All pull requests

Pull requests

• [WS1][Qwen-Image] Add timestep embedding MLP with sm80 evidence and sm90 validation entrypoint

  #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

• Add CUDA conditioning noise mix for H3

  #476·ThreeMonth03 opened Oct 5, 2026·

  ·
  1

• [DSv4][P2-T06] MQA attention and grouped o-proj

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

  ·
  16

• 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·

  ·
  4

• [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·

  ·
  1

• [WS1][CUDA][Qwen3-Next] Restore CUDA dispatch for rms_norm (RFC #428 stack 1/4)

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

  ·
  5

• fix(ratio_kl): isolate inactive logits before normalization

  bug
  #465·MichaelCao0 opened Oct 3, 2026Contributor·

  ·
  2

• fix(sampling): preserve per-row parameters in active chunks

  bug
  #464·MichaelCao0 opened Oct 3, 2026Contributor·

  ·
  2

• feat(gemma): add fused softcapped selected logprob

  #463·Jungle430 opened Oct 3, 2026·1 task·

  ·
  7

• feat(p6): T05 fixed-order fused MoE combine forward

  deepseek-P6DSv4
  #462·bitborne opened Oct 2, 2026Collaborator·5 tasks done·

  ·
  1

• feat(nemotron): prototype fused add RMSNorm forward and backward

  #461·Jungle430 opened Oct 2, 2026·4 tasks·

  ·
  1

• WIP(waiting in process): [DSv4][P3-T06] CUDA router backward core

  #460·KKSK-DON opened Oct 2, 2026·

  ·
  1

• [WS1][kernels] attn_out_bias_gemm for Qwen-Image (python + triton + cuda)

  new-modelplatform: cudaplatform: tritonqwenWS1
  #459·yanbao1217 opened Oct 2, 2026·

  ·
  7

• fix(cli): honor --profile when bootstrapping the runtime

  bug
  #457·Jim-jimu opened Oct 1, 2026Contributor·

  ·
  3

• fix(grpo): stabilize group variance for offset rewards

  bug
  #455·MichaelCao0 opened Sep 30, 2026Contributor·

  ·
  5

• [WS1][kernels] Add fixed-order joint-attention softmax for Qwen-Image

  #453·BPbruce opened Sep 30, 2026·6 of 8 tasks·

  ·
  6

• fix(triton): bound deterministic GEMM reduction grid on CUDA

  Nemotronplatform: triton
  #450·joielab opened Sep 28, 2026·

  ·
  2

• [DSv4][P5-8] Shared Expert TP column/row, SP reduce-scatter, shared-once

  deepseek-P5DSv4
  #449·jyizheng opened Sep 28, 2026·

  ·
  3

• feat(p6): T01 operator contracts, recorded providers and sealed start kit

  deepseek-P6DSv4
  #448·Flink-ddd opened Sep 27, 2026Collaborator·

  ·
  1

• feat(p2): T01 contract, sealed start kit and H100 conformance

  deepseek-P2DSv4
  #447·inaniloquentee opened Sep 27, 2026Collaborator·

  ·
  1

• [DSv4][P1-2] fp32_gemm_rms

  deepseek-P1DSv4platform: cudaplatform: triton
  #443·Junyifannnn opened Sep 24, 2026·

  ·
  2