TheRock (The HIP Environment and ROCm Kit)

The s3_management toolset consists of two primary scripts used to maintain Python package indices on S3:

1. update_dependencies.py: Mirrors dependency packages from PyPI to the project's S3 buckets.
2. manage.py: Generates PEP 503-compliant index.html files for each subdirectory in the S3 buckets, including built packages (like rocm or torch) and dependencies.

Managed S3 Buckets:

• therock-dev-python: Used for development releases.
• therock-nightly-python: Used for nightly/production-like releases.

TheRock is a lightweight, open-source build platform for HIP and ROCm. It is designed for ROCm contributors, developers, researchers, and advanced users who need access to the latest ROCm capabilities without the complexity of traditional package-based installations.

Key features include:

• Nightly releases of ROCm and PyTorch.
• A CMake super-project for building HIP and ROCm from source.
• Support for building PyTorch and JAX with ROCm from source.
• Operating system support for multiple Linux distributions and native Windows.
• Tools for developing individual ROCm components.
• Comprehensive CI/CD pipelines for building, testing, and releasing supported components.

Note: The project is currently in an early preview state.

TheRock provides the build configuration for several ROCm debug tools. Note that TheRock does not host the source code for these tools directly; instead, it manages them via git submodule pointers to external repositories.

Supported debug tools include:

• ROCgdb: A debugger for ROCm.
• ROCdbgapi: An API for debugging.
• ROCr debug agent: A debug agent for ROCr.

To access the actual source code, you must ensure the submodules are initialized according to the project structure.

The ROCm Media Libraries provide high-performance video and image decoding for AMD GPUs on Linux. They leverage VA-API for hardware acceleration.

Supported libraries:

• rocDecode: High-performance video decoding.
• rocJPEG: High-performance JPEG decoding.

The ROCM BLAS suite implements BLAS (Basic Linear Algebra Subprograms) APIs specifically optimized for AMD GPUs. The suite consists of both low-level implementation libraries and high-level user-facing API libraries.

If you are developing applications that require linear algebra operations on AMD hardware, you should target the following user-level API libraries:

• hipBLASLt: A lightweight, high-performance version of hipBLAS.
• hipBLAS: The standard hipBLAS API.
• hipSOLVER: For solving linear systems and eigenvalue problems.
• hipSPARSE: For sparse matrix operations.
• hipSPARSELt: A lightweight, high-performance version of hipSPARSE.

The ROCm SDK package is designed to provide developers with the necessary components for working with ROCm. It enables three primary capabilities:

1. Runtime Library Access: Provides the packages required to use ROCm libraries during application execution.
2. GPU-Specific Library Selection: Allows users to choose specific subsets of libraries tailored to particular GPU architectures.
3. Development Integration: Facilitates development with ROCm and HIP using the CMake build system.

The PyTorch Smoke Tests repository provides a suite of basic tests designed to verify fundamental PyTorch tensor operations.

Important Disclaimer: These tests are intended as quick sanity checks for basic operations and are not exhaustive. They do not replace a comprehensive PyTorch testing suite.

The repository is organized into a hierarchy where sub-projects reside at the leaves. Common sub-project categories include:

• base/: Core, minimally co-dependent utility sub-projects.
• compiler/: Compilers like AMD-LLVM and hipcc.
• core/: Low-level runtimes like ROCR-Runtime and HIP runtimes.
• comm-libs/: Communication libraries (e.g., rccl).
• storage-libs/: Storage libraries (e.g., hipFile).
• math-libs/: Math libraries (e.g., BLAS, hipBLASLt).
• ml-libs/: Machine learning libraries (e.g., MIOpen).
• media-libs/: Media decoding (e.g., rocDecode, rocJPEG).
• profiler/: Profiler sub-projects.

The rocm-profiler Python package provides console script wrappers for various ROCm profiling tools. These scripts act as dispatchers that resolve the packaged ROCm installation root and execute the corresponding platform-specific binaries located in the bin/ directory of the wheel.

Available console entry points include:

• rocprof-compute
• rocprof-sys-avail
• rocprof-sys-causal
• rocprof-sys-instrument
• rocprof-sys-python
• rocprof-sys-run
• rocprof-sys-sample

This directory contains tooling specifically designed to build PyTorch with ROCm Python wheels.

Note: If you prefer to install prebuilt PyTorch packages rather than compiling them from source, refer to the RELEASES.md file instead.

The ROCm/quartz repository is organized into functional areas:

• .github/workflows/: Contains the three main entry points: receive-therock-data.yml (ingestion), notify-downstream.yml (pushing status), and receive-downstream-data.yml (ingesting downstream results).
• config/: Contains subscriber.yml (notification targets) and allow-list/ (mapping repositories to GitHub App IDs for reporting back).
• scripts/: Python-based business logic for validation, schema checking, and database insertion.
• release-nightly/ & prerelease/: Directories containing static status.json artifacts for polling.
• templates/: Example workflow files for downstream projects (pull, push, and reporting models).
• docs/: Implementation guides for different subscription and reporting models.