EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity
2,728 Commits 8 Branches 0 Tags
aac531c53b0166bc3883d1f6491f7f0fbb928197
Commit Graph

92 Commits

Author SHA1 Message Date
kk
5a144a8ab9 Fix run time error in ROCm platform (#5147) 
Co-authored-by: wunhuang <wunhuang@amd.com>
Co-authored-by: root <root@dell300x-pla-t10-17.pla.dcgpu>
 2025-04-07 22:49:40 -07:00
Hubert Lu
afb752bcbe [AMD] Fix missing per_token_group_quant_fp8 for ROCm (#5140) 2025-04-07 22:38:25 -07:00
Chang Su
f04c80dc42 Add Llama4 support (#5092) 
Co-authored-by: Cheng Wan <cwan39@gatech.edu>
Co-authored-by: fzyzcjy <ch271828n@outlook.com>
Co-authored-by: ispobock <ispobaoke@163.com>
 2025-04-07 00:29:36 -07:00
Xiaoyu Zhang
924ca7c92c Add DeepSeek V3/R1 shared experts fusion (#4918) 2025-04-04 01:59:29 -07:00
fzyzcjy
6ff9c6a5e7 Cleanup unused resources after DeepEP operation (#4996) 2025-04-04 00:36:04 -07:00
fzyzcjy
77e929a1a2 Support async DeepEP by splitting into two stages (#4995) 2025-04-04 00:32:27 -07:00
fzyzcjy
febe21ce03 Small refactor DeepEPDispatcher into subclasses (#4994) 2025-04-04 00:24:18 -07:00
Tommy Yang
31035dda44 Add H20 fused MoE kernel tuning configs for DeepSeek V3/R1 (#5057) 2025-04-03 22:14:28 -07:00
AniZpZ
d95269f9b3 [2/3] fix dsv3 awq issue (#4625) 
Co-authored-by: 晟海 <huangtingwei.htw@antgroup.com>
Co-authored-by: laixinn <xielx@shanghaitech.edu.cn>
 2025-04-03 17:36:39 -07:00
fzyzcjy
8e10fec9a8 Small refactor DeepEPMode to clean up code a bit (#4992) 2025-04-03 02:56:44 -07:00
saltyfish66
e41549c3d6 fix: fix illegal cuda memory access at fused_moe_kernel (#4727) 
Co-authored-by: yuethe <yuethe@tencent.com>
 2025-04-03 00:07:32 -07:00
Jinyan Chen
23c764b18a [Feature] Support DeepEP Low Latency (#4767) 
Co-authored-by: sleepcoo <sleepcoo@gmail.com>
Co-authored-by: laixinn <xielx@shanghaitech.edu.cn>
Co-authored-by: ch-wan <cwan39@gatech.edu>
 2025-04-01 09:23:25 -07:00
Qingquan Song
044c315970 Make torch compile configurable for biased_grouped_topk (#4749) 2025-03-28 10:57:52 -07:00
Lianmin Zheng
74e0ac1dbd Clean up import vllm in quantization/__init__.py (#4834) 2025-03-28 10:34:10 -07:00
Xiaoyu Zhang
04e3ff6975 Support compressed tensors fp8w8a8 (#4743) 2025-03-26 13:21:25 -07:00
yuhsaun-t
199bb01d00 Add endpoints to dump selected expert ids (#4435) 
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
 2025-03-24 21:34:19 -07:00
fzyzcjy
ca75741e86 Support async in DeepEP (#4610) 
Co-authored-by: Cheng Wan <cwan39@gatech.edu>
 2025-03-22 22:39:56 -07:00
fzyzcjy
c6d549e773 Multiple tiny code cleanups (#4608) 2025-03-22 22:39:11 -07:00
xutizhou
c2bd094d6e Optimize Permute Kernel in DeepEP (#4643) 
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
 2025-03-22 14:30:34 -07:00
Jinyan Chen
f44db16c8e [Feature] Integrate DeepEP into SGLang (#4232) 
Co-authored-by: Cheng Wan <cwan39@gatech.edu>
Co-authored-by: Xuting Zhou <xutingz@nvidia.com>
 2025-03-19 08:16:31 -07:00
Ke Bao
3ded4b215d Revert "feat: update grouped_topk to support softmax and sigmoid" (#4505) 2025-03-17 11:30:26 -07:00
Wenbo Yang
75b656488a Support serving DeepSeek-R1-Channel-INT8 with 32 L40S. (#4418) 2025-03-17 00:03:43 -07:00
Mick
0f52fb55ec config: Update fused moe config (#4493) 2025-03-16 23:51:58 -07:00
Mick
8ec2ce0726 perf: update fused moe config (#4459) 2025-03-15 21:23:57 -07:00
Yineng Zhang
ad1ae7f7cd use topk_softmax with sgl-kernel (#4439) 2025-03-14 15:59:06 -07:00
Yineng Zhang
977d7cd26a cleanup deps 1/n (#4400) 
Co-authored-by: sleepcoo <sleepcoo@gmail.com>
 2025-03-14 00:00:33 -07:00
Lianmin Zheng
c6d7f8d370 Add some fused elementwise kernels for grok-1 (#4398) 
Co-authored-by: dhou-xai <dhou@x.ai>
Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com>
 2025-03-13 13:39:10 -07:00
Cheng Wan
2f6bacee03 [moe] fix: correct the cache size in the last chunk (#3679) 
Co-authored-by: Abatom <abzhonghua@gmail.com>
 2025-03-12 22:22:13 -07:00
Stefan He
e0917e6bd0 Remove vllm ops scaled fp8 quant and accelerate per token quant by 20-28% (#4215) 
Co-authored-by: Stefan He <bhe@linkedin.com>
 2025-03-12 00:08:03 -07:00
lambert0312
7140ba3573 Add A800 tuning configs for DeepSeek R1/V3 channel-wise INT8 (#4323) 2025-03-11 18:25:56 -07:00
Yineng Zhang
d1da58e275 unify is_cuda and is_hip (#4321) 2025-03-11 18:12:56 -07:00
Ximingwang-09
0f2a2e3c19 Add H20 tuning configs support DeepSeek V3/R1 INT8(block-wise) (#4220) 
Co-authored-by: ximing.wxm <ximing.wxm@antgroup.com>
 2025-03-11 12:32:33 -07:00
yych0745
6a02b32d07 Add A100 tuning configs for DeepSeek R1/V3 channel-wise INT8 (#4287) 
Co-authored-by: HandH1998 <1335248067@qq.com>
 2025-03-11 00:49:06 -07:00
lambert0312
d3ecd63204 Add A800 tuning configs support DeepSeek V3/R1 BF16 and INT8(block-wise) (#4136) 2025-03-11 00:32:25 -07:00
Lianmin Zheng
00d25a7f5e Fix quantization and nightly tests (#4258) 2025-03-10 03:06:21 -07:00
HandH1998
c7f254468f [Feature] DeepSeek V3/R1 INT8 Quantization (channel-wise) (#3888) 
Co-authored-by: yych0745 <1398089567@qq.com>
Co-authored-by: sleepcoo <sleepcoo@gmail.com>
Co-authored-by: b0urnee <2769086541@qq.com>
 2025-03-06 20:54:52 -08:00
HAI
13bc39c5d6 ROCm: enable trillion-parameter MoE models with INT4-FP8 single node (#4152) 2025-03-06 15:33:02 -08:00
HAI
71ab0dabe0 Fix the moe padding conditional logic (#4081) 2025-03-05 10:56:51 -08:00
HAI
51d25405a7 ROCm: update aiter and its usage to fused moe (bloat16, fp8, fp8 block-quant) (#4053) 2025-03-04 03:00:46 -08:00
kk
11eea69e70 Fix assert options.num_stages != 0 error in the latest ROCm build image (#4049) 
Co-authored-by: wunhuang <wunhuang@amd.com>
 2025-03-03 20:37:03 -08:00
Lianmin Zheng
66301e124f Improve code styles (#4021) 2025-03-03 03:20:23 -08:00
Lianmin Zheng
ac2387279e Support penalty in overlap mode; return logprob with chunked prefill; improve benchmark scripts (#3988) 
Co-authored-by: SangBin Cho <rkooo567@gmail.com>
Co-authored-by: dhou-xai <dhou@x.ai>
Co-authored-by: Hanming Lu <hanming_lu@berkeley.edu>
 2025-03-03 00:12:04 -08:00
laixin
b0df5d240b Tuning Script for Feature DeepSeek V3/R1 INT8 Quantization (block-wise) (#3922) 
Co-authored-by: sleepcoo <sleepcoo@gmail.com>
 2025-02-27 10:59:46 +00:00
lukec
21463e321a Expert Parallelism (EP) Support for DeepSeek V3/R1 (#3602) 
Co-authored-by: laixin <xielx@shanghaitech.edu.cn>
Co-authored-by: HandH1998 <1335248067@qq.com>
Co-authored-by: laixin <q865809639@gmail.com>
 2025-02-26 02:29:37 -08:00
laixin
1a6e97577a Feature DeepSeek V3/R1 INT8 Quantization (block-wise) (#3730) 
Co-authored-by: HandH1998 <1335248067@qq.com>
 2025-02-24 05:43:35 -08:00
zixuanzhang226
0c227ee373 feat: update grouped_topk to support softmax and sigmoid (#3680) 2025-02-21 16:30:15 +08:00
HAI
5c54ef0352 AMD/ROCm: update AITER repo to ROCm/aiter (#3747) 2025-02-21 00:18:08 -08:00
chenxiaobing
d5d80ab477 [Bugfix] Fix scores mask for moe topk (#3705) 2025-02-21 02:17:23 +08:00
Cheng Wan
6b0aeb58fd [moe] optim: reduce memory consumption in fused_moe (#3692) 2025-02-20 02:25:05 +08:00
yigex
ddf39d3fce [ROCm] Optimal MOE Tuning for AMD Radeon Graphics (#3567) 2025-02-17 17:54:10 -08:00