IMPORTANT: For the detailed explanation on what is this and how to use it see the instructions for building the older version: https://github.com/Efenstor/PyTorch-ROCm-gfx1010
Prebuilt wheels (2026-03-15): see the prebuilt directory.
These instructions are for building Torch 2.10, TorchVision 0.25 and TorchAudio 2.10 for Python 3.13 and ROCm 7.2 in Debian 13 (trixie) specifically for the gfx101x (RDNA 1) arch (e.g. Radeon RX 5600/5700). RDNA 2 & 3 will probably work out of the box from the official wheel.
If you experience occasional kernel panics when using PyTorch (for example, in ComfyUI those may occur during switching nodes), first of all try to increase the reserved VRAM limit. For example, I had to increase it to the absurd value of 10 GB for the kernel panics to disappear, and amazingly enough it didn't affect performance at all.
IMPORTANT: It seems that more recent kernels from the backports (6.17 and later) are causing occasional freezes with PyTorch no matter the reserved VRAM limit, so I highly recommend to stick to the stable 6.12 kernels.
Also try adding the following parameters to the kernel mode line:
amd_iommu=off amdgpu.cwsr_enable=0 amdgpu.gttsize=8192 ttm.pages_limit=32768000 ttm.page_pool_size=32768000 amdttm.pages_limit=32768000 amdttm.page_pool_size=32768000
Add it to the GRUB_CMDLINE_LINUX_DEFAULT line in /etc/default/grub, execute update-grub and reboot.
ROCm 7.2 (see the section below)
build-essential
clang
cmake
python3
git
IMPORTANT: The version of ROCm included with Debian 13 is both incomplete and outdated. You have to install ROCm 7.2.0 from the AMD's official Linux repositories.
The easiest way is to download and install amdgpu-install as described here, then execute apt install rocm as root. Do not install the AMDGPU driver, the one included with Debian is totally enough. Also instead of amdgpu-install you can add the AMD repos manually.
Download and unpack the missing RocBLAS libraries into /opt/rocm/lib/rocblas/library (as root):
wget https://github.com/Efenstor/PyTorch-ROCm-gfx1010-Debian13/raw/refs/heads/main/files/rocblas_library_gfx1010.tar.gz
tar xv -f rocblas_library_gfx1010.tar.gz -C /opt/rocm/lib/rocblas/library
IMPORTANT: After any upgrade of ROCm you will likely lose those additional libraries so keep them at hand in case you'll have to add them again. Theoretically you can disable upgrading ROCm completely by using the apt-mark hold rocm command. Anyway I recommend to stick to the 7.2 branch of ROCm, so make sure that in /etc/apt/sources.list.d/rocm.list you have the version 7.2 specified directly like this:
deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.2 noble main
deb [arch=amd64,i386 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/graphics/7.2/ubuntu noble main
After that install all the rest of the requirements using the usual apt install.
python -m venv pytorch-2.10
cd pytorch-2.10
source bin/activate
git clone https://github.com/pytorch/pytorch.git --branch=release/2.10 --recurse-submodules pytorch-release-2.10-git
cd pytorch-release-2.10-git
wget -P third_party/composable_kernel/include/ck https://github.com/Efenstor/PyTorch-ROCm-gfx1010-Debian13/raw/refs/heads/main/files/ck.hpp
wget -P third_party/composable_kernel/include/ck_tile/core https://github.com/Efenstor/PyTorch-ROCm-gfx1010-Debian13/raw/refs/heads/main/files/config.hpp
pip install -r requirements.txt
python tools/amd_build/build_amd.py
MAX_JOBS=$(nproc --all) USE_STATIC_MKL=1 USE_ROCM_CK_GEMM=1 USE_ROCM_CK_SDPA=1 USE_FLASH_ATTENTION=OFF USE_MEM_EFF_ATTENTION=OFF PYTORCH_ROCM_ARCH=gfx1010 python3 setup.py bdist_wheel
pip install dist/torch-2.10.0a0+git911aa98-cp313-cp313-linux_x86_64.whl
The resulting wheel file will be in the dist directory.
NOTE 1: USE_FLASH_ATTENTION=OFF and USE_MEM_EFF_ATTENTION=OFF are needed because both flash_attention and mem_eff_attention use aotriton, and aotriton does not support the gfx101x arch. Without these options compilation fails. A possible alternative is to change the value in cmake/External/aotriton.cmake from -DAOTRITON_TARGET_ARCH:STRING=${PYTORCH_ROCM_ARCH} to -DAOTRITON_TARGET_ARCH:STRING=gfx1100 but the consequences of forcing a partially-compatible architecture are unknown and I didn't try it.
NOTE 2: Instead of the wget commands you can clone this repository and use the files from the files directory. Those files were taken from the develop branch of composable_kernel and include the needed support for the gfx10xx arch while the original files don't. Although it's trivial to make the patches it's even easier to replace the whole files.
Execute from the same venv as was used for building torch (or any other with the torch wheel installed).
git clone https://github.com/pytorch/vision.git --branch=release/0.25 --recurse-submodules vision-release-0.25-git
cd vision-release-0.25-git
pip install setuptools==81.0.0
MAX_JOBS=$(nproc --all) python3 setup.py bdist_wheel
The resulting wheel file will be in the dist directory.
git clone https://github.com/pytorch/audio.git --branch=release/2.10 --recurse-submodules audio-release-2.10-git
cd audio-release-2.10-git
pip install -r requirements.txt
MAX_JOBS=$(nproc --all) python3 setup.py bdist_wheel
No building required, version 0.49.2 supports ROCm 7.2.0 out of the box:
pip install -U bitsandbytes
No build for ROCm 7.2.0 + Python 3.13 is available officially from AMD (see here) and I didn't try yet to compile it from the sources.
For now you may try the official wheel:
pip install -U onnxruntime
I cannot prove definitely that it utilizes the GPU but judging by the GPU load graph and the noises my graphics card makes it does, probably that's because I have migraphx installed, which goes together with ROCm 7.2.
Probably won't work for gfx10xx (see the NOTE 1 in the Torch compilation instructions above for explaination) but still worth a try.
pip install -U https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/triton-3.6.0%2Brocm7.2.0.gitba5c1517-cp313-cp313-linux_x86_64.whl