[Feature] optimizer deepep buffer clean

fastdeploy/model_executor/graph_optimization/cudagraph_piecewise_backend.py

@@ -20,7 +20,6 @@
 
 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
@@ -93,7 +92,10 @@ def __init__(self, fd_config: FDConfig, runnable: Callable):
         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()
+            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()
 
         self.cuda_graph_manager = None

fastdeploy/model_executor/layers/moe/fused_moe_backend_base.py

@@ -161,11 +161,11 @@ def apply(
         """
         if layer.ep_size > 1:
             if layer.fd_config.parallel_config.moe_phase.phase == "prefill":
-                if layer.fd_config.parallel_config.splitwise_role == "mixed":
+                if layer.layer_idx == 0:
                     self.ep_prefill_runner.clean_low_latency_buffer()
                 return self.apply_ep_prefill(layer, x, gate)
             else:
-                if layer.fd_config.parallel_config.splitwise_role == "mixed":
+                if layer.layer_idx == 0:
                     self.ep_decoder_runner.clean_low_latency_buffer()
                 return self.apply_ep_decode(layer, x, gate)
         else: