Tensor.inverse() fatal error

I’m not able to calculate the inverse of Tensors without throwing Intel MLK fatal error. In general, all other modules I use seem to be working fine.

>>> import torch
>>> a = torch.rand(5, 5)
>>> a.sqrt()

 0.7312  0.9105  0.5279  0.7337  0.8724
 0.6445  0.9235  0.5381  0.9239  0.8973
 0.7807  0.8782  0.3331  0.6407  0.2607
 0.7194  0.9982  0.4491  0.7978  0.7362
 0.7902  0.4041  0.9476  0.5784  0.7240
[torch.FloatTensor of size 5x5]

>>> a.inverse()
Intel MKL FATAL ERROR: Error on loading function mkl_lapack_ps_avx2_sgetrf_small.

CudaFloatTensor has the same problem.

Any ideas?

did you install pytorch from source (instead of binaries)?
If not, how did you install the binaries? through pip wheel or conda?

Pip install, Python 2, cuda 8.

pip install http://download.pytorch.org/whl/cu80/torch-0.1.12.post2-cp27-none-linux_x86_64.whl 

Re-installing with conda resolved this for me.

It seems the issue is with pip package dependencies.

looks like a screw-up on my side when packaging binaries (sorry about that).

I’ve made a note of it and will fix it in the next pip wheels (I didn’t ship the AVX2 extensions of MKL)

1 Like

Cool thanks for checking!