][MiniMax-H3] generation gate: 50-block generation compare and resume check (RFC #420 stack 10c/10)

  #588·fusheng-ji opened Oct 25, 2026Contributor·

  ·
  1

• [WS1][CUDA][MiniMax-H3] full-model qualification: thresholds, ensemble rule and reference manifest (RFC #420 stack 10b/10)

  #587·fusheng-ji opened Oct 25, 2026Contributor·

  ·
  1

• [WS1][CUDA][MiniMax-H3] ws1_one_h3_block: block 0 node-by-node replay and diagnostics (RFC #420 stack 10a/10)

  #586·fusheng-ji opened Oct 25, 2026Contributor·

  ·
  1

• [DSV4][ROCm] Reuse upstream kernels for recent-128 KV cache

  deepseek-P2DSv4platform: rocm
  #579·inaniloquentee opened Oct 24, 2026Collaborator·

  ·
  5

• [Ascend] Fix deterministic attention backward compilation

  Ascend
  #578·fangfangssj opened Oct 23, 2026·

  ·
  2

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

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

  ·
  1

• [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·

  ·
  1

• [WS1][CUDA][MiniMax-H3] runtime modes, ablation overlay, Diffusers adapter, training provider, LoRA gate (RFC #420 stack 6/7)

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

  ·
  1

• [WS1][CUDA][MiniMax-H3] ws1_full_h3_transformer and ws1_dual_trajectory: 33B forward/backward and rollout-vs-train control (RFC #420 stack 5/5)

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

  ·
  1

• [WS1][CUDA][MiniMax-H3] token_refiner_block: two refiner blocks over the strict operator rows (RFC #420 stack 4/5)

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

  ·
  1

• [WS2][CUDA][MiniMax-H3] ws2 tensor / sequence / context parallel and the hierarchical cross-node collective (RFC #420 stack 3/5)

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

  ·
  1

• [WS1][CUDA][MiniMax-H3] projection rows: audio in/out, text context and QKV on a fixed-order GEMM (RFC #420 stack 2/5)

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

  ·
  4

• [WS1][MiniMax-H3] h3_arch_fingerprint, h3_operator_trace, h3_trajectory_artifact: pinned architecture, strict-run ledger and sealed trajectory (RFC #420 stack 1/5)

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

  ·
  1

• H3 sigma schedule

  MiniMax-H3new-model
  #566·fxyan opened Oct 21, 2026Contributor·

  ·
  10

• [DPSK-V4-P3]: T05 unified RoutePlan assembler, fingerprints and sealed handoffs

  deepseek-P3DSv4new-model
  #565·hihaluemen opened Oct 21, 2026Contributor·

  ·
  1

• [WS1][CUDA][MiniMax-H3] Add FP32 video input projection

  MiniMax-H3new-modelplatform: cudastatus: needs-reviewWS1
  #562·walker-ai opened Oct 19, 2026·

  ·
  10

• [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·

  ·
  13

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

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

  ·
  12

  

• [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·

  ·
  3

• [DSv4][P7-T05] Megatron training readback, fail-closed gate and cross-rank evidence (CUDA SM90)

  deepseek-P7DSv4platform: cuda
  #556·Yuteng-Chen opened Oct 15, 2026·

  ·
  8