A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 14:45 UTC
[Bug] Add e_score_correction_bias to SKIP_TENSORS (#38746)
Signed-off-by: ahao-anyscale <ahao@anyscale.com>
Aaron Hao committed
Apr 2, 2026 at 21:15 UTC
4729b90838a3160505ccced0d96d82d3e7069277
2 files changed
+4
vllm/model_executor/model_loader/reload/layerwise.py
+3
@@ -14,6 +14,7 @@ from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBa
14
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
15
16
from .meta import (
17
+ SKIP_TENSORS,
18
capture_layer_to_meta,
19
get_numel_loaded,
20
materialize_layer,
@@ -124,6 +125,8 @@ def initialize_online_processing(layer: torch.nn.Module):
125
# Wrap each parameter's weight loader
126
# Note that nested wrapping will occur for shared tensors
127
for name, tensor in get_layer_tensors(layer).items():
128
+ if name in SKIP_TENSORS:
129
+ continue
130
if _get_weight_loader(tensor).__name__ != "online_process_loader":
131
tensor.weight_loader = make_online_process_loader(layer, name)
132
vllm/model_executor/model_loader/reload/meta.py
+1
@@ -27,6 +27,7 @@ SKIP_TENSORS: set[str] = {
27
"expert_global_to_physical",
28
"expert_physical_to_global",
29
"expert_local_to_global",
30
+ "e_score_correction_bias",
31
}
32
33