pannous
(Pannous)
May 20, 2022, 8:15am
1
Feature suggestion: enable automatic fallback for layers where mps implementations are not available (yet).
Case in point: I load a random pytorch model, do model.to(“mps”) and upon running the model get an error such as:
‘aten::_slow_conv2d_forward’ is only available for these backends: SparseCPU … …
opened 02:16AM - 20 May 22 UTC
triaged
module: mps
### 🐛 Describe the bug
I was training resnet50 (code from torchvision) on m1 pr… o Macbook.
`NotImplementedError: Could not run 'aten::_slow_conv2d_forward' with arguments from the 'MPS' backend.`
I tried every network from torchvision which consists of convolution operator, same error.
### Versions
# Name Version Build Channel
bzip2 1.0.8 h3422bc3_4 conda-forge
ca-certificates 2022.5.18 h4653dfc_0 conda-forge
certifi 2022.5.18 pypi_0 pypi
charset-normalizer 2.0.12 pypi_0 pypi
cycler 0.11.0 pypi_0 pypi
fonttools 4.33.3 pypi_0 pypi
idna 3.3 pypi_0 pypi
kiwisolver 1.4.2 pypi_0 pypi
libffi 3.4.2 h3422bc3_5 conda-forge
libzlib 1.2.11 h90dfc92_1014 conda-forge
matplotlib 3.5.2 pypi_0 pypi
ncurses 6.3 h07bb92c_1 conda-forge
numpy 1.22.3 pypi_0 pypi
openssl 3.0.3 ha287fd2_0 conda-forge
packaging 21.3 pypi_0 pypi
pillow 9.1.1 pypi_0 pypi
pip 22.1 pyhd8ed1ab_0 conda-forge
pyparsing 3.0.9 pypi_0 pypi
python 3.9.12 h14b404e_1_cpython conda-forge
python-dateutil 2.8.2 pypi_0 pypi
python_abi 3.9 2_cp39 conda-forge
pytorch-ignite 0.4.9 pypi_0 pypi
readline 8.1 hedafd6a_0 conda-forge
requests 2.27.1 pypi_0 pypi
setuptools 62.3.1 py39h2804cbe_0 conda-forge
six 1.16.0 pypi_0 pypi
sqlite 3.38.5 h40dfcc0_0 conda-forge
tk 8.6.12 he1e0b03_0 conda-forge
torch 1.12.0.dev20220519 pypi_0 pypi
torchaudio 0.11.0 pypi_0 pypi
torchvision 0.12.0 pypi_0 pypi
typing-extensions 4.2.0 pypi_0 pypi
tzdata 2022a h191b570_0 conda-forge
urllib3 1.26.9 pypi_0 pypi
wheel 0.37.1 pyhd8ed1ab_0 conda-forge
wrapt 1.14.1 pypi_0 pypi
xz 5.2.5 h642e427_1 conda-forge
zlib 1.2.11 h90dfc92_1014 conda-forge
albanD
(Alban D)
May 20, 2022, 2:21pm
2
Hi!
We added such a feature in master already!
It will be in the next nightly build tomorrow (Saturday May 20th) and is already available if you build from source!
Hey @albanD , is there an ETA on the implementation of this operator? I’m curious how any of the CNN benchmarking was conducted without the presence of this operator? Thanks in advance.
albanD
(Alban D)
August 5, 2022, 6:50pm
4
Hi,
The absence of this operator is linked to some binary incompatibilities with torchvision which confuses some systems.
This should work fine in the upcoming 1.12.1