Error Compiling T5 PyTorch Model

[mariecwhite]

What happened?

When compiling the PyTorch model T5-base, I run into this error:

torch_mlir.compiler_utils.TorchMlirCompilerError: Lowering TorchScript IR -> Torch Backend IR failed with the following diagnostics:
error: 'func.call' op operand type mismatch: expected operand type '!torch.float', but provided '!torch.number' for operand number 0
note: see current operation: %2092 = "func.call"(%1008, %2088, %2089, %2090, %2091) {callee = @__torch_mlir_shape_fn.aten.arange} : (!torch.number, !torch.optional<int>, !torch.optional<int>, !torch.optional<Device>, !torch.optional<bool>) -> !torch.list<int>

Steps to reproduce your issue

Run:

torch-mlir-opt -pass-pipeline='builtin.module(torchscript-module-to-torch-backend-pipeline{backend-legal-ops=torch.aten.flatten.using_ints})' ./t5.mlir

MLIR file: t5.zip

Version information

Torch-MLIR version 20230222.757
Torch version 2.0.0.dev20230212+cu117
IREE version 20230221.437

[mariecwhite]

FYI @monorimet

[ramiro050]

This is an issue with how the shape library handles Scalars. I should have a fix for it in a couple of days

[ramiro050]

Ended up taking less than expected. This should fix the !torch.number isue: #1978. The error I get next is due to there currently being no support for aten.min.other