Using ComfyUI with ROCm 10

ComfyUI's official AMD docs recommend using ROCm 7.2. However, I tested Comfy a month back with the then new ROCm 7.14 on an AMD Strix Halo (see my howto), and it flew (~18-24% faster depending on the model).

Since ROCm 10.0 has been out for a couple of weeks now, I ran the nightly version (10.1.0a20260823) to see if it's any better.

Spoiler: With one exception, the same or slightly faster. In any case, both v7.14 and v10 are significantly faster than v7.2, so upgrade.

¶Why nightly though?

Because the stable ROCm 10.0.x wheels carry the same idle CPU busy-spin bug as 7.14. The fix has landed in AMD's nightly 10.1.0a20260823 build, so nightly it is until a 10.0.x or 10.1.x stable release fixes it in the future.

Now let's set it up.

This guide assumes you are starting from scratch. If you already have ComfyUI set up, skip to creating the venv.

¶Get the appropriate Python version

ComfyUI devs recommend Python 3.12, 3.13 or 3.14 with a disclaimer that some custom nodes might not work properly with Python 3.14.

I use Comfy with Python 3.14 and all of my custom nodes work just fine. If something doesn't work on your system, try Python 3.12 or 3.13.

¶Find your GPU architecture

Run rocm-smi --showproductname, match the GFX Version to the table below, and note it down. You will need this while installing PyTorch:

¶Install ComfyUI

¶Clone the ComfyUI repo and cd into it

git clone https://github.com/Comfy-Org/ComfyUI
cd ComfyUI

¶Create the venv

Inside the ComfyUI folder, run:

python3 -m venv venv-rocm10
source venv-rocm10/bin/activate
pip install -U pip wheel

If you already have a venv set up, this will create a parallel environment with ROCm 10 PyTorch. This way, you can switch between the two environments if needed.

¶Install ROCm 10 PyTorch

Important: device-gfx1151 will work only with Ryzen AI Max+ 395 / 390 (Strix Halo) systems. Replace it with your GPU's tag.

pip install \
  --extra-index-url https://nightly.repo.amd.com/rocm/pytorch/whl-next/ \
  --extra-index-url https://nightly.repo.amd.com/rocm/whl-next/ \
  "torch[device-gfx1151]==2.13.0+rocm10.1.0a20260823" \
  "torchvision[device-gfx1151]==0.28.0+rocm10.1.0a20260823" \
  "torchaudio==2.11.0.2+rocm10.1.0a20260823"

The ROCm version is pinned to a known-good nightly build. Feel free to experiment with new nightlies.

¶Install remaining ComfyUI dependencies

pip install -r requirements.txt
pip check

Verify:

python -c "import torch; print(torch.__version__, torch.cuda.is_available(), torch.cuda.get_device_name(0))"

This should print 2.13.0+rocm10.1.0a20260823 True your-GPU-name.

¶Custom nodes

Some custom nodes will probably whine about unmet dependencies. You can go to each node's folder and run pip install -r requirements.txt, or ask your agent to start Comfy using your preferred flags, monitor its output, and install the missing deps for each node.

Tell your agent to make sure torch, torchvision, and torchaudio aren't replaced by Nvidia alternatives during installation, or your setup will break.

¶Run ComfyUI

We're ready to go. Strix Halo and other high-VRAM machines can use the --highvram flag.

python main.py --bf16-vae --disable-mmap --highvram

For other GPUs (or to conserve VRAM on Strix Halo), use these settings (based on community feedback):

python main.py --bf16-vae --disable-mmap --enable-dynamic-vram --disable-pinned-memory --reserve-vram 1.0

¶Switching between ROCm environments

If you have an existing venv (e.g. ROCm 7.14 from my previous guide), you can switch back to it by deactivating the current one and activating the other:

deactivate
source venv-rocm714/bin/activate
LD_PRELOAD="./libhsa_shim.so /lib64/libhsa-runtime64.so.1.18.0" \
  python main.py --bf16-vae --disable-mmap --highvram

Now run ComfyUI as usual.

¶Benchmark results

Everything runs on a Strix Halo (Ryzen AI Max+ 395) with the --bf16-vae --disable-mmap --highvram flags. Differences under ~1.5 s are just normal gen-to-gen variations.

gen 1 is from a cold start (freshly started ComfyUI process), while gen 2 is warm (the models already in VRAM), which is why first and second gen times differ.

I also slightly varied the prompt between runs (e.g., different hair/glasses for the image gen models, different color-scheme for the Flux edit model).

Basically, ROCm 10 is the same or faster than ROCm 7.14 across the board, except 1.4+ MP images with Anima Aesthetic, which is somewhat slower on ROCm 10.

Weird, but it is what it is.