All pull requests

Pull requests

• [Spec Decode] Keep the trailing prefix-cache block for hybrid models

  #61911·qimcis opened Oct 15, 2026·2 of 6 tasks·

  ·
  1

• [Bugfix][Models] Fix Olmo-Hybrid GDN fused conv1d weight loader and register mapper (#61849)

  bug
  #61910·patrickswedish opened Oct 15, 2026·

  ·
  1

  1

• [Frontend][Perf] Tokenize long prompts in parallel pieces

  #61909·maeehart opened Oct 15, 2026Contributor·6 tasks done·

  ·

• feat(quantization): add AWQ-v2 64KB DMA page-aligned zero-repack loader and specification

  quantization
  #61908·NeonMaw opened Oct 15, 2026·

  ·
  1

• [Bugfix] Return HTTP 400 when local media is disabled without --allowed-local-media-path

  bugmulti-modality
  #61907·patrickswedish opened Oct 15, 2026·5 of 6 tasks·

  ·
  1

  1

• [Bugfix][LoRA] Fix garbage/NaN lora_expand output on SM90 with fp32 i…

  bug
  #61906·BeichenHuang opened Oct 15, 2026·6 tasks done·

  ·
  1

  1

• [CI] Shard the AMD mirrors of Model Executor, Basic Models (Other), Batch Invariance and PCP automatically

  ci/buildrocm
  #61904·Thangnguyenvn98 opened Oct 15, 2026Contributor·

  ·
  2

• [Perf][Kernel] Add a BF16x3 router GEMM for fp32 router weights on Hopper

  ci/build
  #61903·ischencheng opened Oct 15, 2026Contributor·6 tasks done·

  ·
  1

• [ROCm][Perf] Fuse Gemma add+RMSNorm into the AITER MXFP4 GEMM's activation quant

  documentationquantizationrocmtorch.compile
  #61902·fyun89 opened Oct 15, 2026·1 of 6 tasks·

  ·
  2

• [Bugfix][LoRA] Validate dtype at config time for Triton LoRA kernels

  bug
  #61901·VrajPatel105 opened Oct 15, 2026·5 of 6 tasks·

  ·
  1

  1

• [Perf] Make DeepEP high-throughput SM count configurable

  #61900·ibrahimnd2000 opened Oct 15, 2026·

  ·
  1

• [Bugfix][Metrics] Aggregate DP engine samples in ORCA load header

  bugfrontend
  #61899·samayp01 opened Oct 15, 2026Contributor·6 tasks done·

  ·
  1

• [Feature][MoE] Support routed experts capture (R3) with pipeline parallelism

  documentationmrv2needs-rebaseray
  #61897·chenjr97 opened Oct 15, 2026·5 of 6 tasks·

  ·
  4

• [ROCm][Perf] Any-shape low-latency MoE router GEMM on MI355

  ci/buildrocm
  #61895·AndreasKaratzas opened Oct 15, 2026Member·6 tasks done·

  ·
  2

• [CI][ROCm] Split MI300 AITER MoE combinations by FP8 quant type

  ci/buildquantizationrocm
  #61894·aarushjain29 opened Oct 15, 2026Collaborator·

  ·

• [KV Connector] Per-connector "enable_kv_events" switch for KV cache events

  documentationkv-connectormooncakescheduler
  #61893·zhewenl opened Oct 15, 2026Collaborator·2 of 6 tasks·

  ·
  1

• [Bugfix][Rust Frontend] Use xgrammar for gRPC structured outputs

  bugrust
  #61892·alec-flowers opened Oct 15, 2026Contributor·6 tasks done·

  ·

• Fix GPTQ quantizations loading zero point parameters in symmetric configs

  quantization
  #61891·xiaotia1 opened Oct 15, 2026·6 tasks·

  ·
  1

• [Bugfix] Fix FP8 NaN from per-call unit scales in ChannelWiseTorchFP8ScaledMMLinearKernel

  bug
  #61890·nevertmr opened Oct 15, 2026·6 tasks done·

  ·
  2

• [ROCm] Fix MiniMax-M3 sparse PA KV block count

  minimaxrocm
  #61889·ykamiset opened Oct 15, 2026Contributor·6 tasks·

  ·

• [Distributed] Opt-in FP16 partial sums for tensor-parallel row-parallel layers

  nvidia
  #61888·ValerioDolci opened Oct 15, 2026·6 tasks done·

  ·
  1

• [ROCm] Rebuild rocprofiler-sdk with the split-lock fix

  ci/buildrocm
  #61887·amd-mghanimi opened Oct 15, 2026Contributor·6 tasks done·

  ·
  1

• [torch.compile] Migrate ObservabilityConfig to hash_factors

  torch.compile
  #61885·ahujan456 opened Oct 15, 2026·

  ·
  1

• [Bugfix] Reject LoRA at config time for models without SupportsLoRA

  bug
  #61884·YaashuDave opened Oct 15, 2026·6 tasks done·

  ·
  1

  1

• [Chore] Use upstream Transformers InklingProcessor

  inkling
  #61883·hmellor opened Oct 15, 2026Member·6 tasks done·

  ·