enable fused_shared_experts by default

Author: Claude Code

lightllm/models/deepseek2/layer_weights/transformer_layer_weight.py

@@ -40,8 +40,8 @@ def _parse_config(self):
         self.num_attention_heads = self.network_config_["num_attention_heads"]
         self.kv_lora_rank = self.network_config_["kv_lora_rank"]
         self.num_fused_shared_experts = 0
-        if get_env_start_args().enable_fused_shared_experts and self.is_moe:
-            # MOE_MODE 处于 TP 模式下才能使能 enable_fused_shared_experts
+        if not get_env_start_args().disable_fused_shared_experts and self.is_moe:
+            # MOE_MODE 处于 TP 模式下才能使能 fused_shared_experts
             moe_mode = os.getenv("MOE_MODE", "TP")
             assert moe_mode == "TP"
             self.num_fused_shared_experts = self.network_config_.get("n_shared_experts", 0)

lightllm/server/api_cli.py

@@ -466,9 +466,9 @@ def make_argument_parser() -> argparse.ArgumentParser:
         help="""Whether to update the redundant expert for deepseekv3 model by online expert used counter.""",
     )
     parser.add_argument(
-        "--enable_fused_shared_experts",
+        "--disable_fused_shared_experts",
         action="store_true",
-        help="""Whether to enable fused shared experts for deepseekv3 model. only work when MOE_MODE=TP """,
+        help="""Whether to disable fused shared experts for deepseekv3 model. only work when MOE_MODE=TP """,
     )
     parser.add_argument(
         "--mtp_mode",