ROCm-DS 25.10 Release notes

ROCm-DS 25.10 Release notes#

4 min read time

Applies to Linux

AMD is pleased to announce the general availability (GA) release of the ROCm Data Science Toolkit (ROCm-DS), a comprehensive open-source software collection designed to accelerate data science and machine learning workloads on AMD GPUs. This GA release marks a significant milestone for ROCm-DS as hipDF and hipMM transition to production status. Additionally, it introduces two new production components: hipRAFT and hipVS. The GA release provides an end-to-end GPU-accelerated data science and AI ecosystem.

Note

hipGRAPH remains early access (EA) in ROCm-DS 25.10.

ROCm-DS release highlights#

ROCm-DS 25.10 is a general availability release providing a production-ready foundation for GPU-accelerated data science on AMD hardware, with continued development focused on expanded capabilities and performance optimization in future versions. This release includes the following major updates and highlights:

  • Transition of hipDF and hipMM from early access to production-ready status.

  • Introduction of hipRAFT and hipVS for advanced data science, machine learning, and vector similarity workloads.

  • Alignment of all ROCm-DS components with RAPIDS 25.02, ensuring compatibility and feature parity.

  • Enhanced performance, memory efficiency, and API stability across all components.

  • Continued support for both Python and C++ APIs, enabling integration with popular frameworks such as Pandas and other RAPIDS-compatible tools.

ROCm-DS components#

The following table lists ROCm-DS components versions for the 25.10 release, including any version changes for the components. Click to go to the component’s source code on GitHub.

Name Version
hipDF 1.0.0b1 ⇒ 2.0.0
hipMM 1.0.0b1 ⇒ 3.0.0
hipGRAPH 1.0.0b1
hipRAFT 0.1.0
hipVS 0.1.0

Detailed component Changelogs#

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

hipDF (2.0.0)#

Added#

  • Major upgrade aligning hipDF APIs with RAPIDS cuDF 25.02 APIs

  • ROCm 7.0.2 support

Known limitations and notes#

  • DEBUG builds with -O0 optimization are not currently supported. Use -Og or higher for DEBUG builds (default setting). Support for -O0 is planned in a future toolchain update.

  • When using the cudf.pandas acceleration layer with XNACK enabled and workloads that significantly exceed physical GPU VRAM (oversubscription), some systems might exhibit instability or reduced performance under heavy memory pressure.

  • Using the cudf.pandas acceleration layer with XNACK disabled (HSA_XNACK=0) can trigger instabilities.

hipMM (3.0.0)#

Added#

  • Major upgrade aligning hipDF APIs with RAPIDS cuDF 25.02 APIs

  • ROCm 7.0.2 support

hipRAFT (0.1.0)#

Added#

  • Major upgrade aligning hipDF APIs with RAPIDS cuDF 25.02 APIs

  • ROCm 7.0.2 support

Changes#

  • Fixed uninitialized shared memory and bounds checking issues

  • Optimized kernel launch parameters and LDS memory access for AMD GPUs

  • Additional thread safety

Known limitations and notes#

  • Multi-Node/ Multi GPU is experimental

  • Raft-dask is unsupported

hipVS (0.1.0)#

Added#

  • Major upgrade aligning hipDF APIs with RAPIDS cuDF 25.02 APIs

  • ROCm 7.0.2 support

Changes#

  • Fixed uninitialized shared memory and bounds checking issues

  • Optimized kernel launch parameters and LDS memory access for AMD GPUs

  • Relaxed recall thresholds to account for hardware differences

Known limitations and notes#

  • Multi-GPU neighbors API with ROCm RCCL integration is experimental