Introduce prepare for eval, fix evaluation bug

[Niccolo-Ajroldi]

Description

This pull request introduces a prepare_for_eval function and updates the code to support it.

The implementation follows the blueprint of @fsschneider in #719 (comment) and fixes the bug of giving a free evaluation to a submission that goes out of max_runtime (again #719 (comment)).

Function signature

The arguments of prepare_for_eval are the same as update_params, except for batch. I believe that prepare_for_eval should indeed be agnostic to the last batch used during training. The return type is the same as update_params.

def prepare_for_eval(workload: spec.Workload,
                     current_param_container: spec.ParameterContainer,
                     current_params_types: spec.ParameterTypeTree,
                     model_state: spec.ModelAuxiliaryState,
                     hyperparameters: spec.Hyperparameters,
                     loss_type: spec.LossType,
                     optimizer_state: spec.OptimizerState,
                     eval_results: List[Tuple[int, float]],
                     global_step: int,
                     rng: spec.RandomState) -> spec.UpdateReturn:
  return (optimizer_state, current_param_container, model_state)

List of changes

In submission_runner.py:

• add timed call to prepare_for_eval
• add profiler
• move del batch before prepare_for_eval (instead than before evaluation)
• update accumulated_submission_time after prepare_for_eval
• compute is_time_remaining after prepare_for_eval
• proceed to eval iff is_time_remaining
• add prep_eval_rng

Minor changes:

• add PrepareForEvalFn to spec
• add prepare_for_eval to submission template
• add prepare_for_eval to all pytorch and jax submissions
• update the docs

Fixes #719 and #758 .

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅