Cholesky decomposition with Cuda

1

I want to Cholesky decompose a matrix. It works fine on my CPU but when I move the matrix to my GPU the function fails with the following error:

import torch
                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          
cov = torch.eye(2).to("cuda")                                                                                                                                                                                                                                                        
torch.cholesky(cov) 

THCudaCheck FAIL file=/pytorch/aten/src/THC/generic/THCTensorMathPairwise.cu line=196 error=11 : invalid argument
Traceback (most recent call last):
  File "cuda_mvn_test.py", line 35, in <module>
    torch.cholesky(cov)
RuntimeError: cuda runtime error (11) : invalid argument at /pytorch/aten/src/THC/generic/THCTensorMathPairwise.cu:196

My system:

Nvidia driver version: 410.79
Cuda version: 10.0
torch version 1.0.1.post2

I had my mate test it on his system and it worked, so it’s something nasty going on.

2

The issue was that my pip’s torch installation default to Cuda 9. With correct torch package it works fine.

1 Like