Contents:
- Pre-built wheels
- Introduction
- Requirements
- Prepare
- Build
- Troubleshooting
- TorchVision
- TorchAudio
- Extras
For usage examples (chaiNNer, ComfyUI, etc.) see the `usage` directory.
Pre-built wheels of PyTorch, TorchVision and TorchAudio (built for ROCm 6.3.4 and Python 3.11) are located in the `prebuilt` directory.
There is no guarantee that those will work with your particular configuration of video card, ROCm version, Python version, PyTorch version, kernel version, etc.
For best results use the following instructions to build those wheels yourself.
These instructions are the result of the efforts to make PyTorch work with unofficially-supported AMD gfx1010 GPUs (Radeon RX 5000 series) on Debian 12.
Unfortunately none of the pre-compiled binaries work for gfx1010 GPUs. Anecdotally they did work in the past but no more, even with not-so-fresh versions of PyTorch.
Using various tricks, such as setting the HSA_OVERRIDE_GFX_VERSION=10.3.0 environment variable or using older versions of ROCm, PyTorch or kernel do not work as well causing either:
- `Compile with 'TORCH_USE_HIP_DSA' to enable device-side assertions` message
- Crashing X11 every time calculations begin
- Calculations going on forever with 100% GPU load and no characteristic noises coming out of the GPU, which means no actual work is being done
The only working solution seems to be is to build PyTorch and TorchVision from the source code with the support for gfx1010 architecture explicitly on.
The resulting Python wheels were tested with chaiNNer and ComfyUI and both seem to work fine.
build-essential
clang
cmake
python3
git
ROCm (for ROCm build)
To begin with you have to install ROCm. Unfortunately you have to trust AMD here: although ROCm is open-source building it from the source is another enormous task. Good news is that despite the instructions from AMD there is no actual need to install the proprietary amdgpu driver (amdgpu-dkms) and that the upcoming Debian 13 (trixie) will have ROCm in its repositories out of the box.
Download and install amdgpu-install as described here. Then execute apt install rocm as root.
Download and unpack the missing RocBLAS libraries into /opt/rocm/lib/rocblas/library:
wget https://github.com/Efenstor/PyTorch-ROCm-gfx1010/blob/main/prebuilt/rocblas_library_gfx1010.tar.gz
tar xv -f rocblas_library_gfx1010.tar.gz -C /opt/rocm/lib/rocblas/library
Then install all the rest of the requirements using the usual apt install.
The following instructions are given for PyTorch 2.7, when a new version arrives just replace 2.7 with the new numbers.
IMPORTANT: always create a new clean venv to build a new version of pytorch to avoid dependency conflicts.
Create and activate a venv:
python -m venv pytorch-2.7
cd pytorch-2.7
source bin/activate
Later to exit the venv execute deactivate.
Clone the repo:
git clone https://github.com/pytorch/pytorch.git --branch=release/2.7 --recurse-submodules pytorch-release-2.7-git
Prepare:
cd pytorch-release-2.7-git
pip install -r requirements.txt
python tools/amd_build/build_amd.py
Build:
MAX_JOBS=$(nproc) CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) PYTORCH_ROCM_ARCH=gfx1010 python3 setup.py bdist_wheel
The resulting wheel will be in the dist directory. Be patient, build takes a very long time.
If build fails see the Troubleshooting section below, then execute the build command again.
To install the built wheel execute inside a venv:
pip install pytorch-release-2.7-git/dist/torch-2.7.0a0+git1341794-cp311-cp311-linux_x86_64.whl
CMAKE_POLICY_VERSION_MINIMUM=3.5 have to be added to the enviromnent.
Compile with this command instead:
MAX_JOBS=$(nproc) CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) CMAKE_POLICY_VERSION_MINIMUM=3.5 PYTORCH_ROCM_ARCH=gfx1010 python3 setup.py bdist_wheel
C and CXX flags have to be added to lessen the error policy.
To add execute:
echo "set(CMAKE_C_FLAGS \"\${CMAKE_CXX_FLAGS} -Wno-error=maybe-uninitialized -Wno-error=uninitialized -Wno-error=restrict\")" >> third_party/fbgemm/CMakeLists.txt
echo "set(CMAKE_CXX_FLAGS \"\${CMAKE_CXX_FLAGS} -Wno-error=maybe-uninitialized -Wno-error=uninitialized -Wno-error=restrict\")" >> third_party/fbgemm/CMakeLists.txt
Apply patches (read here about this issue):
git apply --directory=third_party/composable_kernel patches/5465fcc9e25ab9828b9d34ce5d341a127ff8ea9e.patch
git apply --directory=third_party/composable_kernel patches/88952b6d4e6bea810aaa4c063bdaf5b8252acb1c.patch
git apply --directory=third_party/composable_kernel patches/3e23090b5b29d0eea3bbec0ee1b03a182894c831.patch
Remember that `patches` is a subdirectory in this repository, so replace it with your path.
If a patch cannot be applied it is probably not needed for your version of Torch (in my case 3e23090b5b29d0eea3bbec0ee1b03a182894c831.patch could not be applied but the build still completed successfully).
This built should be done in a venv with the previously built pytorch wheel installed.
git clone https://github.com/pytorch/vision.git --branch=release/0.22 --recurse-submodules vision-release-0.22-git
cd vision-release-0.22-git
MAX_JOBS=$(nproc) CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) python3 setup.py bdist_wheel
The resulting wheel will be in the dist directory.
This built should be done in a venv with the previously built pytorch wheel installed.
git clone https://github.com/pytorch/audio.git --branch=release/2.7 --recurse-submodules audio-release-2.7-git
cd audio-release-2.7-git
MAX_JOBS=$(nproc) CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) python3 setup.py bdist_wheel
The resulting wheel will be in the dist directory.
STATUS: can be built but does not work
Support for ROCm HIP for bitsandbytes is in active development, in other words you can make it work if you're really tenacious.
The build instructions as of April 2025 (or you can use the pre-build wheel from the prebuilt dir):
git clone https://github.com/bitsandbytes-foundation/bitsandbytes.git --branch=multi-backend-refactor
cd bitsandbytes
git reset --hard a0a95fd
mkdir build; cd build
cmake .. -DCOMPUTE_BACKEND=hip
make -j$(nproc)
cd ..
python3 setup.py bdist_wheel
The resulting wheel will be in the dist directory.
After installing the wheel you also have to install triton v3.1.0:
bin/pip install triton==3.1.0
STATUS: cannot be built
So far the attempts to build xformers v0.0.29.post3 for ROCm gfx1010 were unsuccessful.
The main culprit is composable_kernel_tiled, which has no support for gfx1010, and although I manually added it to include/ck/ck.hpp and include/ck_tile/core/config.hpp I got stuck with a completely different errors, likely related to the versions of gcc/clang that are available in Debian 12.
STATUS: can be built but untested
Torch can be successfully built with the Vulkan backend which in theory would allow it to run without ROCm, but in practice unlike PyTorch for ROCm it is not a drop-in replacement for PyTorch for CUDA and should be explicitly supported by the end-user software. Therefore the following instructions are more like a bonus for those who one day may find such a software.
Additional requirements:
glslc
All the instructions are the same as for ROCm, except for the build command:
MAX_JOBS=$(nproc) CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) CMAKE_POLICY_VERSION_MINIMUM=3.5 USE_VULKAN=1 USE_VULKAN_SHADERC_RUNTIME=1 USE_VULKAN_WRAPPER=0 python3 setup.py bdist_wheel
If you get `is not a member of ‘torch::jit::cuda’` errors then apply patch:
git apply patches/a72d9c3b7554b78a91a3d61b9041b3d1d7acf319.patch
Additional environment variables to try (untested):
USE_VULKAN_FP16_INFERENCE=ON
USE_VULKAN_RELAXED_PRECISION=ON
Later versions of ROCm no longer include the RocBLAS TensileLibrary files for gfx1010. You can find those separately in the `prebuilt` directory as rocblas_library_gfx1010.tar.gz. Just unpack it into /opt/rocm/lib/rocblas/library (as root).