chaiNNer PyTorch with ROCm for AMD GPU (Linux)

[mrintrepide]

Hi,

chaiNNer use different backend, but PyTorch is the most used.
PyTorch have an official ROCm version (and AMD build)

EDIT: Stop working with Kernel 6.18.4 and newer, need to use newer ROCm 7.11 build from nightly AMD repo
Use torch==2.7.1+rocm7.11.0a20260114 and torchvision==0.22.1+rocm7.11.0a20260114 with https://rocm.nightlies.amd.com/v2/gfx-XXXX

You can just install it without rebuild chaiNNer from built-in Dependency Manager
Switch to "Advanced" in Python and select "Manual/Copy"

Click on "Install" PyTorch, it's copy command line to install manually from terminal with pip the PyTorch dependency.

You need to change package name and add AMD PyTorch repo.

For RDNA 2 (GFX1030) ROCm 6.3

/home/$USER/.config/chaiNNer/python/python/bin/python3.11 -m pip install --upgrade torch==2.7.1+rocm6.3 torchvision==0.22.1+rocm6.3 facexlib==0.3.0 einops==0.6.1 safetensors==0.4.0 spandrel==0.4.1 spandrel_extra_arches==0.2.0  --extra-index-url https://download.pytorch.org/whl/rocm6.3/

For RDNA 3 (GFX1100 / GFX1101) ROCm 7.10

/home/$USER/.config/chaiNNer/python/python/bin/python3.11 -m pip install --upgrade torch==2.7.1+rocm7.10.0 torchvision==0.22.1+rocm7.10.0 facexlib==0.3.0 einops==0.6.1 safetensors==0.4.0 spandrel==0.4.1 spandrel_extra_arches==0.2.0  --extra-index-url https://download.pytorch.org/whl/rocm6.3/ --extra-index-url https://repo.amd.com/rocm/whl/gfx110X-dgpu/

For RDNA 3.5 (GFX1150 - Strix Point - Radeon 8X0M) ROCm 7.10

/home/$USER/.config/chaiNNer/python/python/bin/python3.11 -m pip install --upgrade torch==2.7.1+rocm7.10.0 torchvision==0.22.1+rocm7.10.0 facexlib==0.3.0 einops==0.6.1 safetensors==0.4.0 spandrel==0.4.1 spandrel_extra_arches==0.2.0  --extra-index-url https://download.pytorch.org/whl/rocm6.3/ --extra-index-url https://repo.amd.com/rocm/whl/gfx1150/

For RDNA 3.5 (GFX1151 - Strix Halo - Radeon 80X0S) ROCm 7.10

/home/$USER/.config/chaiNNer/python/python/bin/python3.11 -m pip install --upgrade torch==2.7.1+rocm7.10.0 torchvision==0.22.1+rocm7.10.0 facexlib==0.3.0 einops==0.6.1 safetensors==0.4.0 spandrel==0.4.1 spandrel_extra_arches==0.2.0  --extra-index-url https://download.pytorch.org/whl/rocm6.3/ --extra-index-url https://repo.amd.com/rocm/whl/gfx1151/

For RDNA 4 (GFX1200 / GFX1201) ROCm 7.10

/home/$USER/.config/chaiNNer/python/python/bin/python3.11 -m pip install --upgrade torch==2.7.1+rocm7.10.0 torchvision==0.22.1+rocm7.10.0 facexlib==0.3.0 einops==0.6.1 safetensors==0.4.0 spandrel==0.4.1 spandrel_extra_arches==0.2.0  --extra-index-url https://download.pytorch.org/whl/rocm6.3/ --extra-index-url https://repo.amd.com/rocm/whl/gfx120X-all/

For other AMD GPU, check more from AMD repo: https://repo.amd.com/rocm/whl/

[Bl4ckH4nd-87]

Sadly, this didn't work for me.

[Peganeo]

Probably you applied wrong things. Delete everything and try again. Cause it worked for me! Now ROCm 7.2 supporting Windows native. I want to use ROCm in Windows too.

[Bl4ckH4nd-87]

I have a 9070XT, I applied this one:
/home/$USER/.config/chaiNNer/python/python/bin/python3.11 -m pip install --upgrade torch==2.7.1+rocm7.10.0 torchvision==0.22.1+rocm7.10.0 facexlib==0.3.0 einops==0.6.1 safetensors==0.4.0 spandrel==0.4.1 spandrel_extra_arches==0.2.0 --extra-index-url https://download.pytorch.org/whl/rocm6.3/ --extra-index-url https://repo.amd.com/rocm/whl/gfx120X-all/

I'm guessing it's because I'm using Windows, should have mentioned it in my first comment, apologies.

[mrintrepide]

Yes, no Windows package for torch < 2.9 with ROCm and support RDNA 3.5 and 4
ChaiNNer use torch 2.7

[Bl4ckH4nd-87]

That's disappointing. At least the ComfyUI that comes with the new AMD drivers is working well so far, seems to be faster than using NCNN in chaiNNer. I just setup a workflow, and I'm going to compare speeds with a batch of 500 images.

[mrintrepide]

Can you try with:

python3.11 -m pip install --upgrade torch==2.9.1+rocm7.11.0 torchvision==0.24.0+rocm7.11.0 facexlib==0.3.0 einops==0.8.2 safetensors==0.7.0 spandrel==0.4.1 spandrel_extra_arches==0.2.0 --extra-index-url https://download.pytorch.org/whl/rocm6.4 --extra-index-url https://repo.amd.com/rocm/whl/gfx120X-all

I upgraded to 2.9.1 and it's work on Linux. 2.9 have windows version.