This document is relevant for: Inf1, Inf2, Trn1, Trn2

Announcing end of support XLA_USE_BF16 and XLA_DOWNCAST_BF16 environment variables starting next release

Neuron Release 2.24 will be the last release to support the following environment variables:

• XLA_USE_BF16

• XLA_DOWNCAST_BF16

I currently utilize these environment variables in my model code. What do I do?

Customers are recommended to migrate to automatic mixed-precision or use model.to(torch.bfloat16) command to convert their model to BF16 format. For detailed migration guidance, please refer to Migration From XLA_USE_BF16/XLA_DOWNCAST_BF16.

This document is relevant for: Inf1, Inf2, Trn1, Trn2