Commit Graph

60 Commits

Author SHA1 Message Date
Lianmin Zheng
b8574f6953 Clean up eagle code (#2756) 2025-01-06 14:54:18 -08:00
Lianmin Zheng
ad20b7957e Eagle speculative decoding part 3: small modifications to the general scheduler (#2709)
Co-authored-by: kavioyu <kavioyu@tencent.com>
2025-01-02 02:09:08 -08:00
Lianmin Zheng
b0524c3789 Eagle speculative decoding part 2: Fix cuda graph + DP attention hanging (#2684)
Co-authored-by: yukavio <kavioyu@gmail.com>
2024-12-31 02:25:05 -08:00
Lianmin Zheng
f44d143949 Support target model verification in the attention backend (#2678)
Co-authored-by: yukavio <kavioyu@gmail.com>
2024-12-30 22:58:55 -08:00
mobicham
a29dd9501d Add GemLite caching after each capture (#2669) 2024-12-30 05:27:29 -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
Ke Bao
e835a50021 Reorg moe code (#2563) 2024-12-24 01:10:22 +08:00
Lianmin Zheng
21e9e63ad5 Print progress bar during cuda graph capture (#2502) 2024-12-17 06:33:46 -08:00
Lianmin Zheng
7a1aecb938 Simplify pytorch sampling kernel and logit processor (#2491) 2024-12-16 14:11:09 -08:00
Lianmin Zheng
641b7d0ae0 [Minor] Improve code style (#2422) 2024-12-09 06:30:35 -08:00
Lianmin Zheng
a6ca736c8e Simplify stream_output (#2398) 2024-12-08 12:27:13 -08:00
Lianmin Zheng
71e2a27753 Fix the cuda graph capture range for small #max-running-requests (#2359) 2024-12-06 14:13:57 +08:00
Lianmin Zheng
2b0fc5941d [Minor] Code style improvements (#2355) 2024-12-04 19:02:08 -08:00
Lianmin Zheng
07ec07ad1f Improve torch compile for fused moe (#2327) 2024-12-03 01:58:25 -08:00
Lianmin Zheng
3c5538f781 Update CI threshold (#2186) 2024-11-25 15:24:17 -08:00
Lianmin Zheng
c4336b2b60 Use custom allreduce w/ torch.compile (#2185) 2024-11-25 14:55:01 -08:00
Lianmin Zheng
5652c56535 Update CI threshold & Improve code style (#2159) 2024-11-24 06:29:38 -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
Yineng Zhang
766192610e feat: update torch 2.5.1 (#2069) 2024-11-18 21:29:13 +08:00
Ke Bao
62832bb272 Support cuda graph for DP attention (#2061) 2024-11-17 16:29:20 -08:00
Lianmin Zheng
a509552087 [minor] Improve code style and compatibility (#1961) 2024-11-08 02:19:41 -08:00
Lianmin Zheng
86e0dde555 Improve the user control of new_token_ratio (#1811) 2024-10-26 16:39:41 -07:00
Lianmin Zheng
2b80978859 Provide an argument to set the maximum batch size for cuda graph (#1809) 2024-10-26 15:09:33 -07:00
Lianmin Zheng
e646c5901e Fix logprob in the overlapped mode (#1795) 2024-10-25 11:06:57 -07:00
yizhang2077
def55bc876 Qwen2vl support cuda graph and disable radix cache (#1780) 2024-10-25 10:45:17 -04:00
Lianmin Zheng
86a2c473b7 [Fix] Fix seq_lens_sum for cuda graph runner in padded cases (#1789) 2024-10-24 21:26:05 -07:00
Lianmin Zheng
384d85ba35 Re-introduce get_cuda_graph_seq_len_fill_value (#1783) 2024-10-24 13:30:11 -07:00
Lianmin Zheng
fc82f5a743 [Fix] Fix cuda graph padding for triton attention backend (#1782) 2024-10-24 12:33:15 -07:00
Lianmin Zheng
0089c4bc96 [Fix] Fix NaN issues by fixing the cuda graph padding values for flashinfer (#1779) 2024-10-24 04:16:59 -07:00
Lianmin Zheng
05b3bf5e8e Crash the server on warnings in CI (#1772) 2024-10-23 16:27:13 -07:00
Lianmin Zheng
ad4125d1a9 Fuse more ops & Simplify token mapping (#1758) 2024-10-22 23:20:43 -07:00
Liangsheng Yin
94cde10920 Llama3.2 vision model support (#1551) 2024-10-21 15:01:21 -07:00
Lianmin Zheng
09603c6dc9 Maintain seq_lens_sum to make more FlashInfer operations non-blocking (#1741) 2024-10-21 01:43:16 -07:00
Lianmin Zheng
f0f8a7699b Simplify the nan detection and greedy check in sampler (#1709) 2024-10-18 20:21:24 -07:00
Lianmin Zheng
36d5acfca5 Rename InputMetadata -> ForwardBatch (#1543) 2024-09-30 02:41:11 -07:00
Lianmin Zheng
3f0fe08d37 Let ModelRunner take InputMetadata as input, instead of ScheduleBatch (#1541) 2024-09-29 20:28:45 -07:00
Ke Bao
8d4ed42ad5 MoE torch compile (#1497) 2024-09-24 01:46:59 -07:00
Lianmin Zheng
39bb49d156 Update dockerfile to include datamodel_code_generator (#1492) 2024-09-22 04:49:16 -07:00
Lianmin Zheng
2d346a57c2 Fix padding in the cuda graph (#1469) 2024-09-19 01:52:15 -07:00
Ke Bao
76524b70d1 Fix torch compile for deepseek-v2 (#1442) 2024-09-17 00:52:08 -07:00
Lianmin Zheng
9ba1f09760 [Fix] Fix logprob and normalized_logprob (#1428) 2024-09-15 06:36:06 -07:00
Liangsheng Yin
70b6802982 Optimize conflicts between CUDA graph and vocab mask tensors (#1392) 2024-09-13 20:27:53 -07:00
Lianmin Zheng
3efa798116 Support cuda graph in the triton attention backend (#1401) 2024-09-12 00:36:55 -07:00
Lianmin Zheng
fec185ce0c Refactor attention backend (#1381) 2024-09-11 11:44:26 -07:00
Liangsheng Yin
144bc70fcc Organize flashinfer indices update (#1378) 2024-09-10 17:38:59 -07:00
Liangsheng Yin
a5a134f39f Fix bugs in sampler with CUDA graph / torch.compile (#1306) 2024-09-02 23:18:48 +00:00
Liangsheng Yin
381dd57bd6 Sampler cudagraph (#1253) 2024-08-28 18:58:52 -07:00
Yineng Zhang
f25f4dfde5 hotfix: revert sampler CUDA Graph (#1242) 2024-08-28 21:16:47 +10:00
Lianmin Zheng
6cc38b2bf3 [Minor] Add more type annotations (#1237) 2024-08-28 00:54:26 -07:00