Monai whole body segmentation unable to run on gpu in colab

[Alisoltan82]

Good day,

Trying to run whole-body segmentation model (Model_lowers.pt) on colab gpu and getting this error:

error message:
NotImplementedError: Could not run 'aten::slow_conv3d_forward' with arguments from the 'CUDA' backend.

device is detected by torch.device = [Tesla 4]
cuda version = 12.1
pytorch installed = 2.1.0+cu121
inference json device line = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

Any advice to resolve this?

[tangy5]

Looks like a CPU-GPU device tensor convert issue. Are you using only GPUs for inference. Any issues/inconsistencies of CPU/GPU input/output conversions?

[Alisoltan82]

I defined the device = torch.device (‘cuda’)
Same in inference.json file

and noticeably, when I change the to “cpu” it works

[vrassouli]

I have the same problem here! Any progress on the issue?

[diazandr3s]

I have the same problem here! Any progress on the issue?

Hi @vrassouli,

I assume this question is about segmenting CT scans, right?
Have you tried starting the server using a terminal following these instrutions?. Once started, you could use the REST APIs available to interact with the server. Is there a reason for not using MONAI Label and 3DSlicer?

Hope this makes sense