All pull requests

Pull requests

• [Bugfix][Rust Frontend] Prevent stalled engines from delaying auto-aborts

  bugrust
  #61623·reidliu41 opened Oct 14, 2026Contributor·6 tasks·

  ·

• [Bugfix] Read step-start seq_lens without holding the CUDA driver context lock

  bugnvidia
  #61622·maorv opened Oct 14, 2026·6 tasks done·

  ·
  1

• [Feat][Quantization] Online linear NVFP4 to MXFP4

  documentationquantization
  #61620·yma11 opened Oct 14, 2026Contributor·6 tasks·

  ·
  1

• [Bugfix][NIXL] Fix handshake crash for SSM-only models

  bugkv-connector
  #61619·zhouyou9505 opened Oct 14, 2026Contributor·

  ·
  1

• [Bugfix][LoRA] Keep failed activations retryable

  bug
  #61618·LeslieWylie opened Oct 14, 2026·6 tasks done·

  ·
  3

• [Performance] Enable small-head FlashMLA sparse prefill for DS V4.1 on SM90

  deepseekDSv4.1
  #61615·chaunceyjiang opened Oct 14, 2026Contributor·6 tasks done·

  ·

• [Perf][GLM-5.3-Flash] Produce dense QKV directly in prefill convolution

  ci/buildcpuglm
  #61614·qianlihuang opened Oct 14, 2026Contributor·6 tasks·

  ·

• [Bugfix] LLM.__repr__ crashes on non-torch worker models

  bug
  #61613·Asthenia0412 opened Oct 14, 2026·

  ·
  1

• [Bugfix] Guard FlashInfer seq_lens_cpu_upper_bound shortcut with async_scheduling

  bugnvidia
  #61612·Asthenia0412 opened Oct 14, 2026·2 tasks done·

  ·
  1

• [Bugfix] Honor code_revision in config-based tokenizer loading

  bug
  #61611·dskwe opened Oct 14, 2026Contributor·6 tasks done·

  ·
  4

• [Attention][DCP] Support DCP on the FlashMLA sparse bf16 KV path

  #61609·isztldav opened Oct 14, 2026·6 tasks done·

  ·
  1

• [XPU] Add Qwen4Exp (Qwen3.8-Flash-Next) model architecture support

  intel-gpumrv2quantizationqwenspeculative-decoding
  #61607·zhaohb opened Oct 14, 2026·

  ·
  3

• [Bugfix][MoE] Handle empty routing in moe_permute

  bug
  #61606·BESzbing opened Oct 14, 2026·5 of 6 tasks·

  ·
  1

  1

• [Bugfix][LoRA] Reject inconsistent checkpoint factor ranks

  bug
  #61605·LeslieWylie opened Oct 14, 2026·5 of 6 tasks·

  ·
  2

• [Quantization][INC][XPU] Add online NVFP4 to AutoRound MXFP4 conversion

  intel-gpuquantization
  #61602·jl9876 opened Oct 14, 2026Contributor·

  ·

• [Perf][Qwen4Exp] Internal prefill checkpoints for the PLE short conv

  ci/buildqwen
  #61601·ZJY0516 opened Oct 14, 2026Member·6 tasks done·

  ·
  1

• [CPU][Bugfix] Respect expert map in unquantized MoE

  bugcpuquantization
  #61599·AnkitNakhawa opened Oct 14, 2026Contributor·6 tasks done·

  ·
  1

• Fix shutdown order: drain engines before stopping API servers with --api-server-count > 1

  frontend
  #61598·Asthenia0412 opened Oct 14, 2026·

  ·
  1

  2

• [ROCm] Default AITER on for CDNA3+ when the package is installed

  rocm
  #61595·ChuanLi1101 opened Oct 14, 2026Collaborator·5 tasks·

  ·
  1

• Enable gelu_tanh activation support for Gemma4 MoE

  cpu
  #61593·akote123 opened Oct 14, 2026·6 tasks·

  ·
  2

• [Core][Spec Decode] Lookahead block hashes for EAGLE prefix caching

  k3kimikv-cache-managerkv-connectorscheduler
  #61589·ZJY0516 opened Oct 14, 2026Member·2 of 6 tasks·

  ·
  5

• [Security] Reject local verifier references in Speculators metadata

  documentation
  #61587·jperezdealgaba opened Oct 14, 2026Contributor·6 tasks done·

  ·
  1

• [Bugfix] Reserve construction-time scratch across workspace slots

  bugnvidiarocm
  #61584·maithilijoshi20 opened Oct 14, 2026Contributor·6 tasks done·

  ·

• [Bugfix][XPU] Skip Triton allocator registration for TD scaled_mm on XPU

  bugintel-gpuquantization
  #61583·zufangzhu opened Oct 14, 2026Contributor·

  ·
  2

• [DO NOT MERGE][ROCm] Narrow down the elastic EP IPC failure on MI355

  ci/buildreadyrocm
  #61582·stefankoncarevic opened Oct 14, 2026Collaborator·6 tasks·

  ·
  11