[Fix] MLA only supports decode-only full CUDAGraph capture. Make sure all...

[Fix] MLA only supports decode-only full CUDAGraph capture. Make sure all cudagraph capture sizes <= max_num_seq.

vllm/config.py

@@ -4745,7 +4745,10 @@ class VllmConfig:
             batch_size_capture_list = []
             if self.model_config is not None and \
                 not self.model_config.enforce_eager:
-                cuda_graph_sizes = self.scheduler_config.cuda_graph_sizes
+                if self.model_config.use_mla and self.compilation_config.full_cuda_graph:
+                    cuda_graph_sizes = [256]
+                else:
+                    cuda_graph_sizes = self.scheduler_config.cuda_graph_sizes 
                 if len(cuda_graph_sizes) == 1:
                     batch_size_capture_list = [1, 2, 4] + [
                         i for i in range(8, cuda_graph_sizes[0] + 1, 8)