-
Notifications
You must be signed in to change notification settings - Fork 31.3k
Fix slicing for 0-dim param #36580
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.
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 slicing for 0-dim param #36580
Conversation
|
Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. When it is ready for review, please click the |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
src/transformers/modeling_utils.py
Outdated
| else: | ||
| param = param[:] | ||
| if is_safetensors_shard: | ||
| param = param[:] |
There was a problem hiding this comment.
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.There was a problem hiding this comment.
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)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
okay it's the same as here #36372 (comment)
Co-authored-by: fxmarty-amd <fxmarty-amd@users.noreply.github.com>
Co-authored-by: fxmarty-amd <felmarty@amd.com>
What does this PR do?
This PR fixes a slicing issue that we have with 0-dim tensors (both with param from bin file and safetensors pyslice). We can do [:] but [...] works instead. Thx fxmarty-amd for finding this solution.
cc @DN6