thank you for your job

[xgocn]

I have amd mi50 cards and nvidia 3070 cards on the same workstation，could I use your plan？

[RafalSiwek]

Hi @xgocn! Thanks so much for the kind words — really appreciate it!

Regarding your setup: yes, the proposed solution could work, but it’ll take some extra effort. For the AMD MI50 (GCN 5.1, gfx906), you'll likely need to build your own container images and rebuild PyTorch to match the architecture. The container images I shared were built for gfx1011 cards, which aren’t compatible with yours. I had to compile libraries from source with patches and set --offload-arch=gfx1011. In your case, you’d want to set --offload-arch=gfx906 instead. Feel free to follow the PYTORCH_ROCM_ARCH trail to see where it is used.

Even though the MI50 is listed as deprecated in the ROCm 6.2.2 compatibility matrix, you might still get things working without fully rebuilding everything — definitely worth a shot.

As for your NVIDIA setup: my recipes target the Turing architecture, while your card is from the Blackwell generation. You’ll need to install the matching CUDA libraries and build the UCC library with the appropriate NVCC flags — something like NVCC_GENCODE="-gencode=arch=compute_50,code=sm_50" for your RTX 3070.

Hope that clears the path somehow. You can also follow my troubleshooting journal for my use case, and feel free to reach out if you hit any roadblocks!