All pull requests

Pull requests

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

  #562·walker-ai opened Oct 19, 2026·

  ·
  2

• docs: sync blog sidebar navigation

  #561·Ziyi0723 opened Oct 18, 2026·

  ·
  1

  2

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

  #560·ryankert01 opened Oct 16, 2026Collaborator·

  ·
  3

  

• feat(adaln): add shared-profile modulation backends and validation

  #559·hongleng opened Oct 16, 2026Contributor·

  ·
  1

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

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

  ·
  10

  

• [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

• Feat/nemotron h lm head logprob cuda

  Nemotronnew-modelplatform: cuda
  #555·effintell opened Oct 15, 2026·

  ·
  1

• [MiniMax-H3][CUDA] Add gate/up and attention output projection candidates

  MiniMax-H3new-modelplatform: cuda
  #554·zzf2024 opened Oct 14, 2026·

  ·
  1

• [WS1][CUDA][Qwen3-Next] Comparison benchmarks and plots for the Qwen3-Next stack: MoE, attention, logprob, EP (RFC #428 stack 11/11)

  #551·fusheng-ji opened Oct 12, 2026Contributor·

  ·
  1

• [WS1][CUDA][Qwen3-Next] TP distributed logprob parity: fixed reduction order, selected-token logprob exact across NCCL settings (RFC #428 stack 9/11)

  new-modelplatform: cudaqwenWS1
  #549·fusheng-ji opened Oct 12, 2026Contributor·

  ·
  9

• feat(ppo): add Qwen3 BF16 adapters and vLLM TP qualification

  component: executors
  #548·fusheng-ji opened Oct 11, 2026Contributor·

  ·
  1

  20

• feat(cuda): enable recursive deterministic LogP autograd on B200

  platform: cuda
  #547·fusheng-ji opened Oct 11, 2026Contributor·

  ·
  1

• Add Qwen3-8B single-node 1P1D support

  PD Disaggregationqwen
  #546·inaniloquentee opened Oct 11, 2026Collaborator·

  ·
  7

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

  #545·ThreeMonth03 opened Oct 11, 2026·

  ·
  1

• [WS1][CUDA][Qwen3-Next] Batch-invariance matrix for the full model, 97/97 cells bitwise (RFC #428 stack 8/11)

  new-modelplatform: cudaqwenWS1
  #544·fusheng-ji opened Oct 11, 2026Contributor·

  ·
  1

• [WS1][CUDA][Qwen3-Next] Full-model chain: vLLM rollout vs VIME recompute, raw-bit through two updates and a reload (RFC #428 stack 7/11)

  new-modelplatform: cudaqwenWS1
  #543·fusheng-ji opened Oct 11, 2026Contributor·

  ·
  1

• [WS1][Qwen-Image][CUDA/ROCm] Qwen-Image MLP up projection + tanh GELU (mlp_up_gemm_gelu)

  new-modelplatform: cudaqwenWS1
  #541·luoyueyuguang opened Oct 11, 2026Contributor·

  ·
  1

• [WS1][CUDA][MiniMax-H3] ws1_one_h3_block + full-model qualification: block 0 node-by-node and 50-block/generation gates (RFC #420 stack 10/10)

  MiniMax-H3new-modelplatform: cudaWS1
  #539·fusheng-ji opened Oct 11, 2026Contributor·

  ·
  1

• [WS1][Qwen-Image][CUDA] Final AdaLN output (final_adaln_out)

  new-modelplatform: cudaqwenWS1
  #538·fusheng-ji opened Oct 11, 2026Contributor·

  ·
  1

• [WS1][Qwen-Image][CUDA] AdaLN modulation projection (adaln_mod_projection)

  new-modelplatform: cudaqwenWS1
  #537·fusheng-ji opened Oct 11, 2026Contributor·

  ·
  1

• feat(platforms): persistent kernel auto-tuning in the HAL (#50)

  component: kernelsplatform: cudatype: performance
  #536·fusheng-ji opened Oct 11, 2026Contributor·

  ·
  1

  20

• [MiniMax-H3][CUDA] Add deterministic FFN down projection candidate

  MiniMax-H3new-modelplatform: cuda
  #535·zzf2024 opened Oct 10, 2026·

  ·
  4

• [WS1][CUDA][Qwen3-Next] TP4 GDN, convolution and D=256 attention mixers on real weights (RFC #428 stack 6/11)

  new-modelplatform: cudaqwenWS1
  #507·fusheng-ji opened Oct 10, 2026Contributor·

  ·
  1

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

  component: kernelsnew-modelqwenWS1
  #506·yanbao1217 opened Oct 10, 2026·

  ·
  1