I want to Install a PyTorch wheel built for ROCm 7.0 that matches my Python ABI (cp310) so ComfyUI / AnimateDiff can use my AMD GPU (Radeon RX 9070 XT, gfx1201) for either text to video, or text to image then image to video. What we tried (brief list of failed attempts)
-
Probed the PyTorch ROCm index with curl/grep; index returned HTTP 403 Access Denied so no wheels could be listed or fetched.
-
Let pip try to auto-select a wheel from the ROCm index; pip either found nothing or installed incompatible builds (CUDA or cp312) and/or produced huge terminal output that closed the shell.
-
Created and switched to a Python 3.10 conda env (comfyui-rocm-py310) to match expected ROCm wheel ABI, then attempted pip install from rocm index — index blocking prevented success.
-
Ran mamba/conda installs; solver downloaded ~3GB but installed CUDA-targeted PyTorch (cuda12.4) into the env instead of a ROCm build. Removed those CUDA packages repeatedly.
-
Tried to list cp310/cp311 wheels via scripted parsing; returned nothing because the index is inaccessible.
-
Attempted browser access to the ROCm index from the same machine; browser also failed with Access Denied.
-
Repeated uninstall/cleanup cycles to remove wrong CUDA artifacts; ended with a clean env but still no ROCm wheel available.
Current status (key facts)
-
My GPU is detected by ROCm: AMD Radeon RX 9070 XT — gfx1201, Chip ID 0x7550.
-
comfyui-rocm-py310 exists and is cleaned of previous CUDA/pytorch packages.
-
The PyTorch ROCm index (https://download.pytorch.org/whl/rocm7.0/) is blocked (HTTP 403) from your network, preventing pip from fetching +rocm wheels automatically.