better AMD support

[titansonico10]

Summary

The current requirements-rocm.txt uses torch==2.9.1+rocm6.3 which only has
wheels for manylinux_2_28_x86_64 (Linux). On Windows with an RDNA4 GPU (gfx1200,
e.g. RX 9070 XT), installation fails completely because no win_amd64 wheel exists
for that ROCm version.

Environment

• GPU: AMD RDNA4 (gfx1200 architecture, e.g. RX 9070 / 9070 XT)
• OS: Windows 11
• ROCm build detected by ComfyUI: rocm 7.13.0a20260318 (TheRock nightly)
• Install method: Stability Matrix (uses uv instead of pip / install.bat)

Error

× No solution found when resolving dependencies:
╰─▶ Because torch==2.9.1+rocm6.3 has no wheels with a matching platform
    tag (e.g., `win_amd64`) and you require torch==2.9.1+rocm6.3, we can
    conclude that your requirements are unsatisfiable.

hint: Wheels are available for `torch` (v2.9.1+rocm6.3) on the following
platform: `manylinux_2_28_x86_64`

Root Cause

AMD does not publish ROCm 6.x wheels for Windows. Windows ROCm support for
RDNA4 (gfx1200) is only available through TheRock nightly builds, hosted at:

https://therock-nightly-python.s3.us-east-2.amazonaws.com/gfx120X-all/index.html

Additionally, onnxruntime with ROCm EP has been removed since ORT 1.23 —
ROCm 7.0 was the last supported version. On ROCm 7.x only the CPU runtime is
available via standard pip.

Suggested fix for requirements-rocm.txt

# AMD ROCm - TheRock Nightly (gfx1200 / RDNA4 - Windows)
# https://github.com/ROCm/TheRock

--find-links https://therock-nightly-python.s3.us-east-2.amazonaws.com/gfx120X-all/index.html
--pre
torch
torchvision
torchaudio

# ROCm EP removed in ORT 1.23+, last supported on ROCm 7.0
onnxruntime

# optimizers
# TODO

Notes

• For other RDNA generations, the bucket URL changes:
  • RDNA2 (gfx103X): …/gfx103X-all/index.html
  • RDNA3 (gfx110X): …/gfx110X-all/index.html
  • RDNA4 (gfx120X): …/gfx120X-all/index.html
• When installing via Stability Matrix, the .bat files are bypassed entirely.
  SM uses uv and reads requirements-rocm.txt directly, so the fix must be
  applied to that file before reinstalling.
• It would be ideal if OneTrainer could auto-detect the gfx architecture and
  select the correct TheRock index URL automatically.

[titansonico10]

Additional finding: triton missing on Windows (also affects Linux)\n\nAfter getting OneTrainer to launch on Windows with the TheRock nightly fix described above, the console shows:\n\n\nNo module named 'triton', continueing without triton\n\n\ntriton is not listed in any of the requirements files (requirements-global.txt, requirements-rocm.txt, requirements-cuda.txt). It fails silently and OneTrainer continues without it, but it should ideally be listed as an optional dependency or at least documented.\n\n### Notes\n- triton for ROCm on Windows is not available via standard pip — it requires the ROCm/triton fork\n- This reportedly also affects Linux installs\n- The missing training_concepts/concepts.json and training_samples/samples.json warnings on first launch are expected (no config loaded yet) and not an error\n\n### Current workaround\nNo workaround needed for basic functionality — OneTrainer launches and the UI is fully accessible without triton. Training may be slower without it on some operations.\n

[dxqb]

also @Joydn90 regarding #1379 :
whenever we upgrade dependencies, we can only speculate what works for AMD because there currently isn't a contributor who has AMD hardware. see here for last time: #1189 (comment)

if you want OneTrainer to work with AMD, please consider creating a PR that makes these changes and ask others to test it. it's best to join Discord #dev-general for that

[titansonico10]

Thanks @O-J1 for the heads up about #1244. I wasn't aware of the ongoing pixi-based installation work.

Here's the exact info you asked for, in case it's useful for the helper script or documentation:

GPU detection (Windows):
wmic path win32_videocontroller get name or rocminfo

Install PyTorch per architecture (Windows):

Architecture GPU series Command
Architecture GPU series Command
RDNA3 RX 7000 pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ --pre torch torchaudio torchvision
RDNA4 RX 9000 pip install --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/ --pre torch torchaudio torchvision
Alternatively, AMD provides stable ROCm 7.2 wheels directly (Python 3.12 required):

text
pip install --no-cache-dir https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torch-2.9.1%2Brocmsdk20260116-cp312-cp312-win_amd64.whl
https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torchaudio-2.9.1%2Brocmsdk20260116-cp312-cp312-win_amd64.whl `
https://repo.radeon.com/rocm/windows/rocm-rel-7.2/torchvision-0.24.1%2Brocmsdk20260116-cp312-cp312-win_amd64.whl
Additional notes:

onnxruntime should not be pinned ≥ 1.23 for ROCm (EP dropped). Use unpinned.

triton can be skipped on Windows — no ROCm wheels exist, OneTrainer handles it gracefully.

Tested on Windows 11, RX 9060 XT (RDNA4/gfx1200), ROCm 7.13.0a20260318. Happy to help with the helper script or docs if needed.

[titansonico10]

Workaround for Stability Matrix users (tested on Windows 11, RX 9060 XT / RDNA4 / gfx1200)

SM uses uv internally and bypasses install.bat / update.bat completely. The first install attempt will fail — that's expected. Here are the exact steps that worked:

Let SM attempt the install — it will fail (expected)

Open File Explorer and navigate to your OneTrainer venv Scripts folder:
<SM_path>\Packages\OneTrainer\venv\Scripts\

Click the address bar, clear it, type cmd and press Enter

In the terminal, type activate and press Enter to activate the venv

Install PyTorch and ROCM inside venv on onetrainer

https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installrad/windows/install-pytorch.html

Install remaining dependencies:
pip install -r requirements-global.txt

⚠️ Every time SM updates OneTrainer, repeat steps 3–6.

Tested with torch 2.10.0+rocm7.13.0a20260321. Looking for testers with RDNA2/RDNA3 on Windows!

[Joydn90]

You can install Triton by running "pip install triton-windows" in your environment to enable Torch compile. Works with 7900xt. Gives a nice speed boost.