flax
option of forcing the input of softmax to be fp32 for better numerical stability in mixed-precision training.
#3874

Merged

option of forcing the input of softmax to be fp32 for better numerical stability in mixed-precision training. #3874

copybara-service merged 1 commit into main from test_626731402

copybara-service force pushed 2 years ago

option of forcing the input of softmax to be fp32 for better numerica…

e6379c1b

copybara-service force pushed to e6379c1b 2 years ago

copybara-service merged e6379c1b into main 2 years ago

copybara-service deleted the test_626731402 branch 2 years ago

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Milestone

No milestone