Failing to automatically load checkpoints#

When adding a custom checkpoint into model_dir, the run command does not automatically pick that up in my model_dir and runs with it. Instead, it reports that no checkpoint is found.

Explanation#

Cerebras ModelZoo PyTorch runs have a feature (enabled by default) to auto-load the last available checkpoint in the model_dir if a --checkpoint_path is not explicitly provided. It is important to note that only a specific checkpoint naming scheme is checked to find the latest checkpoint. All files in the format checkpoint_<step>.mdl are checked in the model_dir. If one or more are found, the file with the highest value of <step> is chosen and model weights are initialized with that checkpoint. This feature can be turned off by setting runconfig.autoload_last_checkpoint to False in the params yaml file.

Work around#

You can either

  • Provide a checkpoint inside model_dir with the naming format checkpoint_<step>.mdl, or

  • Specify checkpoint path by using the --checkpoint_path flag.