1.x pre PyPi release of Library

[Appa002]

Hi, before the release of Gemma 3 one would install the Flax inference code through pointing pip at a link directly which was given in the docs. Presumably, that was an 1.x version of the library. It appears now that this documentation page has been scrubbed from the internet, and I am not able to install this older (pre-2025) version of the library.

Unfortunately, some research code depends on that version of the library, and this newer version here breaks our existing setup.

Would it be possible for someone to point me to where I can find this older version of the library?
Naturally, a major version bump incurs breaking changes, but it be great if older version could still be accessed.

Cheers!

[Appa002]

Particularly, we are loading an existing gemma checkpoint like this:

def load_gemma(gemma_path, gemma_variant):
    gemma_chkpt_path = os.path.join(gemma_path, gemma_variant)
    gemma_config_params = params_lib.load_and_format_params(gemma_chkpt_path)

    gemma_tokenizer_path = os.path.join(gemma_path, "tokenizer.model")
    vocab = spm.SentencePieceProcessor()
    vocab.Load(gemma_tokenizer_path)

    transformer_config = TransformerConfig.from_params(params=gemma_config_params, cache_size=1024)

    gemma_params = {"params": gemma_config_params["transformer"]}
    gemma_transformer = Transformer(transformer_config)
    return gemma_transformer, gemma_params, transformer_config, vocab

Where Transformer is a modified version of the Class with the same name in the original library. This was the idiomatic way of loading the gemma model in the older version, now with a Kaggle checkpoint this fails with:
FileNotFoundError: No structure could be identified for the checkpoint at /Users/albert/.cache/kagglehub/models/google/gemma-2/flax/gemma2-2b-it.

I'm not too interested in debugging this per-se if we could get access to the older version instead :)

[Conchylicultor]

Yes, the legacy API was never truly released on PyPI and is being removed.

In the long term, I would encourage you to migrate to the new API which supports Gemma3, in addition of being simpler while more features complete.

If you want to install an old version of the library, this is still possible by choosing an old commit. For example, the following should work for you:

pip install git+https://github.com/google-deepmind/gemma.git@cc13ebe966cef38c3c0357e51225815e6f0cd67d

[Appa002]

Thank you so much, this is great!