kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-14 06:27:45 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity
3,026 Commits 25 Branches 59 Tags 2.1 GiB
560ded5450
Commit Graph

3026 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
QI JUN
946ffcd2eb
[None][ci] optimize test cases of dgx b200 (#7948) 
Signed-off-by: junq <22017000+QiJune@users.noreply.github.com>
 2025-09-24 00:39:45 -07:00
Cao Dong
2f8dc6feb0
[None][feat] Return topk logprobs in torch backend (#7756) 
Signed-off-by: Dong Cao <docao@nvidia.com>
 2025-09-24 15:30:39 +08:00
xinhe-nv
62563760fb
[None][chore] update chunked prefill cases (#7921) 
Signed-off-by: Xin He (SW-GPU) <200704525+xinhe-nv@users.noreply.github.com>
 2025-09-24 15:14:49 +08:00
qsang-nv
929ef4c474
[None][chore] remove cubins for ci cases (#7902) 
Signed-off-by: Qidi Sang <200703406+qsang-nv@users.noreply.github.com>
 2025-09-24 14:56:31 +08:00
Pengbo Wang
b890d7fea4
[None][infra] Skip failed test for nvbugs 5537738 (#7946) 
Signed-off-by: Pengbo Wang <221450789+pengbowang-nv@users.noreply.github.com>
 2025-09-23 23:48:50 -07:00
xiweny
276d83c898
[https://nvbugs/5532225] [fix] MoE use stream-dependent workspace (#7940) 
Signed-off-by: Xiwen Yu <13230610+VALLIS-NERIA@users.noreply.github.com>
 2025-09-24 14:44:27 +08:00
Yueh-Ting (eop) Chen
cf100933cc
[TRTLLM-6341][feature] Support SWA KV cache reuse (#6768) 
This merge request attempts to support more SWA KV cache functionality
inside the KV cache manager. Before this merge request, the KV cache for
sliding window attention (SWA) only holds "window size" number of blocks
and reuse them in a cyclic manner. We will not be able to utilize more
GPU memory with this design, leading to a limited max batch size
throughput. Additionally, we will not be able to support KV cache reuse
with this design.

In this MR, we change such behavior to let the manager write blocks in
a linear manner. With a linear block writing behavior, as the attention
window moves on, the out-of-window (OOW) blocks will be detached. Right
now for the sake of a correct feature first, we directly offload the
OOW block from the primary block pool (GPU memory) to the secondary
block pool (host memory). We will improve this in the future by
delegating the block movement to the eviction policy.

KV cache reuse for SWA is not developed in this merge request and will
be amended in a follow-up merge request.

Writing the blocks linearly, the maximum number of blocks allocated for
a sequence(`GenerationRequest`) is the "max sequence length" specified.
The `GenerationRequest` that stores the cache block bookkeeping
structure will now keep "max sequence length" tokens of blocks.

Given the above, main changes are (more context in the MR):
- Remove "cyclic" concept under the kv cache manager, such concept
  originally guards the block reuse under kv cache manager.
- Add detach mechanism and have it under `KVCacheManager::addToken`.
  Please note that detach is still guarded off for SWA when reuse
  is enabled. A follow-up merge request will proceed to improve this.
- Enforce "max sequence length" to be a non-optional parameter to
  the `KVCacheManager`/`BlockManager`
- Let all window size resource pool get identical proportion of memory
- Fix free memory calculation under `resource_manager.py`

Signed-off-by: eopXD <yuehtingc@nvidia.com>
Co-authored-by: Tomer Asida <tasida@nvidia.com>
 2025-09-24 14:28:24 +08:00
Daniel Cámpora
5ccb2dea33
[None][chore] Make sampler type beta. (#7934) 
Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>
 2025-09-23 20:51:39 -07:00
Yuan Tong
70c3b100eb
[#7692][fix] recognize RequestError as per-request error in background handler (#7726) 
Signed-off-by: Yuan Tong <13075180+tongyuantongyu@users.noreply.github.com>
 2025-09-24 11:11:17 +08:00
Yuan Tong
f050b8d871
[None][fix] refine backend option handling for commands (#7829) 
Signed-off-by: Yuan Tong <13075180+tongyuantongyu@users.noreply.github.com>
 2025-09-24 10:54:33 +08:00
Lizhi Zhou
e4f1f90202
[https://nvbugs/5477404][chore] unwaive test_disaggregated_single_gpu.py::test_disaggregated_llama_context_capacity (#7857) 
Signed-off-by: Lizhi Zhou <1432185+reasonsolo@users.noreply.github.com>
 2025-09-24 10:31:35 +08:00
Ziyi Xiong
31ef03fd82
[https://nvbugs/5528405][fix] Set up draft_tokens before scheduling (#7903) 
Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
 2025-09-24 09:56:17 +08:00
Venky
6ff0fad75e
[TRTLLM-7015] [feat] Enable prompt_logprobs in pytorch backend (#7580) 
Signed-off-by: Venky Ganesh <23023424+venkywonka@users.noreply.github.com>
 2025-09-23 18:48:10 -07:00
Lizhi Zhou
7550251988
[TRTLLM-7182][test] add multi-nodes test for disagg-serving (#7470) 
Signed-off-by: Lizhi Zhou <1432185+reasonsolo@users.noreply.github.com>
 2025-09-24 08:31:56 +08:00
mpikulski
9970345919
[TRTLLM-7728][feat] batched sampling by strategy (supersedes enable_mixed_sampler, cf. TRTLLM-7156) (#7294) 
Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com>
 2025-09-23 16:05:05 -07:00
Jhao-Ting Chen
220dc01372
[None][feat] support JIT mha.cu for SPEC_DEC in runtime (#6078) 
Signed-off-by: Jhao-Ting Chen <jhaotingc@nvidia.com>
 2025-09-23 14:56:17 -07:00
Zheng Duan
e3c1a9409f
[TRTLLM-6549][fix] add kv cache time output back (#7798) 
Signed-off-by: zhengd-nv <200704041+zhengd-nv@users.noreply.github.com>
 2025-09-23 14:12:42 -04:00
Yilin Fan
7d4d6cc9e0
[TRTLLM-7292][feat] Support multi-threaded tokenizers for trtllm-serve (cherry-pick) (#7776) 
Signed-off-by: Yilin Fan <206948969+nv-yilinf@users.noreply.github.com>
 2025-09-23 09:39:47 -07:00
Tracin
1f2761e67b
[None][feat] Enable gpt oss on DGX H100. (#6775) 
Signed-off-by: Tracin <10434017+Tracin@users.noreply.github.com>
 2025-09-23 09:35:19 -07:00
Daniel Cámpora
9f1d9b7b18
[None][feat] Use list instead of torch tensor for new tokens in update requests (#7730) 
Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>
 2025-09-23 10:40:08 -04:00
Yanchao Lu
6a36349964
[None][test] Waive another intermittent OOM test (#7930) 
Signed-off-by: Yanchao Lu <yanchaol@nvidia.com>
 2025-09-23 22:34:09 +08:00
Zheyu Fu
34963ec39c
[None][fix] Assign [] to req.py_draft_tokens instead of None when spec decode is off (#7511) 
Signed-off-by: Zheyu Fu <zheyuf@NVIDIA.com>
 2025-09-23 06:54:18 -07:00
Zero Zeng
16bb76c31d
[None][chore] Update benchmark script (#7860) 
Signed-off-by: Zero Zeng <38289304+zerollzeng@users.noreply.github.com>
Co-authored-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>
 2025-09-23 03:15:42 -07:00
ChristinaZ
dd5fb2857a
[None][fix] Re-add the import for allgather that was mistakenly removed. (#7920) 
Signed-off-by: Christina Zhang <83400082+ChristinaZ@users.noreply.github.com>
 2025-09-23 03:09:48 -07:00
Yan Chunwei
3ba19b6ff1
[https://nvbugs/5532023][fix] executor with-statement bug (#7895) 
Signed-off-by: chunweiy <chunweiy@nvidia.com>
 2025-09-23 02:05:39 -07:00
Perkz Zheng
bb64e7462c
[None][fix] fix a bug with trtllm-gen kernels + attention sinks (#7919) 
Signed-off-by: Perkz Zheng <67892460+PerkzZheng@users.noreply.github.com>
 2025-09-23 00:32:04 -07:00
Enwei Zhu
f882fb86db
[https://nvbugs/5367180][fix] Fix xgrammar import before loading tensorrt_llm binary (#7906) 
Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
 2025-09-23 00:29:57 -07:00
Yan Chunwei
40820e6711
[None][fix] CHERRY-PICK trtllm-serve yaml loading (#7551) (#7897) 
Signed-off-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com>
 2025-09-23 14:56:52 +08:00
ruodil
05bec3bf0f
[None][test] rename llm_perf_full to llm_perf_core and add missing cases (#7899) 
Signed-off-by: Ruodi Lu <ruodil@users.noreply.github.com>
Co-authored-by: Ruodi Lu <ruodil@users.noreply.github.com>
 2025-09-22 23:04:34 -07:00
Pengbo Wang
a4b4ed4535
[None][fix] Fix and add test for TRTLLM MoE backend (#7755) 
Signed-off-by: Pengbo Wang <221450789+pengbowang-nv@users.noreply.github.com>
 2025-09-23 11:26:25 +08:00
Pengbo Wang
5792464d37
[None][fix] Read eos_token_id from generation_config for kimi_k2 (#7120) 
Signed-off-by: Pengbo Wang <221450789+pengbowang-nv@users.noreply.github.com>
 2025-09-23 10:47:03 +08:00
Pengbo Wang
08cc7a041f
[https://nvbugs/5355128][fix] Add missing wgmma intrinsic for starcoder (#7643) 
Signed-off-by: Pengbo Wang <221450789+pengbowang-nv@users.noreply.github.com>
 2025-09-23 10:38:58 +08:00
yunruis
126cd707e3
[None][opt] Add batch waiting when scheduling (#7416) 
Signed-off-by: yunruis <205571022+yunruis@users.noreply.github.com>
Co-authored-by: Tao Li @ NVIDIA <tali@nvidia.com>
 2025-09-23 10:27:37 +08:00
Chang Liu
998857bcde
[TRTLLM-7328][feat] E-PD Disagg Support via llmapi (3/N) (#7577) 
Signed-off-by: Chang Liu (Enterprise Products) <9713593+chang-l@users.noreply.github.com>
 2025-09-22 19:07:18 -07:00
jianweiwu
9da4203e2e
[None][feat] Add Tencent HunYuanDenseV1 model support (#7081) 
Signed-off-by: sorenwu <sorenwu@tencent.com>
Signed-off-by: jianweiwu <sorenwu@tencent.com>
Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
Co-authored-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
 2025-09-23 09:27:29 +08:00
Tailing Yuan
740340dd17
[https://nvbugs/5522847][fix] Disable GC on disagg server and client (#7858) 
Signed-off-by: Tailing Yuan <yuantailing@gmail.com>
 2025-09-23 09:16:55 +08:00
Enwei Zhu
8330d5363a
[TRTLLM-8209][feat] Support new structural tag API (upgrade XGrammar to 0.1.25) (#7893) 
Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
 2025-09-23 09:10:09 +08:00
xxi
d471655242
[TRTLLM-7831][feat] Cherry-pick from #7423 Support fp8 block wide ep cherry pick (#7712) 2025-09-23 08:41:38 +08:00
Enwei Zhu
59f57598a7
[https://nvbugs/5504086][fix] Fix MTP vanilla (#7904) 
Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
 2025-09-23 08:38:28 +08:00
ChristinaZ
be576a3152
[None] [feat] Enable run_post_quant_allgather for MoE TRTLLM backend (#6794) 
Signed-off-by: Christina Zhang <83400082+ChristinaZ@users.noreply.github.com>
 2025-09-23 08:24:21 +08:00
Jin Li
b5391b4ac6
[https://nvbugs/5516665][fix] Fix CUTLASS moe fake impl errors (#7714) 
Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com>
 2025-09-22 11:08:39 -07:00
Linda
b1738c3f18
[https://nvbugs/5477359][fix] Removing test waivers (#7877) 
Signed-off-by: Linda-Stadter <57756729+Linda-Stadter@users.noreply.github.com>
 2025-09-22 08:59:13 -07:00
Wanli Jiang
2a30f11d63
[None][chore] Upgrade transformers to 4.56.0 (#7523) 
Signed-off-by: Wanli Jiang <35160485+Wanli-Jiang@users.noreply.github.com>
Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
Co-authored-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
 2025-09-22 22:20:16 +08:00
Yan Chunwei
fadce99af4
[https://nvbugs/5351244][fix] CHERRY-PICK test_mpi_session (#7501) (#7900) 
Signed-off-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com>
 2025-09-22 20:03:30 +08:00
Emma Qiao
324301ccba
[None][infra] Skip failed test for nvbugs 5532023 (#7905) 
Signed-off-by: qqiao <qqiao@nvidia.com>
 2025-09-22 03:49:44 -07:00
Yechan Kim
f77aca9f2c
[TRTLLM-7385][feat] Optimize Qwen2/2.5-VL performance (#7250) 
Signed-off-by: yechank <161688079+yechank-nvidia@users.noreply.github.com>
 2025-09-22 03:40:02 -07:00
HuiGao-NV
0dac1ddb74
[https://nvbugs/5525849][fix] Cherry-pick to fix mismatch of max seq len between kv cache manager and dummy requests (#7855) 
Signed-off-by: Hui Gao <huig@nvidia.com>
 2025-09-22 18:07:47 +08:00
Bo Deng
8cf95681e6
[TRTLLM-7989][infra] Bundle UCX and NIXL libs in the TRTLLM python package (#7766) 
Signed-off-by: Bo Deng <deemod@nvidia.com>
 2025-09-22 16:43:35 +08:00
Emma Qiao
d330d0005c
[None][infra] Waive a failed case on main (#7901) 
Signed-off-by: qqiao <qqiao@nvidia.com>
 2025-09-22 00:37:01 -07:00
xinhe-nv
9c1b75e978
[TRTLLM-7070][feat] add gpt-oss chunked prefill tests (#7779) 
Signed-off-by: Xin He (SW-GPU) <200704525+xinhe-nv@users.noreply.github.com>
 2025-09-22 00:12:43 -07:00