tree 983aac3f6efe191a01c154e4302498d50850b286
parent 630be619044fa38164cb3e4527503ca43003a66b
author Reed Wanderman-Milne <reedwm@google.com> 1660938610 -0700
committer TensorFlower Gardener <gardener@tensorflow.org> 1660938914 -0700

Add XlaReducePrecision TensorFlow op.

See https://www.tensorflow.org/xla/operation_semantics#reduceprecision.

This is somewhat useful for testing fp8. Unfortunately the usefulness is currently quite limited due to the fact certain fp8 representations, such as the ones described in this paper https://arxiv.org/abs/2206.02915, have different representations of Inf/NaN compared to the standard IEEE floating-point dtypes.

PiperOrigin-RevId: 468765094