AMD Data Science 26.07 release notes

AMD Data Science 26.07 release notes#

3 min read time

Applies to Linux

AMD is pleased to announce the AMD Data Science toolkit 26.07 release with the introduction of rocOpt for decision optimization on AMD Instinct GPUs. rocOpt is the AMD ROCm port of the open-source GPU-accelerated optimization engine (based on cuOpt 26.02.00, branch-26.02). rocOpt solves mixed integer linear programming (MILP), linear programming (LP), quadratic programming (QP), and vehicle routing problems (VRP) on AMD GPUs.

AMD Data Science release highlights#

The 26.07 release introduces rocOpt for decision optimization:

  • Initial ROCm/HIP port of the cuOpt engine running natively on AMD Instinct GPUs.

  • Multi-architecture fat binary built by default for gfx942 (MI300X / CDNA3) and gfx950 (MI355X). HIP selects the correct code object at load time, so a single image runs on either part.

  • Full C, C++, Python, and Server API surface carried over from cuOpt.

  • The container image builds on ROCm 7.1.1 and runs on the ROCm 7.2.3 runtime, avoiding a known code generation issue on gfx942 while providing the updated runtime to enable use with other AMD Data Science components.

System requirements#

For the 26.07 release, the AMD Data Science components support the ROCm 7.2.3 release. Refer to the Compatibility matrix to verify supported AMD Instinct GPUs, operating systems, and ROCm versions.

AMD Data Science components#

rocOpt is the only AMD Data Science component in the 26.07 release. Click to go to the component’s source code on GitHub.

Name Version GitHub
rocOpt 1.0.0

Detailed component changelog#

The following sections describe key changes to AMD Data Science components.

rocOpt (1.0.0)#

Supported APIs#

  • C++ API — native API, supporting all features

  • C API — LP, MILP, QP

  • Python API — Routing (TSP, VRP, PDP); LP, MILP, and QP via the algebraic modeling interface

  • Server API — LP, MILP, Routing (TSP, VRP, PDP)

Build & packaging#

  • Ported the RAPIDS/cuOpt build onto AMD Data Science: rapids-cmake, rmm/hipmm, libraft/hipRaft, rocthrust, libhipcxx, and the ROCm-DS logger are sourced from the open-source AMD-Ecosystem forks and the rocm-7.2.3 PyPI index.

  • Two-stage container build (dockerfile.rocm): compile on ROCm 7.1.1, run on ROCm 7.2.3.

  • MPS parser dependency repinned to rocopt-mps-parser (1.0.0).

  • Benchmark tooling: multi-GPU dispatcher, an image benchmark build option, and short LP/MILP pre-flight scripts with idempotent dataset download.

  • rocOpt version set to 1.0.0, CUOPT_VERSION (26.02.00) and CUOPT_BRANCH (branch-26.02) track the upstream base.

Example notebooks#

Jupyter notebooks under examples_notebook/ show end-to-end rocOpt workflows on AMD GPUs and are ready to run in the rocOpt container:

  • Diet optimization (LP) — diet_optimization/diet_optimization_lp.ipynb: the classic diet problem solved as a linear program with the Python modeling API, minimizing cost subject to nutritional requirements.

  • Diet optimization (MILP) — diet_optimization/diet_optimization_milp.ipynb: the same problem with integer decisions, demonstrating mixed integer linear programming.

  • Last-mile delivery (CVRPTW) — last_mile_delivery/cvrptw_benchmark_rocopt.ipynb: a capacitated vehicle routing problem with time windows on the Gehring & Homberger benchmark, focused on time-to-solution and routing performance.

The docker image does not include the example notebooks, so fetch them from the repo into /rocopt-release/examples_notebook. Each notebook includes an environment-setup step for installing prerequisites so it can be run directly from the prebuilt image.

Known limitations#

The Barrier (interior-point) LP solver depends on the cuDSS sparse direct solver, and has no ROCm equivalent in this release. Use PDLP for large LPs and dual simplex for smaller problems.

Contents