fix: update requirements-rocm.txt for Windows support via AMD TheRock nightly

[titansonico10]

Summary

The current requirements-rocm.txt specifies torch==2.9.1+rocm6.3 which only has wheels for manylinux_2_28_x86_64 (Linux). This makes installation completely impossible on Windows, regardless of the install method.

This PR replaces the Linux-only wheels with AMD's TheRock nightly index, which is the only source of ROCm PyTorch wheels for Windows.

Problem

× 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`

Changes

• Replace --extra-index-url https://download.pytorch.org/whl/rocm6.3 + pinned versions with --find-links pointing to AMD's TheRock nightly S3 bucket
• Add commented-out URLs for each RDNA generation (RDNA2/gfx103X, RDNA3/gfx110X, RDNA4/gfx120X) so users can uncomment the one matching their GPU
• Replace onnxruntime==1.23.2 with plain onnxruntime — the ROCm Execution Provider was removed in ORT 1.23 (ROCm 7.0 was the last supported version). See ORT ROCm EP docs
• Use --pre flag to allow nightly/pre-release builds from TheRock

Tested on

• OS: Windows 11
• GPU: AMD RX 9060 XT (gfx1200, RDNA4)
• ROCm build: 7.13.0a20260318 (TheRock nightly)
• Install method: Stability Matrix (uses uv) and standalone
• Result: OneTrainer launches successfully and UI is fully functional

TheRock nightly index URLs

Architecture	GPU series	Index URL
RDNA2 (gfx103X)	RX 6000	https://therock-nightly-python.s3.us-east-2.amazonaws.com/gfx103X-all/index.html
RDNA3 (gfx110X)	RX 7000	https://therock-nightly-python.s3.us-east-2.amazonaws.com/gfx110X-all/index.html
RDNA4 (gfx120X)	RX 9000	https://therock-nightly-python.s3.us-east-2.amazonaws.com/gfx120X-all/index.html

Notes

• On Linux, the existing --extra-index-url https://download.pytorch.org/whl/rocm6.3 approach still works. A future improvement could auto-detect the OS and use the appropriate index.
• triton is also missing on Windows (no ROCm Windows wheels exist for it) — OneTrainer already handles this gracefully with continueing without triton.
• Closes better AMD support #1381

[titansonico10]

Closing this PR in favor of a narrower Windows-only follow-up.

The current version changes requirements-rocm.txt globally, and reviewer feedback is correct that this could affect Linux ROCm users.

I’ll open a new PR that:

• keeps Linux requirements-rocm.txt unchanged
• adds a separate requirements-rocm-windows.txt
• limits the change to Windows AMD / TheRock nightly only

[O-J1]

Please dont, there is already ongoing modernisation of the installation #1244 because Windows + ROCM installation cannot be automated it requires user intervention. It will be wasted effort on your part.

If you can describe exactly how and AMD user should determine what GPU they have, then exactly what to install, we can create a small helper script in the installer that detects if you are Win+AMD and links to documentation on exactly how to do that

[O-J1]

@titansonico10 oops forgot to ping, please see above message

[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.