All pull requests

Pull requests

• [AITER_MXFP4_BF16 MOE backend] Temporarily fix correctness issues with Swiglu + no bias code path

  quantizationrocm
  #59766·fxmarty-amd opened Oct 2, 2026Contributor·

  ·

• [Test] Re-enable Ovis2.5 and Ovis2.6-MoE vLLM tests on Transformers v5

  multi-modality
  #59763·AbhayShuklaIIT opened Oct 2, 2026·6 tasks done·

  ·
  2

• [Misc] Remove code paths for Transformers < 5.16.1

  mistralmulti-modalityreadyspeculative-decodingtool-calling
  #59762·hmellor opened Oct 2, 2026Member·5 of 6 tasks·

  ·
  5

• [Perf][MoE] Fold routed_scaling_factor into output transform weights

  #59761·shantipriya-amd opened Oct 2, 2026Contributor·6 tasks·

  ·

• [ROCm] Optimize MiniMax-M3 SwiGLU activation with prefill-gated clamp…

  minimaxrocm
  #59760·amd-bishwoadhikari opened Oct 2, 2026·

  ·
  1

• [Bugfix][Core] Retire stale Mamba spec blocks in prefill checkpoint steps

  bugkv-cache-manager
  #59759·simon-veitner-redhat opened Oct 2, 2026Contributor·6 tasks done·

  ·

• [KVConnector][NIXL] Pair push-mode requests by router transfer_id

  documentationkv-connector
  #59758·YannikHinteregger opened Oct 2, 2026Contributor·

  ·
  1

  1

• [MoE Refactor] Call monolithic MoE kernels from one place in FusedMoEMethodBase

  quantization
  #59757·aoshen02 opened Oct 2, 2026Contributor·

  ·

• [Bugfix][Tool Parser] Emit FunctionGemma tool calls that arrive in a single streaming delta

  bugtool-calling
  #59754·stanleys12 opened Oct 2, 2026·

  ·
  1

• [Perf][Qwen4Exp] Add SM121 TP=1 LM-head skinny-GEMM plans

  qwen
  #59753·stecasta opened Oct 2, 2026Contributor·

  ·

• [Bugfix][Quark] Pass grouped-routing arguments to OCP MX monolithic kernels

  bugquantizationready
  #59752·aoshen02 opened Oct 2, 2026Contributor·

  ·
  2

• [Bugfix][Attention] Restrict FlashInfer sparse MLA FULL graphs to decode

  bugdflashnvidia
  #59751·gty111 opened Oct 2, 2026Collaborator·5 of 6 tasks·

  ·
  1

• [Bugfix] Validate tool-parser/tokenizer compatibility at startup

  bugtool-calling
  #59749·www6v opened Oct 2, 2026·

  ·
  1

  1

• [Bug fix]: Fix streaming reasoning detokenization for DeepSeek R1 (by…

  bugdeepseektool-calling
  #59748·sudhanshu112233shukla opened Oct 2, 2026·

  ·
  1

• [Bugfix][NIXL] Reject push handshakes between different P/D block sizes

  bugkv-connector
  #59747·lucamotz opened Oct 2, 2026Contributor·

  ·

• [Bugfix][NIXL] Pair P/D blocks by the last token with different block sizes

  bugkv-connector
  #59746·lucamotz opened Oct 2, 2026Contributor·

  ·

• [Rust Frontend] Remove the PyO3 tool-parser bridge

  ci/buildrusttool-calling
  #59744·BugenZhao opened Oct 2, 2026Member·2/2·6 tasks done·

  ·

• [Frontend] Port the MiniMax M3 tool parser to the parser engine

  minimaxtool-calling
  #59743·BugenZhao opened Oct 2, 2026Member·1/2·6 tasks done·

  ·
  2

• [Bugfix][Frontend] Validate tool parser against tokenizer at startup

  bug
  #59742·lalit10 opened Oct 2, 2026Contributor·

  ·
  1

  1

• [MRV2][Spec Decode] Reduced draft vocabulary for MTP drafters

  documentationmrv2qwenspeculative-decoding
  #59740·stecasta opened Oct 2, 2026Contributor·6 tasks done·

  ·
  1

• [ROCm][Attention] MXFP4 KV cache with native FP8 x FP4 sparse MLA attention for GLM-5.3-Flash

  glmrocm
  #59739·amd-dlimpus opened Oct 2, 2026·5 of 6 tasks·

  ·
  1

• [Perf] Enable PDL for fused MoE top-k gating

  #59737·wolfcomos opened Oct 2, 2026·6 tasks done·

  ·
  1

• [Bugfix][Rust Frontend] Fail token-less streams in vllm-bench openai-chat

  bugrust
  #59736·ZhenchengLin opened Oct 2, 2026Contributor·3 of 4 tasks·

  ·
  1

• [Perf] Keep non-speculative GDN decode on the standard path

  #59735·DimensionSTP opened Oct 2, 2026·6 tasks done·

  ·
  6

• [Doc] Document NVLS non-determinism on Hopper NVSwitch nodes

  documentation
  #59734·www6v opened Oct 2, 2026·

  ·
  1

  2