Commit Graph
22059 Commits
Author SHA1 Message Date
MOVZX 76af434852 Merge remote-tracking branch 'upstream/main' into patches 2026-09-28 12:06:52 +07:00
af7f9488c2 [Benchmark] Add Responses API backend to vllm bench serve (#54628)
Signed-off-by: QHarshil <harshil_c@hotmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
2026-09-28 05:04:11 +00:00
Kevin H. LuuandClaude Opus 5.5 0da126676e [Bugfix] Fix the two multimodal root tests that fail on main (OpenPangu-VL embed merge, MiMo sink test fixture) (#58900)
Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-28 04:52:51 +00:00
dd3bc9c72f [Perf][MRV2] Reuse Mamba/GDN metadata across KV cache groups (#58762)
Signed-off-by: Luca Motz <luca.motz@icloud.com>
Co-authored-by: Xin Yang <xyangx@amazon.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
2026-09-28 04:36:58 +00:00
Thien TranandCodex 55de40a2fc [Minimax-M3] Add Encoder CUDA graph support (#58673)
Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: Codex <codex@openai.com>
2026-09-28 04:17:36 +00:00
Flora Feng 6f65fde419 [Bugfix][Frontend] Apply Harmony adjust_request in batched chat completions (#58958)
Signed-off-by: sfeng33 <4florafeng@gmail.com>
2026-09-28 04:12:04 +00:00
jz-yoloandjz-yolo 4a190dc518 [Model] Enable LoRA support for RobertaForSequenceClassification (#58884)
Signed-off-by: jz-yolo <jz-yolo@users.noreply.github.com>
Signed-off-by: Jane Zhu <jane.zhu@slack-corp.com>
Co-authored-by: jz-yolo <jz-yolo@users.noreply.github.com>
2026-09-28 03:50:12 +00:00
Andreas Karatzas 77fbd9e225 [Test][ROCm] Stabilize the mixed OLMoE LoRA test (#58945)
Signed-off-by: Andreas Karatzas <andreas.karatzas@protonmail.com>
2026-09-28 11:47:41 +08:00
zhao, zhenhuiandjiang1.li 5eaa50f151 [CPU] Build CPU wheels on Ubuntu 22.04 with AMX-FP8 support (#58515)
Signed-off-by: zhejiangxiaomai <zhenhui.zhao@intel.com>
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Co-authored-by: jiang1.li <jiang1.li@intel.com>
2026-09-28 11:46:02 +08:00
Tianyu Guo 53d2e16a97 [Bugfix][EPD] Skip sampling for encoder-only async steps (#58490)
Signed-off-by: Tianyu Guo <guoty@inferact.ai>
2026-09-27 20:35:09 -07:00
Flora Feng 027b6f3a22 [Bugfix][Frontend] Use a fresh parser per choice in non-streaming chat completions (#58939)
Signed-off-by: sfeng33 <4florafeng@gmail.com>
2026-09-28 02:01:35 +00:00
Flora Feng 4d07e90653 [Bugfix][Frontend] Sample batched chat completions from the adjusted requests (#58929)
Signed-off-by: sfeng33 <4florafeng@gmail.com>
2026-09-28 01:46:55 +00:00
Ashraf Bhuiyan b9e2903b89 [Mypy] Fix mypy typing for Ultravox and Unlimited-OCR models (#58239)
Signed-off-by: Ashraf Bhuiyan <mbhuiyan@redhat.com>
2026-09-27 21:19:04 -04:00
Thien Tran 39d49539de [KimiViT][Perf] Fuse per-layer QK RoPE into one in-place kernel (#58651) 2026-09-27 18:18:06 -07:00
Flora Feng 3184226984 [Bugfix][Frontend] Count Responses reasoning tokens per tool round (#58927)
Signed-off-by: sfeng33 <4florafeng@gmail.com>
2026-09-28 00:34:38 +00:00
31afa4f1bb [ROCm] Bump AITER to v0.1.23 (#58867)
Signed-off-by: fai <fangzhouai@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
2026-09-27 19:04:58 -05:00
Anton Peganovmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>OpenAI Codex
73859fec58 [Frontend] Switch Python Harmony dependency to oss-harmony (#55128)
Signed-off-by: Anton Peganov <apeganov@nvidia.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
2026-09-28 00:43:56 +04:00
Chao Lei 386ac2573e [Bugfix][KV Connector] Retry Mooncake bootstrap registration on timeout (reopens #55763) (#58919) 2026-09-27 19:47:36 +00:00
Fangzhou Ai 187c81b32d [ROCm][Bugfix] Fall back to default GEMM for CPU tensors on ROCm builds (#58923)
Signed-off-by: fai <fangzhouai@gmail.com>
2026-09-27 14:47:07 -05:00
Jiangyun Zhu 231fdb83cc [Perf][MoE] Use fused MiniMax2 routing with non-unit routed scaling (#58880)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
2026-09-27 19:37:33 +00:00
Xun SunandIsotr0py e7900156e1 [Fast Start] Add /health endpoint for the weight cache daemon (#58552)
Signed-off-by: Xun Sun <UNIDY2002@outlook.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
2026-09-27 17:48:54 +00:00
2b9b55c7f1 [Bugfix] Support repsonse_format + tool_choice=auto (#56086)
Signed-off-by: Artem Perevedentsev <aperevedents@nvidia.com>
Co-authored-by: Yufeng He <40085740+he-yufeng@users.noreply.github.com>
Co-authored-by: pablopupo <145598901+pablopupo@users.noreply.github.com>
Co-authored-by: hubunt <150658615+hubunt@users.noreply.github.com>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
2026-09-27 21:17:34 +04:00
Frank WangMisha Goinmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>Vadim Gimpelson
0376f81530 [Perf][Pooling] Avoid blocking seq_lens GPU-to-CPU copy for pooling in FlashInfer metadata builder (#57214)
Signed-off-by: frankwang28 <frank.wbb@hotmail.com>
Co-authored-by: Misha Goin <mgoin64@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
2026-09-27 16:06:38 +00:00
Shanshan Shen 44af287ebe [ROCm][Perf] Enable layer-aware CSA2 multi-stream overlap for DeepSeek-V4.1-Flash (#57407)
Signed-off-by: Shanshan Shen <87969357+shen-shanshan@users.noreply.github.com>
2026-09-27 23:48:01 +08:00
924707f1bf [CI] Split (H200 MIG 35GB) Spec Decode Speculators + MTP into 4 named jobs (#57237)
Signed-off-by: Thang Nguyen <thangnguyenvn647@gmail.com>
Co-authored-by: Thang Nguyen <thang.nguyen@inferact.ai>
Co-authored-by: Kimi Code <noreply@moonshot.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-27 05:01:31 -07:00
Nicolò Lucchesi c8d7a7dd13 [Perf][Attention] Remove D2H sync from FlashInfer SM90 sparse MLA plan under async scheduling (#58684)
Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
2026-09-27 13:06:23 +02:00
Rakul-Chauhanmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>Cursor
24c9772d19 [Attention][CPU] Use zentorch SDPA for CPU MLA prefill (#54967)
Signed-off-by: Rakul Chauhan <rakul.chauhan@amd.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-27 08:30:55 +00:00
Juntian LiuandClaude Opus 5.5 a9eafde59c [Perf][DSv4.1] Fuse small-batch WO-A with inverse RoPE and MXFP8 quant on SM100/SM103 (#58634)
Signed-off-by: Juntian Liu <Juntianl777@gmail.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-09-26 21:16:12 -07:00
MOVZX 5da84df6b7 [BugFix] Honor drop_eagle_block in MambaManager
On a hybrid model (full-attention + Mamba/GDN layers, e.g. Qwen3.6-35B-A3B),
enabling MTP/EAGLE speculative decoding together with prefix caching corrupts
the recurrent state, and the server serves wrong answers. #43559 reports
tool-eval accuracy dropping from ~90% to ~50%. Nothing raises an exception
and nothing shows up in the logs, so an operator has no signal that anything
is wrong, and prefix caching then spreads it, because the corrupted state is
hash-reachable and every later request sharing that prefix restores from it.

The root cause is the find_longest_cache_hit method on MambaManager,
at vllm/v1/core/single_type_kv_cache_manager.py:1049 on main, which takes
drop_eagle_block in its signature and never acts on it. Its body searches
right-to-left for the last cached block, builds [null]*i + [cached], and
breaks, referencing neither drop_eagle_block nor .pop(.

Signed-off-by: MOVZX <movzx@yahoo.com>
2026-09-27 10:45:10 +07:00
VarshithandClaude fba47397f3 [Bugfix] V1: clear stale allowed_token_ids mask in InputBatch.condense (#43931)
Signed-off-by: Varshith <kvarshithgowda@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
2026-09-27 03:44:54 +00:00
MOVZX 5e6656e7a0 [Bugfix][Spec Decode] Add get_top_tokens() to Qwen4ExpMTP draft model
use_local_argmax_reduction hard-fails at worker startup when the draft
model lacks get_top_tokens(): the V2 speculator probes for the method
in _validate_local_argmax_reduction and raises, with no fallback to
the full-vocab logits path. Qwen4ExpMTP (nvidia and amd variants) has
compute_logits but no get_top_tokens, so any Qwen4Exp MTP drafter with
the flag enabled crashed both TP ranks during load_model.

Qwen4ExpMTP owns a flat self.lm_head and its compute_logits is exactly
self.logits_processor(self.lm_head, hidden_states), the same shape as
Qwen3_5MTP, so LocalArgmaxMixin applies unchanged: vocab-parallel
argmax with a (value, index) gather, O(2*tp_size) vs O(vocab_size)
communication per draft token. No D2T remapping needed since the
draft vocabulary equals the target vocabulary.

Signed-off-by: MOVZX <movzx@yahoo.com>
2026-09-27 10:31:47 +07:00
YannikHintereggerandxijiade.aihemaiti fff03267da [Bugfix][NIXL] Release a dead peer's NIXL state without waiting for TTL (#50047)
Signed-off-by: Yannik Hinteregger <37209495+YannikHinteregger@users.noreply.github.com>
Co-authored-by: xijiade.aihemaiti <3146335281@qq.com>
2026-09-26 20:30:09 -07:00
MOVZXandHinata 43fbd6510b [Qwen4Exp] Serve PLE rows from the loader's mapped checkpoint shards (PR #58815)
Adds an opt-in backend, --engram-config '{"host_file_gather": true}', that
serves Qwen4Exp PLE rows from the checkpoint files through the page cache
instead of allocating the 47.7 GiB table on GPU or in pinned host memory.

- Qwen4ExpPLEFileGatherEmbedding allocates no table rows; load_weights keeps
  the default loader's zero-copy safetensors views of each shard. At bind
  (post-load hook) every view must be file-backed per /proc/self/maps (HF
  cache symlinks work; heap/pinned/anonymous memory is rejected) and the
  shards must cover the whole table.
- Each step, Qwen4ExpModelState.prepare_inputs computes the PLE n-gram ids,
  copies them to pinned host memory, syncs the stream once, issues
  POSIX_FADV_WILLNEED over the coalesced page ranges of the step's rows,
  gathers the rows on the host, and copies them into a fixed device staging
  buffer. The forward reads only the staging buffer, so FULL CUDA graph
  capture is unaffected.
- Rows stay in the checkpoint dtype (FP8 table stays FP8, dequantized on the
  GPU); dummy loads serve zeros; weight reload is rejected; takes precedence
  over cpu_offload; rejected on ROCm and with elastic EP.

Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-09-27 09:54:42 +07:00
MOVZXandHinata 5af48b498c [Qwen3.5] Build MTP head unquantized when checkpoint ships it dense (PR #58813)
Fixes #58807. A compressed-tensors pack-quantized checkpoint that ships
its MTP head dense (plain mtp.*.weight names in a separate bf16 shard)
crashes at weight load: Qwen3_5MultiTokenPredictor builds its Linears
from the target quant config, so the loader expects packed names
(weight_packed) the checkpoint never had, and the weight-map walk dies
with "no module or parameter named fc.weight".

Detect the dense MTP head from the checkpoint itself: read the
safetensors index via the same repo-file channel used for recipe.yaml
(local cache first, Hub download as fallback). If any mtp.* tensor
carries a quantization artifact suffix the head is properly quantized
and nothing changes; if every mtp.* entry is a plain .weight the whole
predictor builds unquantized (fc and layers), mirroring the existing
modelopt_fp4 and GPTQ dynamic carve-outs. Single-file checkpoints have
no index and keep current behavior; NVFP4 checkpoints are unaffected
(their mtp.layers.* carry artifacts).

Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-09-27 09:54:24 +07:00
MOVZXandHinata 835c44af15 [Qwen4Exp] Reinterpret FP8 E4M3 PLE as int8 for SM86 pinned-host lookup (PR #58856)
The pinned-host PLE lookup Triton kernel only gathers and stores bytes,
but passing an fp8e4nv tensor to it fails on architectures without
native FP8 pointer support (e.g. SM86) with an error saying type
fp8e4nv is not supported (supported fp8 dtypes are fp8e4b15, fp8e5).

Since FP8 E4M3 and int8 both occupy one byte per element, reinterpret
the PLE weight view and the output tensor as int8 when the weight dtype
is float8_e4m3fn; the non-FP8 path is unchanged.

Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-09-27 09:54:05 +07:00
MOVZXandHinata 758cc4f9e1 [Qwen3.5] Pass quant_config and prefix to embed_tokens in Qwen3_5Model
Apply the qwen3_5_quantized_embedding patch: construct the embedding
through the model's quant config so quantized (e.g. fp8/int8)
embeddings are supported, and give it a stable prefix for weight
loading.

Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-09-27 09:53:21 +07:00
Hongxia Yang 3137ff0773 [ROCm][Kimi-K3] Make VLLM_ROCM_USE_AITER_MOE_SITUV2 select a4w4/a8w4/a16w4 (#58201)
Signed-off-by: Hongxia Yang <hongxia.yang@amd.com>
2026-09-27 02:26:49 +00:00
Lio Einaudiandmergify[bot] <37929162+mergify[bot]@users.noreply.github.com> 0c87a197b8 [Bugfix] Disable sequence parallelism / async TP under batch invariance and add a TP regression test (#56377)
Signed-off-by: LioEinaudi <zhao3024667639@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
2026-09-26 22:19:08 -04:00
Thien Tran 7a877ae34d [Qwen3.8-Flash-Next] Avoid memory fragmentation in QSA indexer logits workspace (#57105)
Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
2026-09-27 10:11:52 +08:00
eb0f2ca37f [mooncake] support CUSTOM_MEM_POOL in vllm (#49300)
Signed-off-by: bruce.xu <bruce.xb@alibaba-inc.com>
Signed-off-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: bruce.xu <bruce.xb@alibaba-inc.com>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: OpenAI Codex <codex@openai.com>
2026-09-27 00:26:05 +00:00
mevince cd1daf58a9 [Metrics][KV Offload] Add Prometheus metrics for SimpleCPUOffloadConnector (#57251)
Signed-off-by: Vincent <vincexxchan@gmail.com>
2026-09-26 17:17:11 -07:00
4be061c5af [Kernel] Bump FlashKDA to keep the recurrent state in fp32 (#58846)
Signed-off-by: Simon Veitner <sveitner@redhat.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
2026-09-26 22:20:44 +00:00
927c87b347 [PCP][DCP] Support DCP target model with non-DCP Dspark (#56723)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Codex <noreply@openai.com>
2026-09-26 18:12:35 -04:00
Yongye ZhuandClaude Opus 5.5 77871126f9 [Kernel][DSV4.1] Fuse MoE finalize into the TP all-reduce + mHC boundary (#58586)
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-27 03:08:36 +08:00
Cyrus Leung 379e9a1ea8 [Security] Harden message sanitization (#58832)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
2026-09-26 16:28:32 +00:00
Kevin H. Luu 7d8c5fe9a9 [Bugfix][DSV4.1] Avoid host sync in ViT CUDA graph replay metadata (#58499) 2026-09-26 11:37:36 -04:00
Andreas Karatzas 0239bd2549 [CI] Stabilize batch submission in full CUDA graph tests (#58810) 2026-09-26 11:35:56 -04:00
Wentao Ye 3576691c42 [GLM5.3 Bug] Fix sparse indexer attn topk backend selection (#58594)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
2026-09-26 11:35:28 -04:00
Misha Goin 79ac28c5bd [CI] Reduce CUDA graph mode test overhead (#58749) 2026-09-26 11:22:33 -04:00
Ashraf Bhuiyan 5113cf9d16 [Mypy] Fix mypy typing for Qwen and Qianfan models (#58046)
Signed-off-by: Ashraf Bhuiyan <mbhuiyan@redhat.com>
2026-09-26 14:31:51 +00:00