支持pd分离:P做跨机pp2tp8，D做单机的tp8.

使用方式P在connector的参数添加："enable_asymmetric_p2p":true,"remote_tp_size":8,"remote_pp_size":1 这3个参数。例：
"kv_connector_extra_config":{"enable_asymmetric_p2p":true,"remote_tp_size":8,"remote_pp_size":1,"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20001"}

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py

@@ -90,6 +90,8 @@ class P2pNcclConnector(KVConnectorBase_V1):
             if role == KVConnectorRole.WORKER else 0
         self._local_rank = get_world_group().local_rank \
             if role == KVConnectorRole.WORKER else 0
+        self._tp_rank = get_tp_group().rank_in_group \
+            if role == KVConnectorRole.WORKER else 0
 
         self.p2p_nccl_engine = P2pNcclEngine(
             local_rank=self._local_rank,
@@ -105,9 +107,19 @@ class P2pNcclConnector(KVConnectorBase_V1):
         self.pp_size = self.parallel_config.pipeline_parallel_size
         self.tp_size = self.parallel_config.tensor_parallel_size
         self.num_card = self.pp_size * self.tp_size
-        self.multiple_machines = 1 if self.num_card > 8 else 0
 
-        if self.is_producer and self.multiple_machines == 1:
+        self.remote_tp_size = self.config.get_from_extra_config(
+            "remote_tp_size", self.tp_size)
+        self.remote_pp_size = self.config.get_from_extra_config(
+            "remote_pp_size", self.pp_size)
+        self.enable_asymmetric_p2p = self.config.get_from_extra_config(
+            "enable_asymmetric_p2p", False)
+        
+        self.remote_num_card = self.remote_tp_size * self.remote_pp_size
+        self.multiple_machines_d = 1 if self.remote_num_card > 8 else 0
+        self.multiple_machines_p = 1 if self.num_card > 8 else 0
+
+        if self.is_producer and self.multiple_machines_p == 1:
             self.ip_map = {}
             self.duplicate_keys = []
             config_file = os.getenv('IP_CONFIG_FILE')
@@ -417,7 +429,7 @@ class P2pNcclConnector(KVConnectorBase_V1):
 
                 pp_rank = (self.parallel_config.rank // self.parallel_config.tensor_parallel_size
                     ) % self.parallel_config.pipeline_parallel_size
-                if (self.multiple_machines):
+                if (self.multiple_machines_p and self.multiple_machines_d):
                     ip_second = self.get_ip_value(ip)
                     if (self.pp_size == 1):
                         if self._rank < 8:
@@ -433,8 +445,16 @@ class P2pNcclConnector(KVConnectorBase_V1):
                             self.p2p_nccl_engine.send_tensor(request_id + "#" + layer_name,
                                                         kv_cache, str(ip_second) + ":" + str(port + self._rank))
                     else:
-                        print("Error: only suppprt pp1 pp2 !!!!!!")
+                        logger.error("Error: multiple machines only suppprt pp1tp16 and pp2tp8!!!!!!")
+                elif (self.multiple_machines_p and not self.multiple_machines_d):
+                    if (self.pp_size == 2):
+                        remote_address = ip + ":" + str(port + self._tp_rank)
+                        self.p2p_nccl_engine.send_tensor(request_id + "#" + layer_name,
+                                                        kv_cache, remote_address)
                     else:
+                        logger.error("Error: P multiple machines D machine only suppprt P:pp2tp8 D:tp8 !!!!!!")
+
+                elif (not self.multiple_machines_p and not self.multiple_machines_d):
                     if (self.pp_size == 1):
                         self.p2p_nccl_engine.send_tensor(request_id + "#" + layer_name,
                                                     kv_cache, remote_address)
@@ -454,8 +474,9 @@ class P2pNcclConnector(KVConnectorBase_V1):
                             self.p2p_nccl_engine.send_tensor(request_id + "#" + layer_name,
                                                         kv_cache, ip + ":" + str(port + i))
                     else:
-                        print("Error: only suppprt pp1 pp2 pp8!!!!!!")
-
+                        logger.error("Error: only suppprt pp1 pp2 pp8!!!!!!")
+                else:
+                        logger.error("Error: not support!!!!!!")
     def wait_for_save(self):
         pass
         # if self.is_producer: