Remove dependency on HuggingFace token

[tengyifei]

Right now we need a HuggingFace token in order to run tests/run the model locally. This token is used to fetch the model config and fetch tokenizers.

It should be possible to drop in a tokenizer model as a file. It should also be possible to specify a model config as a dict as opposed to a str like Meta-Llama/Llama-3.1-8B, which would cause transformers to look up the model using a token. Once we can do these, we'll no longer need a token to do dev work and testing.

[yaoshiang]

The way to do this is probably to require all TP scripts use a cloud cache location, and always use the cache_dir param.

A separate script is required to load that cache. The separate script obviously needs a HF token.

We will have to see if cache_dir accepts gs://... paths or if we need to mount the gcs bucket to filesystem with gcsfuse.

[vlasenkoalexey]

+1 if we store tokenizer and dataset on GCS, we should be able to get along without HF token