don't double eplb prefix

Signed-off-by: Patryk Saffer <[email protected]>

Author: Patryk999

vllm/distributed/eplb/eplb_state.py

@@ -297,7 +297,7 @@ def build(
             device=device,
         )
 
-        eplb_load_path = parallel_config.eplb_config.eplb_load_path
+        eplb_load_path = parallel_config.eplb_config.load_path
         eplb_save_dir = parallel_config.eplb_config.save_dir
 
         eplb_step_interval = parallel_config.eplb_config.step_interval

vllm/model_executor/models/deepseek_v2.py

@@ -246,7 +246,7 @@ def __init__(
             else self.routed_scaling_factor,
             e_score_correction_bias=self.gate.e_score_correction_bias,
             enable_eplb=self.enable_eplb,
-            eplb_record_metrics=eplb_config.eplb_record_metrics,
+            eplb_record_metrics=eplb_config.record_metrics,
             num_redundant_experts=self.n_redundant_experts,
             is_sequence_parallel=self.is_sequence_parallel,
             n_shared_experts=config.n_shared_experts

vllm/model_executor/models/mixtral.py

@@ -140,7 +140,7 @@ def __init__(
             dp_size=dp_size,
             prefix=f"{prefix}.experts",
             enable_eplb=self.enable_eplb,
-            eplb_record_metrics=parallel_config.eplb_config.eplb_record_metrics,
+            eplb_record_metrics=parallel_config.eplb_config.record_metrics,
             num_redundant_experts=self.n_redundant_experts,
         )
 

vllm/model_executor/models/qwen3_moe.py

@@ -169,7 +169,7 @@ def __init__(
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
             enable_eplb=self.enable_eplb,
-            eplb_record_metrics=eplb_config.eplb_record_metrics,
+            eplb_record_metrics=eplb_config.record_metrics,
             num_redundant_experts=self.n_redundant_experts,
             is_sequence_parallel=self.is_sequence_parallel,
         )

vllm/model_executor/models/qwen3_next.py

@@ -168,7 +168,7 @@ def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
             quant_config=quant_config,
             prefix=f"{prefix}.experts",
             enable_eplb=self.enable_eplb,
-            eplb_record_metrics=eplb_config.eplb_record_metrics,
+            eplb_record_metrics=eplb_config.record_metrics,
             num_redundant_experts=self.n_redundant_experts,
             is_sequence_parallel=self.is_sequence_parallel,
         )

vllm/v1/worker/gpu_model_runner.py

@@ -1992,11 +1992,11 @@ def eplb_step(self, is_dummy: bool = False, is_profile: bool = False) -> None:
         """
         if not self.parallel_config.enable_eplb:
             return
-        if is_profile and self.parallel_config.eplb_config.eplb_load_path is not None:
+        if is_profile and self.parallel_config.eplb_config.load_path is not None:
             return
         if (
-            self.parallel_config.eplb_config.eplb_load_path is not None
-            and self.parallel_config.eplb_config.eplb_save_dir is None
+            self.parallel_config.eplb_config.load_path is not None
+            and self.parallel_config.eplb_config.save_dir is None
         ):
             return
 
@@ -2930,9 +2930,9 @@ def load_model(self, eep_scale_up: bool = False) -> None:
                 old_global_expert_indices,
                 rank_mapping,
             )
-            if self.parallel_config.eplb_config.eplb_load_path is not None:
+            if self.parallel_config.eplb_config.load_path is not None:
                 self.eplb_state.rearrange(self.model)
-                if self.parallel_config.eplb_config.eplb_save_dir is None:
+                if self.parallel_config.eplb_config.save_dir is None:
                     self.eplb_state = None
 
         if (