🐛 Describe the bug

In the following Python program, x + 2 gets traced as torch.ops.aten.add.Tensor instead of torch.ops.aten.add.Scalar which would be more technically correct (and is the op that torch.jit.script gives). I noticed this because in Torch-MLIR we verify that the operands match the schema exactly. Is it possible to have aot_autograd trace the more technically correct overload? Or can you provide guidance about how to correctly emulate the promotion semantics in appropriate generality?

Also, to add onto that, torch.ops.aten.add.Tensor is missing an argument. I need special-case code in my importer to handle this (by querying the schema for the default value). One possibility could be that PyTorch provides a utility that normalizes the graph of aten ops to "typecheck" correctly against the schemas of the constituent ops including reifying all default values.

from typing import List

import torch
import torch._dynamo as dynamo
from torch._dynamo.optimizations.training import aot_autograd


def my_backend(gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor]):
    gm.print_readable()
    return gm


@dynamo.optimize(aot_autograd(fw_compiler=my_backend))
def f(x):
    return x + 2


example_inputs = (torch.randn(3, 4),)
f(*example_inputs)

Output:

class <lambda>(torch.nn.Module):
    def forward(self, arg0_1: f32[3, 4]):
        # File: /tmp/repro2.py:15, code: return x + 2
        add: f32[3, 4] = torch.ops.aten.add.Tensor(arg0_1, 2);  arg0_1 = None
        return (add,)     

Versions

PyTorch version: 2.0.0.dev20221213+cpu
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Debian GNU/Linux rodete (x86_64)
GCC version: (Debian 12.2.0-3) 12.2.0
Clang version: 14.0.6-2
CMake version: version 3.25.0
Libc version: glibc-2.35

Python version: 3.10.8 (main, Nov 3 2022, 15:17:13) [GCC 12.2.0] (64-bit runtime)
Python platform: Linux-5.19.11-1rodete1-amd64-x86_64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

Versions of relevant libraries:
[pip3] numpy==1.24.0rc2
[pip3] torch==2.0.0.dev20221213+cpu
[pip3] torchvision==0.15.0.dev20221213+cpu
[conda] Could not collect

cc @chauhang @penguinwu @avikchaudhuri @zhxchen17 @tugsbayasgalan @angelayi @suo @ydwu4 @bdhirsh @bobrenjc93 @aorenste @ezyang @anijain2305 @gmagogsfm @zou3519 @msaroufim @wconstab @soumith @ngimel