[CUDAGraph]CUDA Graph support unique memory pool (#4230)

* cuda graph use unique memory pool

* fix custom device import bug

* refine code

* refine code

* refine code

fastdeploy/config.py

@@ -841,8 +841,13 @@ class GraphOptimizationConfig:
         Now don't support capture both decode-only and prefill-only"""
         self.full_cuda_graph: bool = True
 
+        """ Maximum CUDA Graph capture size """
         self.max_capture_size: int = None
+        """ Record maps mapped from real shape to captured size to reduce runtime overhead """
         self.real_shape_to_captured_size: dict[int, int] = None
+        """ Whether to use shared memory pool for multi capture_size """
+        self.use_unique_memory_pool: bool = False
+
         # CINN Config ...
         if args is not None:
             for key, value in args.items():

fastdeploy/model_executor/graph_optimization/cudagraph_piecewise_backend.py

@@ -96,6 +96,13 @@ class CudaGraphPiecewiseBackend:
         self.cudagraph_capture_sizes = fd_config.graph_opt_config.cudagraph_capture_sizes
         self.warm_up_size = fd_config.graph_opt_config.cudagraph_num_of_warmups
         self.real_shape_to_captured_size = fd_config.graph_opt_config.real_shape_to_captured_size
+        self.unique_memory_pool_id = None
+        if self.fd_config.graph_opt_config.use_unique_memory_pool:
+            # TODO(gongshaotian): Optimize code
+            if paddle.is_compiled_with_cuda():
+                from paddle.base.core import CUDAGraph
+
+                self.unique_memory_pool_id = CUDAGraph.gen_new_memory_pool_id()
 
         self._create_entry_dict()
 
@@ -169,7 +176,7 @@ class CudaGraphPiecewiseBackend:
             input_addresses = [x.data_ptr() for (_, x) in kwargs.items() if isinstance(x, paddle.Tensor)]
             entry.input_addresses = input_addresses
 
-            new_grpah = graphs.CUDAGraph()
+            new_grpah = graphs.CUDAGraph(pool_id=self.unique_memory_pool_id)
             paddle.device.synchronize()
 
             # Capture