Release Notes

Contents

Release Notes#

This section includes the release information for PyTorch-Plugin-FL.

v0.1.0#

Note

This is a preview release. The version number shown is a pre-release identifier and may change upon final release. Content in this preview is for reference only and does not constitute a commitment or warranty for the final product.

  • Added features:

    • Initial release of PyTorch-Plugin-FL as part of FlagOS.

    • PrivateUse1-based custom device plugin registering flagos as a first-class PyTorch device.

    • Automatic FlagGems Triton operator registration for the flagos backend.

    • Per-operator configurable backend routing via backends.conf with environment variable overrides.

    • Multi-platform support: NVIDIA CUDA, MetaX MACA, Huawei Ascend.

    • Complete device management API: stream, event, RNG, AMP, memory allocator, DeviceGuard.

    • Lightweight C++ dispatch stub replacing PyTorch’s heavier DispatchStub.

    • C++ stub-only mode (FLAGOS_DISABLE_FLAGGEMS_PY=1) for minimal overhead.

    • Distributed training support via torch_fl.distributed (DDP/FSDP patch).

    • MACA cu-bridge ABI shim for symbol version compatibility.

    • Ascend NPU support with ACL NN API kernels.

    • Integration test suite with factory ops, dispatch routing, CPU fallback tracing, and Qwen3 inference/training tests.

Contents