Commit Graph

Select branches

Hide Pull Requests

main

patches

sm86

Mono Color

• 295c062db8 Remove sudo from dockerfile command (#60488) Tiezhen WANGandKevin H. Luu 2026-10-08 19:24:56 -07:00
• 4611c2c3d2 [BugFix][DiffusionGemma][CPU] Copy async output snapshots instead of aliasing reused buffers (#59829) Mia OjedaandLucas Wilkinson 2026-10-08 20:16:51 -06:00
• 201e537d98 [Bugfix][LoRA] Validate incompatible MoE adapter formats (#59924) 4 people 2026-10-08 18:49:13 -07:00
• 581692cc0f [Bugfix][Spec Decode] Fix multimodal embedding merge in Llama 4 and Mistral Large 3 EAGLE drafters (#57633) Nishant RevurandClaude 2026-10-08 18:44:50 -07:00
• 4274ae956c [Bugfix][EC Connector] Release shared memory on NIXL setup failure (#59875) DGlan 2026-10-08 20:38:16 -04:00
• ab905a885d [CI] Run GH200 test on the shared arm64 CI image (#60253) 3 people 2026-10-08 18:22:20 -04:00
• 8d10c38b73 [Perf] Reduce fallback decode CPU by 69% at 32K (#60497) KamranandCodex 2026-10-09 03:12:56 +05:30
• 240785b82c [Bugfix] Keep the GLM-5.3 kpool tail and Qwen4 QSA ring out of the null block (#59528) Yifan QiaoandClaude Opus 5.5 2026-10-08 13:44:04 -07:00
• 2ddde19e8e [Frontend] Support min_p in the Responses API (#48084) sungbin1015andFlora Feng 2026-10-09 05:38:17 +09:00
• c5c1161382 [ROCm][Build] Fail fast when setup.py develop deps aren't pip-preinstalled (#59447) Divakar Verma 2026-10-08 16:28:38 -04:00
• ae9adec905 [Frontend] Expose /v1/systemone without an opt-in flag (#60677) 3 people 2026-10-08 16:21:05 -04:00
• 72d59adc2c [Bugfix][KV Connector][NIXL] Transfer the whole PLE short-conv page in disaggregated P/D (#59997) 4 people 2026-10-08 21:58:13 +02:00
• 1afb029839 [Bugfix][HiSparse] Route by swap-row capacity, not the index-group workspace (#60682) Matthew BonanniandClaude Opus 5.5 2026-10-08 15:45:44 -04:00
• c23ca06b6d [Bugfix][Structured Output] Feed the token that implicitly ends reasoning to the grammar (#59627) 4 people 2026-10-08 21:38:33 +02:00
• bab0ac61f8 [Bugfix][Frontend] Forward the reasoning wiring to the engine for batch chat completions (#59263) dingdangmaoandFlora Feng 2026-10-09 03:32:46 +08:00
• e3a29aa9b3 [Bugfix][MoE] Size the permute scratch from its inputs (#60447) Misha GoinandClaude 2026-10-08 15:32:07 -04:00
• fff10d2072 [Attention] Allow an NVFP4 KV cache on SM90 with fa2 (#60620) SeongJun LeeandMisha Goin 2026-10-09 04:16:22 +09:00
• fc494822ea [Test][Quantization] Gate CUTLASS W4A8 kernel tests to SM90 (#58878) Kevin LiandMisha Goin 2026-10-08 14:14:35 -05:00
• 18751bc6a7 [ROCm][CI] Add coverage for AiterExperts token padding: inf/nan garbage rows (#59334) Divakar VermaandAndreas Karatzas 2026-10-08 15:11:45 -04:00
• 9bbc581560 [Bugfix] Close MessageQueue zmq resources via finalizer to avoid exit (#60472) Aarushi JainandAndreas Karatzas 2026-10-08 13:53:00 -05:00
• 56e6da2c47 [Bugfix][Frontend] Honor stop_token_ids on the Responses API (#56592) Yeonwoo SungandFlora Feng 2026-10-09 03:47:30 +09:00
• c4a5ff5c18 [Frontend] shutdown engine core and its sub processes asap when initializing (#52299) Ning XieandSimon Mo 2026-10-09 02:46:01 +08:00
• 2274e86842 [Spec Decode][Model] Support DFlash2 draft models with GLM-5.3-Flash (#56983) Shijin Zhang 2026-10-09 02:41:10 +08:00
• 30c7405890 [MoE] MXFP4 oracle: add CUTLASS W4A4 backend and BF16 activation fallback (#60079) Misha GoinandClaude 2026-10-08 14:39:09 -04:00
• d8fa463e2b [Bugfix][Prefix caching] Fix inconsistent hybrid Mamba prefix cache boundaries (#60533) 6 people 2026-10-08 14:33:48 -04:00
• 1664c4f444 [Bugfix][MiniMax-M3] Guard sparse decode sentinel indices (#54597) 4 people 2026-10-08 11:27:57 -07:00
• 972195f4f5 [CI][Build] build macos wheels using abi3 (#60312) TialoandMisha Goin 2026-10-08 21:19:20 +03:00
• 705e124c05 [Bugfix] Followup to #59299: preserve Unicode in structured decision state prompts (#60651) Francisco Javier ArceoandMisha Goin 2026-10-08 14:00:03 -04:00
• 242e4213fc [Doc] Require GCC 13 for CUDA source builds (#60259) Doug Smith 2026-10-08 12:43:39 -04:00
• 8674a794c7 [Perf] Enable Kimi K3 shared expert sharding with DeepEP v2 (#56875) Elvir CrnčevićandMisha Goin 2026-10-08 18:42:15 +02:00
• f4dde3132c [CI][ROCm] Reuse AITER's prebuilt MoE module in the modular-kernel harness (#60633) stefankoncarevic 2026-10-08 18:18:05 +02:00
• 9f3d70e081 [Quantization] Enable shared expert fusion compatibility with online shared_expert quantization (showcase: along Quark MXFP4 routed experts) (#55686) 3 people 2026-10-08 18:04:46 +02:00
• e2414ec1db [ROCm][Qwen4Exp] Use fused PLE Triton kernels on the AMD backend (#60021) Mikko TukiainenandCursor 2026-10-08 18:47:56 +03:00
• 0fc0cefb4e [Bugfix] Don't int-sort attention-type names in --kv-cache-dtype-skip-layers for packed KV dtypes (#60364) Justin BarlowandHarry Mellor 2026-10-08 09:34:55 -06:00
• e41d5d5543 [Bugfix] Don't warn about missing watermark config for default requests (#60412) 3 people 2026-10-08 17:29:20 +02:00
• 0751f5272f [Bugfix] Scope deprecated CLI argument warnings to the parser that defines them (#59917) Haokai MaandClaude 2026-10-08 08:29:04 -07:00
• 194bd72be7 [Compile] Remove torch.compile-based sequence parallelism and AsyncTP (#60517) Woosuk Kwon 2026-10-08 08:28:23 -07:00
• 072d49bb6e [Platform] Support platform-specific MLA prefill backend selection (#56959) wangxiyuanandmergify[bot] <37929162+mergify[bot]@users.noreply.github.com> 2026-10-08 23:25:21 +08:00
• fd928e5f3c [Bugfix] Check batch-invariance support for every attention backend (#60271) 3 people 2026-10-08 08:19:57 -07:00
• 994b8b77be [Bugfix][ROCm] RDNA3 W4A16 MoE: handle the N-first CT weight layout (#60388) JartX 2026-10-08 17:06:39 +02:00
• d80ddaac73 [ROCm][Perf] Kimi-K3 Store only the current rank's shards in latent MoE up-proj (#59591) kliuae 2026-10-08 22:55:20 +08:00
• d33872dfeb [Feat] Structured decisions endpoint (/v1/systemone) (#59299) 6 people 2026-10-08 08:48:12 -06:00
• 273e099693 [Bugfix][ROCm] Select page-aligned kernel blocks for pooled indexers so the block table addresses storage pages (GLM-5.3-Flash) (#59412) yhcheong (vllmellm) 2026-10-08 22:44:58 +08:00
• 2f7f05c256 [Bugfix][MoRIIO] Prevent remote KV reload after decoder preemption (#53250) zzaebok 2026-10-08 22:28:16 +08:00
• daa9085143 [KV Connector][NIXL] Register HiSparse host pool under single TP Rank (#60161) 3 people 2026-10-08 19:55:02 +05:30
• 4ed4c4433b [Bugfix][HiSparse] Drop the unaccounted, write-only prefill mirror staging buffer (#60429) Matthew BonanniandClaude Opus 5.5 2026-10-08 10:00:49 -04:00
• 194da61de1 [Bugfix] Reject negative --device-ids indices instead of selecting the wrong device (#49053) Vineeth Sai VarikuntlaandHarry Mellor 2026-10-08 06:01:25 -07:00
• 458ba2edf8 [ROCm][Kimi-K3] Fuse AttnRes output with per-token FP8 input quantization (#59069) 3 people 2026-10-08 20:01:47 +08:00
• 73c742b09c [Bugfix] Make disable_any_whitespace actually disable whitespace on xgrammar (#58067) 3 people 2026-10-08 04:41:14 -07:00
• b49c921a27 [Bugfix][Rust Frontend] Always include token_ids in completion stream choices (#60286) devshah-cohereandBugen Zhao 2026-10-08 07:11:55 -04:00
• ba77c4c130 [Bugfix][Attention] Preserve small FP8 softmax weights in Triton attention (#60156) Andy Lo 2026-10-08 11:35:22 +01:00
• e161a0798c [Perf][Core] Reduce free KV cache queue overhead (#60269) Nicolò Lucchesi 2026-10-08 12:17:11 +02:00
• 459764104d [Frontend] Move offline_utils.py to entrypoints/common. (#58052) wang.yuqi 2026-10-08 17:56:05 +08:00
• e73895a5df [Bugfix] Make suppress_stdout redirect fd 1 instead of sys.stdout.fileno() (#59336) Muhammad Usman MateenandMisha Goin 2026-10-08 10:52:50 +01:00
• dfaabf91ac [XPU] [CI] Fix stale ignore path for weight_transfer_metrics tests on Intel CI (#60558) Chaojun Zhang 2026-10-08 17:48:47 +08:00
• 24be8c0bca [CI] Remove dead in-tree TPU (torch_xla) CI scripts (#60591) Kevin H. LuuandClaude Opus 5.5 2026-10-08 02:42:07 -07:00
• e265b77982 [CI][Mamba] Use well-conditioned prompts for FlashInfer ReplaySSM parity tests (#60515) Jiangyun ZhuandClaude Opus 5.5 2026-10-08 17:39:56 +08:00
• 7816afdb25 [Bugfix][Core] Drain async KV loads in pause(mode="wait"), and pause the drain test only once a request is in flight (#60572) 3 people 2026-10-08 11:33:41 +02:00
• c4cd88d9fe [CI] Make (H100) Distributed DP + EP optional (#60540) Kevin H. LuuandClaude Opus 5.5 2026-10-08 02:31:13 -07:00
• 21d9758459 Remove Python 3.10 support now that it is EOL (#60402) Harry MellorandClaude Opus 5.5 2026-10-08 10:30:52 +01:00
• 0004536dd0 [XPU][Graph] Replace boolean-mask assignment with torch.where (#60529) liuzhenweiandKunshang Ji 2026-10-08 17:13:48 +08:00
• 70bdce6133 [Model][GLM-5.3-Flash] Warm up the DCP top-k merge kernels for the kpool indexer (#60032) Nicolò Lucchesi 2026-10-08 11:05:52 +02:00
• 7d47ac2ddb [CI] Enroll five steps in automatic sharding (#60492) 3 people 2026-10-08 04:52:59 -04:00
• 4cf1307e72 [XPU] Add tuned Mamba SSU configs for 2 more B70 shapes (#57565) Paweł Mańczak 2026-10-08 10:36:05 +02:00
• 0f112d1800 [Bugfix][Pooling] Preserve cache_salt for chunked embeddings (#47696) Ting SUN 2026-10-08 09:33:28 +02:00
• 29ae81fe83 [Rust Frontend] Preserve explicit zero sampling values in gRPC requests (#60115) Alec 2026-10-08 00:27:47 -07:00
• fd0216d3b5 [Bugfix][CPU] Fix fused Gumbel-max wrapping-window bias (#60221) Rehan KhanandBob 2026-10-08 12:45:53 +05:30
• e58f8dde7f [Bugfix] Use FusedMoEWithLoRA for shared MoE LoRAs on 3D models (#60098) ARAVINDHAN Tandlinitra24 2026-10-08 00:04:32 -07:00
• 032d39e9b4 [XPU][CI] Split Intel jobs to improve B50 utilization (#58817) xiangdong 2026-10-08 14:53:37 +08:00
• 814301b268 [Bugfix][Quant] Read safetensors headers of checkpoints with other file names (#60411) stefankoncarevicandAndreas Karatzas 2026-10-08 08:38:41 +02:00
• ee406adf0c [Bugfix] Fix ZeroDivisionError in Qwen3-VL video sampling when the source fps is unknown (#60501) Vadim GimpelsonandClaude 2026-10-08 10:28:44 +04:00
• f226af0e11 [Bugfix] Eliminate frontend ZMQ port TOCTOU with inherited listeners (#54113) 3 people 2026-10-08 14:25:27 +08:00
• fb2ac82458 [ROCm][DSv4.1][Perf] Replace the Torch prefill combine with the shared combine_topk_swa_indices kernel (#56638) 5 people 2026-10-07 23:15:18 -07:00
• 95307ac6d4 [Doc] Update ViT CUDA graph design doc (#60519) Shanshan Shen 2026-10-08 14:00:58 +08:00
• 816173ca9d [Kernel] Remove WAR for LL CuTeDSL BF16 GEMM (#60434) Jiahan Chang (Cyrus) 2026-10-08 13:38:07 +08:00
• 81198e97ba [Refactor] Abstract DeepGemm MegaMOE backend (#58053) 3 people 2026-10-08 13:05:39 +08:00
• 4ac0179f8a [ROCm][Perf] Use native merge_attn_states kernel on gfx942/gfx950 (#60159) 3 people 2026-10-08 06:49:49 +02:00
• 775792f2d5 [Bugfix] Reland #53020: tie lm_head.weight for Nemotron Parse when checkpoint omits it (#60272) 3 people 2026-10-07 21:40:55 -07:00
• 4187ade7b8 [Frontend] Consolidate rl entrypoints. (#57849) 3 people 2026-10-08 12:37:11 +08:00
• 1fbbfa1657 [Bugfix][Kimi-K3] Declare max_tp_shards on the DSpark MLA KV cache spec (#59733) 4 people 2026-10-08 13:33:32 +09:00
• 5dd4a628dd [ROCm][Bugfix] Map sleep-mode cuMem memory for the host (#60367) MattandClaude Opus 5.5 2026-10-07 21:26:12 -07:00
• 8352b2427f [Multimodal] Optimize encoder cudagraph padding logic (#58925) Isotr0pyandShanshan Shen 2026-10-08 11:38:09 +08:00
• 87d9996abe [ROCm][Perf][GLM-5.3-Flash] BF16 splitk kernel for sparse MLA decode (#58584) Simon DanielssonandDouglas Lehr 2026-10-08 04:37:51 +02:00
• 3722645539 [Bugfix][Frontend] Accept successful 2xx responses for batch output uploads (#60507) bnkaiteamandFlora Feng 2026-10-08 11:32:35 +09:00
• caff2b44cb [BugFix] Pick the DP world-group port at bind time via the coordination store (#51018) 4 people 2026-10-08 10:30:35 +08:00
• a4b14e57b6 [Platform] Allow BackendEnum dynamic register (#41091) wangxiyuanandmergify[bot] <37929162+mergify[bot]@users.noreply.github.com> 2026-10-08 09:56:19 +08:00
• 9dba339d07 [Bugfix][ROCm] Keep MLA derived weights in place on weight reload (#60389) MattandClaude Opus 5.5 2026-10-07 18:38:06 -07:00
• 8733a21d55 [Core] Bound execute_dummy_batch RPC waits by the execute-model timeout (#60398) Nicolò Lucchesi 2026-10-08 03:22:04 +02:00
• fdfc171a74 [Bugfix][ToolParser] Stream every tool call when one delta completes several (#60354) Junyi Yaoandsfeng33 2026-10-07 18:07:35 -07:00
• 2dd887426e [Bugfix][Frontend] Do not modify caller params in LLM.score() (#59959) 4 people 2026-10-08 09:03:08 +08:00
• 9e3e37cb3c [TEST][XPU][CI] disable tests of unsupported quantization (#60383) 3 people 2026-10-08 02:56:28 +02:00
• 1995c0fcd0 [Bugfix][Frontend] Apply output_config and thinking in Anthropic count_tokens (#59180) Kang NahoonandFlora Feng 2026-10-08 09:24:20 +09:00
• 1283362e76 [CI] Add headroom and selection coverage for H200 fast lanes (#56775) Kevin H. LuuandSherlock 2026-10-07 17:17:56 -07:00
• a58bdd0d4e [Perf] JIT monitor: hook TileLang JITImpl.compile instead of __call__ (#58685) Nicolò Lucchesi 2026-10-08 02:11:59 +02:00
• 8faac3dd2a [Bugfix] [Spec Decode] DeepSeek V4/V4.1 NVFP4: keep ignored MTP/DSpark experts on MXFP4 (#59655) 3 people 2026-10-07 17:10:32 -07:00
• ed825aafc0 [Bugfix][Frontend] Fix Responses namespace aliases in named tool calling (#59723) YueshenZandsfeng33 2026-10-08 08:07:52 +08:00
• 38dc8ee500 [Bugfix] Validate tool-parser/tokenizer compatibility at startup (#59749) www6vandsfeng33 2026-10-08 07:46:41 +08:00
• d6fe5dca68 [Bugfix][Parser] Emit buffered post-reasoning text when no tool parser is configured (#58911) Dhairyashil R Gandsfeng33 2026-10-08 04:51:51 +05:30
• 4ac5a36a8e [CI/Build][NVIDIA] Build NIXL with NIXL EP from source in Rubin image (#60090) Artem Perevedentsev 2026-10-08 01:44:17 +03:00
• ef63c23d35 [Bugfix][DSv4.1] Skip SWA bounded replay for KV loads that carry the window (#59197) 3 people 2026-10-07 15:28:42 -07:00