22 Commits
Author SHA1 Message Date
MOVZX 76df77c04e Merge branch 'main' into patches
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-10-15 21:16:39 +07:00
bfdbfa1632 [Bugfix][PCP] Align MoE routing padding with gathered tokens (#60946)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: wangyicong <wangyicong@bytedance.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
2026-10-14 21:04:00 -04:00
Ayushman Singh 9b5c6707a3 [Bugfix][MRV2] Capture DCP FULL cudagraphs with KV context (#61496)
Signed-off-by: Ayushman Singh <40520701+ayush1399@users.noreply.github.com>
2026-10-14 19:20:48 +00:00
MOVZX ac305c44cb Merge branch 'main' into patches
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-10-09 17:41:13 +00:00
Giancarlo Delfin 6ee9be6b3c [Spec Decode] Rename AR speculators to StandaloneAR / TargetDependentAR (#60335)
Signed-off-by: Giancarlo Delfin <gdelfin@inferact.ai>
2026-10-07 14:01:59 -04:00
Wentao Ye 74c5cbcd7b [Bugfix][MRV2][Spec Decode] Honor dynamic K in autoregressive speculators, 1.2~1.3x kernel perf improvement (#57053)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
2026-10-05 14:54:11 -04:00
MOVZX f83ecfac16 Merge remote-tracking branch 'upstream/main' into patches 2026-10-05 08:52:31 +07:00
aoshen02andClaude Opus 5.5 5e56e9af2a [Feature] Release the CUDA graph pool on sleep (#59160)
Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-04 01:13:10 +00:00
MOVZX 2877121b2d Merge remote-tracking branch 'upstream/main' into patches
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-10-03 21:59:03 +07:00
Qiu ChunshuoandLucas Wilkinson 8486f1ebfe [Perf][PCP] Shard decode requests across PCP ranks (#52162)
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
2026-10-01 23:50:13 -04:00
Harry MellorandClaude Opus 5.5 fc2c801ad9 Disable mypy arg-type and assignment checks in tests (#59428)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-30 15:18:50 +00:00
5dd693122c [MyPy][2/N] Fix mypy errors in small tests/ dirs (#51043)
Signed-off-by: Martin Hickey <martin.hickey@ie.ibm.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-29 21:27:45 +01:00
MOVZXandHinata ae0350c0de [Spec Decode][MRV2] RecoverSSM for Qwen GDN and PLE short conv (PR #58863)
Intake of open upstream PR vllm-project/vllm#58863. Active only with
VLLM_USE_V2_MODEL_RUNNER=1; inert otherwise.

Speculative decoding on Gated DeltaNet snapshotted the full recurrent
state after every draft position and reserved one Mamba block per draft
position. On a 24 GB card that is reserved VRAM we never use, plus
per-step copy traffic.

RecoverSSM, the protocol Kimi-K3 KDA already uses, replaces this:
verify runs from a read-only checkpoint and stores a small per-token
replay record [correction (V), normalized key (K), gate (1)] in fp32;
commit replays only the accepted tokens and writes the recurrent state
once, plus the block-boundary state in align mode. The PLE dilated
short conv compacts its accepted window with the same plan.

Conflict in vllm/models/qwen4_exp/nvidia/mtp.py: our get_top_tokens()
work put LocalArgmaxMixin on Qwen4ExpMTP, the PR adds SupportsReplaySSM
to the same base list. Kept both.

Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-09-29 13:39:09 +07:00
Almog Tavor 171d1b9ef6 [Elastic EP] Support Model Runner V2 (#53934)
Signed-off-by: almogtavor <almogtavor@gmail.com>
Signed-off-by: Almog Tavor <70065337+almogtavor@users.noreply.github.com>
2026-09-29 00:20:40 +00:00
Jürgen SchmiedandClaude a80e41d690 [Mamba] Share RecoverSSM commit kernels and helpers across models
RecoverSSM's commit-plan and conv-window compaction kernels are model independent.
Move them from models/kimi_k3 into model_executor/layers/mamba/ops/recoverssm.py
(Kimi-K3 keeps its private names as import aliases; kernel source unchanged), add
mamba/recoverssm_utils.py (activation predicate, argument validation) and the
generic spec-row/request mapping helpers to v1/attention/backends/recoverssm_metadata.py.

CacheConfig.use_kda_recoverssm becomes use_recoverssm: the switch now selects RecoverSSM
for any model whose recurrent layers implement it, not only KDA.

No functional change for Kimi-K3.

Signed-off-by: Jürgen Schmied <juergenschmied70@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
2026-09-27 00:07:04 +02:00
a4eb3f25d6 [Spec Decode] Enable Gemma4 DSpark adaptive verification with FlashInfer (#57263)
Signed-off-by: zixi-qi <zixi@inferact.ai>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
2026-09-26 17:01:17 +08:00
Andreas KaratzasandOpenAI Codex 23110a0c9e [Bugfix] Capture prefill kernels for mixed FULL graphs (#58275)
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
2026-09-24 02:05:07 -07:00
Wentao Ye 07950d4734 [Kimi Bug] Fix kimi k3 startup cuda graph issue with recoverSSM (#55774)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
2026-09-08 10:29:53 -04:00
Rahul ChalamalaJanelle CaiDevin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>Devin
d3d79ffc1e [Bugfix][Spec Decode] Capture the widest uniform decode batch by default (#50488)
Signed-off-by: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com>
Signed-off-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Rahul Chalamala <rahul@modal.com>
Co-authored-by: Janelle Cai <janelle.cai@modal.com>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Devin <devin@cognition.ai>
2026-08-28 22:38:13 -07:00
Qiu Chunshuo b3af042abd Bugfix: use PCP slot mappings for PIECEWISE capture (#53869)
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
2026-08-27 05:39:36 +00:00
c74dd2d208 [Bugfix][MRV2] Dispatch uniform decode to a padded FULL cudagraph (#53407)
Signed-off-by: Xiaohu Guo <Xiaohu.Guo@amd.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: xiaohuguo2023 <149615094+xiaohuguo2023@users.noreply.github.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
2026-08-25 06:59:38 +00:00
Ilya MarkovandMarkov Ilya f890e1dbe2 [BugFix] Set graph_pool_id before FULL CUDA graph capture in ModelRunner V2 (#48843)
Signed-off-by: Markov Ilya <markovilya19@gmail.com>
Co-authored-by: Markov Ilya <markovilya19@gmail.com>
2026-07-21 08:59:26 -04:00