See our collection for all versions of Gemma.

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

kerasformers/gemma-7b

Paper: Gemma: Open Models Based on Gemini Research and Technology (arXiv:2403.08295) · HF Papers

Gemma is Google's open decoder-only LLM family (RMSNorm, GeGLU, RoPE, multi-query attention). Base checkpoints are for completion; -it / 1.1 variants are instruction-tuned for chat.

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

Pure-Keras 3 conversion of google/gemma-7b 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.gemma import GemmaGenerate, GemmaTokenizer

model = GemmaGenerate.from_weights("kerasformers/gemma-7b")
tokenizer = GemmaTokenizer.from_weights("kerasformers/gemma-7b")

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

Load any Gemma v1 variant the same way with from_weights("kerasformers/<variant>"):

Variant	Hub	Type
`gemma-2b`	`kerasformers/gemma-2b`	base
`gemma-2b-it`	`kerasformers/gemma-2b-it`	instruct
`gemma-1.1-2b-it`	`kerasformers/gemma-1.1-2b-it`	instruct (1.1)
`gemma-7b`	`kerasformers/gemma-7b`	base
`gemma-7b-it`	`kerasformers/gemma-7b-it`	instruct
`gemma-1.1-7b-it`	`kerasformers/gemma-1.1-7b-it`	instruct (1.1)

Tips

Set KERAS_BACKEND before importing Keras / kerasformers.
Prefer GemmaTokenizer.from_weights(...) so the chat template matches.
Larger checkpoints: try load_dtype="bfloat16" or quantization="int8".
See Gemma docs and Loading Weights.
Community / upstream safetensors still work via the hf: prefix, e.g. GemmaGenerate.from_weights("hf:google/gemma-7b").