fix DP&&TP (#3872)

Author: lizhenyun01

fastdeploy/config.py

@@ -1284,10 +1284,6 @@ def check(self):
             f"be less than or equal to max_num_partial_prefills: {self.max_num_partial_prefills}"
         )
         assert self.splitwise_role in ["mixed", "prefill", "decode"]
-        # TODO(@wufeisheng): TP and EP need to be supported simultaneously.
-        assert (self.parallel_config.tensor_parallel_size == 1 and self.parallel_config.expert_parallel_size >= 1) or (
-            self.parallel_config.tensor_parallel_size >= 1 and self.parallel_config.expert_parallel_size == 1
-        ), "TP and EP cannot be enabled at the same time"
 
         if not self.cache_config.enable_chunked_prefill:
             if not envs.ENABLE_V1_KVCACHE_SCHEDULER:

fastdeploy/worker/worker_process.py

@@ -655,7 +655,9 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
         num_experts_per_rank = num_experts // parallel_config.expert_parallel_size
         num_experts_start_offset = expert_parallel_rank * num_experts_per_rank
         max_chips_per_node = 16 if current_platform.is_iluvatar() else 8
-        parallel_config.local_data_parallel_id = expert_parallel_rank % max_chips_per_node
+        parallel_config.local_data_parallel_id = parallel_config.data_parallel_rank % (
+            max_chips_per_node // parallel_config.tensor_parallel_size
+        )
 
         parallel_config.expert_parallel_rank = expert_parallel_rank
         parallel_config.num_experts_per_rank = num_experts_per_rank