nanopar

Like tinypar but with a different license

Usage

./deps.sh
./download.sh
python create_data.py
# LLaMA inference test
MODEL_TYPE=llama torchrun cache_logprobs.py --model_dir model_dir/llama-2-13b ---params_file params.json --test_inference=1
# NeoX inference test
MODEL_TYPE=neox torchrun cache_logprobs.py --model_dir model_dir/neox --params_file config.json --test_inference=1

# training is WIP
MODEL_TYPE=llama torchrun train_dpo.py --model_dir model_dir/llama-2-13b ---params_file params.json