🚀 The feature, motivation and pitch

Hey 👋 from the Hugging Face Open-Source team,

We're seeing the following issue over and over again across libraries

RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'

or:

RuntimeError: "addmm_impl_cpu_" not implemented for 'Half'

E.g.: https://github.com/runwayml/stable-diffusion/issues/23

The problem here is that a PyTorch model has been converted to fp16 and the user tried to run it on CPU, e.g. the following:

from torch import nn
import torch

linear = nn.Linear(2,2, dtype=torch.float16)

tensor = torch.ones((2,), dtype=torch.float16)

linear(tensor)

yields:

"addmm_impl_cpu_" not implemented for 'Half'

Could we maybe catch such errors in the forward of https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module

and return a simpler error message that just says "Float16 cannot be run on CPU"?

Alternatives

No response

Additional context

No response

cc @malfet