Skip to content

-1e9 constants in T5 implementation #17215

Description

@marhlder

These -1e9 constants are too large for fp16 training

extended_attention_mask = (1.0 - extended_attention_mask) * -1e9

encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9

Maybe they should be made configurable?

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions