AMD Data Science 26.10 updates the toolkit to ROCm 10.0.0 and brings the core libraries to their RAPIDS 26.06 upstream baselines. The release adds modernized memory-resource interfaces, improved distributed GPU communication, and expanded vector-search and machine-learning functionality.
rocOpt is not updated with the 26.10 release. For more information, see the {doc}`Compatibility matrix <compatibility-matrix>`.
- AMD Data Science release highlights
- System requirements
- AMD Data Science components
- Detailed component changelogs
- Important compatibility notes
The 26.10 release includes the following major updates and highlights:
- Support for upstream RAPIDS 26.06 components.
- Supports AMD Instinct GPUs MI355X (GPU target gfx950), MI325X / MI300X (GPU target gfx942), MI250 / MI210 (GPU target gfx90a)
- ROCm 10.0.0 support across updated components.
- Support for Python 3.11, 3.12, and 3.13.
- Improved hipBLASLt fallback behavior and numerical stability for randomized SVD on
gfx942in hipRAFT.
Refer to the {doc}Compatibility matrix <compatibility-matrix> to verify supported AMD Instinct GPUs, operating systems, and ROCm versions.
The following table lists AMD Data Science component versions for the 26.10 release, including any version
changes. Click {fab}github to go to the component's source code on GitHub.
| Name | Version | GitHub |
|---|---|---|
| hipDF | 3.0.0 ⇒ 4.0.0 | |
| hipMM | 4.0.0 ⇒ 5.0.0 | |
| hipRAFT | 1.0.0 ⇒ 2.0.0 | |
| hipVS | 1.0.0 ⇒ 2.0.0 | |
| dask-hip | 1.0.0 ⇒ 2.0.0 | |
| hip-ucxx | 0.1.0 ⇒ 0.2.0 |
rocOpt is not updated with the 26.10 release. For more information, see the {doc}`Compatibility matrix <compatibility-matrix>`.
The following sections describe key changes to AMD Data Science components.
hipDF now follows the RAPIDS cuDF 26.06 API baseline, improving portability
for applications that use cuDF-compatible DataFrame and Python interfaces.
This release adds ROCm 10.0.0 and gfx950 support and expands AMD PyPI
packaging across the core C++ and Python libraries and the Dask integration.
hipDF continues to provide GPU-native Series and DataFrame operations, including joins, aggregations, window operations, string and list processing, and CSV, JSON, Parquet, ORC, HDF5, Feather, and Avro I/O.
hipMM adopts the upstream RMM 26.06 memory-resource model and the unified CCCL memory-resource design. This provides a more consistent foundation for allocation, pooling, tracking, and interoperability across AMD Data Science components.
This release also modernizes device-buffer and allocator interfaces, updates
the Python bindings, and provides AMD PyPI packages for both the C++ library
and Python API. Applications using older device_memory_resource APIs might
need migration updates.
hipRAFT upgrades to the RAFT 26.06 baseline and adds PCA, truncated SVD, and sparse randomized SVD primitives. It also exposes NumPy serialization helpers and adopts CCCL-based memory-resource handling, including support for tracking resources.
This release improves hipBLASLt fallback behavior, numerical stability for
randomized SVD on gfx942, gfx90a builds, histogram initialization, and
stream synchronization. Python packages are built with scikit-build-core
and are published for Python 3.11, 3.12, and 3.13.
hipVS upgrades from cuVS 25.10 to cuVS 26.06 and makes multi-GPU, multi-node, and wave32 execution official rather than experimental. The release brings several capabilities that make vector-search workloads more portable and more useful on AMD GPUs:
- End-to-end FP64 brute-force k-nearest-neighbor support across the C++, C, and Python APIs.
- ROCm-native JIT-LTO support through hipRTC and hipModule for CAGRA, IVF-Flat, IVF-PQ, IVF-SQ, and user-defined metrics.
- A Composable Kernel backend for fused FP32 L2, cosine, and inner-product distance calculations, including skinny-K optimization.
- Improved FP32 NN-Descent recall through a native HIP local-join kernel.
- IVF-SQ, PCA, KDE, and user-defined-distance APIs, plus expanded CAGRA, KMeans, NN-Descent, HNSW, and all-neighbors functionality.
- Rust
hipvsbindings updated for the 26.06 API, including Vamana and DiskANN support. - Manylinux wheels for Python 3.11 and later and a new
HIPVS_ENABLE_CK_BACKENDbuild option.
The release also hardens index deserialization against malicious inputs and improves correctness for wave32, RDNA3 CK distance paths, sparse metrics, multi-GPU JIT/UDF execution, and graph-search workloads.
dask-hip improves deployment of Dask workloads across AMD GPU systems. It
provides one-worker-per-GPU placement with HIP_VISIBLE_DEVICES, CPU
affinity, UCX communication, GPU-memory spilling, hipMM pool integration, and
command-line worker management.
This release updates the dask-cuda 26.06 baseline, removes deprecated upstream components, and replaces CUDA-specific compatibility dependencies with HIP-aware Python interoperability. This allows existing Dask GPU workflows to be adapted to AMD systems with fewer application changes.
hip-ucxx updates its UCXX baseline to 0.50.01 and provides the communication layer for distributed AMD GPU workloads. It supports tag matching, streams, active messages, remote memory access, and GPU-direct transfers over ROCm-enabled UCX transports.
This release includes asynchronous Python bindings, the
distributed-ucxx Dask backend, and support for hipMM device buffers. It
also updates the AMD PyPI packages and build requirements for ROCm 10.0.0.
- Use isolated AMD Data Science environments when installing the AMD PyPI packages.
- hipVS INT8 IVF-PQ coarse search is unsupported and FP64 brute-force search uses the tiled, non-fused kNN path.
- hipDF has known limitations around some
cudf.pandasconfigurations under heavy XNACK memory pressure. - Some advanced hipRAFT distributed and randomized-SVD APIs remain unsupported or experimental.