riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-31 02:02:21 +06:00
Code Issues Actions 23 Packages Projects Releases Wiki Activity

Revert #37031 (#37178)

Browse Source 
Update modeling_utils.py
This commit is contained in:
Cyril Vallez 2025-04-01 19:48:15 +02:00 committed by GitHub
parent bc2dea3f54
commit 41f5c3216c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 0 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
src/transformers/modeling_utils.py
View File

@ -761,9 +761,6 @@ def _load_state_dict_into_meta_model(
if is_meta_state_dict:
file_pointer = safe_open(shard_file, framework="pt", device=tensor_device)
# Used to fix the issue mentioned in #37031: when loading a model with tied weights in state_dict + `tie_word_embeddings = False`,
# we need to make sure they are not loaded as tied weights!
data_ptrs = set()
for param_name, empty_param in state_dict.items():
if param_name not in expected_keys:
continue
@ -833,14 +830,8 @@ def _load_state_dict_into_meta_model(
if is_fsdp_enabled():
param_device = "cpu" if is_local_dist_rank_0() else "meta"
# avoid tied weights
if param.data_ptr() in data_ptrs:
param = param.clone()
_load_parameter_into_model(model, param_name, param.to(param_device))
# Add `data_ptr` of `model.state_dict()[param_name]` to avoid tied weights
data_ptrs.add(model.state_dict()[param_name].data_ptr())
else:
hf_quantizer.create_quantized_param(
model, param, param_name, param_device, state_dict, unexpected_keys