Llama 3.2 now available in Keras via keras_hub — on-the-fly HF checkpoint support

AI Impact Summary

Llama-3.2 is now usable in Keras via keras_hub, with on-the-fly conversion from Hugging Face checkpoints such as meta-llama/Llama-3.2-1B-Instruct. The integration supports safetensors and cross-backend execution (JAX, PyTorch, TensorFlow) through Keras' multi-backend architecture, enabling easier experimentation and scalable deployment on TPU/GPU with optional model parallelism. This lowers integration friction for teams looking to fine-tune or deploy Llama-3.2 directly from HF, without separate conversion steps, and showcases Keras as a viable path for scalable LLM workloads.