Legacy ROCm build: a component index that points at its own successor

The top of the README announces this repository's own end

The first thing on the page is a warning rather than a description. It says the repository will be deprecated soon, and that ROCm/TheRock is where to go from now on, including for issues and discussions. That is an unusual position for a hub to take: the index points away from itself. The last push was on 2026-10-14, the repository is now archived, and the release tags carry version numbers up to rocm-7.14.0. The README and the tags stay readable, while the same README tells you new work belongs somewhere else.

A map of components, not the implementation

Almost nothing here is code. The top level holds README.md, CHANGELOG.md, CONTRIBUTING.md, GOVERNANCE.md, LICENSE, a CMakeLists.txt, and directories named cmake/, docs/ and tools/, next to housekeeping files for Read the Docs, markdownlint and spellcheck. The primary language is recorded as Shell, which describes the build glue rather than the stack. Every entry in the core component list resolves somewhere else: the libraries sit in ROCm/rocm-libraries, the systems layer in ROCm/rocm-systems, each split by domain. What this repository adds up to is the inventory, the changelog and the release tags, not a build you can run from here.

HIP sits underneath every other component

ROCm is described as an open-source software stack of drivers, development tools and APIs that take GPU programming from low-level kernels up to end-user applications, aimed at GPU-accelerated HPC, AI, scientific computing and CAD. The layer the rest depends on is HIP, a C++ runtime API and kernel language for portable GPU code with an interface similar to NVIDIA CUDA. OpenMP and OpenCL are supported alongside it, and the stack ships the compilers, debuggers and libraries needed to build and run a workload. Machine learning frameworks attach at the top: the introduction names PyTorch and TensorFlow as the integrations.

Most of the library list is a hip and a roc pair

Read the math and compute section as pairs: hipBLAS and rocBLAS, hipFFT and rocFFT, hipRAND and rocRAND, hipSOLVER and rocSOLVER, hipSPARSE and rocSPARSE. Each of those appears twice, once under the hip prefix and once under roc. Sitting beside them are names that appear once: Composable Kernel, hipCUB, MIOpen, rocPRIM, rocThrust and rocWMMA. Communication has only two entries, RCCL and rocSHMEM, and those are the ones that start to matter once more than one GPU is in play. The rest of the inventory is grouped by job: HIPIFY, LLVM, ROCr Runtime and SPIRV-LLVM-Translator under runtimes and compilers; rocprofiler-compute, rocprofiler-systems, ROCprofiler-SDK, ROCdbgapi, ROCgdb and the ROCR Debug Agent for profiling and debugging; AMD SMI, the ROCm Data Center Tool and rocminfo for control and monitoring; rocDecode and rocJPEG for media; hipFile for storage.

Support is defined in a matrix this page does not copy

Hardware and operating system support is delegated rather than stated. The instruction is to use the compatibility matrix in the ROCm documentation for official support across ROCm versions, operating system kernels and GPU architectures, covering CDNA and Instinct, RDNA and Radeon, plus Radeon Pro. None of it is reproduced here: no kernel versions, no minimum driver, no per-distribution package names. What the text does put a name to is the range. Recent releases cover Ubuntu, RHEL, SLES, Oracle Linux, Debian and Rocky Linux, and the GPU targets named are CDNA4, CDNA3, CDNA2, RDNA4 and RDNA3. If your card or your kernel is not on that matrix, this repository will not tell you, and neither will anything else in it.

Release tags stop two months before the last commit

Three releases are visible: rocm-7.14.0 on 2026-07-16, rocm-7.2.4 on 2026-05-29 and rocm-7.2.3 on 2026-05-04. Each is named after the ROCm version it carries, so a tag here is a stack-wide version rather than the version of one package. Two things follow from that. The numbering jumps from 7.2.4 to 7.14.0 with nothing in between on the release page, and the newest tag is roughly three months older than the last push on 2026-10-14, so commits sitting on the default branch are not covered by any release listed here.

Installation is a link to the documentation portal

There is no installation command in this repository, and there is not meant to be. The quick start comes down to two references: follow the ROCm installation guide to install ROCm on your system, and for machine learning work use the PyTorch or JAX install pages for ROCm. Everything executable lives in the documentation portal and in the component repositories. One mismatch is worth noticing: the introduction names PyTorch and TensorFlow as the frameworks ROCm integrates with, while the quick start links only PyTorch and JAX. If TensorFlow is why you are here, the place to start is the AI Ecosystem documentation portal rather than those two framework links.

TheRock is where the issues and the work move

The successor is named in the warning at the top: use ROCm/TheRock from now on, including for issues and discussions. That matters for planning because the licence split follows the repository split. This repository is MIT, which covers the documentation, changelog and release plumbing here, while the components it points at live in their own repositories under their own terms. Three checks are worth doing before a build depends on this hub: whether TheRock already carries the component you need, which repository your CI should track, and what the compatibility matrix says about your GPU architecture. Following a hub that has announced its own deprecation costs time later.

Editorial conclusion

Read this repository as a map and a changelog, then build against the documentation portal and the component repositories. It earns a look if you are new to ROCm and want the component inventory in one page, or if you are moving a CUDA codebase and want to know which layers have a HIP equivalent. Skip it if you need install steps, version requirements or a list of supported hardware, because none of that is kept here. Before planning anything, open ROCm/TheRock and confirm the component you need has moved, then check the compatibility matrix for your GPU architecture.

Frequently asked questions

Official sources

1. License: MIT
2. Project website
3. README
4. Releases
5. ROCm/legacy-rocm-build on GitHub