diff --git a/vllm/model_executor/model_loader/reload/layerwise.py b/vllm/model_executor/model_loader/reload/layerwise.py
index 5d4af7d1f..2934b8b5a 100644
--- a/vllm/model_executor/model_loader/reload/layerwise.py
+++ b/vllm/model_executor/model_loader/reload/layerwise.py
@@ -14,6 +14,7 @@ from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBa
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from .meta import (
+    SKIP_TENSORS,
     capture_layer_to_meta,
     get_numel_loaded,
     materialize_layer,
@@ -124,6 +125,8 @@ def initialize_online_processing(layer: torch.nn.Module):
     # Wrap each parameter's weight loader
     # Note that nested wrapping will occur for shared tensors
     for name, tensor in get_layer_tensors(layer).items():
+        if name in SKIP_TENSORS:
+            continue
         if _get_weight_loader(tensor).__name__ != "online_process_loader":
             tensor.weight_loader = make_online_process_loader(layer, name)
 
diff --git a/vllm/model_executor/model_loader/reload/meta.py b/vllm/model_executor/model_loader/reload/meta.py
index 82bf9ce3d..91fce6f57 100644
--- a/vllm/model_executor/model_loader/reload/meta.py
+++ b/vllm/model_executor/model_loader/reload/meta.py
@@ -27,6 +27,7 @@ SKIP_TENSORS: set[str] = {
     "expert_global_to_physical",
     "expert_physical_to_global",
     "expert_local_to_global",
+    "e_score_correction_bias",
 }