Commit Graph

Select branches

Hide Pull Requests

main

patches

sm86

Mono Color

• d169c5a903 [Docs] Add disagg PD flow-example guidance to /pr-checklist skill (#60111) Nicolò LucchesiandMistral Vibe 2026-10-08 00:16:05 +02:00
• f38c6778ce [quantization] Add per-token NVFP4 quantization MoE support for ReLU2 (#56740) 3 people 2026-10-08 06:01:09 +08:00
• b803edbc9f [Model] Streamline EmbeddingGemma 2 config resolution and harmonize Triton prefill attention (#60289) 3 people 2026-10-07 18:56:40 -03:00
• 947dd62c07 [CI] Fix failing test_deepseek_v4_rocm_dspark* tests (#60462) Simon Danielsson 2026-10-07 23:18:19 +02:00
• b47ec440bc [CI] Stamp provenance labels on the torch-nightly image so Initialized Snapshot E2E can run (#59487) Thang NguyenandKimi 2026-10-07 17:15:39 -04:00
• a982c81e3e [Bugfix] Limit GlmMoeDsaForCausalLM fp8 KV cache default to SM100 (#60281) Misha GoinandClaude Opus 5.5 2026-10-07 16:37:13 -04:00
• 08567505b0 [Bugfix][KV Offload] Restore KVCR construction through the secondary-tier factory (#58088) Rui 2026-10-08 03:36:57 +08:00
• e572e02d21 [UX] Skip checkpoint shards an MTP head does not load (#60152) 3 people 2026-10-07 15:16:19 -04:00
• 17c70c8953 [KV Offload] Update KVCR adapter to pool-and-index API (#59899) Moein Khazraee 2026-10-07 12:14:07 -07:00
• b9767299db [Bugfix] Respect Inductor deterministic mode in combo-kernel defaults (#59074) 3 people 2026-10-08 00:38:11 +05:30
• 32fbfa15e8 [Bugfix][Quantization] Accept folded NVFP4 scales for Humming A16 (#60337) Jinzhen LinandKevin H. Luu 2026-10-08 02:04:43 +08:00
• 6ee9be6b3c [Spec Decode] Rename AR speculators to StandaloneAR / TargetDependentAR (#60335) Giancarlo Delfin 2026-10-07 11:01:59 -07:00
• fba462e008 [Bugfix]Fix FlashInfer warmup crash: autotune(tuning_buckets=...) must round M up, not down (Invalid MXFP8 split-K tactic with spec-decode) (#58165) 3 people 2026-10-07 14:01:21 -04:00
• 554340f3d3 [Attention] Support NVFP4 KV cache on SM8x and SM12x with FlashInfer (#46963) 4 people 2026-10-08 02:24:21 +09:00
• 3d6853b32c [Bugfix] Seed hybrid mamba state index with mamba_block_size on prefix-cache hits (#55601) 4 people 2026-10-07 19:09:50 +02:00
• a8260cbc5c [CI] Bump Transformers version to 5.19.0 (#60381) Harry MellorandClaude Opus 5.5 2026-10-07 18:01:41 +01:00
• d64a18832b [ROCm][Triton] Migrating MiniMax-M3 kernels from make_block_ptr to tensor descriptors (#59285) 5 people 2026-10-07 12:49:59 -04:00
• 240ee8de9e [Bugfix][KV Connector][ROCm] Fix MoRI-IO KV block-offset for speculative decoding (MTP) (#55053) MohitAMDandCursor 2026-10-07 09:33:30 -07:00
• c741bfca70 [Docs] Add review areas for andylolu2 (#60416) Andy Lo 2026-10-07 17:05:29 +01:00
• f7f62d1916 [XPU][MoE] Tune Triton fused MoE for Intel XPU (#53065) pmanczak 2026-10-07 18:02:13 +02:00
• 9367d8b969 [Bugfix][MLA] Scope DSpark non-causal capability to builder layers (#57992) xaguilar-amdandmergify[bot] <37929162+mergify[bot]@users.noreply.github.com> 2026-10-07 17:22:32 +02:00
• 71f9a74920 [CI][ROCm] Re-enable the DSv4.1 decoder replay graph test on ROCm (#60392) stefankoncarevic 2026-10-07 16:57:52 +02:00
• 14e3902d51 [Frontend][Rust] Add gRPC forbidden token sequences and cache usage (#59837) Alec 2026-10-07 07:46:44 -07:00
• 548597367e [ROCm][Perf] Add AITER FlyDSL GDN prefill backend (#57560) Mikko TukiainenandCursor 2026-10-07 17:44:22 +03:00
• f6e3944e31 [ROCm][DSv4] AITER MegaMoEV2 Integration For DeepSeek V4 (#59685) Micah WilliamsonandFangzhou Ai 2026-10-07 09:35:53 -05:00
• c83935b300 [Bugfix][Model Loader] Support reload_weights with runai_streamer load format (#58124) 3 people 2026-10-07 17:04:35 +03:00
• 5281e49908 [ROCm][Perf] Enable medium-skinny dispatch in RDNAHybridW4A16LinearKernel (#52619) 00CC 2026-10-07 21:21:43 +08:00
• efb8d85f24 [ROCm][CI] Add FP32 rounding slack to accuracy test near-tie upper bounds (#60292) Divakar VermaandAndreas Karatzas 2026-10-07 09:19:22 -04:00
• 2a54f6b625 [ROCm][Perf][GLM-5.3-Flash] Enable AITER fused shared experts on gfx950 (#59221) 3 people 2026-10-07 15:53:40 +03:00
• d9503dc503 [Bugfix][Qwen4Exp] Derive the QSA head_dim from the unsharded head count (#59945) 3 people 2026-10-07 21:24:40 +09:00
• 3ca00a8261 [CI] Skip EmbeddingGemma2 pooling tests needing transformers >= 5.19.0 (#60371) stefankoncarevic 2026-10-07 12:22:31 +02:00
• 6b75dfb8a1 [Misc] Report a missing DeepSelect extension only when it is requested (#60242) SeongJun Lee 2026-10-07 18:53:17 +09:00
• 02d8a94e61 [CI][ROCm] Skip DSv4.1 decoder replay graph test on ROCm (#60378) Kevin H. LuuandClaude Opus 5.5 2026-10-07 02:38:31 -07:00
• df417f780a [Bugfix][MoRIIO] Exclude synchronous READ destinations from KV zeroing (#59164) 4 people 2026-10-07 16:35:56 +08:00
• 42ecc8eebe [CI][RL] Consolidate RL entrypoint tests under tests/entrypoints/rl (#59948) 3 people 2026-10-07 16:29:09 +08:00
• bc8dd4ad27 [Rust Frontend] Share SchemaRoot between argument coercion and grammars (#59408) Bugen ZhaoandClaude Opus 5.5 2026-10-07 16:28:27 +08:00
• 90f4936d9e [CI/Build] Widen the ragged prefill scoring tolerance, drop batch invariance (#60192) stefankoncarevic 2026-10-07 10:21:47 +02:00
• 7436a7f119 [Rust Frontend] Backtrack from safe text at a complete marker (#59563) Bugen ZhaoandClaude Opus 5.5 2026-10-07 15:45:15 +08:00
• e6fc81bc78 [Bugfix][Scheduler] Let pooling chunked prefill use full context (#48039) 3 people 2026-10-07 03:32:00 -04:00
• cbee82324c [Docs][RL] Move pause/resume and batch invariance examples to examples/rl (#59949) 3 people 2026-10-07 14:22:10 +08:00
• 43b4aaea3e [Bugfix][Frontend] Return 4xx for client errors in /cohere/v2/chat (#60309) SrirajandClaude 2026-10-07 10:25:34 +05:30
• aecb717f10 [CI] Allowlist-shrink batch 2: wire 9 tests/models/ files into Buildkite, fix 2 stale bugs found (#59050) Turner JabbourandClaude Sonnet 5 2026-10-06 22:16:53 -06:00
• ef014bdf06 [Bugfix][Frontend] Return logprobs in Cohere Chat v2 responses (#59072) Mohammad HijjawiandCyrus Leung 2026-10-07 05:13:52 +01:00
• 2342343746 [ROCm][DCP] Default MLA DCP verify to the round-robin asm decode where it applies (#59965) Hyukjoon LeeandClaude Opus 5.5 2026-10-07 12:10:54 +09:00
• 8a26869bca [Bugfix] Index Mamba align state copies by request slot in MRV2 (#60210) 3 people 2026-10-07 10:32:46 +08:00
• 956960f85f [Perf][MRV2] Avoid stream syncs in NonUvaBuffer host-to-device copies (#60215) 3 people 2026-10-07 07:21:12 +05:30
• 283d76d56a [Bugfix] Widen ep_gather output offset to int64 (#58373) Robert ShawandClaude Code 2026-10-06 21:45:04 -04:00
• fe51194e2d [Bugfix][MiniMax-M3] Fence decode indexer shared-memory stage reuse (#60194) 4 people 2026-10-07 03:35:12 +02:00
• 4b4eb6277a [CI] Remove AMD mirror of CPU-only Rust frontend cargo steps (#60323) Kevin H. LuuandClaude Opus 5.5 2026-10-06 18:30:52 -07:00
• 13f8c453ee [CI][Releases] Trigger perf-eval Through Release Pipeline (#52213) Micah WilliamsonandKevin H. Luu 2026-10-06 20:27:27 -05:00
• 703c29179c [TEST][XPU][CI] Generalize model runner prefill tail classification tests to all platforms (#60201) Miłosz Grunwald 2026-10-07 02:59:20 +02:00
• db6e3cd8c4 [Bugfix][Frontend] Log the full request body when max_log_len is unset (#60310) SrirajandFlora Feng 2026-10-07 06:12:17 +05:30
• 6c2925fe5f [Bugfix][Core] Preserve speculative metrics when coalescing outputs (#55708) jxjandOpenAI Codex 2026-10-07 08:35:14 +08:00
• d77cf81156 [Bugfix][Frontend] Honor default penalties in batch chat completions (#60307) SrirajandFlora Feng 2026-10-07 05:51:00 +05:30
• 5ab4e4445b [Perf][MoE] Skip padded work in block-FP8 DeepGEMM experts (#59128) 3 people 2026-10-06 18:49:57 -05:00
• b6d8e8afd9 [Bugfix][MRV2][Spec Decode] Accept num_speculative_tokens in NgramGPUSpeculator.propose (#60300) Cheese CakeandClaude Opus 5.5 2026-10-07 03:32:04 +05:30
• ae19af883a [CI] Require transformers 5.19 for the EmbeddingGemma2 registry test (#60295) 3 people 2026-10-06 14:50:10 -07:00
• 385d86a6cf [Attention] Add UltraQuant 4-bit KV cache backend (FlyDSL D=256) (#57057) aditi-amd 2026-10-06 13:42:52 -07:00
• 819852df6b [ModelRunner V2] Speculative Decoding NGram GPU Implementations (#40704) 5 people 2026-10-07 04:32:59 +08:00
• dac35c5b66 [Misc] Add chat-completion options to gsm8k_eval.py CLI (#60114) Benjamin ChislettandClaude Opus 5.5 2026-10-06 16:13:44 -04:00
• e43db1f5e2 Pad to 64 not 128 Q heads in flashMLA sparse (#60029) PatrykSafferandMisha Goin 2026-10-06 22:03:05 +02:00
• d1f3d8b870 [Quantization] Prefer Humming before Marlin backends on SM90 (#56997) 4 people 2026-10-07 03:31:35 +08:00
• bb87d227d4 [Bugfix][Model] Fix mypy safe-super error in EmbeddingGemma2Model (#60280) Roger WangandKimi Code 2026-10-06 15:10:00 -04:00
• e37e51dd24 [Metrics] Expose cached prompt tokens by cache tier (#56318) 4 people 2026-10-06 13:52:05 -05:00
• ad0f67a7ec [Bugfix][Spec Decode] Keep heterogeneous-vocab draft models on Model Runner V1 (#59541) Nishant RevurandClaude 2026-10-06 11:03:44 -07:00
• 3403e0f176 [Bugfix] Persist FlashInfer autotune cache per rank to fix the rank-0-only cache-hit deadlock (#57635) 3 people 2026-10-06 13:51:01 -04:00
• 02b83919aa [Model] Support EmbeddingGemma2 multimodal pooling architecture (#60254) Luciano MartinsandLuciano Martins 2026-10-06 14:43:38 -03:00
• 7b665f7a58 [Frontend] Integer token IDs for generate output logprobs (GenerateLogProbs) (#58181) 3 people 2026-10-06 21:22:47 +04:00
• e82b80099a [Perf][HiSparse] Cache per-request residency instead of rescanning every page each step (#60083) Matthew BonanniandClaude Opus 5.5 2026-10-06 13:04:55 -04:00
• fba31f31a9 [GLM5.3 Perf] Switch to fp8 kv cache by default for GLM, 2.3%~5.5% E2E Throughput Improvement (#60140) Wentao Ye 2026-10-06 12:54:20 -04:00
• f8d92cae41 [Quantization] Layer re-quantization for linear layers through online quantization API (MXFP8 -> FP8 PTPC showcase) (#55684) 4 people 2026-10-06 18:38:18 +02:00
• 29f955cce0 [ROCm] Fuse MLA dual RMSNorm + FP8 group quant for DeepSeek-R1 (#54857) eky-amdandClaude Fable 5 2026-10-06 12:26:38 -04:00
• 4516896626 [CI/Build] Warm the DP engines before measuring load balance in test_load (#60241) stefankoncarevic 2026-10-06 18:23:37 +02:00
• fcf53d240c [Spec Decode] Remove eager metadata rebuild during MTP fused multi-step decode (#58463) Giancarlo Delfin 2026-10-06 09:17:30 -07:00
• 4ea0c28bc4 [Bugfix] Warn only on per-request do_normalize/do_rescale overrides (#60240) SeongJun Lee 2026-10-07 00:55:27 +09:00
• aad4579bbe Upgrade tpu-inference to v0.31.0 (#60190) Jeff-Tseng-dev 2026-10-06 15:44:56 +00:00
• cc73cca9c8 [Build] Migrate vendored DeepGEMM from pybind to TORCH_LIBRARY (abi3) (#48962) Chris LeonardandShengqi Chen 2026-10-06 11:39:04 -04:00
• cbe1f9740d [NIXL] Bump NIXL version to 1.5.0 (#56907) ovidiusm 2026-10-06 16:55:03 +02:00
• 0719975cee [CI][Rust Frontend] Pass --locked to cargo binstall (#60182) Bugen ZhaoandClaude Opus 5.5 2026-10-06 22:53:17 +08:00
• 8bd7737a03 [Docs] Clarify that vLLM does not isolate tenants sharing a server (#59811) 3 people 2026-10-06 08:39:59 -06:00
• 4a99ca40d6 [Perf][MoE] Eliminate staging from NCCL symmetric reduce-scatter (#49194) Samuel NordmannandMisha Goin 2026-10-06 16:37:01 +02:00
• bf049ddb8f [Docs] Run pr-checklist before opening a PR (#60235) Jiangyun ZhuandClaude Opus 5.5 2026-10-06 22:12:18 +08:00
• b267fe4881 [ROCm][Perf] W4A16: pad gfx11 weight and activation strides (#56301) Matthias Gehre 2026-10-06 15:22:33 +02:00
• 049507aa76 [Bugfix][Structured Output] Flag multi-branch allOf as unsupported for xgrammar (#59061) Artem PerevedentsevandShashank Varma 2026-10-06 16:01:04 +03:00
• 1a86733c7b [Docs] Add model recipes skill (#57786) Roy Wang 2026-10-06 20:40:10 +08:00
• 3faf214051 [vllm-bench, feature] Added support for Mooncake-style, timed-traces replay. (#55937) Rodrigo Garcia 2026-10-06 14:39:35 +02:00
• 25385102fe [Bugfix][KV Offload] Honor speculative cacheability in SimpleCPU (#60071) Sherif Waly 2026-10-06 13:25:08 +01:00
• 3e182185aa [Deprecation] Deprecate sonet dataset as scheduled (#60129) Wentao Ye 2026-10-06 08:15:12 -04:00
• 3313825983 [Bugfix][DiffusionGemma] Read the causal mask as int32 and compile the sample step (#59992) Mia Ojeda 2026-10-06 05:25:51 -06:00
• 09e0ce3d82 [Bugfix][Frontend] Apply model-default reasoning parser in GPU-less render server (#54835) 3 people 2026-10-06 03:46:24 -07:00
• 68088ed392 Fix grammar in BlockPool.reset_prefix_cache docstring: "to invalid" → "to invalidate" (#59886) zhangchengzhucufe-dev 2026-10-06 18:19:54 +08:00
• 888074b42d [CPU][Recipes] Detect model head constraints for automatic tensor parallel selection in vLLM Recipes Tool (#60170) Louie Tsai 2026-10-06 03:15:40 -07:00
• 6bbad6acdf [Rust Frontend] Pass tool defer_loading through to chat templates (#60203) Bugen ZhaoandClaude Opus 5.5 2026-10-06 18:00:15 +08:00
• 31e2443c90 [Model] LongCat-Flash: scale MLA norms while loading and drop the post-load sweep (#60080) aoshen02andClaude Opus 5.5 2026-10-06 16:47:02 +08:00
• e11962fc1c [Feat][Mamba2] Enable internal prefill checkpoints (#57329) roikoren755andzjy0516 2026-10-06 10:53:51 +03:00
• edfccc9469 [ROCM][Perf] Add ROCM_SEGMENTED_ATTN attention-backend works across RDNA3/3.5/4 (#59132) 4 people 2026-10-06 15:53:47 +08:00
• 15c1cc12ac [Bugfix][EPLB] Use group-local ranks for torch P2P transfers (#55804) 5 people 2026-10-06 15:44:57 +08:00
• 6ad2e06c4a fix: exclude flashinfer-jit-cache from wheel install_requires (#60052) Tarun Kumar 2026-10-06 12:01:13 +05:30
• 1e5d0ea888 [CI/Build] Fix incorrect CLI arg validation tests in test_cli_args.py (#46779) sungbin1015andClaude 2026-10-06 14:58:47 +09:00
• 40c86ae1ce [Kernel] Enable OAI Triton MXFP4 MoE on SM12x (#58877) 3 people 2026-10-06 13:49:13 +08:00