flax
ed42d067 - Align layernorm dtype handling with batchnorm (i.e., use requested dtype for layernorm outputs, even though intermediate computations are f32).

Commit

5 years ago

Align layernorm dtype handling with batchnorm (i.e., use requested dtype for layernorm outputs, even though intermediate computations are f32). PiperOrigin-RevId: 317602129

References

#331 - Align layernorm dtype handling with batchnorm (i.e., use requested dtype for layernorm outputs, even though intermediate computations are f32).

Author

jekbradbury

Committer

a-googler

Parents

1c509ed4

flax ed42d067 - Align layernorm dtype handling with batchnorm (i.e., use requested dtype for layernorm outputs, even though intermediate computations are f32).

flax
ed42d067 - Align layernorm dtype handling with batchnorm (i.e., use requested dtype for layernorm outputs, even though intermediate computations are f32).