 Vadim GimpelsonandClaude
|
ec8b20fbf0
|
[Bugfix] Fix Gemma 4 encoder CUDA graph capture OOM at startup (#61365)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
|
2026-10-17 11:04:49 +00:00 |
|
|
Vadim GimpelsonandClaude
|
783380af39
|
[Bugfix] Fix NaN output of the Mamba2 chunk scan kernel on sm90 (#61718)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
|
2026-10-16 17:24:51 -07:00 |
|
|
|
818f70c295
|
[Bugfix] Mask special added tokens in xgrammar JSON, regex and EBNF grammars (#60749)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Artem Perevedentsev <aperevedents@nvidia.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Artem Perevedentsev <aperevedents@nvidia.com>
|
2026-10-13 14:38:23 +03:00 |
|
|
Vadim GimpelsonandClaude
|
cd7517fb17
|
[Bugfix] Fix ReplaySSM ring shape assertion when TP does not divide Mamba2 n_groups (#60704)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
|
2026-10-13 11:31:53 +00:00 |
|
|
Vadim GimpelsonandClaude
|
183625d4f9
|
[Bugfix] Fall back to auto for unquantized MoE layers when --moe-backend is marlin (#60219)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
|
2026-10-13 08:48:13 +00:00 |
|
|
|
04462732a4
|
[Bugfix] Reject max_num_scheduled_tokens=0 at config time instead of hanging (#60714)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Hui Li <40770106+untrall@users.noreply.github.com>
|
2026-10-09 19:15:10 +00:00 |
|
|
|
9824159d94
|
[Bugfix] Fix InternLM2 tool parser dropping characters from streamed arguments (#60709)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>
|
2026-10-08 21:20:49 -07:00 |
|
|
Vadim GimpelsonandClaude
|
ee406adf0c
|
[Bugfix] Fix ZeroDivisionError in Qwen3-VL video sampling when the source fps is unknown (#60501)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
|
2026-10-08 06:28:44 +00:00 |
|
|
|
b0b23a48d4
|
[Bugfix] Fix Qwen4Exp FP8 PLE pinned lookup kernel failing to compile below SM89 (#60142)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
|
2026-10-06 01:56:44 +00:00 |
|
|
|
0eac152707
|
[Bugfix] Route every speculation-capable row through the speculative path so recurrent state stays consistent (#56531)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
|
2026-10-06 03:03:44 +04:00 |
|
|
Vadim GimpelsonandClaude
|
d0d6e5f3a2
|
[Bugfix] Fix Mamba page size AssertionError with spec decoding on GraniteMoeHybrid, FalconH1 and Zamba2 (#59975)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
|
2026-10-05 12:49:47 +00:00 |
|
|
Vadim GimpelsonandClaude
|
4f52fa35ef
|
[Bugfix] Fix Qwen4Exp PLE embedding rejecting INC (AutoRound) checkpoints (#59990)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
|
2026-10-05 01:52:06 +00:00 |
|
|
Vadim Gimpelson
|
ca7e1f2c43
|
Move CI failure diagnosis docs into ci-fails-buildkite skill (#45975)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-06-19 20:12:40 +00:00 |
|
|
Vadim Gimpelsonandmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
|
fa27d4e9cf
|
[PERF] [Qwen3.5] Split mixed prefill+decode batches: route decodes to the recurrent kernel (#44700)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
|
2026-06-06 22:13:50 +08:00 |
|
|
Vadim Gimpelson
|
67d3792d99
|
[Bugfix] Fix Qwen3.5-FP8 nightly fail. Guard fused_add_rms_norm input/weight dtype mismatch in RMSNorm + quant fusion (#44694)
|
2026-06-06 08:46:14 -04:00 |
|
|
Vadim GimpelsonandCursor
|
4765f0f189
|
[Bugfix] Fix sequence_parallel_chunk_impl custom op aliasing its input (#44130)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
|
2026-06-05 23:56:36 +00:00 |
|
|
Vadim GimpelsonandJee Jee Li
|
8b3b71ee9d
|
[CI/Build] Bump flashinfer to v0.6.12 (#44036)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
|
2026-06-02 15:19:05 -07:00 |
|
|
Vadim GimpelsonandCursor
|
bf18d7e0b4
|
[Misc][NUMA] Auto-bind to PCT priority cores on DGX B300 + widen EngineCore across shard NUMA nodes (#43270)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Cursor <noreply@cursor.com>
|
2026-05-29 10:07:44 +08:00 |
|
|
Vadim Gimpelson
|
5d126dd155
|
[Bugfix] Exclude Ray DP from #42585's deferred port allocation (#43864)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-05-28 15:55:14 +00:00 |
|
|
Vadim Gimpelson
|
c1c4db8b4b
|
Log dummy DP step in iteration details (#41406)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
|
2026-05-28 12:18:39 +00:00 |
|
|
Vadim GimpelsonandNick Hill
|
d6b48f928f
|
[BugFix] Fix hard-coded timeout for multi-API-server startup (#43768)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
|
2026-05-28 00:09:13 -07:00 |
|
|
Vadim Gimpelsonandgemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
|
812e7e7364
|
[Bugfix][V1] Fix TOCTOU race causing intermittent EADDRINUSE on multi-API-server DP startup (#42585)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
|
2026-05-26 14:06:00 -07:00 |
|
|
Vadim Gimpelson
|
9c78c99995
|
[MISC] Fix symm_mem cap-equal gate; log AR backend selection (#42993)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-05-20 08:50:24 -07:00 |
|
|
Vadim GimpelsonandNick Hill
|
f60c6b33a5
|
[V1][DP][LB] Publish request counts at the start of each engine step (#41626)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
|
2026-05-14 15:39:24 +00:00 |
|
|
Vadim Gimpelson
|
a2e776d716
|
[Bugfix] Accept canonicalized modelopt_* quant_method in _extract_modelopt_quant_algo (#42181)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-05-11 11:10:57 -04:00 |
|
|
Vadim Gimpelson
|
6d85b36a9f
|
Revert #38730 and #38791 (#40032)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
|
2026-04-21 11:44:11 -04:00 |
|
|
Vadim Gimpelson
|
771913e4a0
|
[Bugfix] Fix NVFP4+MTP crash: force unquantized mtp.fc for Qwen3.5 (#38832)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-04-03 04:45:57 +04:00 |
|
|
Vadim Gimpelson
|
bb51d5b40d
|
Add @vadiklyutiy as committer (#38589)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-03-31 07:50:04 +08:00 |
|
|
Vadim Gimpelson
|
52069012fe
|
[Bugfix] Fix DeepGemm E8M0 accuracy degradation for Qwen3.5 FP8 on Blackwell (#38083)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-03-26 01:21:47 -07:00 |
|
|
Vadim Gimpelson
|
4f16ebbbd3
|
[Bugfix] Disable monolithic TRTLLM MoE for Renormalize routing (#37591) (#37605)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-03-20 12:19:26 -07:00 |
|
|
Vadim GimpelsonandPavani Majety
|
6c1cfbad32
|
Support non-contiguous KV cache in TRTLLM fp8 dequant kernel (#36867)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
Co-authored-by: Pavani Majety <pavanimajety@gmail.com>
|
2026-03-16 17:48:42 -07:00 |
|
|
Vadim Gimpelson
|
8374387bd8
|
[FlashInfer] Revert block_size 16 + head_size 256 workaround on Blackwell (#36987)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-03-16 09:04:29 +00:00 |
|
|
Vadim Gimpelson
|
4ff8c3c8f9
|
[BUGFIX][Mamba][Qwen3.5] Zero freed SSM cache blocks on GPU (#35219)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-03-10 03:32:20 -07:00 |
|
|
Vadim Gimpelson
|
33a0d43c71
|
[BUGFIX][Qwen3.5] Hardcode mlp.gate as not quantizable (#35156)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-02-23 19:42:24 -08:00 |
|
|
Vadim Gimpelson
|
74d90b1ce4
|
[Model Bash][DSR1] Add selective dynamic shape marking for CustomOp (#34900)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-02-21 19:28:01 -05:00 |
|
|
Vadim Gimpelson
|
59965affbd
|
[BUGFIX] Fix _dummy_run missing prepare_inputs_event synchronization (#34866)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-02-20 05:54:27 -08:00 |
|
|
Vadim Gimpelson
|
604b9eaec5
|
[BUGFIX] Fix accuracy regression for NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 with TP>1 (#34476)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-02-14 23:25:17 -08:00 |
|
|
Vadim Gimpelson
|
000214c4bb
|
[BUGFIX] Fix accuracy bugs in Qwen3-Next MTP (#34077)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-02-10 10:57:11 -05:00 |
|
|
Vadim Gimpelson
|
824058076c
|
[PERF] Change GDN Attention State Layout from [N, HV, K, V] to [N, HV, V, K] (#33291)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-02-04 11:20:52 +00:00 |
|
|
Vadim Gimpelson
|
a372f3f40a
|
[MISC] Fix Tensor Parallelism for Quantized Mamba Models with n_groups=1 (#33257)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-02-03 15:10:31 -05:00 |
|
|
Vadim Gimpelson
|
70917b1c55
|
[MISC] Add .cursor to .gitignore (#32868)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-01-22 17:27:13 +00:00 |
|
|
Vadim Gimpelson
|
0727cc9ecf
|
[BUGFIX] Fix test_mla_backends.py. Scale MLA projection weights to prevent numerical instability (#32529)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-01-19 13:49:29 -05:00 |
|
|
Vadim Gimpelson
|
6101a26dc9
|
[BUGFIX] Fix degenerate strides in TRTLLM query tensors for FlashInfer backend. Fixes issue #32353 (#32417)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-01-18 16:57:32 -08:00 |
|
|
Vadim Gimpelson
|
9f430c94bd
|
[BUGFIX] Add missed remaping of the names of fp8 kv-scale (#32199)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-01-12 20:42:06 +00:00 |
|
|
Vadim Gimpelson
|
e15a5ff07b
|
[MISC] Add strict contiguity check for FlashInfer attention tensors (#32008)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
|
2026-01-10 12:40:05 -08:00 |
|
|
Vadim Gimpelson
|
22dffca982
|
[PERF] Speed-up of GDN attention decode part (Qwen3-Next) (#31722)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2026-01-06 17:32:46 +00:00 |
|
|
Vadim Gimpelson
|
bc0a5a0c08
|
[CI] Add Qwen3-Next-FP8 to Blackwell model tests (#31049)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-12-23 17:21:50 -08:00 |
|
|
Vadim Gimpelson
|
717ac33d9c
|
[PERF] Qwen3-next. Add fp8 cutlass MoE tuned configs. chmod -x *MI308X.json (#29553)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-12-18 13:16:04 +08:00 |
|
|
Vadim Gimpelson
|
173b356abf
|
[PERF] Remove TRTLLM Gen attn kernel limitation max_seq_len <=131072 (#28755)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-11-15 15:43:41 +05:30 |
|
|
Vadim Gimpelson
|
f948ab6945
|
[CI Failure] nm-testing/Qwen2-0.5B-Instruct-FP8-SkipQKV was removed from HF. Skip it in tests (#28170)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-11-06 01:22:13 +00:00 |
|
|
Vadim Gimpelson
|
b6a248bdd7
|
[PERF] Decouple projections from GDN custom op. Attempt 2 (#28083)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-11-05 17:01:12 -08:00 |
|
|
Vadim Gimpelson
|
d4e547bb7e
|
Revert "[PERF] Decouple projections from GDN custom op" (#28080)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-11-04 15:58:23 -08:00 |
|
|
Vadim Gimpelson
|
5fd8f02ea9
|
[PERF] Decouple projections from GDN custom op (#27512)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-11-04 08:11:41 -08:00 |
|
|
Vadim Gimpelson
|
785d8b6410
|
[PERF] Qwen3-next MTP speedup (change bool mask indexing to index_select / index_copy to reduce d2h) (#26437)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-10-16 12:18:31 +08:00 |
|
|
Vadim Gimpelsonandgemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
|
82e64c7a20
|
[PERF] [Qwen3-next] Speed up gated RMSNorm (#26207)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
|
2025-10-12 08:27:50 +00:00 |
|
|
Vadim Gimpelson
|
e94cfd51da
|
[BUG] Qwen3-next MTP. Fix attn metadata build bug (#26564)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-10-10 14:59:03 -04:00 |
|
|
Vadim Gimpelson
|
072d7e53e5
|
[PERF] Add conv1d metadata to GDN attn (#25105)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-09-18 14:27:49 +00:00 |
|
|
Vadim Gimpelson
|
7a70a71892
|
[Qwen3-Next] Add B200 MoE configs for Qwen3-next (#24698)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
|
2025-09-11 15:34:58 -07:00 |
|
|
Vadim Gimpelson
|
58eee5f2e0
|
[PERF] Use faster way of decode in tokenizer: avoid useless list-to-list conversion (#20000)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@centml.ai>
|
2025-08-02 01:43:52 -07:00 |
|
|
Vadim Gimpelson
|
f73d02aadc
|
[BUG] Fix #20484. Support empty sequence in cuda penalty kernel (#20491)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@centml.ai>
|
2025-07-05 19:38:02 -07:00 |
|
|
Vadim Gimpelson
|
9a3b88328f
|
[PERF] Speedup of MRoPE prepare inputs (#19939)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@centml.ai>
|
2025-06-23 23:01:26 -07:00 |
|
|
Vadim Gimpelson
|
5d6d1adf15
|
[KERNEL] Sampler. CUDA kernel for applying repetition penalty (#18437)
|
2025-06-03 21:13:01 -07:00 |
|
|
Vadim Gimpelson
|
67da5720d4
|
[PERF] Speed up Qwen2.5-VL model by speed up rotary position embedding (#17973)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@centml.ai>
|
2025-05-15 23:31:02 -07:00 |
|
|
Vadim Gimpelson
|
90df7f23aa
|
[Doc] Add guidance for using ccache with pip install -e . in doc (#14901)
|
2025-03-16 23:10:04 +00:00 |
|