feat: vmap optimizer (#166)

feat: scan layers + gradient checkpointing (#161)

Merge branch 'main' of https://github.com/borisdayma/dalle-mini into main

feat(text): support emojis (#154)

fix: smelu

fix: sinkformer

fix: support smelu

feat: allow relative position (#156)

feat: sinkhorn in lse mode (#155)

fix: sinkformer gradient

feat(model): allow bias (#152)

feat: add sinkformer + custom final ln + pre-ln (#151)

feat: placeholders for more config

feat: force final ln in encoder

feat: allow more configurations

fix: DeepNet doesn't scale weights of embedding/output layers (#150)

feat: remove unecessary LN

feat: add cogview

fix(textnormalizer): consider utf8 on windows (#148)

feat: implement transformer variants (#144)

feat(data): super conditioning (#141)

feat: support pod (#139)

feat: handle gradient checkpointing

feat: load from bucket

feat: reduce artifact space + offset step

feat: restore weights on CPU

fix: position embedding for generate method

fix: typo

fix: load from checkpoint

feat(modeling): simplify abstract_init

feat: handle model parallel

style(tokenizer): remove unused variables

feat: use fast tokenizer

feat(train): distributed_shampoo with pjit

feat(model): clean way to load on cpu

feat(train): no batch dimension with pjit

feat: use_artifact if run existing

Style (isort).

Tokenizer, config, model can be loaded from wandb.

Change import order again.

Fix import order to make isort happy.

Accept changes suggested by linter.

Never consider local dirs as remote wandb references.

Store resolved path after loading model.

Override from_pretrained to support wandb artifacts.

feat: support pypi