[Bug] Unsloth detected signs of an AMD ROCm GPU, but your current PyTorch build has no usable HIP accelerator

[guinmoon]

In Unsloth Studio, starting with package version 2026.4.5.
When loading any model or starting finetune, I've tried different versions of Pytorch with ROCm support, and the same thing happens everywhere.

"Error loading model: Subprocess error: Failed to import ML libraries: Unsloth detected signs of an AMD ROCm GPU, but your current PyTorch build has no usable HIP accelerator.
This usually means torch/torchvision/torchaudio were installed from default PyPI wheels instead of ROCm wheels.\nDetected ROCm version hint: 7.1\nTry reinstalling PyTorch wheels with:
uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.1/ --upgrade --force-reinstall\nVerified index URL is reachable: https://download.pytorch.org/whl/rocm7.1/", "exc_info": true}

My Env:

HIP_VISIBLE_DEVICES="1"
FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE"

With the package unsloth[amd]==2026.4.4, everything works fine.

Moreover, if I simply launch finetune not from the studio but from a python script, everything works fine even with version 2026.4.8.