PR #17306 Fix -1e4 as attn mask

Fix -1e4 as attn mask #17306

ydshieh merged 19 commits into huggingface:main from ydshieh:no_-1e4_for_attn_mask

ydshieh commented on 2022-05-17

ydshieh force pushed 4 years ago

ydshieh commented on 2022-05-17

ydshieh changed the title ~~[WIP] Fix -1e4 as attn mask~~ Fix -1e4 as attn mask 4 years ago

ydshieh requested a review from

sgugger 4 years ago

ydshieh requested a review from

patil-suraj 4 years ago

ydshieh requested a review from

LysandreJik 4 years ago

ydshieh requested a review from

patrickvonplaten 4 years ago

ydshieh marked this pull request as ready for review 4 years ago

sgugger commented on 2022-05-17

patrickvonplaten commented on 2022-05-17

patil-suraj commented on 2022-05-18

ydshieh force pushed 4 years ago

sgugger approved these changes on 2022-05-25

patrickvonplaten commented on 2022-05-25

patrickvonplaten approved these changes on 2022-05-25

ydshieh force pushed 4 years ago

stephenroller commented on 2022-05-31

ydshieh force pushed 4 years ago

LysandreJik approved these changes on 2022-06-03

ydshieh commented on 2022-06-03

ydshieh force pushed 4 years ago

ydshieh marked this pull request as draft 4 years ago

ydshieh marked this pull request as ready for review 3 years ago

Use torch.finfo(self.dtype).min

0322ea5e

for GPTNeoX

eda30548

for Albert

78cb386d

For Splinter

5aaa4f93

Update src/transformers/models/data2vec/modeling_data2vec_audio.py

47cc6af5

better style

1cc0b232

fix -inf used in Bart-like models

3861fe04

Fix a few remaining -inf

217a1df6

more fix

2e5c050b

clean up

51a577d1

For CLIP

3201a0f1

For FSMT

3da95f95

clean up

b018a517

fix test

4faab4ae

Add dtype argument and use it for LayoutLMv3

a2b419c1

fix

bcf7536a

fix

73b5e654

fix conflict

5dc2a3f7

ydshieh force pushed to 5dc2a3f7 3 years ago

update FlaxLongT5Attention

25eac912

ydshieh merged d3cb2888 into main 3 years ago

Reviewers

LysandreJik

sgugger

patrickvonplaten

stephenroller

patil-suraj

Assignees

No one assigned

Labels

None yet

Milestone

No milestone

transformers Fix -1e4 as attn mask #17306 Merged

Fix -1e4 as attn mask #17306

transformers
Fix -1e4 as attn mask
#17306

Merged