ROCm-DS 26.03 Release notes

ROCm-DS 26.03 Release notes#

6 min read time

Applies to Linux

AMD is pleased to announce the ROCm Data Science Toolkit (ROCm-DS) 26.03 release with major component upgrades, broader hardware enablement, and new distributed computing capabilities for AMD GPU environments. ROCm-DS is an open-source data science toolkit designed to improve the performance of data preparation, analytics, machine learning, and vector search workloads on AMD GPUs.

With the 26.03 release, ROCm-DS continues to strengthen its software stack while improving portability for users bringing existing GPU data science workflows to AMD hardware. The 26.03 release introduces dask-hip and hip-ucxx as new components, adding foundational support for distributed and multi-GPU workflows on ROCm. In addition, core libraries, including hipDF, hipMM, hipRAFT, and hipVS, have been updated to incorporate newer upstream functionality and support for ROCm 7.2.3.

Note

hipGRAPH remains early access (EA) in ROCm-DS 26.03.

ROCm-DS release highlights#

The 26.03 release includes the following major updates and highlights:

  • Updated core libraries, with key components aligned to newer upstream 25.10 APIs and functionality.

  • New distributed computing components:

    • dask-hip, AMD’s port of dask-cuda, for deploying and managing Dask workers on HIP-enabled GPU systems.

    • hip-ucxx, adding UCXX-based communication support for distributed GPU workflows on ROCm.

  • Major hipDF upgrade to 3.0.0, including API updates, a new pylibhipdf Python wrapper compatible with upstream pylibcudf, and expanded GPU architecture support.

  • Major hipMM upgrade to 4.0.0, introducing a precompiled shared library model (librmm.so), improved logging integration, enhanced diagnostics, and updated Python packaging.

  • hipRAFT 1.0.0 and hipVS 1.0.0 updated with new primitives, performance improvements, API changes, and expanded multi-GPU capabilities.

  • Expanded hardware support, including support for the gfx950 GPU architecture across major libraries, with additional experimental gfx11xx/gfx12xx RDNA support in hipDF.

  • ROCm 7.2.3 support across updated components.

  • Python 3.13 support added in selected components.

  • Continued support for Python and C++ APIs, helping developers integrate ROCm-DS into analytics, AI, and GPU-accelerated data science workflows.

This release advances ROCm-DS toward a more complete platform for distributed, multi-GPU, and large-scale data science workloads on AMD GPUs.

System requirements#

For the 26.03 release, the ROCm-DS components support the ROCm 7.2.3 release. Refer to the Compatibility matrix to verify supported AMD Instinct GPUs, operating systems, and ROCm versions.

ROCm-DS components#

The following table lists ROCm-DS components versions for the 26.03 release, including any version changes. Click to go to the component’s source code on GitHub.

Name Version
hipDF 2.0.0 ⇒ 3.0.0
hipMM 3.0.0 ⇒ 4.0.0
hipGRAPH 1.0.0b1
hipRAFT 0.1.0 ⇒ 1.0.0
hipVS 0.1.0 ⇒ 1.0.0
dask-hip 1.0.0
hip-ucxx 0.1.0

Detailed component Changelogs#

The following sections describe key changes to ROCm-DS components.

hipDF (3.0.0)#

Added#

  • API alignment with upstream 25.10 functionality

  • Support for ROCm 7.2.3

  • New pylibhipdf wrapper module with Python bindings compatible with upstream pylibcudf

  • Support for the gfx950 GPU architecture

  • Experimental support for gfx11xx and gfx12xx RDNA architectures:

    • gfx1100

    • gfx1101

    • gfx1200

    • gfx1201

  • Added hipcomp zstd GPU decompression support for ORC and Parquet

Known limitations#

  • DEBUG builds with -O0 are not currently supported. Use -Og or higher.

  • Some cudf.pandas acceleration layer configurations may show instability under heavy memory pressure, particularly with XNACK-related settings.

hipMM (4.0.0)#

Added#

  • Updated to newer upstream 25.10 functionality

  • Support for ROCm 7.2.3

  • Support for the gfx950 GPU architecture

  • Transition from a header-only library to a shared library model, with core implementations now delivered in librmm.so

  • Improved downstream build times through precompiled library delivery

  • New non-blocking HIP stream support via the cuda_stream class

  • Enhanced out-of-memory diagnostics with clearer failure reporting

  • Integration with ROCmDS-Logger

  • Python package support through the new amd-hipmm package and hipmm namespace

  • Python 3.13 support

Changes#

  • Downstream projects must now link against librmm.so using find_package(rmm) and rmm::rmm

  • The logging API has migrated from spdlog to rocmds-logger

hipRAFT (1.0.0)#

Added#

  • Updated to newer upstream 25.10 functionality

  • ROCm 7.2.3 support

  • Support for gfx950

  • Added RAFT_CUDA_TRY error-checking support for HIP runtime API calls

  • Ported NN-Descent graph construction kernel with wavefront-size-64 compatibility

  • Improved hipBLASLt GEMM device pointer mode handling on AMD GPUs

Changes#

  • OpenMP cleanup and HIP TLS exit behavior

  • Half-precision construction under HIP

  • GEMM test stability

  • MST correctness on 64-wide wavefronts

  • Synchronization and race conditions

  • RCCL header inclusion

  • Eigenvalue ordering and histogram initialization

Known limitations#

  • Multi-node / multi-GPU is experimental

  • raft-dask is experimental

hipVS (1.0.0)#

Added#

  • Updated to newer upstream 25.10 functionality

  • ROCm 7.2.3 support

  • Support for gfx950

  • Enabled multi-GPU algorithms and RCCL configurations

  • New Composable Kernel (CK) pairwise distance support for:

    • L2

    • cosine

    • inner-product distance kernels

  • Improved support for large datasets, including batch processing enhancements for NN-Descent

  • Added benchmark visualization tooling for interactive comparison of benchmark results

Changes#

  • Wavefront-width compatibility (wf32 and wf64)

  • ANN algorithm correctness

  • HIP race conditions

  • Spectral clustering stability

  • Graph build robustness

Known limitations#

  • Multi-node / multi-GPU remains experimental

  • wf32 support remains experimental

dask-hip (1.0.0)#

Added#

  • Initial release of dask-hip

  • ROCm 7.2.3 support

  • Based on upstream 25.10 functionality

  • Ships as the amd-dask-hip Python package on AMD PyPI

  • Introduces a pynvml-to-amdsmi translation layer so existing dask-cuda logic can operate on AMD GPUs with minimal changes

  • ROCm auto-detection at import time

  • Support for HIP_VISIBLE_DEVICES, automatically mapped to CUDA_VISIBLE_DEVICES when needed

  • Runtime dependencies for HIP interoperability, GPU telemetry, and numba-hip support

  • Optional UCXX-related dependencies for ROCm communication testing

Known limitations#

  • Only AMD GPUs are supported

  • UCXX-based multi-node communication on ROCm is experimental

  • TCP transport is fully supported

  • GPUDirect Storage (GDS) is not available on AMD in this release

  • Proxy-based spilling behavior may differ from NVIDIA GPU platforms

hip-ucxx (0.1.0)#

Added#

  • hip-ucxx is introduced in ROCm-DS 26.03 as a new communication component for distributed GPU workflows on AMD platforms.

  • ROCm 7.2.3 support

Note

UCXX-based multi-node communication is early access (EA) in ROCm-DS 26.03.

Contents