update cuda_graph_sizes

vllm/config/__init__.py

@@ -3787,7 +3787,7 @@ class VllmConfig:
             batch_size_capture_list = []
             if self.model_config is not None and \
                 not self.model_config.enforce_eager:
-                if self.model_config.use_mla and self.compilation_config.full_cuda_graph and self.scheduler_config.max_num_seqs<=256:
+                if self.model_config.use_mla and self.compilation_config.full_cuda_graph and self.scheduler_config.max_num_seqs<=512:
                     cuda_graph_sizes = [self.scheduler_config.max_num_seqs]
                 else:
                     cuda_graph_sizes = self.scheduler_config.cuda_graph_sizes