Document AMD ROCm install for Windows in README#1559
Conversation
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
| **Windows:** manual install: | ||
|
|
||
| 1. Find your GPU's GFX architecture from [AMD's table](https://github.com/ROCm/TheRock/blob/main/RELEASES.md#supported-python-device--install-extras) (e.g. `gfx1100`). | ||
| 2. Create and activate a venv, then (replace `gfx####`): |
There was a problem hiding this comment.
give the commands. not all users know how to do that
There was a problem hiding this comment.
It should be the same procedure as described in the section on LN52-54. Can just refer to Step 1 thru 4 above
NeuralFault
left a comment
There was a problem hiding this comment.
Just about covers installing on AMD in Windows. Couple things though, the venv creation/activation can just reference the steps given at the beginning of the section to cover on giving the commands to do so.
Also, giving correct full URL path to the bnb whl file, its install command, and 1 (potientially 2) ways to actually set the environment variables as a lot of the less technically inclined users may not know how. Can give example command like I've shown and possibly a blurb about setting a shortcut that sets the variables and launches OT in one go.
|
as long as it's not possible to run AMD without setting 5 environment variables, this is too much for the main README |
Summary
Documents AMD ROCm installation in the README. Linux already works out of the box (the installer auto-detects the GPU and installs the ROCm PyTorch build), but the Windows path was undocumented. Windows needs a manual install because the ROCm wheels are architecture-specific and come from AMD's multi-arch index rather than the PyTorch ROCm index.
The Windows steps are consolidated from community testing (thanks to NeuralFault and others). I have not run the Windows install myself, so the procedure still needs confirmation from an AMD/Windows user — hence draft.
Test plan
OneTrainer has no automated test suite. Manual verification is expected.
pre-commit run --all-filespassesAI assistance
🤖 Generated with Claude Code