DLRM model to torch mlir

Incubator torch-mlir
1

I am trying to lower a DLRM model (GitHub - facebookresearch/dlrm: An implementation of a deep learning recommendation model (DLRM)) to torch-mlir, but am running into the issue below:

module = torch_mlir.compile(
model_traced,
(dense_x, lS_o, lS_i),
output_type=“torch”,
# output_type=“linalg-on-tensors”,
use_tracing=False,
verbose=False,
)

code/torch.py:23:23: error: unsupported by backend contract: Unimplemented operator ‘aten.unbind.int’

Just wanted to make sure I didn’t miss anything obvious: has anyone successfully converted a DLRM model () to torch-mlir?

Cheers,
Jim

2

I don’t know if people have gotten DLRM to work, but there is a patch up right now that will add some basic support for aten.unbind.int. I would ask on the Discord if people have been able to run DLRM.

3

Hi, checkout this dlrm