Cannot use Pytorch on A100-PCIE-40GB x 8

1

I’m trying to run PyTorch on a server with eight A100-PCIE-40GB GPUs,
but torch.cuda.is_available() returns False.
The server doesn’t have an NVSwitch, so I don’t believe the issue is related to nvidia-fabricmanager.

Running lspci | grep -i nvswitch produces no output.
nvcc --version shows CUDA 12.4, and nvidia-smi appears normal.

image
image749×909 26.3 KB

image
image950×122 6.11 KB

2

Disable MIG or make sure it’s properly configured and the Python process uses the desired MIG slice.