gemma

The Gemma architecture transformer variant.

Supports both the Gemma 1 and Gemma 2 architectures. Based on the Flax reference implementation at google-deepmind/gemma.

See the Gemma technical reports for more information:

• Gemma 1: https://arxiv.org/abs/2403.08295

• Gemma 2: https://arxiv.org/abs/2408.00118

Functions

gemma_from_pretrained_checkpoint(ckpt_params)

Builds a Gemma model from a pretrained checkpoint.