[aarch64] nn.Linear(20, 1) inference fails #114750
Description
🐛 Describe the bug
Hi,
As the title indicates, I have a model created with TorchScript and saved with the .pt extension under torch==1.13.1 but when I load it back in another env using torch==2.1.1 it fails to infer.
The error is the following:
I'm also having errors when running a pt fail saved with TorchScript that was implemented with torch==1.13.1, but torch==2.1.1 raises the following:
249, in forward
x = self.fc1(x)
x = torch.nn.functional.relu(x)
x = self.fc2(x)
~~~~~~~~ <--- HERE
return x
File "/Users/panagiotisvardanis/miniconda3/envs/mac-pytorch/lib/python3.8/site-packages/torch/nn/modules/linear.py", line 114, in forward
def forward(self, input: Tensor) -> Tensor:
return F.linear(input, self.weight, self.bias)
~~~~~~~~ <--- HERE
RuntimeError: could not create a primitive descriptor for a matmul primitive
Versions
Pip requirements from the environment where the model was created:
[pip3] torch==1.13.1
[pip3] torchvision==0.14.0
Pip requirements from the environment where the model is failing to infer:
[pip3] torch==2.1.1
[pip3] torchvision==0.16.1
cc @ezyang @gchanan @zou3519 @kadeng @EikanWang @jgong5 @wenzhe-nrv @sanchitintel @albanD @mruberry @jbschlosser @walterddr @mikaylagawarecki @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @mingfeima @ashokei @jingxu10 @min-jean-cho @yanbing-j @Guobing-Chen @Xia-Weiwen @malfet @snadampal