Unable to download `gs://big_vision/paligemma_tokenizer.model`

[abhaybd]

Does the big_vision bucket no longer exist? If I run gsutil ls gs://big_vision/paligemma_tokenizer.model I get:

ServiceException: 401 Anonymous caller does not have storage.objects.get access to the Google Cloud Storage object. Permission 'storage.objects.get' denied on resource (or it may not exist).

This is causing the code to crash when instantiating models, as the download of the paligemma tokenizer parameters fails.

[Abhiram824]

Also running into this issue

[F-Fer]

Same issue here. Was anyone able to solve this?

[F-Fer]

I looked into this. The GCS bucket appears to have had its anonymous access revoked. I had a copy of paligemma_tokenizer.model already cached, so I uploaded it here: https://github.com/F-Fer/paligemma_tokenizer

You can simply place it inside the openpi cache directory:

mkdir -p ~/.cache/openpi/big_vision/
cp /path/to/your/paligemma_tokenizer.model ~/.cache/openpi/big_vision/paligemma_tokenizer.model

The maybe_download function in src/openpi/shared/download.py checks this cache path before attempting the GCS download, so once the file is in place, no network access to Google is needed.

A proper fix would probably involve hosting the tokenizer somewhere else and updating the URL in src/openpi/models/tokenizer.py.

[vertix]

We hit this too while running inference validation across multiple servers.

As a workaround, we've hosted a public mirror of the tokenizer file:

mkdir -p ~/.cache/openpi/big_vision/
curl -o ~/.cache/openpi/big_vision/paligemma_tokenizer.model \
  https://storage.eu-north1.nebius.cloud/positronic-public/assets/paligemma_tokenizer.model

The file is 4.1 MB, no authentication required.

We maintain https://github.com/Positronic-Robotics/positronic – an open-source Python toolkit for ML-driven robotics that integrates with OpenPI (among other policy frameworks). If you're working with pi0 policies and want
streamlined dataset capture, training, and inference deployment, check it out.

[abhaybd]

gs://big_vision/paligemma_tokenizer.model seems to be back up! Closing.