bigscience-workshop/Megatron-DeepSpeed

Pull Requests Commits

Add eval-only arg (#188)

SaulLu committed 4 years ago

Verified 05851eb5

removed regular package for megatron model (#192)

stas00 committed 4 years ago

Verified 4989bcde

Merge remote-tracking branch 'origin/main' into log-grad-norm

stas00 committed 4 years ago

f1c9ff89

stas00 committed 4 years ago

11ef669e

add layernorm in Embedding (#191)

stas00 committed 4 years ago

Verified b1359ee0

reworking the state dumper

stas00 committed 4 years ago

5edefea4

Support skip iteration flag (#177)

jaketae committed 4 years ago

Verified 106a9a6f

rework to integrate real iterations

stas00 committed 4 years ago

6b11ad70

Merge remote-tracking branch 'origin/main' into log-grad-norm

stas00 committed 4 years ago

ab0f78ed

experimental forward tb logger

stas00 committed 4 years ago

e4445dab

Full seqlen eval for CL+PP (#187)

conglongli committed 4 years ago

Verified 2105ff38

[BNB] integrate `StableEmbeding` into `VocabParallelEmbedding` logic (#182)

stas00 committed 4 years ago

Verified a34ca7f2

[PrefixLM] Figuring out why prefix lm is doing poorly on short context (#169)

thomasw21 committed 4 years ago

Verified 6d146b5f

[CI] improvements (#185)

stas00 committed 4 years ago

Verified 590f3e27

Alternative fix to TP > 1 (#178)

thomasw21 committed 4 years ago

Verified 2d9744f2

Fixed TP > 1 issue with new validation scheme

TevenLeScao committed 4 years ago

b982e040

simplifying tests

TevenLeScao committed 4 years ago

b1fc4927

Fixed merge oversight in tensorboard logs

TevenLeScao committed 4 years ago

18b704e1

Adding language specific validation sets for Multilingual model training (#97)

hadyelsahar committed 4 years ago

Verified 846c0879

Fix prefix lm offsets (#167)

thomasw21 committed 4 years ago

Verified 5e1f2101

chore: clean up print format

Jake Tae committed 4 years ago

934c0b30

docs: add comment on why `.clone()` was added

Jake Tae committed 4 years ago

1cce6594

Update main.yml (#172)

stas00 committed 4 years ago

Verified 5df29b5d

[CI] fix ci / update packages (#170)

stas00 committed 4 years ago

Verified b39dd1cf

make it work with any torch version

stas00 committed 4 years ago

f301ffd9

fix: clone input for backward hook

Jake Tae committed 4 years ago

daa2b630

fix: check if `grad_output` is `None`

Jake Tae committed 4 years ago

dd3bef26

fix: grad_outputs instead of inputs, iterate over tuple

Jake Tae committed 4 years ago

4f42e591

fix: use named_modules, add min max logging

Jake Tae committed 4 years ago

eabb8c5a

[checkpoint] only one latest file (#164)

stas00 committed 4 years ago

Verified 45dcd528

Newer Older