Commit Graph

Select branches

Hide Pull Requests

main

patches

sm86

Mono Color

• e7a5bdb454 [ROCm][CI] Move Spec Decode Speculators + MTP to MI355 (#62053) djramic 2026-10-16 20:22:38 +02:00
• ae8f41238c [Bugfix][Parser] Honor a prompt-closed reasoning block in non-streaming parse and reasoning token counts (#61857) alberto 2026-10-16 17:54:42 +01:00
• cffc0cfa31 [CI] Fix B200 FusedMoE Layer Kernels hang on NVSHMEM IB setup (#61695) 3 people 2026-10-16 12:00:12 -04:00
• 706791d516 [CI][AMD] Move 17 test groups from MI300 to MI355 and test large DeepSeek/Nemotron evals on AITER (#61955) Andreas KaratzasandClaude 2026-10-16 10:58:59 -05:00
• 76081b9ee4 [Bugfix][Model] Honor quantization config in Inkling MTP (#61921) Sawyer BowermanandCodex 2026-10-16 11:49:26 -04:00
• 1a5b187066 Add @bnellnm to CODEOWNERS (#62025) bnellnm 2026-10-16 11:16:03 -04:00
• 8e25bbca19 [Refactor] Remove dead code multiple places (#61843) Wentao Ye 2026-10-16 11:01:07 -04:00
• 8a6c0449b3 [Bugfix][ROCm][GLM-5.2] Stop forcing skip_final_all_reduce on AMD DSA model path (#58589) 4 people 2026-10-16 19:08:55 +05:30
• de7eeaa87b [Perf] Share Triton JIT warmup compilation pool across registered kernels (#61422) Ayushman Singh 2026-10-16 08:59:42 -04:00
• be95edb7f6 Merge branch 'main' into patches MOVZX 2026-10-16 17:35:42 +07:00
• 64175d3827 [Bugfix][Frontend] Keep the parser's tool call IDs in batch derender (#61852) Martin Hickey 2026-10-16 11:27:17 +01:00
• 4ab99ff136 [Bugfix][Frontend] Honor the constructor chat template in offline scoring (#61990) alpandmottopanikeiku 2026-10-16 03:20:46 -07:00
• 024c5188bf [Frontend] Resolve logprobs on the parsed streaming derender path (#60437) 4 people 2026-10-16 03:04:15 -07:00
• 9a16d43ef0 [Bugfix][Mooncake] Fix MultiConnector stats and proxy SSE responses (#61218) Hexiang Wang 2026-10-16 17:54:48 +08:00
• 8dc54f35dd [Offloader] Stream every multimodal tower parameter to host RAM MOVZXandHinata 2026-10-16 16:07:53 +07:00
• 07f3857fb9 [XPU][Linear][MXFP4] enable torch as mxfp4 linear backend on xpu (#51965) zofia 2026-10-16 17:07:05 +08:00
• 18c06ed406 [Offloader] Stream multimodal towers from host RAM with bulk-copy MOVZXandHinata 2026-10-16 14:32:14 +07:00
• fa75d3456d [Feature][PCP] Add DeepSeek-V4.1-Flash support (#59857) 5 people 2026-10-16 14:17:51 +08:00
• 074c83a09a [CI][Intel] Skip ROCm-only test_topk_gating_enabled_by_cdna on XPU (#61932) Chaojun Zhangandvllm-agent 2026-10-16 13:44:32 +08:00
• e365792997 [Dependency] Upgrade TokenSpeed MLA to 2.0.3 (#61689) Wei Zhao 2026-10-16 01:28:59 -04:00
• 49cb2bbf71 [Bugfix] Honor code_revision in config-based tokenizer loading (#61611) dskwe 2026-10-16 13:21:12 +08:00
• 4ac1f8d2ac [Bugfix][Structured Output] Mask grammar-invalid implicit reasoning ends (#61574) 3 people 2026-10-16 07:53:41 +03:00
• 92dc3dd393 [Bugfix] Honor OPT token embedding width for prompt embeddings (#61736) maithilijoshi20andCyrus Leung 2026-10-16 13:41:50 +09:00
• b466fbacd2 [ROCm][Perf] Any-shape low-latency MoE router GEMM on MI355 (#61895) Andreas KaratzasandClaude 2026-10-15 23:22:31 -05:00
• 642c18d64c [Perf] Kimi K3 Specialized Decode: CuTeDSL Optimized GB300 (1.5-2x) (#61867) 3 people 2026-10-15 22:50:41 -04:00
• 3ddc6702a8 [XPU][Docker] Bump nixl to 1.5.0 (#60356) 3 people 2026-10-15 19:35:29 -07:00
• 05ad0a8ace [Perf][DSv4.1] Split fused_wqa_wkv by token across TP ranks on SM90 (#58872) Shuolei Wang 2026-10-16 10:31:34 +08:00
• dd9e07a25f [Feature] Add External LB Mode Support for Elastic EP Scaling (#43202) 9 people 2026-10-16 09:29:02 +08:00
• fb3fec970e [Bugfix][Kernel] silu_and_mul_quant: run the vectorized loop (#61716) MonishverandKevin H. Luu 2026-10-15 21:16:53 -04:00
• 29345f42a6 [ROCm][Qwen4Exp] Support HC and MoE sequence parallelism on ROCm (#61471) Andreas KaratzasandClaude 2026-10-15 20:11:17 -05:00
• a956a8126e [Bugfix][MoE] Keep the gate's hash_indices_table out of RoutedExperts parameters (#61930) Kevin H. LuuandClaude Opus 5.5 2026-10-15 18:00:17 -07:00
• bee95149e2 Update DeepGEMM commit (#61744) Jee Jee Li 2026-10-16 08:54:40 +08:00
• 8f93c64866 [Bugfix] remove duplicate hash-routing capability method (#61924) Tessa PhamandBenjamin Chislett 2026-10-15 20:40:23 -04:00
• a8be1500ba [Bugfix][MoE] Restore TP-sharded routed-expert capture for DP=1 (#61388) 4 people 2026-10-16 08:21:24 +08:00
• ef38e8d006 [Model][Hardware][AMD][Kernel] Part 2/2 -> Enable e2e QK Norm + RoPE + KV Cache runtime fusion for Qwen3-30B-A3B on ROCM_ATTN (#42755) Jack HuandDouglas Lehr 2026-10-15 20:10:14 -04:00
• af28c171f5 [Dependency] Upgrade FlashInfer to 0.7.2 and restore KDA auto backend (#61913) Anerudhan GopalandOpenAI Codex 2026-10-15 17:04:24 -07:00
• 19b0efce30 [Perf][Core] Skip the null head of sliding-window block tables (#61724) Juntian LiuandClaude Opus 5.5 2026-10-15 16:14:51 -07:00
• 4a89316efa [Core] Drop the Ray driver-node GPU pin left over from the legacy executor (#61713) Woosuk KwonandClaude Opus 5.5 2026-10-15 16:00:18 -07:00
• 3f1e57225a [ROCm][Perf][Kimi-K3] Use AITER Gluon KDA kernels for decode and spec decode (#57392) 3 people 2026-10-16 01:54:24 +03:00
• de7d70676d [Perf] Low-latency CPU all-reduce for DP metadata syncs (#61767) 3 people 2026-10-15 15:29:13 -07:00
• 784d1fff0c [MRV2][Spec Decode] Quantized draft lm_head for drafters that share the target head (#59973) Stefano CastagnettaandKevin H. Luu 2026-10-15 15:06:40 -07:00
• d92a4f25f2 [watermarking] Add DFlash support to dual-key gumbel-max watermarking (#59554) 5 people 2026-10-16 00:06:06 +02:00
• e711723352 [ROCm] gfx1250 optimizations and model enablement (#58800) 7 people 2026-10-15 16:51:57 -05:00
• 9b99ccb56b [CI] Shard the AMD mirrors of the DP, RayExecutorV2 and batch invariance steps automatically (#61861) Thang NguyenandClaude Opus 5.5 2026-10-15 17:48:19 -04:00
• 9edcc7fa7e [Bugfix][Spec Decode] Avoid reusing random draws across block verification rounds (#61161) DimensionSTPandKevin H. Luu 2026-10-16 06:46:38 +09:00
• 0c25849292 [ROCm][Perf][Bugfix] Probe the preshuffled FP4 GEMM config by logical K (#59406) Mikko Tukiainen 2026-10-16 00:11:33 +03:00
• f8460ef236 [Kernel][DSv4.1] Bump FlashMLA to the 2026.09.30 release and its interleaved V4.1 KV cache (#61371) 4 people 2026-10-15 13:58:25 -07:00
• 33212db2c2 [KV Connector][NIXL] Reject NVFP4 KV cache with heterogeneous TP head split (#61865) 3 people 2026-10-15 22:57:37 +02:00
• c414bf0295 [CI] Shard four AMD mirrors automatically (#61817) Thang NguyenandClaude Opus 5.5 2026-10-15 16:57:35 -04:00
• 54d9b1cb48 [GLM5.3 Perf] Inplace update draft decode metadata, 1.7% E2E Throughput Improvement (#61873) Wentao Ye 2026-10-15 16:20:12 -04:00
• ee8c47961b [CI] Use an ungated model as the attention test config default (#61732) SeongJun LeeandWei Zhao 2026-10-16 05:19:09 +09:00
• 21f4b6dc84 [KV Offload] Split the CPU offload region across shm files (#59546) 4 people 2026-10-15 23:17:40 +03:00
• 16532876e7 [CI] Shard four more AMD mirrors automatically (#61860) Thang NguyenandClaude Opus 5.5 2026-10-15 16:13:51 -04:00
• d12ea8e309 [Bugfix][MRV2] Keep input_ids on later PP stages for raw-token models (#61016) 3 people 2026-10-15 20:54:21 +01:00
• d486219fc2 [Bugfix][Core] Uncount stop-trimmed spec tokens from num_computed_tokens (#58668) 3 people 2026-10-15 12:50:06 -07:00
• 49f944be3e [Perf][HiSparse] Reuse pinned host pool after graph profiling (#61668) 3 people 2026-10-15 22:33:03 +03:00
• 105605b8e9 [ROCm] Support DeepSeek-V4 hash routing in AITER Triton W4A16 MoE (#60282) Rohan PotdarandClaude Opus 5.5 2026-10-15 14:10:12 -05:00
• 82b9ccaa79 [Perf][MoE] Skip the TRTLLM fp8-block activation-scale transpose copy (#61773) Thien Tran 2026-10-16 03:09:07 +08:00
• 4d58fa254a [MoE] Unify DeepGEMM MegaMoE experts behind one feature-matrix layer; add NVFP4 (#61538) 3 people 2026-10-15 12:02:01 -07:00
• 4d9a8d6a41 [Bugfix] Make Humming DeviceInfo lookups traceable under torch.compile (#61858) Misha GoinandClaude Opus 5.5 2026-10-15 14:12:21 -04:00
• 47fd76e4ce [ROCm][Perf] Route softmax MoE top-k, including fused shared-expert sigmoid, through AITER topk_gating (#56226) Mikko TukiainenandCursor 2026-10-15 20:53:18 +03:00
• 4f49b55c1f [MoE] Shape-based padding for FlashInfer TRT-LLM MXFP8 experts (gated and non-gated) (#61293) 3 people 2026-10-15 10:44:11 -07:00
• 7cd7982dfd [Spec Decode] Don't reserve DFlash/DSpark draft slots in the scheduler token budget (#59468) 3 people 2026-10-15 10:23:58 -07:00
• 338dab077c [Bugfix] Fix Apertus streaming scientific notation (#61812) bzsuniandFlora Feng 2026-10-16 00:49:57 +08:00
• 957d512e96 [Bugfix][Frontend] Preserve streaming chunk boundaries in RequestOutputCollector (#58941) 3 people 2026-10-15 09:38:28 -07:00
• a9e90b9d1d [Bugfix][Frontend] Reject Cohere REQUIRED tool choice without tools (#61825) Jovan 2026-10-16 00:37:08 +08:00
• d6f0b7a416 [CI] Skip exited descendants when finding the worker in the FT e2e test (#61648) 3 people 2026-10-15 10:08:14 -06:00
• ed3050add4 [Core] Apply prefill_schedule_interval outside data parallelism (#54627) 4 people 2026-10-15 09:03:13 -07:00
• c825e1d235 [Perf] Fuse the mamba align block-table gather into one Triton kernel, 7 kernels -> 1, 6x speedup on all shapes for this kernel (#58737) 3 people 2026-10-15 09:02:53 -07:00
• 5d6754ff6b fix: keep the refusal-direction projection dynamo-safe MOVZXandHinata 2026-10-15 22:56:06 +07:00
• a97d19d792 [UMBP][1/10] Add shared connector, runtime contracts, and request lifecycle (#61789) CharlesWuandYichao Zhu 2026-10-15 23:12:39 +08:00
• 3dcfaeb4e8 [Frontend] Apply each endpoint's top_logprobs cut in derender (#59518) 3 people 2026-10-15 19:04:04 +04:00
• 0076adc78b [Qwen3.5] Project the refusal direction out of the residual stream at runtime MOVZXandHinata 2026-10-15 21:40:16 +07:00
• 011fa68ca9 [ROCm][Perf] Pad fp16/bf16 GEMM weight row stride for L2 cache (#55090) Matthias Gehre 2026-10-15 16:26:32 +02:00
• 76df77c04e Merge branch 'main' into patches MOVZX 2026-10-15 21:16:39 +07:00
• d12d5460a5 [Bugfix][MoE] Keep routing rows during DeepSeek V4.1 decoder replay (#61643) Ronaldandaoshen02 2026-10-15 22:02:06 +08:00
• 24f632ca7e [Bugfix] Engage indexer TP row sharding on total rows, not per-rank rows (#61647) 3 people 2026-10-15 21:54:14 +08:00
• 69425846f8 [Bugfix][CPU] Report the real cause when the worker outgrows its memory reservation (#59472) mosafariuk 2026-10-15 14:52:05 +01:00
• b7b20b1b8f [Bugfix][Frontend] Leave the final stop token out of chat derender text (#61757) 3 people 2026-10-15 06:51:34 -07:00
• 71536bd5f8 [MM][CG] Support ViT full CUDA graph for DeepSeek-VL2 (#46005) littlecircle0730andShanshan Shen 2026-10-15 06:38:52 -07:00
• 0951fc183c [ROCm][DSv4.1] Disambiguate ROCm attention class names (#58689) 3 people 2026-10-15 16:38:39 +03:00
• 1463501553 [Bugfix] Terminate streaming sessions on ABORT/ERROR and prevent zombie sessions (#58942) 3 people 2026-10-15 06:32:19 -07:00
• 766c837bd5 [Model] Support every Decision 3.0 size (d3-flash, d3-mini, d3-nano, d3-lite, d3-edge) (#61778) Xunzhuo 2026-10-15 21:29:24 +08:00
• 9cede34a67 [Bugfix] Create cuBLAS handles before memory profiling to avoid OOM (#61044) Laith SakkaandNick Hill 2026-10-15 05:59:22 -07:00
• 252be28c18 [Frontend] Replace return_mm_kwargs with an mm_payload enum on render requests (#61427) Martin Hickey 2026-10-15 13:41:46 +01:00
• c6b772cf6b [Bugfix][Frontend] Emit kv/ec transfer params at the top level of render output (#61241) Martin Hickey 2026-10-15 13:34:15 +01:00
• e8e1e96196 [ROCm][DSv4] AITER MegaMoE Fixed-Slot Dispatch For DeepSeek-V4 (#61262) Micah Williamson 2026-10-15 06:10:59 -05:00
• 8b27d27402 [Bugfix][KV Offload] Always attempt NIXL agent removal during peer teardown (#60279) 3 people 2026-10-15 16:37:23 +05:30
• 81e42c9987 [Feature][Frontend] Control the log level of request JSON body using env var (#60630) Talor Abramovich 2026-10-15 14:07:19 +03:00
• 365aaa77bd [Bugfix][DSv4] Remap Quark MTP layer_quant_config onto draft prefixes (#56054) jimmy-adams 2026-10-15 18:51:57 +08:00
• e10c10a644 [Core] Add PROMOTION_LATENCY histogram metric for tiering offload (#53910) 5 people 2026-10-15 12:47:41 +02:00
• 6fe7401fac [Perf] Read VLLM_COMPUTE_NANS_IN_LOGITS once per IterationStats (#61710) Pei LiandKevin H. Luu 2026-10-15 03:35:54 -07:00
• 716c5225ae [Model] Add AliceAI support (#57963) TimofeyandCodex 2026-10-15 13:15:03 +03:00
• 7020600b1d [KV Connector][NIXL] Register packed MLA KV rows as a single region (#61074) 4 people 2026-10-15 11:18:15 +02:00
• 80dd705c1d [Frontend] Make finish_reason and Responses item status truthful (#59324) Flora FengandChauncey 2026-10-15 05:12:02 -04:00
• 427915bed2 [Build][CPU] Install Run:ai Model Streamer in vllm-openai-cpu image (#61382) Denis KhachyanandClaude Code 2026-10-15 12:11:51 +03:00
• 7f730b4cea [CI] Run 4xH100 steps nightly only (#61782) Kevin H. LuuandClaude Opus 5.5 2026-10-15 01:32:43 -07:00
• 0f0bc4dbb1 [Test] Split the chat_utils tests. (#61765) wang.yuqi 2026-10-15 16:26:22 +08:00
• b965ee564d [Chore] Use upstream Transformers DeepseekV4Config (#61392) 4 people 2026-10-15 09:08:08 +01:00
• 1c4c319792 [Frontend] Advertise --enable-x/--disable-x for boolean CLI flags (#61479) Misha GoinandClaude Opus 5.5 2026-10-15 04:08:02 -04:00