|
## 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 |
|
|
|
```shell |
|
python script.py -h |
|
``` |
|
|
|
to get a list of available options. For instance, here's how you would modify the checkpoint dir: |
|
|
|
```shell |
|
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: |
|
|
|
```shell |
|
python generate.py \ |
|
--checkpoint_path "data/checkpoints/foo/7B/lit-llama.pth" \ |
|
--tokenizer_path "data/checkpoints/foo/tokenizer.model" |
|
``` |
|
|
|
and |
|
|
|
```shell |
|
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. |
|
|