Commit History

  1. feat: reduce artifact space + offset step

    34cf91c

  1. feat: restore weights on CPU

    5f954fc

  1. fix: position embedding for generate method

    ebac379

  1. fix: typo

    68cc185

  1. fix: load from checkpoint

    44b7c3e

  1. feat(modeling): simplify abstract_init

    fa72aa7

  1. feat(train) - handle multiple nodes (#130)

    Unverified signature
    0952927
  2. feat: handle model parallel

    1bb3269

  1. fix: style

    386f839

  1. style(tokenizer): remove unused variables

    605df32
  2. feat: use fast tokenizer

    767d78a

  1. feat(train): improve pjit speed

    f254058
  2. fix(train): consider correct batch size

    b7c7458

  1. feat(train): distributed_shampoo with pjit

    cc34d07

  1. style: unsused import

    7a176b9

  1. feat(model): clean way to load on cpu

    12f323d
  2. feat(train): no batch dimension with pjit

    df1fe19
  3. feat(train): progress on pjit

    49597a2

  1. feat: use_artifact if run existing

    a5ed112
  2. Load from wandb artifact (#121)⁠

    Unverified signature
    f69b21b
  3. Style (isort).

    Pedro Cuenca
    f9d51f7

  1. Tokenizer, config, model can be loaded from wandb.⁠

    Pedro Cuenca
    7e48337

  1. feat(data): support accumulation in non-streaming

    88c8e06
  2. feat: custom gradient accumulation

    2d07559

  1. Change import order again.⁠

    Pedro Cuenca
    2b2be9b
  2. Fix import order to make isort happy.

    Pedro Cuenca
    64d99b2
  3. Accept changes suggested by linter.

    Pedro Cuenca
    9f522b8
  4. Never consider local dirs as remote wandb references.⁠

    Pedro Cuenca
    08dd098
  5. Store resolved path after loading model.⁠

    Pedro Cuenca
    55a631d
  6. Override from_pretrained to support wandb artifacts.

    Pedro Cuenca
    1023afa
  7. fix(data): no shuffling of validation data

    ddcbc6a
  8. fix(train): handle seed_dataset

    8b72ed8

  1. feat: support pypi

    f5dba1e