Add curriculum learning callback

[b-chu]

Curriculum learning callback

Requirements

• Requires StreamingDataset
• Each datamix must have a duration and train_loader
• Duration must be positive and defined in terms of epochs or tokens
• Duration units must match max_duration
• The length of the schedule must be equal to max_duration
• The part of the schedule that has already been trained on cannot be changed in future resumption runs

Features

• Defines schedules by specifying datamixes in terms of training duration
• Supports epochs and tokens for each iteration
• Enables single run curriculum learning
• Swaps the entire dataloader at the start of each iteration

Other

• Refactor build_tokenizer to avoid circular dependencies
• Refactor process_init_dist to avoid circular dependencies

Manual tests

Matches old callback behavior

Resumes correctly in the middle of the schedule

Resumes correctly when new datamix added to schedule

Resumes correctly when callback added after initial training run

API

Old API:

train_loader:
  <some params>
callbacks:
  curriculum_learning:
    dataset_index: 0

Start a new run

train_loader:
  <some params>
callbacks:
  curriculum_learning:
    dataset_index: 1

Start a new run

train_loader:
  <some params>
callbacks:
  curriculum_learning:
    dataset_index: 2

New API:

train_loader:
  <dataloader parameters>
callback:
  curriculum_learning:
  - duration: <number>tok
    train_loader:  # matches top level train_loader
      <dataloader parameters>
  - duration: <number>tok
    train_loader:
      <dataloader parameters>
  - duration: <number>tok
    train_loader:
      <dataloader parameters>

[snarayan21]

@b-chu about the new API, couple questions:

train_loader:
  <some params>
callbacks:
  curriculum_learning:
    duration: 5000000tok
    schedule:
    - duration: 5000000tok
      train_loader:
        <some params>
    - duration: 5000000tok
      train_loader:
        <some params>

1. so I still have to specify train_loader as a top-level entry?
2. the first duration specified is for the top-level train_loader?

[snarayan21]

Also, I'm worried about the loss curves in the plots you shared, they don't look fully deterministic to me. What model size and batch size were you running at, and with which datasets? Longer training runs with a bigger model and small batch size, without shuffling, would be helpful so that we can determine if the loss curves are actually deterministic or not. Just looking at the first few steps most training runs will look pretty similar regardless of the data ordering.

[mvpatel2000]

This needs a composer release first, right?

[mvpatel2000]

First pass, mostly lgtm but some minor comments.

[mvpatel2000]

second pass, lgtm besides a few minor comments. requiring review from @milocress

[milocress]

LGTM, just a couple questions

[dakinggg]

will approve after comments are addressed, overall lgtm