ROCm PyTorch 2.6.0 GPU Docker support for AMD gfx1150 with Python 3.10 and 3.13 compatibility

[spirittechie]

Hey folks,

After intense debugging and a lot of pain, I’m dropping a working ROCm PyTorch 2.6.0 Docker build that supports AMD gfx1150 GPUs (like the Radeon 890M) entirely with GPU acceleration enabled.

Key points:

Base image: rocm6.4.1_ubuntu22.04_py3.10_pytorch_release_2.6.0

Works cleanly with Python 3.13 out-of-the-box, recognizing gfx1150 correctly

Python 3.10 requires environment spoofing via HSA_OVERRIDE_GFX_VERSION=11.0.0 for GPU support.

All pinned dependencies are locked to avoid sympy and other conflicts.

Run the container with --device=/dev/kfd --device=/dev/dri --group-add video to enable GPU access

Tested: torch.cuda.is_available() returns True inside the container

Tested PyTorch GPU tensor operations without HIP kernel errors

This should save anyone struggling to get ROCm PyTorch working in Docker with this hardware and Python setup. I’m not pushing a commit yet; this is live dev/testing but stable enough to share.

Feel free to ask questions or contribute improvements!