In nunif, PyTorch cu128 (CUDA 12.8) is used by default because it is the only version that supports the RTX 50xx series.
However, cu128 has dropped support for GPUs from the Volta (Tesla V100, Titan V), Pascal (GTX 10xx, Tesla P40, Tesla P4), and Maxwell (GeForce 9xx, Titan X) generations.
cu126 still supports those GPUs.
If you want to use cu126, use requirements-torch-cu126.txt instead of requirements-torch.txt.
pip3 install -r requirements-torch-cu126.txt
| PyTorch CUDA Version | Compute Capability |
|---|---|
cu128 |
sm_75, sm_80, sm_86, sm_90, sm_100, sm_120, compute_120 |
cu126 |
sm_50, sm_60, sm_70, sm_75, sm_80, sm_86, sm_90 |