SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

guard size mismatch check to only quantized models (#38397)

Browse Source 
fix
This commit is contained in:
Marc Sun
2025-05-27 11:45:03 +02:00
committed by GitHub
parent 1a5be2f5c0
commit 55f2333366
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/transformers/modeling_utils.py
View File

@@ -1577,7 +1577,8 @@ def _find_mismatched_keys(
# This skips size mismatches for 4-bit weights. Two 4-bit values share an 8-bit container, causing size differences. # This skips size mismatches for 4-bit weights. Two 4-bit values share an 8-bit container, causing size differences.
# Without matching with module type or parameter type it seems like a practical way to detect valid 4bit weights. # Without matching with module type or parameter type it seems like a practical way to detect valid 4bit weights.
if not ( if not (
new_state_dict[key].shape[-1] == 1 is_quantized
and new_state_dict[key].shape[-1] == 1
and new_state_dict[key].numel() * 2 == model_state_dict[key].numel() and new_state_dict[key].numel() * 2 == model_state_dict[key].numel()
): ):
mismatched_keys.append(key) mismatched_keys.append(key)