Commit History

feat(demo): uncomment pip install
094e178

boris commited on

feat: improve inference demo
35fe578

boris commited on

fix: position embedding for generate method
ebac379

boris commited on

feat(train): use compilation cache
da9367c

boris commited on

fix: typo
68cc185

boris commited on

fix: load from checkpoint
44b7c3e

boris commited on

fix: style
d483294

boris commited on

Merge branch 'main' of https://github.com/borisdayma/dalle-mini into main
0a691de

boris commited on

feat: log num_parameters early
7cfe576

boris commited on

fix: distributed shampoo class
696422e

boris commited on

feat: update distributed_shampoo
5996680

boris commited on

feat(modeling): simplify abstract_init
fa72aa7

boris commited on

feat(train) - handle multiple nodes (#130)
0952927
unverified

boris commited on

feat: handle model parallel
1bb3269

boris commited on

feat(train): more custom x-axis
5f28cd2

boris commited on

feat(train): split artifact into model/state (#128)
7c4c287
unverified

boris commited on

fix: style
386f839

boris commited on

fix(train): opt_state_shape for distributed_shampoo
225b6ff

boris commited on

feat(train): split artifact into model/state
fa5b058

boris commited on

style(tokenizer): remove unused variables
605df32

boris commited on

feat: use fast tokenizer
767d78a

boris commited on

feat(train): another 25% faster
14abe8c

boris commited on

Merge pull request #127 from borisdayma/pjit-t5x
e4401dd
unverified

boris commited on

feat(train): overhead from 70% to 1% 🥳
2b7f5f1

boris commited on

feat(pjit): follow t5x style
7b5868f

boris commited on

fix(train): grads spec
00710bc

boris commited on

feat(train): improve pjit speed
f254058

boris commited on

fix(train): consider correct batch size
b7c7458

boris commited on

feat(train): custom start_preconditioning_step
8149924

boris commited on

feat(train): handle distributed_shampoo in pjit
032f623

boris commited on

feat: update distributed_shampoo + fix None spec
8a9e367

boris commited on

feat(train): distributed_shampoo with pjit
cc34d07

boris commited on

feat(train): use pjit (#125)
f5239e1
unverified

boris commited on

style: unsused import
7a176b9

boris commited on

fix style
f044cb8

boris commited on

feat(train): restore opt_state efficiently
1bfc1b5

boris commited on

feat(model): clean way to load on cpu
12f323d

boris commited on

feat(train): load model on CPU
3d43591

boris commited on

feat(train): different rng per node
2d212d8

boris commited on

feat(train): no batch dimension with pjit
df1fe19

boris commited on

feat(train): progress on pjit
49597a2

boris commited on

feat(train): start pjit support
0081723

boris commited on

feat: use_artifact if run existing
a5ed112

boris commited on

Load from wandb artifact (#121)
f69b21b
unverified

boris commited on

Style (isort).
f9d51f7

Pedro Cuenca commited on

feat(train): update sweep config
bbbf7c8

boris commited on

Use DalleBartTokenizer. State restoration reverted to previous method:
ae983d7

Pedro Cuenca commited on

Tokenizer, config, model can be loaded from wandb.
7e48337

Pedro Cuenca commited on

fix(train): variable not defined
4c87adf

boris commited on

feat(train): cleanup args
a2bf605

boris commited on