[bug]: main-rocm image being built with wrong version of PyTorch that supports CUDA and not ROCm

[gcolds]

Is there an existing issue for this problem?

• I have searched the existing issues

Operating system

Linux

GPU vendor

AMD (ROCm)

GPU model

Radeon VII

GPU VRAM

16GB

Version number

v5.11.0rc1

Browser

Firefox 128

Python dependencies

No response

What happened

Image generation only occurs with my cpu and not my GPU. From my log of Invoke AI startup I can see the PyTorch device selected is the CPU

invokeai-start.txt

I ran a quick one liner script to determine what version of PyTorch is installed in the docker image and I get the following:

oot@31d21a497724:/opt/invokeai# python3 -c "print('\n\n'); import torch;
print('torch.version:', torch.version);
print('torch.version.cuda:', torch.version.cuda);
print('torch.version.hip:', torch.version.hip);
print('CUDA available:', torch.cuda.is_available());
print('HIP (ROCm) available:', torch.version.hip is not None);"

torch.version: 2.6.0+cu124
torch.version.cuda: 12.4
torch.version.hip: None
CUDA available: False
HIP (ROCm) available: False
root@31d21a497724:/opt/invokeai#

If I'm interpreting things correctly, that output is telling me that the CUDA version of PyTorch is built into the main-rocm image.

What you expected to happen

I expect main-rocm to be built with the version of PyTorch that supports ROCm.

How to reproduce the problem

Simply spin up a docker container using the v5.11.0rc1 main-rocm image.

Additional context

No response

Discord username

No response