All pull requests

Pull requests

• [ROCm][CI] Disable RCCL RAS for the MI355 DP + EP test

  ci/buildrocm
  #60727·Rohan138 opened Oct 8, 2026Collaborator·6 tasks done·

  ·

• [Test] Allow one fp8 step on K bytes in the kpool decode fuzz test

  #60725·Rohan138 opened Oct 8, 2026Collaborator·6 tasks done·

  ·

• [Test] Allow one bf16 ulp in the fused mHC post/pre residual check

  #60724·Rohan138 opened Oct 8, 2026Collaborator·6 tasks done·

  ·
  4

• [Frontend] Accept each connection on the least loaded API worker

  frontendready
  #60723·hao-aaron opened Oct 8, 2026Contributor·6 tasks done·

  ·
  10

• [Misc] Consolidate csrc shared headers and prune dead code

  ci/buildcpudocumentationk3kiminvidiarocmtorch.compile
  #60721·mgoin opened Oct 8, 2026Member·6 tasks done·

  ·
  1

• Use NCCL2 peer groups and reuse singleton world group

  #60720·tushar00jain opened Oct 8, 2026Contributor·

  ·
  1

• [Bugfix] Reject --convert embed/classify with non-pooling runners at config time

  bugpooling
  #60719·vadiklyutiy opened Oct 8, 2026Member·6 tasks done·

  ·
  1

  1

• [Bugfix] Return vocab rank instead of list position for logprob_token_ids

  bugmrv2rust
  #60718·vadiklyutiy opened Oct 8, 2026Member·2 of 6 tasks·

  ·
  1

  1

• [ROCm][Kimi-K3] Use vLLM's FP8 max in the fused AttnRes quantization

  k3kimiquantizationreadyrocm
  #60717·Rohan138 opened Oct 8, 2026Collaborator·6 tasks done·

  ·
  4

• [Bugfix][Kernel] Restore FP fusion for indexer Q quantization

  bugdeepseekDSv4quantization
  #60716·warrendeng opened Oct 8, 2026·6 tasks done·

  ·
  1

  2

• [CI][Test] Allow 10% ngram acceptance drop under preemption on CUDA in test_async_scheduling_accuracy

  nvidia
  #60715·atalman opened Oct 8, 2026Contributor·

  ·

• [Bugfix] Reject max_num_scheduled_tokens=0 at config time instead of hanging

  bugready
  #60714·vadiklyutiy opened Oct 8, 2026Member·2 of 6 tasks·

  ·
  1

  4

• [Feature] Add Winnow read strategy to the Decisions API

  documentationfrontend
  #60713·Txoka opened Oct 8, 2026·6 tasks done·

  ·
  3

• [Mooncake] Allow Mamba cache mode none for capacity-only workers

  kv-connectormooncake
  #60712·wzhao18 opened Oct 8, 2026Contributor·

  ·

• [Bugfix] Let W4A8 MoE fall back to Humming for non-256-aligned shapes

  bugnvidiaquantization
  #60711·ischencheng opened Oct 8, 2026·6 tasks done·

  ·
  1

  1

• [ROCm][Perf] W4A16: opt-in dense weight cache for RDNA hybrid prefill

  rocm
  #60710·zohaibm-amd opened Oct 8, 2026·6 tasks done·

  ·
  1

• [Bugfix] Fix InternLM2 tool parser dropping characters from streamed arguments

  bugreadytool-calling
  #60709·vadiklyutiy opened Oct 8, 2026Member·6 tasks done·

  ·
  1

  5

• [HiSparse] Write mixed decode/prefill attention outputs in place

  nvidia
  #60707·MatthewBonanni opened Oct 8, 2026Member·

  ·

• [HiSparse] Bound prefill staging temporaries by chunk, not prompt length

  #60706·MatthewBonanni opened Oct 8, 2026Member·

  ·

• [Bugfix][CPU] Avoid Triton JIT stall in _ranks_kernel on CPU

  bugmrv2
  #60705·null-Exception1 opened Oct 8, 2026Contributor·5 of 6 tasks·

  ·

• [Bugfix] Fix ReplaySSM ring shape assertion when TP does not divide Mamba2 n_groups

  bug
  #60704·vadiklyutiy opened Oct 8, 2026Member·6 tasks done·

  ·
  1

  1

• [Bugfix][Benchmark] Support boolean sweep plot filters for equality comparisons (#60629)

  bugperformance
  #60702·VimalN2005 opened Oct 8, 2026·

  ·
  1

  1

• [Bugfix][Benchmark] Reject expired sweep readiness probes

  bugperformance
  #60699·rezacr588 opened Oct 8, 2026·6 tasks done·

  ·
  1

  1

• [Bugfix] Use native memory fence for SHM broadcasts

  bugci/build
  #60697·Jayaragul opened Oct 8, 2026·6 tasks·

  ·
  3

• [Frontend] structured decisions: noul and score question types

  documentationfrontend
  #60696·mmastrac opened Oct 8, 2026Contributor·5 of 6 tasks·

  ·
  5