All pull requests

Pull requests

• [WS1][CUDA][MiniMax-H3] Add batch-invariant full attention

  #553·RichApple123 opened Oct 13, 2026·

  ·
  1

• docs: display blog posts in a responsive card grid

  #552·ninghaolu opened Oct 13, 2026·

  ·
  1

  1

• [WS1][CUDA][Qwen3-Next] Separate logprob comparison benchmarks and plots from TP validation

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

  ·
  1

• [WS1][MiniMax-H3] Add video patch pack/unpack

  #550·walker-ai opened Oct 12, 2026·

  ·
  4

• [WS1][CUDA][Qwen3-Next] TP distributed logprob parity: fixed reduction order, selected-token logprob exact across NCCL settings (RFC #428 stack 9/10)

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

  ·
  8

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

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

  ·
  1

  16

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

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

  ·
  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 8/8)

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

  ·
  1

• [WS1][CUDA][Qwen3-Next] Full-model chain: vLLM rollout vs VIME recompute, raw-bit through two updates and a reload (RFC #428 stack 7/8)

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

  ·
  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

• test(swiglu): match reference layout in Ascend autograd wrapper test

  bug
  #540·fusheng-ji opened Oct 11, 2026·

  ·
  2

• feat(h3): extend block validation to full-model and generation qualification

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

  ·
  1

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

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

  ·
  1

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

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

  ·
  1

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

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

  ·
  1

  20

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

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

  ·
  3

• [WS1][CUDA][Qwen3-Next] TP4 GDN, convolution and D=256 attention mixers on real weights (RFC #428 stack 6/8)

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

  ·
  1

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

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

  ·
  1

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

  new-modelqwenWS1
  #505·zzf2024 opened Oct 10, 2026·

  ·
  1

• [Nemotron][CUDA] Add deterministic moe_router_dispatch

  #504·RichApple123 opened Oct 10, 2026·

  ·
  1

• [WS1][CUDA][Qwen3-Next] MoE route/combine contract with a batch-invariant grouped expert GEMM (RFC #428 stack 5/8)

  new-modelplatform: cudaqwenWS1
  #503·fusheng-ji opened Oct 9, 2026·

  ·
  1

• Support deterministic Megatron sequence parallelism for G11 consistency

  #500·inaniloquentee opened Oct 9, 2026Collaborator·

  ·
  1

• [WS1][Qwen-Image][CUDA/ROCm] Qwen-Image MLP down projection (mlp_down_gemm, CUDA)

  new-modelplatform: cudaqwenWS1
  #499·luoyueyuguang opened Oct 8, 2026Contributor·

  ·
  1