M4 Visualization - NOT UP TO DATE (August 16th). ASK VICTOR

For visualizations, we have a main app which calls multiple child apps to retrieve generations via Gradio API. This allows us to parallelize calls to multiple models at the same time instead of running them sequentially.

How to?

The process of adding a model to the main space:

Use huggingface-cli login to login with an auth token that has a read/write access to the HuggingFaceM4 org on the hub.
Use ./upload_checkpoint_to_hub_gcs.sh script to upload a checkpoint from GCP store to the hub. An example command to upload checkpoint for step 3000 from tr_121ter to the hub: ./m4/visualization/upload_checkpoint_to_hub_gcs.sh gs://hf-science-m4-cold/local_experiment_dir/tr_121ter/opt_step-3000. This will create model repo under the HuggingFaceM4 repo on the hub. If you are on the cluster, use ./upload_checkpoint_to_hub_s3.sh instead. I recommend being on a compute node to avoid disk space issues (uploading to the hub consists in downloading locally the checkpoint, creating a repo on the hub, copying it locally, filling it with the weights and commiting the weights to the hub repo).
[MANUAL] Go to the hub, create a repo of type space with the same name as the model. In the space's settings, add a secret HF_AUTH_TOKEN with a token which has read access to the HuggingFaceM4 repo. This step can be potentially automated in the future.
[MANUAL] Edit m4/visualization/app.py's three dictionary to include your model in the existing formats of those dictionaries.
Run m4/visualization/sync-repo.sh <name_of_the_space_on_the_hub> to sync the repo with the local setting. This will automatically update the space to have the latest code as in the m4/visualization/app.py.
Run m4/visualization/sync-repo.sh main to update the main repo as well with the new model.