First stable release since 0.1.1 (December 2024).
Highlights:
- Fault-tolerant DDP and HSDP support.
- LocalSGD, DiLoCo, streaming recovery, and peer checkpoint transports.
- Reconfigurable Gloo, NCCL, XCCL, and torchcomms process groups.
- FP8 collectives and broader collective coverage.
- Improved Lighthouse coordination, diagnostics, timeouts, and OpenTelemetry tracing.
- Linux wheels for CPython 3.9–3.13 plus a source distribution.
Requires Python 3.9+ and PyTorch 2.7+.
Full changes: 551fc4e...v0.2.0