xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-23 06:55:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix cuda graph sizes when running with speculative decoding (#30330)

Browse Source 
Signed-off-by: Patryk Saffer <patryk.saffer99@gmail.com>
Signed-off-by: PatrykSaffer <patryk.saffer@mistral.ai>
Co-authored-by: Patryk Saffer <patryk.saffer99@gmail.com>
This commit is contained in:
PatrykSaffer 2025-12-10 01:47:07 +01:00 committed by GitHub
parent 03b5f940fd
commit 4c2e10ea19
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/config/vllm.py
View File

@ -1047,8 +1047,14 @@ class VllmConfig:
self.compilation_config.max_cudagraph_capture_size self.compilation_config.max_cudagraph_capture_size
) )
if max_cudagraph_capture_size is None: if max_cudagraph_capture_size is None:
decode_query_len = 1
if (
self.speculative_config
and self.speculative_config.num_speculative_tokens
):
decode_query_len += self.speculative_config.num_speculative_tokens
max_cudagraph_capture_size = min( max_cudagraph_capture_size = min(
self.scheduler_config.max_num_seqs * 2, 512 self.scheduler_config.max_num_seqs * decode_query_len * 2, 512
) )
max_num_tokens = self.scheduler_config.max_num_batched_tokens max_num_tokens = self.scheduler_config.max_num_batched_tokens
max_cudagraph_capture_size = min(max_num_tokens, max_cudagraph_capture_size) max_cudagraph_capture_size = min(max_num_tokens, max_cudagraph_capture_size)