Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix bin state dict param #36580

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix bin state dict param #36580

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f8d1eec
fix
SunMarc Mar 6, 2025
9ea253c
Merge branch 'main' into fix-pointer
SunMarc Mar 7, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 7 additions & 4 deletions src/transformers/modeling_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -799,9 +799,10 @@ def _load_state_dict_into_meta_model(
for submodule in model.modules():
full_tp_plan.update(getattr(submodule, "_tp_plan", {}))

is_safetensors_shard = shard_file.endswith(".safetensors")
file_pointer = None
bin_state_dict = None
if shard_file.endswith(".safetensors"):
if is_safetensors_shard:
file_pointer = safe_open(shard_file, framework="pt", device=tensor_device)
else:
map_location = "cpu"
Expand Down Expand Up @@ -831,7 +832,7 @@ def _load_state_dict_into_meta_model(
# we need to use serialized_param_name as file pointer is untouched
param = (
file_pointer.get_slice(serialized_param_name)
if shard_file.endswith(".safetensors")
if is_safetensors_shard
else bin_state_dict[serialized_param_name]
)

Expand Down Expand Up @@ -900,13 +901,15 @@ def _load_state_dict_into_meta_model(
output_fn = partial(tp_layer._prepare_output_fn, tp_layer.output_layouts, tp_layer.use_local_output)
distribute_module(module_to_tp, device_mesh, None, input_fn, output_fn)
else:
param = param[:]
if is_safetensors_shard:
param = param[:]
if old_param is not None and old_param.is_contiguous():
param = param.contiguous()
module_to_tp.load_state_dict({param_type: param}, strict=False, assign=True)

else:
param = param[:]
if is_safetensors_shard:
param = param[:]
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also can't slice 0-dim tensors. See CI from diffusers with MusicLDM and AudioLDM2 which use ClapModel from transformers. https://github.com/huggingface/diffusers/actions/runs/13714387592/job/38359634948?pr=10997

model = ClapModel(
  (text_model): ClapTextModel(
    (embeddings): ClapTextEmbeddings(
      (word_embeddings): Embedding(100...eatures=16, bias=True)
    (activation): ReLU()
    (linear2): Linear(in_features=16, out_features=16, bias=True)
  )
)
state_dict = {'audio_model.audio_encoder.batch_norm.bias': tensor(..., device='meta', size=(8,)), 'audio_model.audio_encoder.batch_...ice='meta', size=(8,)), 'audio_model.audio_encoder.batch_norm.running_var': tensor(..., device='meta', size=(8,)), ...}
...
>               param = param[:]
E               IndexError: slice() cannot be applied to a 0-dim tensor.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

did you launch the tests if this specific pr ?
For safetensors file, param is a PySlice object. This is why we need to do param = param[:]
param = file_pointer.get_slice(serialized_param_name)

if param_casting_dtype is not None:
param = param.to(param_casting_dtype)
if old_param is not None and old_param.is_contiguous():
Expand Down