All pull requests

Pull requests

• [HiCache] Fix DSA direct-linker assembly with skipped top-k layers

  hicache
  #43773·lihenga opened Oct 12, 2026·3 of 5 tasks·

  ·

• [Core] Add request-level customized_info keys and mark Full prefill capture batches

  run-ci
  #43772·metamergebot opened Oct 12, 2026Collaborator·3 of 5 tasks·

  ·

• [AMD] Compact attention hip drop aiter qkpv

  jit-kernelrun-ci
  #43771·michaelzhang-ai opened Oct 12, 2026Collaborator·5 tasks·

  ·
  2

• [diffusion] Compare ReduceOp by identity so the compiled TP all-reduce stays one graph on ROCm

  diffusion
  #43770·michaelzhang-ai opened Oct 12, 2026Collaborator·5 tasks·

  ·

• [diffusion] progressive resolution: fix Wan2.2 A14B and short-schedule crashes

  diffusion
  #43768·ckvermaAI opened Oct 12, 2026Contributor·5 tasks·

  ·

• [WIP] [dLLM] Fix DiffusionGemma decision prefix reliability (stack 3/3)

  documentation
  #43767·Hangzhi opened Oct 12, 2026Contributor·6 tasks·

  ·

• [AMD][CI] Drop tc_piecewise split-op cases on ROCm and run the breakable ones

  run-ci
  #43766·michaelzhang-ai opened Oct 12, 2026Collaborator·2 tasks done·

  ·
  1

• [Fix] Link cuda_ipc JIT against CUDA driver stubs dir

  jit-kernelrun-ci
  #43764·JustinTong0323 opened Oct 12, 2026Collaborator·

  ·
  1

• [Fix] Mistral: map control tokens back to ids on the multimodal text path

  #43763·matan-arkin opened Oct 12, 2026·5 tasks done·

  ·
  1

• [PD] Reuse startup decode graphs across role switches

  #43760·LLLL114 opened Oct 12, 2026Contributor·5 tasks·

  ·
  1

• [CPU] Supports TP=6 configuration for Muse models in a CPU environment.

  #43759·nzr-niu opened Oct 12, 2026Contributor·

  ·

• [diffusion] bind worker CPUs to the GPU's NUMA node when memory policy is not permitted

  diffusionrun-ci
  #43756·mickqian opened Oct 12, 2026Collaborator·2 tasks done·

  ·
  1

• [AMD] Reject balancedness reporting with the MegaMoE prefill-only EPLB recorder (fixes prefill hang)

  #43755·Duyi-Wang opened Oct 12, 2026Collaborator·2 of 3 tasks·

  ·

• [EPLB] Add --eplb-max-rebalance-count to freeze the expert layout after K rebalances

  documentation
  #43754·Duyi-Wang opened Oct 12, 2026Collaborator·3 tasks done·

  ·

• [Fix] Gemma 4 vision pooler: mask padding patches and scale in float32

  Multi-modal
  #43753·JonSnow1807 opened Oct 12, 2026·4 of 5 tasks·

  ·

• [dLLM] Support DiffusionGemma decision APIs (stack 2/3)

  documentation
  #43752·Hangzhi opened Oct 12, 2026Contributor·

  ·

• [Fix] Make tokenizer request-state initialization atomic

  #43751·ch-wan opened Oct 12, 2026Collaborator·3 tasks done·

  ·

• [diffusion] Fix dynamic-batching bugs: non-output-rank split, OOM memory climb

  diffusion
  #43750·ckvermaAI opened Oct 12, 2026Contributor·5 tasks·

  ·
  3

• [Bugfix] Reject unflushed ReplaySSM cache checkpoints

  memory-poolunified-radix-cache
  #43749·alphabetc1 opened Oct 12, 2026Collaborator·5 tasks done·

  ·

• [NPU] Initialize padded vision attention outputs to prevent MiniCPM NaNs

  Multi-modal
  #43748·McZyWu opened Oct 12, 2026Contributor·4 tasks done·

  ·

• [Diffusion] MiniMax H3: allow the AdaLN cache with a quantized DiT

  diffusiondocumentation
  #43745·SSHdotCodes opened Oct 12, 2026·5 tasks done·

  ·

• [Spec] Add aux hidden-state capture for Granite and GraniteMoe targets

  #43744·ZN1010 opened Oct 12, 2026·4 of 5 tasks·

  ·

• [Spec] Fix NGRAM over-sampling token 0 from padded draft siblings

  jit-kernel
  #43742·samridh10exe opened Oct 12, 2026·5 tasks done·

  ·
  1

• 🚧 [WIP][llm][npu] Split mixed chunked-prefill FIA into prefill/decode calls on Ascend 950

  documentationjit-kernelnpu
  #43740·TallMessiWu opened Oct 12, 2026Contributor·4 of 5 tasks·

  ·

• 🚧 [WIP][llm][npu][quant] Add W4A8 MXFP quantization support for Qwen3.5 MoE on Ascend NPU

  documentationjit-kernelnpuquant
  #43739·TallMessiWu opened Oct 12, 2026Contributor·

  ·