What is the correct strategy and device for mps

[sneiman]

Trying to get some baby stuff working with mps - but keep getting the error:

RuntimeError: Mismatched Tensor types in NNPack convolutionOutput
This occurs as nn.Conv2d() is passed the data during a forward() call.

The sign on from ptl shows mps selected:

GPU available: True (mps), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs

But the data at entry to forward shows device as -1.

I've tried lots of different settings for strategy and device. Any help appreciated.

[alejandroarmas]

Do you have an update on your situation? I'm experiencing the same thing

[Jonathan-Nyquist]

I'm also having this problem.

[nathom]

The issue for me was mixing CPU tensors with GPU tensors. If x is the input to Conv2d, ensure you call x.to(device) before passing it in if the convolution is running on your GPU.