[Graph Optimization] Add the CUDAGraph usage switch for Draft Model

fastdeploy/config.py

@@ -591,6 +591,9 @@ def __init__(
         """ Whether to use shared memory pool for multi capture_size """
         self.use_unique_memory_pool: bool = False
 
+        """ Whether to use cudagraph for draft model."""
+        self.draft_model_use_cudagraph: bool = True
+
         self.max_capture_size: int = None
         self.real_shape_to_captured_size: dict[int, int] = None
         # CINN Config ...

fastdeploy/spec_decode/mtp.py

@@ -82,6 +82,7 @@ def __init__(
         self._init_model_inputs()
 
         # CUDA Graph
+        self.draft_model_use_cudagraph = self.graph_opt_config.draft_model_use_cudagraph
         self.cudagraph_capture_sizes = list(reversed(self.graph_opt_config.cudagraph_capture_sizes))
         self.sot_warmup_sizes = self.graph_opt_config.sot_warmup_sizes
 
@@ -591,7 +592,7 @@ def _initialize_forward_meta(self, step_use_cudagraph: bool = False):
             attn_backend.init_attention_metadata(self.forward_meta)
 
         # TODO(gongshaotian): Use CUDAGraph with Draft Model
-        self.forward_meta.step_use_cudagraph = step_use_cudagraph
+        self.forward_meta.step_use_cudagraph = step_use_cudagraph and self.draft_model_use_cudagraph
 
     def exist_prefill(self):
         """