biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Qwen3Next] Fixes the cuda graph capture conditions under large batch sizes (#24660) (#24667)

Browse Source 
Signed-off-by: Tao He <linzhu.ht@alibaba-inc.com>
This commit is contained in:
Tao He
2025-09-13 06:31:32 +08:00
committed by GitHub
parent 5fe643fc26
commit 8226dd56bf
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/v1/attention/backends/gdn_attn.py
View File

@@ -209,7 +209,8 @@ class GDNAttentionMetadataBuilder(
# prepare tensors for cudagraph # prepare tensors for cudagraph
if (self.use_full_cuda_graph and num_prefills == 0 and num_decodes == 0 if (self.use_full_cuda_graph and num_prefills == 0 and num_decodes == 0
and num_spec_decodes <= self.decode_cudagraph_max_bs): and num_spec_decodes <= self.decode_cudagraph_max_bs
and m.num_actual_tokens <= self.decode_cudagraph_max_bs):
num_total_tokens = self.vllm_config.pad_for_cudagraph( num_total_tokens = self.vllm_config.pad_for_cudagraph(
m.num_actual_tokens) m.num_actual_tokens)
batch_size = num_total_tokens // (self.num_spec + 1) batch_size = num_total_tokens // (self.num_spec + 1)