CHANGELOG.md

April 2, 2026 · View on GitHub

Release History

  • [2026/03] Released v0.11:

    • Enables kernel-based communication on heterogeneous platforms, including NVIDIA and Hygon.
    • Adds support for both host-side and device-side one-sided communication semantics.
    • Introduces adaptor plugin support, enabling dynamic loading of user-defined Device, CCL, and Net adaptor implementations.

  • [2026/02] Released v0.10:

    • Implements 11 chip-decoupled collective communication algorithms in uniRunner mode.
    • Refactors Device Intra-/Inter-node API and integrates NCCL Device API support on NVIDIA platforms.
    • Enhances usability with pip install support for FlagCX and an NCCL wrapper plugin for seamless adoption on NVIDIA platforms.

  • [2026/01] Released v0.9:

    • Adds support for Enflame, including topsAdaptor and ecclAdaptor.
    • Extends flagcxCCLAdaptor to support symmetric operations.
    • Introduces the NCCL Device API in ncclAdaptor to enable customized AllReduce operations.
    • Refactors glooAdaptor to support both TCP and IB transports, with automatic NIC detection.

  • [2025/12] Released v0.8:

    • Enables intra-node zero-copy to improve data transfer efficiency for small messages.
    • Supports a naive AllReduce implementation in uniRunner mode using a CPU-centric, device-assisted algorithm.
    • Adds one-sided communication primitives via the new APIs flagcxHeteroPut and flagcxHeteroPutSignal.

  • [Unreleased] Test infrastructure restructure and bug fixes (PR #413):

    • Fixed NCCL group imbalance in ncclAdaptorGather/ncclAdaptorScatter: errors inside ncclGroupStart()/ncclGroupEnd() no longer skip ncclGroupEnd(), preventing deadlocks.
    • Reduced unit-test buffer allocation from 1GB to 4MB per buffer, cutting memory from 32GB to 128MB for 8-rank runs.
    • Improved collective test correctness by using rank-dependent data patterns, catching rank-ordering and single-rank-copy bugs.
    • Added infinite-loop guard in perfBenchmarkLoop for stepFactor <= 1.
    • Wired PERF_COMMON_SRC into test/perf/host_api/Makefile build.
    • Removed TRACE-level debug logging from CI workflow.

  • [2025/11] Released v0.7:

    • Added support to TsingMicro, including device adaptor tsmicroAdaptor and CCL adaptor tcclAdaptor.
    • Implemented an experimental kernel-free non-reduce collective communication (SendRecv, AlltoAll, AlltoAllv, Broadcast, Gather, Scatter, AllGather) using device-buffer IPC/RDMA.
    • Enabled auto-tuning on NVIDIA, MetaX, and Hygon platforms, achieving 1.02×–1.26× speedups for AllReduce, AllGather, ReduceScatter, and AlltoAll.
    • Enhanced flagcxNetAdaptor with one-sided primitives (put, putSignal, waitValue) and added retransmission support for reliability improvement.

  • [2025/10] Released v0.6:

    • Implemented device-buffer IPC communication to support intra-node SendRecv operations.
    • Introduced device-initiated, host-launched device-side primitives, enabling kernel-based communication directly from devices.
    • Enhanced auto-tuning with 50% performance improvement on MetaX platforms for the AllReduce operations.

  • [2025/09] Released v0.5:

    • Added support for AMD GPUs, including a device adaptor hipAdaptor and a CCL adaptor rcclAdaptor.
    • Introduced flagcxNetAdaptor to unify network backends, currently supporting socket, IBRC, UCX and IBUC (experimental).
    • Enabled zero-copy device-buffer RDMA (user-buffer RDMA) to boost performance for small messages.
    • Supported auto-tuning in homogeneous scenarios via flagcxTuner.
    • Added test automation in CI/CD for PyTorch APIs.

  • [2025/08] Released v0.4:

    • Supported heterogeneous training of ERNIE4.5 (Baidu) on NVIDIA and Iluvatar GPUs with Paddle + FlagCX.
    • Improved heterogeneous communication across arbitrary NIC configurations, with more robust and flexible deployments.
    • Introduced an experimental network plugin interface with extended supports for IBRC and SOCKET. Device buffer registration now can be done via DMA-BUF.
    • Added an InterOp-level DSL to enable customized C2C algorithm design.
    • Provided user documentation under docs/.

  • [2025/07] Released v0.3:

    • Integrated three additional native communication libraries: HCCL (Huawei), MUSACCL (Moore Threads) and MPI.
    • Enhanced heterogeneous collective communication operations with pipeline optimizations.
    • Introduced device-side functions to enable device-buffer RDMA, complementing the existing host-side functions.
    • Delivered a full-stack open-source solution, FlagScale + FlagCX, for efficient heterogeneous prefilling-decoding disaggregation.

  • [2025/05] Released v0.2:

    • Integrated 3 additional native communications libraries, including MCCL (Moore Threads), XCCL (Mellanox) and DUCCL (BAAI).
    • Improved 11 heterogeneous collective communication operations with automatic topology detection and full support to single-NIC and multi-NIC environments.

  • [2025/04] Released v0.1:

    • Added 5 native communications libraries including CCL adaptors for NCCL (NVIDIA), IXCCL (Iluvatar), and CNCL (Cambricon), and Host CCL adaptors GLOO and Bootstrap.
    • Supported 11 heterogeneous collective communication operations using the C2C (Cluster-to-Cluster) algorithm.
    • Provided a full-stack open-source solution, FlagScale + FlagCX, for efficient heterogeneous training.
    • Natively integrated into PaddlePaddle v3.0.0, with support for both dynamic and static graphs.