Publishing Weights

Publishing weights

If you've trained a model that you'd like to share with the community, you can publish any number of checkpoints you've previously saved.

Once published, your checkpoint can be loaded by any tinker user and used to further train a new model or be sampled against.

Publishing

tinker checkpoint publish $TINKER_CHECKPOINT_PATH

where $TINKER_CHECKPOINT_PATH is a checkpoint path in the form of tinker://14bdf3a1-0b95-55c7-8659-5edb1bc870af:train:17/weights/checkpoint_id_to_publish.

You may confirm your checkpoint is published by dumping the checkpoint info and checking the Public property:

tinker checkpoint info tinker://14bdf3a1-0b95-55c7-8659-5edb1bc870af/weights/checkpoint_id_to_publish
                              Checkpoint: weights/checkpoint_id_to_publish
┏━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
 Property         Value                                                                          
┡━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
 Checkpoint ID    weights/checkpoint_id_to_publish                                               
 Type             training                                                                       
 Tinker Path      tinker://14bdf3a1-0b95-55c7-8659-5edb1bc870af/weights/checkpoint_id_to_publish 
 Size             342.4 MB                                                                       
 Public           No                                                                             
 Created          23 minutes ago                                                                 
 Training Run ID  14bdf3a1-0b95-55c7-8659-5edb1bc870af                                           
└─────────────────┴────────────────────────────────────────────────────────────────────────────────┘

Unpublishing

tinker checkpoint unpublish $TINKER_CHECKPOINT_PATH`

Loading public weights

Loading public weights is exactly the same as loading a non-public one:

ckpt_path = ...
training_client = service_client.create_training_client_from_state(ckpt_path)