All pull requests

Pull requests

• [Fix] Match DeepSeek-V4.1 KV capacity to physical layouts

  #41275·ischencheng opened Sep 25, 2026·

  ·

• [sglang-miles] Recover disk-delta checkpoints safely across resets and retries

  #41274·zianglih opened Sep 25, 2026Contributor·

  ·

• [Scheduler] Add --chunked-prefill-reserve-tokens

  documentation
  #41273·francescov1 opened Sep 25, 2026·

  ·

• [diffusion] fix: LoRA-wrapped linears crash Qwen-Image and MiniMax-H3 inference

  diffusionlora
  #41272·Rockdu opened Sep 25, 2026Contributor·

  ·

• [cherrypick from #41235] [PD] Keep the sampling mask of a replayed rebootstrap token

  #41271·nanjiangwill opened Sep 25, 2026Contributor·

  ·

• [cherrypick from #40986] [Sampling] Stream sampling masks as per-request arrays

  #41270·nanjiangwill opened Sep 25, 2026Contributor·

  ·

• [Spec][DP] Skip the per-rank eager-draft fallback for a missing DSA seed under DP attention

  #41269·zidanehuang001 opened Sep 25, 2026·

  ·

• [sgl-router] Add --tokenizer-backend fast and --tokenizer-l1-cache-mb

  deepseekdocumentationnpurouterrun-ci
  #41268·sherlockwu opened Sep 25, 2026Collaborator·4/4·

  ·

• [Fix] Honor temperature in SequenceClassification setwise scoring

  #41267·vmanvs opened Sep 25, 2026·

  ·

• [Diffusion] Reuse bit-exact packed SwiGLU for Ming-Image

  diffusion
  #41266·cuzmi opened Sep 25, 2026·

  ·

• Support MLA DCP with Mooncake HiCache L3

  documentationhicache
  #41265·ByronHsu opened Sep 25, 2026Collaborator·

  ·

• [MLX] Compute prefill logits for the last position only

  apple-silicon
  #41264·luke-edward opened Sep 25, 2026·

  ·

• [mem_cache] Free a chunked prefill's out-of-window SWA slots before sizing its next chunk

  #41263·bill-h-lin opened Sep 25, 2026·

  ·

• [XPU] Install sglang-kernel-xpu from sgl-kernel-xpu main branch

  dependencies
  #41262·arathi-hlab opened Sep 25, 2026Collaborator·

  ·

• [PD] fix: cache resumed decode-radix requests from root instead of an unlocked re-match

  run-ci
  #41261·alphabetc1 opened Sep 25, 2026Collaborator·

  ·
  1

• [Frontend] Parallelize long chat prompt encoding

  #41259·xiaofei-zheng opened Sep 25, 2026·

  ·
  1

• [AMD] Let the GLM DSA NextN draft declare its own shared-expert fusion architecture

  #41258·xiaofei-zheng opened Sep 25, 2026·

  ·
  1

• [Refactor] Choose a dense layer's boundaries under attention DP from both sides' declarations

  bypass-fail-fastparallel-stagesrun-cirun-ci-extra
  #41257·ch-wan opened Sep 25, 2026Collaborator·

  ·

• [Refactor] Pick the two-batch-overlap split's layout moves once and remove execute

  #41256·ch-wan opened Sep 25, 2026Collaborator·

  ·

• [Refactor] Run the LayerNorm SP region's boundary steps in the communicator itself

  #41255·ch-wan opened Sep 25, 2026Collaborator·

  ·

• [Refactor] Declare at construction the layers whose FFN completes its own reduction

  deepseek
  #41254·ch-wan opened Sep 25, 2026Collaborator·

  ·

• [Refactor] Drive an FFN exit's flags and its completion from one selection

  #41253·ch-wan opened Sep 25, 2026Collaborator·

  ·

• [Refactor] Choose prepare_attn / prepare_mlp steps and fused kernels at construction

  deepseek
  #41252·ch-wan opened Sep 25, 2026Collaborator·

  ·

• [Perf] Optimize DeepSeek V4.1 Flash dense, indexer, mHC and MoE paths on Hopper

  deepseekjit-kernelrun-ci
  #41251·BBuf opened Sep 25, 2026Collaborator·

  ·

• Fix crash when top_logprobs exceeds vocabulary size

  #41250·dajiaohuang opened Sep 25, 2026·

  ·
  1