From 4729b90838a3160505ccced0d96d82d3e7069277 Mon Sep 17 00:00:00 2001 From: Aaron Hao Date: Thu, 2 Apr 2026 21:15:05 -0700 Subject: [PATCH] [Bug] Add e_score_correction_bias to SKIP_TENSORS (#38746) Signed-off-by: ahao-anyscale --- vllm/model_executor/model_loader/reload/layerwise.py | 3 +++ vllm/model_executor/model_loader/reload/meta.py | 1 + 2 files changed, 4 insertions(+) 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", }