customize_paths.md

Customize paths

The project is setup to use specific paths to read the original weights and save checkpoints etc.

For all scripts, you can run

python script.py -h

to get a list of available options. For instance, here's how you would modify the checkpoint dir:

python scripts/convert_checkpoint.py --checkpoint_dir "data/checkpoints/foo"

Note that this change will need to be passed along to subsequent steps, for example:

python generate.py \
  --checkpoint_path "data/checkpoints/foo/7B/lit-llama.pth" \
  --tokenizer_path "data/checkpoints/foo/tokenizer.model"

and

python quantize/gptq.py \
  --checkpoint_path "data/checkpoints/foo/7B/lit-llama.pth" \
  --tokenizer_path "data/checkpoints/foo/tokenizer.model"

To avoid this, you can use symbolic links to create shortcuts and avoid passing different paths.