torch-mlir/examples/lazytensor/README.md

1.6 KiB

Future Work for Lazy Tensor Core

In the last part of the section Understand The Metrics Report, it is mentioned that after running the metrics report,

If you see aten:: ops other than nonzero and _local_scalar_dense, that usually means a missing lowering in the accelerator plugin.

Looking at the sample output and the sample output produced by running a ResNet18 model and a MaskRCNN model, respectively, on the Lazy Tensor Core using the TorchScript backend, the following operations are needed and not yet supported by the backend:

  • aten::convolution_overrideable
  • aten::max_pool2d_with_indices
  • aten::mean.out
  • aten::sort
  • aten::arange.start_out
  • aten::bitwise_and.Tensor_out
  • aten::clamp.out
  • aten::exp.out
  • aten::index.Tensor
  • aten::nonzero
  • aten::rsqrt.out
  • aten::sigmoid.out
  • aten::topk.values
  • aten::upsample_nearest2d.out

Note: This list is incomplete because currently the MaskRCNN example crashes halfway through when run on LTC. The output error can also be found in the MaskRCNN sample output.