transformers/docs/source/en/model_doc/gemma2.md

Gemma2

Overview

The Gemma2 model was proposed in Gemma2: Open Models Based on Gemini Technology and Research by Gemma2 Team, Google. Two Gemma2 models are released, with parameters sizes of 9 billion (9B) and 27 billion (27B).

The abstract from the blog post is the following:

Now we’re officially releasing Gemma 2 to researchers and developers globally. Available in both 9 billion (9B) and 27 billion (27B) parameter sizes, Gemma 2 is higher-performing and more efficient at inference than the first generation, with significant safety advancements built in. In fact, at 27B, it offers competitive alternatives to models more than twice its size, delivering the kind of performance that was only possible with proprietary models as recently as December.

Tips:

• The original checkpoints can be converted using the conversion script src/transformers/models/Gemma2/convert_Gemma2_weights_to_hf.py

• Gemma2 uses sliding window attention every second layer, which makes it unsuitable for typical kv caching with [~DynamicCache] or tuples of tensors. To enable caching in Gemma2 forward call, you must initialize a [~HybridCache] instance and pass it as past_key_values to the forward call. Note, that you also have to prepare cache_position if the past_key_values already contains previous keys and values.

This model was contributed by Arthur Zucker, Pedro Cuenca and Tom Arsen.

Gemma2Config

autodoc Gemma2Config

Gemma2Model

autodoc Gemma2Model - forward

Gemma2ForCausalLM

autodoc Gemma2ForCausalLM - forward

Gemma2ForSequenceClassification

autodoc Gemma2ForSequenceClassification - forward

Gemma2ForTokenClassification

autodoc Gemma2ForTokenClassification - forward