fix(ui): enable rocm-smi support by correcting flags and parsing

[Soddentrough]

ai-toolkit runs on systems with AMD GPUs but displays an error about 'nvidia-smi' in the dashboard when doing so.

This patch removes the hard-coded dependency on 'nvidia-smi' allowing ai-toolkit to operate with either 'nvidia-smi' or 'rocm-smi'. It first checks for 'nvidia-smi' and then checks for 'rocm-smi' which may cause an issue if both are installed but it solves a need today.

[dkspwndj]

Thank you for proceeding with the modification!! But it's not running properly.. I look forward to seeing good results in the future!!

[Soddentrough]

Thinking this might be a path issue (where is your rocm-smi installed?) I will update the logic to check for rocm-smi in this order:

1. which rocm-smi
   2, Check for $ROCM_PATH/bin/rocm-smi
2. Check for /usr/bin/rocm-smi
3. Check for /opt/rocm/bin/rocm-smi

[Soddentrough]

There was a parsing issue when handling the JSON output of rocm-smi creating a phantom device.

[dkspwndj]

Forgive me.. I've been laying down the ROCm to the extent that I need the CompyUI with Zluda and then I figured out if the ROCm was properly laid. Now I'll make a separate PyTorch 3.12 folder to lay the ROCm and try it out there..

[dkspwndj]

Now tested ROCm 7.1.1 installed venv.
But.. not work well..

[Soddentrough]

I'm sorry I didn't notice you're testing with a Windows system. rocm-smi is only available on Linux or WSL. Might be able to use hipinfo.exe on Windows to enumerate the devices but I don't think that has dynamic performance statistics for power/utilization/mem, so stats would show "0".

I don't currently have a way of testing this though and I think for Windows maybe using "Get-Counter" for dynamic performance counters could be the way to go.

[Soddentrough]

This now uses amd-smi by default with fallback to rocm-smi. And where amd-smi doesn't fully support a GPU (eg: Strix iGPU) we use the sysfs hwmon metrics. This also allows us to show "VRAM" and "GTT" (shared memory) used by an APU.

[mickabrig7]

Hello and thanks for your PR !
Unfortunately it doesn't seem like it's working on my system running a 9070 XT (RDNA 4, gfx1201):

> ai-toolkit-ui@0.1.0 start
> concurrently --restart-tries -1 --restart-after 1000 -n WORKER,UI "node dist/cron/worker.js" "next start --port 8675"

[WORKER] TOOLKIT_ROOT: /home/mickabrig7/ai-toolkit
[WORKER] Cron worker started with interval: 1000 ms
[UI]    ▲ Next.js 15.1.7
[UI]    - Local:        http://localhost:8675
[UI]    - Network:      http://192.168.0.100:8675
[UI]
[UI]  ✓ Starting...
[UI]  ✓ Ready in 242ms
[UI] (node:8719) Warning: `--localstorage-file` was provided without a valid path
[UI] (Use `node --trace-warnings ...` to show where the warning was created)
[UI] state { loading: true, gpuData: null, error: null, lastUpdated: null }

I'm running CachyOS with the rocm-smi-lib and amdsmi packages, Pytorch is installed in my venv using the following command:

uv pip install --upgrade --index-url https://download.pytorch.org/whl/nightly/rocm7.1 --pre torch torchaudio torchvision

Here's some more information about my setup:

 uname -srv
Linux 6.18.1-2-cachyos #1 SMP PREEMPT_DYNAMIC Sat, 13 Dec 2025 20:30:10 +0000

 which amd-smi
/opt/rocm/bin/amd-smi

 which rocm-smi
/opt/rocm/bin/rocm-smi

[Soddentrough]

The error you are getting should not exist as I replaced it with a vendor neutral "GPUs not detected" style message.

Can you run :

$ git rev-parse HEAD

[mickabrig7]

Ugh sorry, a bit shameful but I was naively cloning your fork without checking the branch and I was just on the default main, nevermind 🫠

[Soddentrough]

As long as I'm not the only one making those sorts of mistakes! :D

[rncz]

Works for me now, thanks a lot!

Training Z Image Turbo right now at around 3.5sec/it. My setup AMD RX7700XT 12 GB VRAM and Ubuntu 24.04 LTS:

What I did:

• Installed ROCm through the AMD docs, so that I have a symlink at /opt/rocm
• amd-smi and rocm-info worked fine on the command line, but I had to make sure to run it as root user. my default user did not have access to amd-smi.
• I compiled the bitsandbytes package from the AMD docs