frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 23:03:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix][Minor] Fix full cuda graph bug when max_num_seqs < 512 (#19171)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-06-04 22:41:25 -07:00
committed by GitHub
parent 0678b52251
commit af7fc84fd2
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/v1/worker/gpu_model_runner.py
View File

@ -1737,7 +1737,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
# has num_tokens in total.
assert num_tokens <= self.scheduler_config.max_num_batched_tokens
max_num_reqs = self.scheduler_config.max_num_seqs
num_reqs = max_num_reqs if num_tokens >= max_num_reqs else num_tokens
num_reqs = min(num_tokens, max_num_reqs)
min_tokens_per_req = num_tokens // num_reqs
num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
num_scheduled_tokens_list[-1] += num_tokens % num_reqs
@ -1765,7 +1765,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
self.kv_cache_config.kv_cache_groups):
attn_metadata_i = (
self.attn_metadata_builders[kv_cache_group_id].build(
num_reqs=num_tokens,
num_reqs=num_reqs,
num_actual_tokens=num_tokens,
max_query_len=num_tokens,
common_prefix_len=0,