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

Signed-off-by: Tao He <[email protected]>

Author: sighingnow

vllm/v1/attention/backends/gdn_attn.py

@@ -209,7 +209,8 @@ def build(  # type: ignore[override]
 
         # prepare tensors for cudagraph
         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(
                 m.num_actual_tokens)
             batch_size = num_total_tokens // (self.num_spec + 1)