Pull Requests vllm-project/vllm

Revert "[Bugfix] Fix RMSNorm kernels to multiply in weight's native dtype" (#42379) bug documentation ready verified

#44088 opened 2026-05-31 02:06 by vllm-agent

fix(spec_decode): ensure draft model uses correct parallel/model config in all proposers speculative-decoding v1

#44086 opened 2026-05-30 23:52 by anmolxlight

[Bugfix] Quark: set moe_quant_config in QuarkW8A8Int8MoEMethod bug

#44085 opened 2026-05-30 23:51 by jashwanth-12

fix: bind structured-output grammar on Responses API when reasoning parser is active frontend

#44084 opened 2026-05-30 23:38 by anmolxlight

[Core] Avoid scheduler `RoutedExpertsManager` import unless needed ready v1

#44083 opened 2026-05-30 23:25 by njhill

[Bugfix] Cache the EAGLE/MTP lookahead block in the SWA prefix-cache mask bug v1 kv-connector

#44082 opened 2026-05-30 22:48 by ivanium

[Perf] Gate AR+RMSNorm fusion cap on the active fast all-reduce threshold

#44080 opened 2026-05-30 21:41 by A1c0r-Z

[MRV2] Remove Eagle's dedicated CUDA graph pool ready v1 nvidia

#44078 opened 2026-05-30 21:15 by LucasWilkinson

refactor(envs): migrate vllm/envs.py to pydantic-settings structured-output intel-gpu v1

#44077 opened 2026-05-30 20:10 by AnshManwani

Bump the minor-update group across 1 directory with 148 updates rocm ci/build nvidia dependencies

#44076 opened 2026-05-30 20:07 by dependabot[bot]

[ROCm][Perf] Fused MoE W4A16 HIP kernel for AMD RDNA3 (gfx1100) rocm ready ci/build verified

#44075 opened 2026-05-30 19:29 by JartX

[Core] Pluggable sleep-mode backend abstraction (RFC #34303) v1

#44074 opened 2026-05-30 18:52 by matteso1

[Bugfix] Add FlashInfer B12x to linear backend selection bug

#44073 opened 2026-05-30 18:34 by mmangkad

[Bugfix] Expand packed module names in GPTQ modules_in_block_to_quantize bug

#44072 opened 2026-05-30 18:17 by ben7am1n

Add env var for FlashInfer autotune cache documentation

#44071 opened 2026-05-30 18:05 by mmangkad

fix(config): reject negative max_logprobs (except -1) and long_prefill_token_threshold

#44070 opened 2026-05-30 17:33 by hclsys

[MRV2][DSV4] Reset FlashMLA tile metadata in prefill capture ready v1 nvidia

#44069 opened 2026-05-30 16:53 by WoosukKwon

fix: skip stale LRU cache order entries

#44068 opened 2026-05-30 15:46 by he-yufeng

docs: fix tokenizer optimization typo documentation

#44066 opened 2026-05-30 14:27 by chunyang-wen

[FlashAttention] Sync FA with upstream ready ci/build

#44065 opened 2026-05-30 12:37 by MatthewBonanni

[KV Connector] Add KDA 4-state support to NixlConnector for Kimi Linear PD disaggregation v1 kv-connector

#44064 opened 2026-05-30 12:11 by JaredforReal

Int4 kivi kv cache v1 nvidia

#44059 opened 2026-05-30 09:06 by a1exxd0

[Bugfix][Tool Parsers] Validate JSON arguments in extract_tool_calls for Kimi K2, DeepSeek V3/V3.1 bug tool-calling deepseek

#44058 opened 2026-05-30 08:59 by rahulsolanki001

[Bugfix] Reject non-positive values for ParallelConfig int knobs bug

#44057 opened 2026-05-30 08:31 by jwzheng96

fix(ngram): match async ngram_gpu acceptance rate to CPU speculative-decoding v1

#44056 opened 2026-05-30 07:49 by shiyangyang2001-lgtm

docs(nixl): document KV Transfer metrics log fields and Prometheus counters documentation kv-connector

#44055 opened 2026-05-30 07:47 by sridhar-3009

fix(ngram): sync ngram_gpu acceptance rate to match CPU v1

#44054 opened 2026-05-30 07:44 by shiyangyang2001-lgtm

[Bugfix][V1][TurboQuant] Reserve workspace before CUDA graph capture bug v1 nvidia

#44053 opened 2026-05-30 06:38 by Bot1822

Add speculative decoding metrics frontend v1

#44052 opened 2026-05-30 05:32 by naomili0924

[CI] Stabilize the multi-audio OpenAI server path frontend multi-modality

#44051 opened 2026-05-30 05:06 by AndreasKaratzas