All pull requests

Pull requests

• [Doc][ROCm] Align Kimi-K3 MI355X recipe with InferenceX #4215

  k3kimirocm
  #62273·chunfangamd opened Oct 18, 2026Contributor·6 tasks done·

  ·
  1

• [Core] Per-request auxiliary readouts in generation (MRV2)

  documentationmrv2scheduler
  #62272·sylvanus4 opened Oct 18, 2026·6 tasks done·

  ·
  2

• [Bugfix][Tool Parser] Fix invalid streamed JSON for control characters in hy_v3 tool args

  bugtool-calling
  #62271·Pranavv-dev opened Oct 18, 2026·6 tasks done·

  ·
  1

• [ROCm][Perf][MiniMax-M3] Enable the fp8 indexer cache on gfx942

  minimaxrocm
  #62270·akii96 opened Oct 18, 2026Contributor·

  ·

• [Bugfix][DSv4.1] Keep the Engram all-to-all buffers in one allocator

  bugdeepseekDSv4.1
  #62269·Juntian777 opened Oct 18, 2026Contributor·6 tasks done·

  ·

• [Bugfix] Guard Conch W4A16 preconditions and pad non-tile-multiple outputs

  bug
  #62268·vllmellm opened Oct 18, 2026Contributor·5 of 6 tasks·

  ·

• [Doc][RL] Document the weight-transfer HTTP protocol for trainers without vLLM

  documentationready
  #62267·aoshen02 opened Oct 18, 2026Contributor·6 tasks done·

  ·
  4

• [Bugfix][Distributed] Register custom AR graph buffers only after a capture that did not raise

  bug
  #62266·kimseunghyun-kr opened Oct 18, 2026Contributor·6 tasks done·

  ·

• [Bugfix][Attention] Let b12x pick its default KV block size

  bug
  #62265·VidyasagarDudekula opened Oct 18, 2026·6 tasks done·

  ·
  1

  1

• [Bugfix][Tool Parser] Coerce only canonical JSON number spellings

  bugtool-calling
  #62263·he-yufeng opened Oct 18, 2026Contributor·

  ·
  1

• [Perf] Avoid quadratic output text copies in the incremental detokenizer

  #62262·chudinovAI opened Oct 18, 2026·6 tasks done·

  ·
  1

• [Tool Parser] Fix truncated/unclosed DSML parameters and add incremental DSML arg streaming

  deepseekDSv4.1tool-calling
  #62259·deep1blue opened Oct 18, 2026·

  ·
  1

  1

• [Benchmark][Spec Decode] Add oracle-drafter benchmark for the speculative decoding speedup ceiling

  performance
  #62258·jiangLLM opened Oct 18, 2026Contributor·6 tasks done·

  ·

• [Docs] Add ROCm decode megakernel (MonoKernel) authoring guide

  documentationrocm
  #62256·mpashkovskii opened Oct 18, 2026Contributor·

  ·
  1

• [Refactor][KV Connector][Mooncake] Move connector metadata and wire structs to their own modules

  kv-connectormooncake
  #62255·CAICAIIs opened Oct 18, 2026·6 tasks done·

  ·
  1

• [Bugfix][Tool Parser] Preserve non-canonical numeric strings when schema allows string

  bugtool-calling
  #62254·ninglinLiu opened Oct 18, 2026·5 of 6 tasks·

  ·
  1

  1

• [Bugfix][Frontend] Preserve scheduling priority for online beam search

  bugfrontend
  #62253·dskwe opened Oct 18, 2026Contributor·6 tasks done·

  ·

• [Bugfix][KV Connector] Support unequal P/D TP for hybrid Mamba models

  bugkv-connectormooncake
  #62251·zyulai opened Oct 18, 2026·6 tasks done·

  ·
  3

• [Perf][Pooling] Batch SPLADE max-pooling projections

  pooling
  #62250·taneem-ibrahim opened Oct 18, 2026Contributor·6 tasks done·

  ·

• [Bugfix] Respect max_model_len for Clef SystemOne reads

  bugdocumentationfrontend
  #62249·webcodes-cz opened Oct 18, 2026·5 of 6 tasks·

  ·
  2

• [Model] Support box and ground decisions with GUI-Actor

  documentationfrontendmulti-modalitynew-modelpooling
  #62248·ningyx25 opened Oct 18, 2026·6 tasks done·

  ·
  1

  2

• [Bugfix][Structured Output] Honor resolved backends in mixed requests

  bugstructured-output
  #62246·webcodes-cz opened Oct 18, 2026·2 of 6 tasks·

  ·
  1

• [XPU] Enable spec-decode adaptive verification without a confidence head

  intel-gpuspeculative-decoding
  #62244·yma11 opened Oct 18, 2026Contributor·

  ·

• [Bugfix][KV Connector] Report every transfer group as loaded in Decod…

  bugkv-connector
  #62242·Juntian777 opened Oct 18, 2026Contributor·

  ·

• [ROCm][Bugfix] Release vLLM's communicators on external-launcher shutdown

  bugrocm
  #62241·AndreasKaratzas opened Oct 18, 2026Member·5 of 6 tasks·

  ·
  2