A UV-managed Python project for running PyTorch with AMD ROCm on Windows, specifically optimized for RX 6800/6900 (gfx1030) GPUs.
| GPU | Nightly Env | Official Env |
|---|---|---|
| RX 6800 | ✅ Working | ❌ Not detected |
| RX 6800 XT | ✅ Should work | ❌ Not detected |
| RX 6900 XT | ✅ Should work | ❌ Not detected |
| RX 6600/6600 XT | ✅ Should work | ❌ Not detected |
Recommendation for RX 6000 series: Use the Nightly environment. The official AMD wheels do not detect gfx1030 GPUs on Windows.
# 1. Setup nightly environment (recommended for RX 6800/6900)
.\scripts\setup-nightly.bat
# 2. Test GPU detection
.\scripts\run-test.bat nightly
# 3. Activate and use
.\activate.bat
python -c "import torch; print(torch.cuda.get_device_name(0))"
# Output: AMD Radeon RX 6800- Tested GPUs
- Overview
- Supported GPUs
- Prerequisites
- Installation
- Testing
- Project Structure
- Environment Variables
- Troubleshooting
- References
This project provides two PyTorch environments for AMD GPUs on Windows:
| Environment | ROCm Version | PyTorch Version | Target GPUs | Windows Status |
|---|---|---|---|---|
| Nightly | 7.12 | 2.10.0 | gfx103x (RX 6800/6900/6600) | ✅ Working |
| Official | 7.2 | 2.9.1 | gfx1100/gfx1101 (RX 7000) |
AMD's official PyTorch wheels for Windows only support RDNA3 (RX 7000) and RDNA4 (RX 9000) GPUs. The RX 6000 series (RDNA2, gfx1030) is supported by the HIP SDK but NOT by the official PyTorch wheels.
This project offers two approaches:
-
Nightly gfx103x ✅ RECOMMENDED: Uses AMD's semi-official nightly builds specifically compiled for gfx103x GPUs. Tested and working on RX 6800.
-
Official + HSA Override ❌: Uses the official wheels with
HSA_OVERRIDE_GFX_VERSION=10.3.0. Does not work on RX 6800 - GPU not detected.
| GPU | Architecture | Support |
|---|---|---|
| RX 7900 XTX | gfx1100 | ✅ Official |
| RX 7900 XT | gfx1100 | ✅ Official |
| RX 7800 XT | gfx1101 | ✅ Official |
| RX 7700 XT | gfx1101 | ✅ Official |
| RX 9070 / 9070 XT | gfx1200/1201 | ✅ Official |
| RX 6800 / 6800 XT | gfx1030 | |
| RX 6900 XT | gfx1030 |
| GPU | Architecture | Support |
|---|---|---|
| RX 6800 / 6800 XT | gfx1030 | ✅ |
| RX 6900 XT | gfx1030 | ✅ |
| RX 6800M | gfx1031 | ✅ |
| RX 6600 / 6600 XT | gfx1032 | ✅ |
| RX 6600M | gfx1032 | ✅ |
-
UV Package Manager (will auto-install Python 3.12)
pip install uv # OR pipx install uv- UV automatically downloads and manages Python 3.12
- No manual Python installation needed!
-
AMD HIP SDK
- Download: https://rocm.docs.amd.com/projects/install-on-windows/en/latest/
- You mentioned this is already installed ✅
-
AMD Graphics Driver
- Minimum: Version 26.1.1
- Download: https://www.amd.com/en/resources/support-articles/release-notes/RN-RAD-WIN-26-1-1.html
- Python 3.12 (if you want system-wide installation): https://www.python.org/ftp/python/3.12.10/python-3.12.10-amd64.exe
- Git (for cloning): https://git-scm.com/download/win
- Visual C++ Redistributable: https://aka.ms/vs/17/release/vc_redist.x64.exe
- Visual Studio Build Tools (for compiling extensions): https://aka.ms/vs/17/release/vs_BuildTools.exe
# 1. Setup the nightly environment (WORKS for RX 6800)
.\scripts\setup-nightly.bat
# 2. Test GPU detection
.\scripts\run-test.bat nightly
# 3. Activate and use
.\activate.bat
python -c "import torch; print(torch.cuda.get_device_name(0))"
# Output: AMD Radeon RX 6800.\scripts\setup-nightly.batThis will:
- Create
.venv-nightlywith Python 3.12 (UV auto-downloads if needed) - Install ROCm 7.12 SDK for gfx103x
- Install PyTorch 2.10.0 built specifically for gfx103x
- Install triton-windows and other dependencies
Tested on: RX 6800 - ✅ GPU detected and working
.\scripts\setup-official.batStatus: ❌ GPU not detected on RX 6800 even with HSA_OVERRIDE_GFX_VERSION
The official AMD wheels (ROCm 7.2) only support gfx1100/gfx1101 (RX 7000 series) and gfx120x (RX 9000 series) on Windows. RX 6000 series users should use the nightly environment.
.\scripts\run-test.bat nightly============================================================
RCOM GPU Detection Test
============================================================
--- System Information ---
OS: Windows 11
Python: 3.12.9
Architecture: AMD64
--- GPU Detection ---
CUDA Available: True
ROCm/HIP Version: 7.2.0
Device Count: 1
Device Name: AMD Radeon RX 6800
[SUCCESS] GPU detected!
--- Tensor Operations Test ---
[OK] Success! Result shape: (1000, 1000), dtype: float32
--- Memory Test ---
[OK] Total GPU Memory: 16368 MB
============================================================
All Tests Passed!
============================================================
RCOM/
├── pyproject.toml # UV project configuration
├── .python-version # Python version pin (3.12)
├── README.md # This file
├── activate.bat # Quick activation script
├── quick-start.bat # Setup both environments
│
├── envs/
│ ├── official/
│ │ └── requirements.txt # Official ROCm 7.2 wheel URLs
│ └── nightly/
│ └── requirements.txt # Nightly gfx103x wheel URLs
│
├── scripts/
│ ├── setup-official.bat # Setup official environment
│ ├── setup-nightly.bat # Setup nightly environment
│ ├── run-test.bat # Test runner for both envs
│ └── test-gpu.py # GPU detection test script
│
├── .venv-official/ # Official environment (created by setup)
└── .venv-nightly/ # Nightly environment (created by setup)
Overrides GPU architecture detection. Required for RX 6000 series with official wheels.
| GPU Series | Value | Architecture |
|---|---|---|
| RX 6800/6800 XT/6900 XT | 10.3.0 |
gfx1030 |
| RX 6600/6600 XT | 10.3.2 |
gfx1032 |
| RX 7900 XTX/XT | 11.0.0 |
gfx1100 |
| RX 7800 XT | 11.0.1 |
gfx1101 |
| RX 7700 XT | 11.0.1 |
gfx1101 |
set HSA_OVERRIDE_GFX_VERSION=10.3.0
python your_script.py$env:HSA_OVERRIDE_GFX_VERSION = "10.3.0"
python your_script.py# ROCm debug logging
set MIOPEN_LOG_LEVEL=5
set MIOPEN_FIND_MODE=5
# Triton cache directory
set TRITON_CACHE_DIR=C:\path\to\cache- Official ROCm 7.2: https://repo.radeon.com/rocm/windows/rocm-rel-7.2/
- Nightly gfx103x: https://rocm.nightlies.amd.com/v2-staging/gfx103X-dgpu/
- Test on your AMD GPU
- Report issues with GPU model, driver version, and error messages
- Submit PRs for improvements