[BUG] Fix bug for pd in fd (#3034)

* Fix bug for pd in fd

* Fix bug for pd in fd

---------

Co-authored-by: Jiang-Jia-Jun <163579578+Jiang-Jia-Jun@users.noreply.github.com>

Author: rainyfly

fastdeploy/cache_manager/cache_messager.py

@@ -142,6 +142,7 @@ def __init__(
 
         self.gpu_id = gpu_id
         self.cache_info = dict()
+        self.dp_rank_id = local_data_parallel_id
 
         layerwise_send_cache_thread = threading.Thread(target=self._prefill_layerwise_send_cache_thread)
         layerwise_send_cache_thread.daemon = True
@@ -159,29 +160,29 @@ def _prefill_layerwise_send_cache_thread(self):
             prefilled_layer_idx_data = np.zeros(shape=[1], dtype=np.int32)
             try:
                 step_shm_value = IPCSignal(
-                    name=f"splitwise_complete_prefilled_step_{self.rank}",
+                    name=f"splitwise_complete_prefilled_step_{self.dp_rank_id}",
                     array=prefilled_step_idx_data,
                     dtype=np.int32,
                     suffix=self.gpu_id,
                     create=True,
                 )
                 layer_shm_value = IPCSignal(
-                    name=f"splitwise_complete_prefilled_layer_{self.rank}",
+                    name=f"splitwise_complete_prefilled_layer_{self.dp_rank_id}",
                     array=prefilled_layer_idx_data,
                     dtype=np.int32,
                     suffix=self.gpu_id,
                     create=True,
                 )
             except:
                 step_shm_value = IPCSignal(
-                    name=f"splitwise_complete_prefilled_step_{self.rank}",
+                    name=f"splitwise_complete_prefilled_step_{self.dp_rank_id}",
                     array=prefilled_step_idx_data,
                     dtype=np.int32,
                     suffix=self.gpu_id,
                     create=False,
                 )
                 layer_shm_value = IPCSignal(
-                    name=f"splitwise_complete_prefilled_layer_{self.rank}",
+                    name=f"splitwise_complete_prefilled_layer_{self.dp_rank_id}",
                     array=prefilled_layer_idx_data,
                     dtype=np.int32,
                     suffix=self.gpu_id,

fastdeploy/cache_manager/transfer_factory/rdma_cache_transfer.py

@@ -45,7 +45,7 @@ def __init__(
             return
         self.messager = rdma_comm.RDMACommunicator(
             splitwise_role,
-            rank,
+            gpu_id,
             str(rdma_port) if splitwise_role == "decode" else "0",
             cache_k_ptr_list,
             cache_v_ptr_list,