MPSNDArray Error: buffer is not large enough

mps
1

The code below exits with this error:
/AppleInternal/Library/BuildRoots/.../Library/Caches/com.apple.xbs/Sources/MetalPerformanceShaders/MPSCore/Types/MPSNDArray.mm:782: failed assertion `[MPSNDArray, initWithBuffer:descriptor:] Error: buffer is not large enough. Must be 1024 bytes

import torch
mps = torch.device("mps")

size = 16
A = torch.rand(size, device=mps)
F = torch.rand(size, size, size, device=mps)
print(A@F)

What happened? Thanks a lot in advance.

2

That’s a bug that the PyTorch team needs to fix. Metal throws a fatal error if the descriptor doesn’t match the allocated buffer’s size. @albanD does the backend compile with Metal API validation on? I can’t remember whether that would trigger such an assertion failure, or if MPS’s validation mechanisms are strictly independent of the underlying Metal framework.

3

Hi,
Thanks for the report. That does look like a bug on our end!
Could you open an issue on github with this code sample so that we can easily track the fix for this?

4
5

Many thanks for creating the issue.

6

I have the error.
/AppleInternal/Library/BuildRoots/8d3bda53-8d9c-11ec-abd7-fa6a1964e34e/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShaders/MPSCore/Types/MPSNDArray.mm:782: failed assertion `[MPSNDArray, initWithBuffer:descriptor:] Error: buffer is not large enough. Must be 32768 bytes
’
I use Macbook Pro M1 (Monterey), PyCharm, homebrew, miniforge3. It starts after the last line of code in script. Everything else runs as it should.

Last line of code:
train(bert_classifier, train_dataloader, val_dataloader, epochs=2, evaluation=True)

7

From now on, could we keep discussion on the GitHub issue instead of posting on this thread?

8

Ok, I were not aware of that pytorch used GitHub for issues. Regards Marie