All pull requests

Pull requests

• [Perf][MRV2][Mamba2] Fuse aligned metadata CUDA graph inputs

  #65182·RishiParsai01 opened Nov 2, 2026·7 tasks done·

  ·
  1

• [Bugfix][Quantization] Respect non-gated shard count in FP4 Marlin

  bugquantization
  #65181·HDCharles opened Nov 2, 2026Contributor·6 tasks done·

  ·

• [Rust Frontend] Implement the scale-out render routes on the main server

  rust
  #65180·vineoy opened Nov 2, 2026·

  ·
  1

  1

• [Perf][DCP] Sparse indexer: read the rank-major DCP top-k gather in place

  #65178·mickaelseznec opened Nov 2, 2026Contributor·2 of 6 tasks·

  ·

• [Bugfix][Quantization] Index INT8 MoE per-channel weight scales independently of act quant - #58570

  bugquantization
  #65177·fxmarty-amd opened Nov 2, 2026Contributor·

  ·
  1

  3

• [Bugfix][Spec Decode] Keep the rejection-sampling residual nonzero in fp32

  bugmrv2speculative-decoding
  #65176·shiannn13 opened Nov 2, 2026·4 of 6 tasks·

  ·
  1

  1

• [Spec Decode] XPress: feed the online acceptance estimator

  documentationmrv2qwenspeculative-decoding
  #65175·tomasruizt opened Nov 2, 2026Contributor·

  ·
  1

• [Bugfix][Kimi-VL] Set pad_token_id explicitly in KimiVLConfig

  bugkimi
  #65174·kazutoiris opened Nov 2, 2026·4 of 6 tasks·

  ·
  1

• [Bugfix][ROCm] Muse Spark: ensure shared RoPE cos_sin_cache is on CUDA for fused kernel

  bugnvidiarocm
  #65170·akii96 opened Nov 2, 2026Contributor·

  ·

• [Bugfix][Reasoning] kimi_k3: return the reply as content when the think channel is skipped

  bugk3kimitool-calling
  #65169·Jasen2201 opened Nov 2, 2026Contributor·

  ·

• [Fast start] Pluggable shared stores for daemon warmup caches

  documentationfrontend
  #65168·Isotr0py opened Nov 2, 2026Member·6 tasks·

  ·
  1

• [V1]_Support_multimodal_decoder_DBO_inputs

  mrv2
  #65167·sudhanshu112233shukla opened Nov 2, 2026Contributor·

  ·
  1

• [Quantization] Extend int8_per_channel_weight_only to LinearBase.

  documentationquantization
  #65165·menogrey opened Nov 2, 2026Contributor·5 of 6 tasks·

  ·
  9

• [Fusion] Residual stream for allreduce manual fusion

  ci/buildk3kimiqwentorch.compile
  #65164·ElizaWszola opened Nov 2, 2026Contributor·

  ·

• [Bugfix][Rust Frontend] Stop accepting gRPC connections when shutdown starts

  bugrust
  #65163·theminjunchoi opened Nov 2, 2026·

  ·
  1

  2

• [Build] Fetch triton_kernels from a source tarball instead of a git clone (avoiding +400 MB pull)

  ci/build
  #65162·fxmarty-amd opened Nov 2, 2026Contributor·3 tasks·

  ·

• [Feature][MooncakeStore] Add automatic deployment cache prefixes

  ci/builddocumentationkv-connectormooncake
  #65160·hj-mistral opened Nov 2, 2026Contributor·5 of 6 tasks·

  ·
  1

• [Bugfix][Scheduler] Drain in-flight stale output on sync KV load-failure recovery (#65131)

  bugscheduler
  #65159·Aravind0403 opened Nov 2, 2026·

  ·
  1

  1

• [CI] Add check-docstring-style pre-commit hook and use single backticks in docstrings

  agentsbeambuild-docscoherecpudeepseekdflashDSv4DSv4.1frontendglmgpt-osshisparseinklingintel-gpuk3kimikv-cache-managerkv-connectorminimaxmistralmooncakemrv2multi-modalityneeds-rebasenixlnvidiaperformancepoolingquantizationqwenrayrocmschedulerspeculative-decodingstructured-outputtool-callingtorch.compilevllm-ir
  #65158·hmellor opened Nov 2, 2026Member·6 tasks done·

  ·
  1

• [Bugfix][Frontend] Close the parent's listening sockets after launching the Rust frontend

  bugrust
  #65156·theminjunchoi opened Nov 2, 2026·

  ·
  1

• [Bugfix][XPU] Make layer_norm batch invariant under VLLM_BATCH_INVARIANT

  bugintel-gpu
  #65154·tzielinski-habana opened Nov 2, 2026Contributor·

  ·

• [Bugfix][Step5] Size the KV block to the CSA tail ring

  bug
  #65153·akii96 opened Nov 2, 2026Contributor·5 of 6 tasks·

  ·

• [Kernel] Enable Step5 SwiGLU in TRTLLM MoE

  documentationnvidiaquantization
  #65152·jeejeelee opened Nov 2, 2026Member·6 tasks·

  ·
  5

• [Bugfix][Rust Frontend] Preserve null DeepSeek tool descriptions

  bugdeepseekdocumentationDSv4.1rust
  #65150·emerardd opened Nov 2, 2026·6 tasks done·

  ·
  1

  2

• [Quantization][Fusion] Extend manual activation fusion to more NVFP4 linear backends

  ci/buildnvidiaquantization
  #65149·jzakrzew opened Nov 2, 2026Contributor·6 tasks·

  ·