[CudaGraph] support cudagraph use shared pool (#4199)

* support cudagraph use shared pool

* add envs

* change CUDAGRAPH_POOL_ID to int

* change CUDAGRAPH_POOL_ID to use_memory_pool

* unify use_unique_memory_pool

* fix use_unique_memory_pool

fastdeploy/config.py

@@ -588,6 +588,9 @@ class GraphOptimizationConfig:
         Thus this flag cannot be used together with splitting_ops."""
         self.full_cuda_graph: bool = True
 
+        """ Whether to use shared memory pool for multi capture_size """
+        self.use_unique_memory_pool: bool = False
+
         self.max_capture_size: int = None
         self.real_shape_to_captured_size: dict[int, int] = None
         # CINN Config ...

fastdeploy/model_executor/graph_optimization/cudagraph_piecewise_backend.py

@@ -20,6 +20,7 @@ from typing import Callable, Dict, List, Optional
 
 import paddle.jit.dy2static.utils as jit_utils
 import paddle.nn.layer
+from paddle.base.core import CUDAGraph
 from paddle.device.cuda import graphs
 
 from fastdeploy import envs
@@ -85,17 +86,14 @@ class Dy2StCudaGraphManager:
 class CudaGraphPiecewiseBackend:
     """Manage the capture and replay of CUDA graphs at the subgraph level."""
 
-    def __init__(
-        self,
-        fd_config: FDConfig,
-        runnable: Callable,
-    ):
+    def __init__(self, fd_config: FDConfig, runnable: Callable):
         self.fd_config = fd_config
         self.runnable = runnable
         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
-
+        if self.fd_config.graph_opt_config.use_unique_memory_pool:
+            self.unique_memory_pool_id = CUDAGraph.gen_new_memory_pool_id()
         self._create_entry_dict()
 
         self.cuda_graph_manager = None
@@ -168,7 +166,11 @@ 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)
+                if self.fd_config.graph_opt_config.use_unique_memory_pool
+                else graphs.CUDAGraph()
+            )
             paddle.device.synchronize()
 
             # Capture