Commit Graph

35 Commits

Author SHA1 Message Date
Lianmin Zheng
82dec1f70b Remove redundant type conversion (#4513) 2025-03-17 05:57:35 -07:00
JieXin Liang
0212d2e288 [Fix] use torch.inference_mode() instead of torch.no_grad() (#4372) 2025-03-16 22:54:16 -07:00
Lianmin Zheng
c76040e31b Support page size > 1 (#4356) 2025-03-12 22:22:39 -07:00
Ying Sheng
d3d4d76758 [Eagle] Refactor eagle speculative decoding (#3986)
Co-authored-by: Ke Bao <ISPObaoke@163.com>
2025-03-05 08:06:07 -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
Jackmin801
5f0e7de339 [Feat] Return hidden states (experimental) (#3364)
Co-authored-by: Chayenne <zhaochen20@outlook.com>
2025-02-10 15:54:37 -08:00
Chunyuan WU
63051738a9 Enable CPU device on SGLang (#2806) 2025-01-16 21:22:53 -08:00
Lianmin Zheng
8b6ce52e92 Support multi-node DP attention (#2925)
Co-authored-by: dhou-xai <dhou@x.ai>
2025-01-16 11:15:00 -08:00
Lianmin Zheng
f65c13b559 Remove normalized_prompt_logprobs from the engine to make code easier to maintain (#2902) 2025-01-15 04:54:14 -08:00
Lianmin Zheng
9c6ba2484f Refactor logprob computation to return the real logprob used in sampling (#2664) 2024-12-30 04:51:38 -08:00
fzyzcjy
fd28640dc5 Add update_weights_from_tensor (#2631) 2024-12-28 13:30:27 -08:00
Lianmin Zheng
0e7409adb6 Fix the overlap for xgrammar (#2377) 2024-12-06 05:49:29 -08:00
Qun Yang
37ee906f61 Add more support for intel Gaudi accelerators (#2357) 2024-12-06 01:16:33 -08:00
Chayenne
983bfcf386 Online weight updates from torch.distributed (#2279) 2024-12-01 23:23:18 -08:00
Chayenne
7d1485d376 Add get weights by parameter name for llama (#2266) 2024-11-29 23:36:38 -08:00
Chayenne
7d5d1d3d29 udate weights from disk (#2265) 2024-11-30 01:17:00 +00:00
Lianmin Zheng
d4fc1a70e3 Crash the server correctly during error (#2231) 2024-11-28 00:22:39 -08:00
Xuehai Pan
62a4a339eb docs: fix module docstrings and copyright headers (#2077) 2024-11-22 22:16:53 +08:00
Lianmin Zheng
722530fa01 Enable overlap scheduler by default for the triton attention backend (#2105) 2024-11-20 02:58:35 -08:00
Lianmin Zheng
3295cd8af2 Allow skipping warmup in bench_offline_throughput.py (#2103) 2024-11-20 01:25:21 -08:00
Lianmin Zheng
7d671e4ad2 Enable overlap by default (#2067) 2024-11-19 22:07:58 -08:00
Lianmin Zheng
ffd20fcd03 Make constrained decoding work for overlap scheduler (#2095) 2024-11-19 15:04:43 -08:00
Lianmin Zheng
b7a065eae3 Use cuda event wait and synchronization instead of busy waiting (#2089) 2024-11-19 00:21:46 -08:00
Lianmin Zheng
b110453802 Simplify logits penalizer (#2086) 2024-11-18 17:48:28 -08:00
Lianmin Zheng
a9e90b4bce [Minor] Fix styles for overlap mode (#2068) 2024-11-17 19:49:20 -08:00
Ke Bao
62832bb272 Support cuda graph for DP attention (#2061) 2024-11-17 16:29:20 -08:00
Lianmin Zheng
38625e2139 Remove monkey_patch_vllm_dummy_weight_loader (#2064) 2024-11-17 15:48:12 -08:00
Lianmin Zheng
edad373135 Fix illegal memory access in overlap mode & Use more fused triton kernels for building meta data (#2051) 2024-11-16 16:14:23 -08:00
Ke Bao
976bc302e5 Support DP MLA (#1970) 2024-11-16 09:01:43 +00:00
Lianmin Zheng
e646c5901e Fix logprob in the overlapped mode (#1795) 2024-10-25 11:06:57 -07:00
Lianmin Zheng
ad4125d1a9 Fuse more ops & Simplify token mapping (#1758) 2024-10-22 23:20:43 -07:00
Lianmin Zheng
7ce3606891 Faster overlap mode scheduler (#1738) 2024-10-21 04:30:52 -07:00
Lianmin Zheng
cf470fea32 Make token mapping non-blocking in the overlapped mode (#1740) 2024-10-20 23:25:14 -07:00
Lianmin Zheng
b121bc03a3 Simplify batch result resolution (#1735) 2024-10-20 19:47:14 -07:00
Lianmin Zheng
b48edff67f Split the overlapped version of TpModelWorkerClient into a separate file (#1726) 2024-10-20 00:29:29 -07:00