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
3,214 Commits 25 Branches 59 Tags 2.1 GiB
46ee7acb33
Commit Graph

469 Commits

Author SHA1 Message Date
John Calderon
46ee7acb33
[TRTLLM-6780][fix] Add multimodal data to dummy requests during memory profiling (#7539) 
Signed-off-by: John Calderon <johncalesp@gmail.com>
Signed-off-by: John Calderon <jcalderon@nvidia.com>
Signed-off-by: john calderon <jcalderon@nvidia.com>
Signed-off-by: John Calderon <jcalderon@nvidia>
 2025-10-16 17:49:22 +02:00
Wangjue Yao
9865d3d770
[None][feat] Support cached tokens for Openai server (#7637) 
Signed-off-by: wjueyao <wyao123@terpmail.umd.edu>
Co-authored-by: Pengyun Lin <81065165+LinPoly@users.noreply.github.com>
 2025-10-16 20:51:37 +08:00
Chuang Zhu
40d129a415
[None][fix] Fix cache buffer size for window (#8320) 
Signed-off-by: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
 2025-10-16 09:01:11 +08:00
HuiGao-NV
e265eb5fe9
[None][feat] reuse cudagraph memory pool in normal forward flow (#8095) 
Signed-off-by: Hui Gao <huig@nvidia.com>
 2025-10-16 07:08:44 +08:00
QI JUN
65ec01b257
[TRTLLM-8532][chore] clean warmup method of ModelEngine (#8264) 
Signed-off-by: junq <22017000+QiJune@users.noreply.github.com>
 2025-10-15 08:40:58 -07:00
QI JUN
616d1df7a0
[None][chore] set the default value of max_num_tokens explicitly (#8208) 
Signed-off-by: junq <22017000+QiJune@users.noreply.github.com>
 2025-10-14 23:03:02 -07:00
Fanrong Li
0d20a8fd61
[TRTLLM-8536][feat] Add the sparse attention framework and one use case--RocketKV support (#8086) 
Signed-off-by: Fanrong Li <23290157+lfr-0531@users.noreply.github.com>
Signed-off-by: yuhangh <58161490+heyuhhh@users.noreply.github.com>
Co-authored-by: yuhangh <58161490+heyuhhh@users.noreply.github.com>
 2025-10-14 08:23:16 -07:00
Yuxian Qiu
3450fe9944
[None][fix] Fix dummy load format for key models. (#7993) 
Signed-off-by: Yuxian Qiu <142763828+yuxianq@users.noreply.github.com>
 2025-10-14 11:18:39 +08:00
Zheyu Fu
bac665e650
[TRTLLM-7412][feat] Turn off spec decode when the rolling average acceptance length drops below threshold. (#7283) 
Signed-off-by: Zheyu Fu <zheyuf@NVIDIA.com>
 2025-10-13 15:51:14 -07:00
Robin Kobus
db8c63b9b1
[TRTLLM-4517] [feat] Additional model outputs (#7206) 
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
 2025-10-13 15:33:18 +02:00
Leslie Fang
8d1b068b1a
[TRTLLM-8477][chore] Replace KvCacheConfigCpp with KvCacheConfig inside PyExecutor (#8259) 
Signed-off-by: leslie-fang25 <leslief@nvidia.com>
 2025-10-13 14:55:36 +08:00
amitz-nv
fac47e2826
[https://nvbugs/5510879][fix] Fix pytorch & TRT-python flows fused LoRA adapter modules weight split with TP>1 (#8063) 
Signed-off-by: Amit Zuker <203509407+amitz-nv@users.noreply.github.com>
 2025-10-12 12:29:52 -07:00
kris1025
a7ea544dbe
[TRTLLM-7384][feat] enable rejection sampling for CDL (#7731) 
Signed-off-by: linquanh <linquanh@nvidia.com>
 2025-10-12 20:38:48 +08:00
Ziyi Xiong
efd4ffa03b
[https://nvbugs/5534705][fix] Skip unnecessary CUDA graph capture (#8050) 
Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
 2025-10-11 13:26:55 +08:00
QI JUN
48c15d805c
[https://nvbugs/5558167][fix] update canceled_req_ids correctly for canceled requests (#8207) 
Signed-off-by: junq <22017000+QiJune@users.noreply.github.com>
 2025-10-10 18:58:26 +08:00
mpikulski
7b6803b6e9
[TRTLLM-7769][chore] document the role of 'd2t' (#8174) 
Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com>
 2025-10-09 13:13:50 -04:00
mpikulski
8298e93bd8
[TRTLLM-8414][chore] BREAKING CHANGE: refine sampling strategy selection (#8132) 
Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com>
 2025-10-08 15:46:50 +02:00
Jonas Yang CN
88ea2c4ee9
[TRTLLM-7349][feat] Adding new orchestrator type -- ray (#7520) 
Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com>
Co-authored-by: Yuan Tong <13075180+tongyuantongyu@users.noreply.github.com>
Co-authored-by: Erin Ho <14718778+hchings@users.noreply.github.com>
 2025-10-04 08:12:24 +08:00
Ziyi Xiong
7bc2d9e993
[https://nvbugs/5537878][fix] Reserve an extra slot for padded batch (#7998) 
Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com>
 2025-10-03 08:42:52 -07:00
Yilin Fan
01423ac183
[None][feat] perf_metrics endpoint functionality improvement (#8005) 
Signed-off-by: Yilin Fan <206948969+nv-yilinf@users.noreply.github.com>
Signed-off-by: nv-yilinf <206948969+nv-yilinf@users.noreply.github.com>
 2025-10-02 17:43:25 -07:00
Daniel Cámpora
ab433b7228
[None][fix] Fix access to new tokens in sampler. (#7958) 
Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>
 2025-10-02 15:41:21 -04:00
Patrice Castonguay
fefa7d8fa3
[None][feat] Support for cancelling requests with disaggregation (#8114) 
Signed-off-by: Shunkang <182541032+Shunkangz@users.noreply.github.co>
Signed-off-by: Patrice Castonguay <55748270+pcastonguay@users.noreply.github.com>
Co-authored-by: Shunkang <182541032+Shunkangz@users.noreply.github.co>
 2025-10-02 11:04:26 -07:00
brb-nv
bd3d0ad233
[TRTLLM-7733][feat] Executor changes to support helix parallelism (#7972) 
Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>
 2025-10-01 22:13:03 -04:00
Izzy Putterman
1ad7bc4c78
[None][feat] Draft: Save state first pass (#7012) 
Signed-off-by: Izzy Putterman <iputterman@nvidia.com>
 2025-10-01 18:40:55 -04:00
Yibin Li
d7581bb551
[TRTLLM-8031][feat] Add chunked return_generation_logits logic (#7831) 
Signed-off-by: Yibin Li <109242046+yibinl-nvidia@users.noreply.github.com>
 2025-10-01 12:47:07 -04:00
Guoming Zhang
b4be0d2e4c
[None][chore] Refine qwen3-next implementation. (#8064) 
Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>
 2025-09-30 15:05:13 -04:00
Yechan Kim
948b8b9569
[None][fix] Fix CUDA graph for Qwen2.5-VL (#8047) 
Signed-off-by: yechank <161688079+yechank-nvidia@users.noreply.github.com>
 2025-09-30 14:40:03 +08:00
Cao Dong
62010c0ab7
[None][feat] Return topk logprobs in torch backend (#7976) 
Signed-off-by: Cao Dong <87467313+dcaox@users.noreply.github.com>
 2025-09-30 09:32:37 +08:00
bhsueh_NV
38d6e4e60b
[None][feat] Support Qwen3 next (#7892) 
Signed-off-by: mengw <12670782+wm2012011492@users.noreply.github.com>
Signed-off-by: bhsueh <11360707+byshiue@users.noreply.github.com>
Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>
Co-authored-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>
 2025-09-29 21:16:07 +08:00
mpikulski
a0d489a8d5
[TRTLLM-7728][perf] improve batched sampling perf for contiguous batches (#7908) 
Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com>
 2025-09-29 13:32:50 +01:00
Zongfei Jing
e9f26feeb6
[None][chore] Cherry-pick from (#7598) Make low_precision_combine as a llm arg (#7898) 
Signed-off-by: Zongfei Jing <20381269+zongfeijing@users.noreply.github.com>
 2025-09-28 22:32:33 -04:00
Yukun He
28b9a81c58
[TRTLLM-4500][feat] Add serialization/deserialization options for AutoTuner profiling cache (#7738) 
To achieve determinism for the AutoTuner profiling cache, serialization and deserialization are introduced to store the cache on disk in JSON format. Use TLLM_AUTOTUNER_CACHE_PATH to indicate the path where the cache file should be stored:

Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
 2025-09-29 07:40:51 +08:00
Xianjie Qiao
c8f98b3065
[None] [feat] Update disagg gen-only benchmark. (#7917) 
Signed-off-by: Xianjie <5410381+qiaoxj07@users.noreply.github.com>
 2025-09-28 09:56:56 +08:00
Iman Tabrizian
33282351a2
[TRTLLM-6106][feat] Add support for KVCache transfer from KVCache reuse path (#6348) 
Signed-off-by: Iman Tabrizian <10105175+tabrizian@users.noreply.github.com>
 2025-09-27 19:29:30 -04:00
YueWeng
a4243f0da5
[TRTLLM-6393][feat] add static tree sampling and verification (#7161) 
Signed-off-by: Yue Weng <25103990+yweng0828@users.noreply.github.com>
 2025-09-26 13:16:16 -04:00
HuiGao-NV
f4d3be4bbc
[None][feat] Add a standalone buffer cache class and reuse buffers between cduagraph and no-graph flow (#7669) 
Signed-off-by: Hui Gao <huig@nvidia.com>
 2025-09-26 07:28:06 -07:00
HuiGao-NV
a9965d84e0
[None][chore] Report NCCL error message but not OOM when NCCL error happens (#8009) 
Signed-off-by: Hui Gao <huig@nvidia.com>
 2025-09-25 23:07:32 -07:00
sunnyqgg
2e5850c28a
[TRTLLM-7330][feat] Eagle3 cuda graph support for the first draft model inference (#7363) 
Signed-off-by: qgai <qgai@nvidia.com>
 2025-09-26 11:28:05 +08:00
QI JUN
1529a6f22d
[None][chore] extract weights loading related logic to model loader (#7579) 
Signed-off-by: junq <22017000+QiJune@users.noreply.github.com>
 2025-09-25 10:19:22 -07:00
Guoming Zhang
202bed4574 [None][chroe] Rename TensorRT-LLM to TensorRT LLM for source code. (#7851) 
Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>
Signed-off-by: Wangshanshan <30051912+dominicshanshan@users.noreply.github.com>
 2025-09-25 21:02:35 +08:00
Leslie Fang
342014069e
[None][chore] Validate features combination (#7630) 
Signed-off-by: leslie-fang25 <leslief@nvidia.com>
 2025-09-25 08:01:13 +08:00
Iman Tabrizian
da30d496b0
[None][fix] Revert "[None][feat] Return topk logprobs in torch backend (#7756)" (#7969) 
Signed-off-by: Iman Tabrizian <10105175+tabrizian@users.noreply.github.com>
 2025-09-24 15:36:38 -07:00
Yuxian Qiu
48fda86c56
[None][fix] Fix dummy load format for DeepSeek. (#7874) 
Signed-off-by: Yuxian Qiu <142763828+yuxianq@users.noreply.github.com>
 2025-09-24 23:03:16 +08:00
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
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
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
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
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
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