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

Announcing end of support for XLA_USE_BF16 and XLA_DOWNCAST_BF16 starting next release

Starting with Neuron Release 2.23, Neuron will begin phasing out support for the XLA_DOWNCAST_BF16 and XLA_USE_BF16 environment variables. In this release, usage of these variables will trigger warnings. Neuron will end support in a subsequent release, aligned with the torch-xla maintenance schedule.

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