kerasformers/gemma-2-2b

See our collection for all versions of Gemma 2.

Run Gemma 2 with Keras 3: JAX, PyTorch, or TensorFlow

kerasformers/gemma-2-2b

Paper: Gemma 2: Improving Open Language Models at a Practical Size (arXiv:2408.00118) · HF Papers

Gemma 2 builds on Gemma with alternating sliding-window attention, attention/logit softcapping, and GQA. Sizes span 2B / 9B / 27B with base and -it chat variants.

For more details on the model, please go to Google's original model card.

Pure-Keras 3 conversion of google/gemma-2-2b for kerasformers. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is a base (pretrained) checkpoint: prompt it with raw text completion, or prefer an -it variant for chat.

✨ Quick start

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from kerasformers.models.gemma2 import Gemma2Generate, Gemma2Tokenizer

model = Gemma2Generate.from_weights("kerasformers/gemma-2-2b")
tokenizer = Gemma2Tokenizer.from_weights("kerasformers/gemma-2-2b")

inputs = tokenizer("The capital of France is")
outputs = model.generate(**inputs, max_new_tokens=32)
print(tokenizer.decode(outputs[0]))

Load any Gemma 2 variant the same way with from_weights("kerasformers/<variant>"):

Variant	Hub	Type
gemma-2-2b	kerasformers/gemma-2-2b	base
gemma-2-2b-it	kerasformers/gemma-2-2b-it	instruct
gemma-2-9b	kerasformers/gemma-2-9b	base
gemma-2-9b-it	kerasformers/gemma-2-9b-it	instruct
gemma-2-27b	kerasformers/gemma-2-27b	base
gemma-2-27b-it	kerasformers/gemma-2-27b-it	instruct

Tips

• Set KERAS_BACKEND before importing Keras / kerasformers.
• Prefer Gemma2Tokenizer.from_weights(...) so the chat template matches.
• Larger checkpoints: try load_dtype="bfloat16" or quantization="int8".
• See Gemma 2 docs and Loading Weights.
• Community / upstream safetensors still work via the hf: prefix, e.g. Gemma2Generate.from_weights("hf:google/gemma-2-2b").

Special Thanks

A huge thank you to the Google Gemma authors for creating and releasing these models.

License: Gemma (gated). Accept the license on the upstream Hub card before downloading.