Stanford Alpaca is a model fine-tuned from the LLaMA-7B.
The inference code is using Alpaca Native model, which was fine-tuned using the original tatsu-lab/stanford_alpaca repository. The fine-tuning process does not use LoRA, unlike tloen/alpaca-lora.
For the Alpaca-7B:
-
Linux, MacOS
-
1x GPU 24GB in fp16 or 1x GPU 12GB in int8
-
PyTorch with CUDA (not the CPU version)
-
HuggingFace Transformers library
pip install git+https://github.com/huggingface/transformers.git
Currently, the Transformers library only has support for LLaMA through the latest GitHub repository, and not through Python package.
-
If run in 8-bit (quantized model), install Bitsandbytes and set
load_in_8bit=true
How to use.
Download model weights from https://huggingface.co/chavinlo/alpaca-native
Change
./checkpoint-1200/
to the directory of your HuggingFace format model files directory.