Accelerate v1.8.0: FSDPv2/FP8, DeepSpeed regional compile, ipex.optimize deprecation
Action Required
Developers using v1.8.0 should adapt to the new FSDPv2 prep flow and DeepSpeed regional compile, plan for ipex.optimize deprecation by upgrading to PyTorch 2.8+ or SwanLab-based tracking, and can expect improved Intel CPU distributed training performance with updated CCL/KMP settings.
AI Impact Summary
This release consolidates FSDPv2 integration with FP8 and enforces a streamlined preparation path, reducing cross-feature configuration errors but requiring users to adopt the new prep workflow. It introduces regional compilation for DeepSpeed (in-place modification via Module.compile) and performance enhancements for Intel CPU distributed training through updated CCL_WORKER_COUNT and KMP parameters, plus expanded XPU support and SwanLab tracking integration. The ipex.optimize deprecation indicates future optimizations will land in PyTorch, so teams should plan migrations to PyTorch 2.8+ or SwanLab-based tracking to avoid gaps; tracker initialization is also deferred to prevent premature distributed setup.
- Date
- Date not specified
- Change type
- deprecation
- Severity
- high