[Bugfix] Skip linear bias in layerwise reload to avoid corruption (#49586)
Signed-off-by: li-jinpeng <3332126450@qq.com> Signed-off-by: xymli <xymli@tencent.com> Co-authored-by: xymli <xymli@tencent.com>
li-jinpeng committed
Jul 24, 2026 at 16:53 UTC
163ecba37766e22036d8456a64d9122520647ab7
1 file changed
+3
vllm/model_executor/model_loader/reload/meta.py
+3
@@ -29,6 +29,9 @@ SKIP_TENSORS: set[str] = {
29
"expert_physical_to_global",
30
"expert_local_to_global",
31
"e_score_correction_bias",
32
+ # Built after create_weights(), so it is not tracked by the layerwise-reload
33
+ # trigger and would be re-materialized into uninitialized memory. Skip it.
34
+ "bias",
35
}
36
37