[Bugfix][EPLB] Prevent user-provided EPLB config from being overwritten with defaults (#29911)

Signed-off-by: Sage Moore <sage@neuralmagic.com>
(cherry picked from commit e6f114ac)

tests/distributed/test_eplb_spec_decode.py

@@ -22,7 +22,14 @@ def get_model_args(
         "num_speculative_tokens": 1,
         "max_model_len": model_max_len,
     }
-
+    eplb_config = {
+        "num_redundant_experts": tp_size,
+        "window_size": 128,
+        "step_interval": 1024,
+        "log_balancedness": False,
+    }
+    if use_async:
+        eplb_config["use_async"] = True
     model_args = {
         "pretrained": model_name,
         "dtype": "auto",
@@ -31,15 +38,10 @@ def get_model_args(
         "gpu_memory_utilization": 0.7,
         "speculative_config": speculative_config,
         "enable_expert_parallel": True,
-        "num_redundant_experts": tp_size,
-        "eplb_window_size": 128,
-        "eplb_step_interval": 1024,
-        "eplb_log_balancedness": False,
+        "eplb_config": eplb_config,
         "enable_eplb": True,
         "max_model_len": model_max_len,
     }
-    if use_async:
-        model_args["eplb_config"] = {"use_async": True}
     return model_args
 
 

vllm/engine/arg_utils.py

@@ -420,10 +420,6 @@ class EngineArgs:
     )
     _api_process_count: int = ParallelConfig._api_process_count
     _api_process_rank: int = ParallelConfig._api_process_rank
-    num_redundant_experts: int = EPLBConfig.num_redundant_experts
-    eplb_window_size: int = EPLBConfig.window_size
-    eplb_step_interval: int = EPLBConfig.step_interval
-    eplb_log_balancedness: bool = EPLBConfig.log_balancedness
     max_parallel_loading_workers: int | None = (
         ParallelConfig.max_parallel_loading_workers
     )
@@ -1581,16 +1577,6 @@ class EngineArgs:
             )
             self.disable_nccl_for_dp_synchronization = True
 
-        # Forward the deprecated CLI args to the EPLB config.
-        if self.num_redundant_experts is not None:
-            self.eplb_config.num_redundant_experts = self.num_redundant_experts
-        if self.eplb_window_size is not None:
-            self.eplb_config.window_size = self.eplb_window_size
-        if self.eplb_step_interval is not None:
-            self.eplb_config.step_interval = self.eplb_step_interval
-        if self.eplb_log_balancedness is not None:
-            self.eplb_config.log_balancedness = self.eplb_log_balancedness
-
         parallel_config = ParallelConfig(
             pipeline_parallel_size=self.pipeline_parallel_size,
             tensor_parallel_size=self.tensor_parallel_size,