[BugFix]Fix ep size (#3092)

* fix ep

* fix num_layer

fastdeploy/cache_manager/prefix_cache_manager.py

@@ -168,7 +168,7 @@ class PrefixCacheManager:
                 + f" --device_id {int(device_ids[i])}"
                 + f" --rank {i}"
                 + f" --splitwise_role {self.splitwise_role}"
-                + f" --num_layers {cache_config.model_cfg.num_layers}"
+                + f" --num_layers {cache_config.model_cfg.num_hidden_layers}"
                 + f" --head_dim {cache_config.model_cfg.head_dim}"
                 + f" --kv_num_head {kv_num_head}"
                 + f" --mp_num {tensor_parallel_size}"

fastdeploy/config.py

@@ -270,6 +270,8 @@ class ParallelConfig:
             if hasattr(self, key):
                 setattr(self, key, value)
 
+        # currently, the expert parallel size is equal data parallel size
+        self.expert_parallel_size = self.data_parallel_size
         self.use_ep = self.expert_parallel_size > 1
         if self.splitwise_role == "mixed":
             self.moe_phase = MoEPhase(phase="prefill")

fastdeploy/engine/engine.py

@@ -1082,6 +1082,7 @@ class LLMEngine:
             f" --splitwise_role {self.cfg.splitwise_role}"
             f" --kv_cache_ratio {self.cfg.cache_config.kv_cache_ratio}"
             f" --expert_parallel_size {self.cfg.parallel_config.expert_parallel_size}"
+            f" --data_parallel_size {self.cfg.parallel_config.data_parallel_size}"
             f" --quantization {self.cfg.model_config.quantization}"
             f" --ori_vocab_size {ori_vocab_size}"
             f" --speculative_config '{self.cfg.speculative_config.to_json_string()}'"

fastdeploy/worker/worker_process.py

@@ -513,6 +513,12 @@ def parse_args():
         default=1,
         help="expert parallel size",
     )
+    parser.add_argument(
+        "--data_parallel_size",
+        type=int,
+        default=1,
+        help="data parallel size",
+    )
     parser.add_argument(
         "--enable_expert_parallel",
         action="store_true",