v1.13.0: AWS Neuron support, IPEX removal, and DeepSpeed/FSDP2 distributed training fixes
Action Required
This release enables AWS Neuron hardware acceleration, reduces dependency surface by removing IPEX, and tightens distributed training with FSDP2/DeepSpeed fixes and FP8 support, improving stability and scalability for large models.
AI Impact Summary
v1.13.0 adds AWS Neuron support for Trainium/Inferentia devices, enabling hardware-accelerated training and inference paths. It removes the IPEX dependency and makes XPU code device-agnostic, reducing external dependencies and simplifying cross-device usage. The release also ships distributed-training improvements across FSDP2 (upcasting, embedding error handling, DCP loading, bf16 stability, and torch < 2.7.0 compatibility) and DeepSpeed Sequence Parallelism, plus FP8 enhancements and Transformer Engine fixes to improve stability and performance for large models.
Affected Systems
- Date
- Date not specified
- Change type
- deprecation
- Severity
- high