kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-02-03 17:52:19 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity
2,289 Commits 28 Branches 64 Tags 2 GiB
dd11e08d26
Commit Graph

60 Commits

Author SHA1 Message Date
Sergey Klevtsov
27fc35175e
[None][feat] CUTLASS MoE FC2+Finalize fusion (#3294) 
Signed-off-by: Sergey Klevtsov <sklevtsov@nvidia.com>
 2025-08-12 15:56:48 +08:00
NVJiangShao
2f2f5cc72c
[TRTLLM-6744][feat] Remove input_sf swizzle for module WideEPMoE (#6231) 
Signed-off-by: Jiang Shao <91270701+StudyingShao@users.noreply.github.com>
 2025-08-08 11:13:42 +08:00
hlu1
8207d5fd39
[None] [feat] Add model gpt-oss (#6645) 
Signed-off-by: Hao Lu <14827759+hlu1@users.noreply.github.com>
 2025-08-07 03:04:18 -04:00
liji-nv
1daa8c3232
[https://nvbugs/5340941][https://nvbugs/5375785] - fix: Wrap attentio… (#6355) 
Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com>
 2025-08-01 07:38:06 -04:00
Jinyang Yuan
97f7e12588
[fix] Fix perf regression caused by MoE autotuner when using DeepEPLowLatency (#6288) 
Signed-off-by: Jinyang Yuan <154768711+jinyangyuan-nvidia@users.noreply.github.com>
 2025-07-28 01:37:11 -04:00
liji-nv
e07fff4f78
[https://nvbugs/5340941] - fix: Correct custom ops used by Qwen3 Moe … (#6285) 
Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com>
 2025-07-25 14:49:45 +08:00
danielafrimi
ff9963978a
Add register_fake for finegrained_mixed_dtype_gemm torch_op (#6255) 
Signed-off-by: Daniel Afrimi <danielafrimi8@gmail.com>
 2025-07-22 16:59:55 +03:00
liji-nv
3e0fb60e50
[TRTLLM-4279] feat: Multistream initial support for torch compile flow (#5847) 
Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com>
 2025-07-21 19:10:22 +08:00
Yuening Li
e8c068b4b1
[TRTLLM-5863][feat] Support Weight-Only-Quantization in PyTorch Workflow (#5850) 
Signed-off-by: Yuening Li <62227368+yueningl@users.noreply.github.com>
Co-authored-by: Yuening Li <62227368+yueningl@users.noreply.github.com>
 2025-07-21 15:17:35 +08:00
danielafrimi
5300a99bd8
W4A8 GEMM (#6005) 
Signed-off-by: Daniel Afrimi <danielafrimi8@gmail.com>
 2025-07-20 17:34:57 +03:00
Dom Brown
afaa388bee [TRTLLM-6100] fix: Nvbug 5356427: autotuned TRTLLM Gen fp8 block scale MoE illegal memory access (#5676) 
Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com>
 2025-07-14 17:17:30 +08:00
Enwei Zhu
bc1d4fb5da
[NvBug 5378370] fix: Fix alltoall for llama4 (apply_router_weight_on_input=True) (#5902) 
Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
 2025-07-12 15:50:31 +09:00
Dom Brown
3e3b1769ad
[TRTLLM-5881] feat: Integrate TRT-LLM Gen FP4 block scale MoE with Pytorch workflow kernel autotuner (#5764) 
Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com>
 2025-07-09 08:21:58 +01:00
DylanChen-NV
5ca2b9bb15
[TRTLLM-5812][feat] support FP8 row-wise dense GEMM in torch flow (#5615) 
Signed-off-by: Dylan Chen <191843203+DylanChen-NV@users.noreply.github.com>
 2025-07-07 18:04:57 +08:00
Daniel Stokes
ec6c7dff1a
feat: Add support for MXFP8xMXFP4 in pytorch (#5535) 
Signed-off-by: Daniel Stokes <40156487+djns99@users.noreply.github.com>
 2025-07-06 15:32:06 -07:00
Jhao-Ting Chen
77082cde38
[https://nvbugspro.nvidia.com/bug/5329655] [feat] Pytorch path add spec dec param to attention op (#5146) 
Signed-off-by: Jhao-Ting Chen <jhaotingc@nvidia.com>
 2025-07-02 04:54:43 -04:00
liji-nv
c345f5876c
[feat] Support torch compile for attention dp (#5086) 
Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com>
 2025-07-01 13:48:52 -04:00
danielafrimi
7a617ad1fe
feat: W4A16 GEMM (#4232) 
Signed-off-by: Daniel Afrimi <danielafrimi8@gmail.com>
 2025-07-01 10:36:05 +03:00
Li Min
6021a439ab
Make moe permute and final as custom op (#5412) 
Signed-off-by: Mindy Li <11663212+limin2021@users.noreply.github.com>
 2025-06-27 15:48:33 -07:00
Daniel Stokes
83a1f60556
feat: Expose bias and FP8_MXFP4 MOE CUTLASS backend features to pytorch (#5410) 
Signed-off-by: Daniel Stokes <40156487+djns99@users.noreply.github.com>
 2025-06-27 12:29:34 +08:00
Yukun He
9ee33605bb
[TRTLLM-6019] feat: Remove cutlass min latency code from AutoTuner. (#5394) 
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
 2025-06-26 13:12:03 +08:00
jmydurant
578dbc8d9a
feat: chunked prefill for MLA (Blackwell) (#4651) 
Signed-off-by: Mingyang Jiang <13463932+jmydurant@users.noreply.github.com>
 2025-06-26 09:01:00 +08:00
Yukun He
3fc57543e2
[5356427] fix: Remove the seq_len of 4096 from FP8 block scale MoE tuning configs. (#5485) 
The seq_len of 4096 will cause some unknown CUDA illegal memory access issue if run with some other tests consecutively.
Put a saturated upper bound for any sequence length larger than it.
 2025-06-26 08:38:35 +08:00
dongxuy04
4f0f17ac8a
feat: Misc Opt for large scale EP (#5374) 
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
 2025-06-20 13:11:31 +08:00
Yukun He
6711ad9cf3
[TRTLLM-5589] feat: Minor optimizations for tunable FP8 batched GEMM op. (#5139) 
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
 2025-06-18 14:33:46 +08:00
Dom Brown
44fb3c1673
[TRTLLM-5770] feat: Integrate TRT-LLM Gen FP8 block scale MoE with Pytorch workflow kernel autotuner (#5207) 
- Adds a new Python custom op (fp8_block_scale_moe_runner) and a FP8BlockScaleMoERunner class for autotuning.
- Updates C++ MoE and batched GEMM kernels to accept a configIndex for workspace sizing and execution.
- Extends the unit test to run both autotuned and non-autotuned code paths.

Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com>
 2025-06-17 21:01:56 +08:00
Enwei Zhu
4b82b8b4c7
[TRTLLM-5330] perf: Optimize MoE supplementary kernels for large-scale EP (#5215) 
Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
 2025-06-17 15:23:24 +08:00
Tracin
ef3fdc8051
feat: Add w4a8_mxfp4_fp8 quantization recipe. (#4867) 
Signed-off-by: Tracin <10434017+Tracin@users.noreply.github.com>
 2025-06-16 11:30:57 +08:00
yunruis
b99c5ce8c1
Feat/ds r1 min latency opt round3, add router gemm, fused a gemm, PDL (#4560) 
Signed-off-by: yunruis <yunruis@nvidia.com>
Signed-off-by: kduan <176893526+Kefeng-Duan@users.noreply.github.com>
Signed-off-by: Kefeng-Duan <176893526+Kefeng-Duan@users.noreply.github.com>
Co-authored-by: kduan <176893526+Kefeng-Duan@users.noreply.github.com>
 2025-06-14 17:36:22 +08:00
Dom Brown
9c012d5bf8
[TRTLLM-5589] feat: Integrate TRT-LLM Gen FP8 Batched GEMM with Pytorch workflow kernel autotuner (#4872) 
Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com>
 2025-06-09 11:02:48 +01:00
Daniel Stokes
3a4851b7c3
feat: Add Mixture of Experts FP8xMXFP4 support (#4750) 
Signed-off-by: Daniel Stokes <40156487+djns99@users.noreply.github.com>
 2025-06-09 13:25:04 +08:00
Shiyu Li
b0d287c9b7
[TRTLLM-4647][fix] Fix the no fusion allreduce hanging (#4594) 
Signed-off-by: Shiyu Li <shili@nvidia.com>
 2025-06-04 18:26:13 -07:00
Yukun He
5fa6fbd989
feat: Enhance AutoTuner inference path and code readability (#4466) 
Fix AutoTuner warmup request generating.
* The current warmup phase creates one request, which is insufficient for the warmup to cover the max_num_tokens. Revise the warmup phase to a batch of requests to cover the max_num_tokens to eliminate potential fallback cases.
Refactor AutoTuner API and reduce host overhead.

Refine (min, opt, max) values of optimization profile setup for get_valid_tactics to achieve the correct canImplement definition.
* Refine cache key assembly process to reduce host overhead and simplify API.
* Fix lru_cache usage to reduce host overhead.
* Move tuning config initialization as a one-time object in tunable runner to reduce host overhead.

Improve tuning config readability.
* Use dataclass to define tuning config.

Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
 2025-06-04 10:53:11 +08:00
Tian Zheng
9832787050
[feat] Enable NVFP4 output for TRTLLM attention kernels (#4737) 
Signed-off-by: Tian Zheng <29906817+Tom-Zheng@users.noreply.github.com>
 2025-06-03 10:00:17 +08:00
tomeras91
bf9cd11fd4
[TRTLLM-4783][feat] Mamba2 kernel updates for Nemotron-H (#4494) 
Signed-off-by: Tomer Asida <57313761+tomeras91@users.noreply.github.com>
 2025-06-01 13:56:44 +03:00
Enwei Zhu
25dde49c28
fix: EP load balancer with MTP layer and route offset by EP rank (#4767) 
Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
 2025-06-01 00:07:44 +08:00
Jinyang Yuan
5339d367ce
[perf] Reduce the workspace size of FP4 activation scales for MoE (#4303) 
Signed-off-by: Jinyang Yuan <154768711+jinyangyuan-nvidia@users.noreply.github.com>
 2025-05-30 09:03:52 +08:00
Yilin Fan
31bb650298
Cherry pick feat/llama4 to main (#4739) 
Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com>
Signed-off-by: Yilin Fan <206948969+nv-yilinf@users.noreply.github.com>
Co-authored-by: Chenfei Zhang <chenfeiz@nvidia.com>
 2025-05-30 05:28:40 +08:00
Mike Iovine
9c0de251db
[feat] Integrate Hopper chunked attention kernels (#4330) 
* Integrate chunked attention kernels

Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com>

* Fix cache key

Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com>

* Fix lint

Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com>

---------

Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com>
 2025-05-22 17:10:57 -04:00
dongxuy04
21aff2e313
feat: large-scale EP(part 2: MoE Load Balancer - core utilities) (#4384) 
* first commit of cpp moe loadbalance code

Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>

* add python bindings for moe load balance

Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>

* add python wrapper, ut and bug fixes

Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>

* add binding for layerId and update binding test

Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>

* add host tensor sharing and ut

Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>

---------

Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
 2025-05-20 17:53:48 +08:00
liji-nv
58e405624a
[https://nvbugs/5123103][fix] Fix torch compile for DeepSeekV3 (#3952) 
Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com>
 2025-05-19 22:12:25 +08:00
Jinyang Yuan
b618e1f55b
perf: Eliminate the need for attention DP padding when possible (#3439) 
Signed-off-by: Jinyang Yuan <154768711+jinyangyuan-nvidia@users.noreply.github.com>
Co-authored-by: raccoonliukai <raccoonliu@tencent.com>
 2025-05-17 13:30:55 +08:00
Daniel Cámpora
df19430629
chore: Mass Integration 0.19 (#4255) 
* fix: Fix/fused moe 0.19 (#3799)

* fix bug of stream init

Signed-off-by: bhsueh <11360707+byshiue@users.noreply.github.com>

* fix bug

Signed-off-by: bhsueh <11360707+byshiue@users.noreply.github.com>

---------

Signed-off-by: bhsueh <11360707+byshiue@users.noreply.github.com>

* fix: Add pre-download of checkpoint before benchmark. (#3772)

* Add pre-download of checkpoint before benchmark.

Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>

* Add missing remote code flag.

Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>

* Move from_pretrained to throughput benchmark.

Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>

* Move download and use snapshot_download.

Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>

* Removed trusted flag.

Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>

* Fix benchmark command in iteration log test.

Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>

---------

Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>

* [https://nvbugspro.nvidia.com/bug/5241495][fix] CUDA Graph padding with overlap scheduler (#3839)

* fix

Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>

* fuse

Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>

* fix

Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>

* fix

Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>

---------

Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>

* TRTLLM-4875 feat: Add version switcher to doc (#3871)

Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>

* waive a test (#3897)

Signed-off-by: Superjomn <328693+Superjomn@users.noreply.github.com>

* docs:fix https://nvbugs/5244616 by removing new invalid links. (#3939)

Signed-off-by: nv-guomingz <37257613+nv-guomingz@users.noreply.github.com>
Co-authored-by: nv-guomingz <37257613+nv-guomingz@users.noreply.github.com>

* fix: remote mpi session abort (#3884)

* fix remote mpi session

Signed-off-by: Superjomn <328693+Superjomn@users.noreply.github.com>

* fix

Signed-off-by: Superjomn <328693+Superjomn@users.noreply.github.com>

---------

Signed-off-by: Superjomn <328693+Superjomn@users.noreply.github.com>

* skip fp8 gemm for pre-hopper (#3931)

Signed-off-by: Ivy Zhang <25222398+crazydemo@users.noreply.github.com>

* [https://nvbugspro.nvidia.com/bug/5247148][fix] Attention DP with overlap scheduler (#3975)

* fix

Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>

* update multigpu list

Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>

* fix namings

Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>

---------

Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>

* Doc: Fix H200 DeepSeek R1 perf doc (#4006)

* fix doc

Signed-off-by: jiahanc <173873397+jiahanc@users.noreply.github.com>

* update perf number

Signed-off-by: jiahanc <173873397+jiahanc@users.noreply.github.com>

---------

Signed-off-by: jiahanc <173873397+jiahanc@users.noreply.github.com>

* Fix the perf regression caused by insufficient cache warmup. (#4042)

Force tuning up to 8192 sequence length for NVFP4 linear op. Also, make this runtime-selectable with UB enabled.

Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>

* doc: Update 0.19.0 release notes (#3976)

Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>

* Optimize the AutoTuner cache access code to reduce host code overhead. (#4060)

The NVFP4 Linear op is very sensitive to the host overhead.
This PR introduces customizable `find_nearest_profile` and `get_cache_key_specifc`, which allow users to override the default method for generating the cache key.

Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>

* Update switcher (#4098)

Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>

* doc: update release notes (#4108)

Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>

* docs:update 0.19 doc. (#4120)

Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>

* docs:add torch flow supported model list. (#4129)

Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>

* doc: Release V0.19 Perf Overview Update (#4166)

Signed-off-by: zpatel <22306219+zbpatel@users.noreply.github.com>

* Fix readme of autodeploy.

Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>

* Update tensorrt_llm/_torch/pyexecutor/llm_request.py

Co-authored-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
Signed-off-by: Daniel Cámpora <961215+dcampora@users.noreply.github.com>

* Revert mgmn worker node.

Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>

* Change to disable_overlap_scheduler.

Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>

---------

Signed-off-by: bhsueh <11360707+byshiue@users.noreply.github.com>
Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>
Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>
Signed-off-by: Superjomn <328693+Superjomn@users.noreply.github.com>
Signed-off-by: nv-guomingz <37257613+nv-guomingz@users.noreply.github.com>
Signed-off-by: Ivy Zhang <25222398+crazydemo@users.noreply.github.com>
Signed-off-by: jiahanc <173873397+jiahanc@users.noreply.github.com>
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>
Signed-off-by: zpatel <22306219+zbpatel@users.noreply.github.com>
Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>
Signed-off-by: Daniel Cámpora <961215+dcampora@users.noreply.github.com>
Co-authored-by: bhsueh_NV <11360707+byshiue@users.noreply.github.com>
Co-authored-by: Frank <3429989+FrankD412@users.noreply.github.com>
Co-authored-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com>
Co-authored-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com>
Co-authored-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com>
Co-authored-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com>
Co-authored-by: nv-guomingz <37257613+nv-guomingz@users.noreply.github.com>
Co-authored-by: Ivy Zhang <25222398+crazydemo@users.noreply.github.com>
Co-authored-by: jiahanc <173873397+jiahanc@users.noreply.github.com>
Co-authored-by: Yukun He <23156053+hyukn@users.noreply.github.com>
Co-authored-by: Zac Patel <22306219+zbpatel@users.noreply.github.com>
 2025-05-16 10:53:25 +02:00
zhhuang-nv
97bc680cd8
feat: support kv cache reuse for MLA (#3571) 
* support kv cache reuse for MLA

load compressed_kv and k_pe and do up-projection
use 192/128 head size MLA context kernel
support Blackwell and Hopper now

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* add CI test

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* fix: set k_pe head_num to 1 for kernel 2 and kernel 2V2

Signed-off-by: Mingyang Jiang <13463932+jmydurant@users.noreply.github.com>

* resolve comments

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* use GPTJ style RoPE for MLA

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* fix rebase error and some docs

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* fix kv_lens

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* tiny fix

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* fix torch compile

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* fix: use normal device memory instead of pinned memory for unit test

Signed-off-by: Mingyang Jiang <13463932+jmydurant@users.noreply.github.com>

* fix L0 tests

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* fix torch compile after rebase

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* resolve comments

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

* resolve comments again

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>

---------

Signed-off-by: Zhen Huang <145532724+zhhuang-nv@users.noreply.github.com>
Signed-off-by: Mingyang Jiang <13463932+jmydurant@users.noreply.github.com>
Signed-off-by: zhhuang-nv <145532724+zhhuang-nv@users.noreply.github.com>
Co-authored-by: Mingyang Jiang <13463932+jmydurant@users.noreply.github.com>
 2025-05-15 15:22:21 +08:00
Barry Kang
20b42912ce
[TRTLLM-3330][feat] Support DeepSeek-R1 W4A8 on Hopper (#4123) 
Support DeepSeek-R1 W4A8 on Hopper

Co-authored-by: Barry Kang <43644113+Barry-Delaney@users.noreply.github.com>
Co-authored-by: Jiang Shao <91270701+StudyingShao@users.noreply.github.com>
Signed-off-by: Barry Kang <43644113+Barry-Delaney@users.noreply.github.com>
 2025-05-14 15:48:07 +08:00
Simeng Liu
286a789549
feat: Add heuristic for GroupRMSNorm kernel selection. (#4047) 
* feat: Add heuristic for GroupRMSNorm kernel selection.

Implements a logistic regression model to dynamically select between:
- GroupRMSNormBaseKernel: Allocates warps proportional to sum of dimensions
  (better SM occupancy in most cases)
- GroupRMSNormLargeBatch: Allocates warps proportional to max dimension
  (better block scheduling in large batch scenarios)

Selection heuristic considers batch size, allocated warps, and scheduling
efficiency on the current GPU architecture. Models for Compute Capability
9.x and 10.x are trained base on nsys kernel runtime data.
The default kernel selection is the base kernel.

The python operator group_rms_norm will use the heuristic by default.
User can pick to use the base or large batch kernels as well.

Signed-off-by: Simeng Liu <simengl@nvidia.com>

* Address the comments.

Signed-off-by: Simeng Liu <simengl@nvidia.com>

---------

Signed-off-by: Simeng Liu <simengl@nvidia.com>
 2025-05-13 08:52:53 +08:00
Yi Zhang
91bf5e6a8e
[TRTLLM-3105][feat] Add Piecewise CUDA Graph Support (#3804) 
Add Piecewise CUDA Graph Support

Signed-off-by: Yi Zhang <187001205+yizhang-nv@users.noreply.github.com>
 2025-05-09 11:04:01 +08:00
Yukun He
5b61486d87
chore: Clean up the legacy DeepseekAllreudceFusionOp. (#4081) 
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
 2025-05-09 10:20:41 +08:00
Simeng Liu
873c7532fd
feat: Add group_rms_norm kernel to normalize multiple inputs in a single operator. (#3438) 
* feat: Add group_rms_norm kernel to normalize multiple inputs in a single operator.

Previously, the RMSNorm implementation only supported a single input tensor. With group_rms_norm, multiple tensors can be normalized together:
```python
input_a, input_b, ... = group_rms_norm([input_a, input_b, ...])
```
All input tensors must share the same batch dimension. The kernel partitions work by dynamically assigning warp groups proportional to the last dimension of each input, improving launch efficiency and reducing overhead.

This MR provides two implementations:
GroupRMSNormKernel: Optimized for small-to-medium batch sizes
GroupRMSNormKernelLargeBatch: Contains additional optimizations for large batch sizes

Both kernels are currently exposed as custom PyTorch ops. A future MR will implement heuristic-based kernel selection and expose a unified interface.

Signed-off-by: Simeng Liu <simengl@nvidia.com>

* Resolve comments and fix typo with IS_FLASHINFER_AVAILABLE

Signed-off-by: Simeng Liu <simengl@nvidia.com>

---------

Signed-off-by: Simeng Liu <simengl@nvidia.com>
 2025-05-02 13:25:30 +08:00
Yukun He
9cc5922a0b
Clean up allreduce op in Deepseek V3 model. (#3829) 
* Replace deepseek_allreduce op with the new unified allreduce op and moe_allreduce op.
* Minor revision of moe_allreduce op argument names.

Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
 2025-05-01 07:56:36 +08:00