Open Nicholas-Freitas opened 1 hour ago
Hi Nicholas,
thx for detailed issue report.
If if read your logs correctly, huggingface first wants to establish a connection to confirm your download, because failure happens during this check:
if not has_file(pretrained_model_name_or_path, safe_weights_name, **has_file_kwargs):
Hello all,
I'm running Chai-1 in a docker container on a node with no internet access (this is a security necessity for my university's compute cluster). I've pre-downloaded the model weights and ESM embedding weights into a directory within the docker container, which I reference using the CHAI_DOWNLOADS_DIR environment variable.
When I run the
predict_structure.py
example in the docker container, it runs to completion and doesn't appear to be downloading anything (I've checked the files in the download directory). However, if I launch the docker container with no internet connection using the--network=none
parameter, I get an error which I've pasted in full below.The error suggests it's unable to find the ESM model, despite it being downloaded in the default location,
$CHAI_DOWNLOADS_DIR/esm/models--facebook--esm2_t36_3B_UR50D
. Could this be becauseesm.py
is expecting the model to be saved under$CHAI_DOWNLOADS_DIR/esm/facebook/esm2_t36_3B_UR50D
, which is different from the actual download location?https://github.com/chaidiscovery/chai-lab/blob/306f53c7f45a6dd082aabd0d82a36d64fcf51f82/chai_lab/data/dataset/embeddings/esm.py#L57
Here's the full error message, thanks for the help!