torchtune icon indicating copy to clipboard operation
torchtune copied to clipboard

no error when lacking hugging face permissions

Open dangbert opened this issue 9 months ago • 3 comments

I briefly changed my hugging face token (via huggingface-cli login) to one that accidentally lacked permissions to download the llama2 model, and model downloads stopped working BUT a successful download was falsely reported:

$ tune download meta-llama/Llama-2-7b-hf  --output-dir /tmp/checkpoints2
Ignoring files matching the following patterns: *.safetensors
Successfully downloaded model repo and wrote to the following locations:

The above was the output (with no locations listed) and running tree /tmp/checkpoints2 showed the directory was still empty.

Expected behavior: Report an error and exit with a non-zero status code when the download fails due to hugging face token permissions

dangbert avatar May 01 '24 22:05 dangbert

It looks like Torchtune defers to huggingface_hub to download checkpoints, and handles errors from huggingface_hub.snapshot_download. Would it be worth trying the method directly to see what happens? I've included an example below for the command you shared, which assumes you've run huggingface-cli login. I haven't managed to reproduce your error unfortunately.

from huggingface_hub import snapshot_download
import os

snapshot_download(
    "meta-llama/Llama-2-7b-hf",
    local_dir="/tmp/checkpoints2",
    local_dir_use_symlinks="auto",
    ignore_patterns="*.safetensors",
    token=os.getenv("HF_TOKEN", None),
)

For me, trying to access meta-llama/Meta-Llama-3-8B, which I don't have access to, outputs:

Cannot access gated repo for url https://huggingface.co/api/models/meta-llama/Meta-Llama-3-8B/revision/main.
Access to model meta-llama/Meta-Llama-3-8B is restricted and you are not in the authorized list. Visit https://huggingface.co/meta-llama/Meta-Llama-3-8B to ask for access.

EDIT: One more thing you could try make sure you do is huggingface-cli logout just incase there's something weird going on there.

SalmanMohammadi avatar May 02 '24 14:05 SalmanMohammadi

@dangbert did you have any luck with this? I found doing a couple things helped when I was having issues downloading models:

  • checking if ~/.cache/huggingface/ has a cache for the model you're trying to download, and deleting if so.
  • passing -hf-token explicitly to the tune download command

SalmanMohammadi avatar May 17 '24 11:05 SalmanMohammadi

thanks @SalmanMohammadi unfortunately I haven't had time to test this again but will report back when I do

dangbert avatar May 20 '24 11:05 dangbert

Closing it since its stale. If its still an issue, please reopen it! thanks :)

felipemello1 avatar Jun 28 '24 16:06 felipemello1