Sylvain Gugger

Use commit hash to look in cache instead of calling head

1

# What does this PR do? This PR tries to limit the calls to requests.head made for cached models every time we try to load them. Currently on the main...

Preserve hub-related kwargs in AutoModel.from_pretrained

1

# What does this PR do? As was reported in #18537, when using `AutoConfig` inside the `AutoModel.from_pretrained` method, some kwargs are deleted and not passed to the `from_pretrained` method of...

Cache non-existence of files or completeness of repo

14

In Transformers, we very often try to download files that do not exist: this is because each tokenizer has a list of optional files so we try to download them...

Add ability to turn on and off progress bars

3

Not everyone wants to see progress bars when downloading models/datasets as tqdm can clog the logs pretty easily. That's why we have a mode in Transformers to deactivate tqdm when...

Add a utility to list cached things

7

In Transformers, we have [this function](https://github.com/huggingface/transformers/blob/ab2006e3d6db88654526a4169e65d4bfc52da2e3/src/transformers/utils/hub.py#L119) that returns the list of cached models. It would be nice to rely on an util in this repository instead (not limited to models)...

When there is an error message with a git command (for isntance git push), the url is shown and that url contains the user token. Users that dutifully copy the...

Fix big model inference for T5 models in float16

1

# What does this PR do? This PR fixes big model inference for large T5 models. The problem is that T5 models have some weights kept in float32, which interferes...

Rework a bit the LLaMA conversion script

1

# What does this PR do? This PR makes sure the LLaMA conversion script stays up to date with `save_pretrained` by having the checkpoint being loaded in an actual model...

Fix various imports

1

# What does this PR do? While building the v2 of the test fetcher, I discovered (I mean the util discovered) that some imports in the source code are wrong....

Adapt find_tied_parameters to handle breaking change in Accelerate

1

# What does this PR do? In the upcoming version of Accelerate, `find_tied_parameters` returns a list of list instead of dictionary. While there is a hack in place to make...