make_simple_dynamo_backend does not work for single argument models.

[ttrouwen-dmatrix]

Description

make_simple_dynamo_backend does not work for single argument models. To support multiple args, it takes in *inputs. In case of a single torch.Tensor input it will unpack it instead of handling it in its entirety.

Specifically for the example below you will get the following error:

`File "/disk1/ttrouwen/software/compiler/build/torch-mlir/python_packages/dmir_compiler/torch_mlir/dynamo.py", line 144, in dynamo_callable
    result = user_callable(*inputs)
TypeError: backend.<locals>.exec() takes 1 positional argument but 64 were given`

Example

import torch
import torch.nn as nn
import functorch.compile
from torch_mlir.dynamo import make_simple_dynamo_backend

@make_simple_dynamo_backend
def backend(gm, sample_inputs):
    assert len(sample_inputs) == 1
    def exec(inp):
        return gm.forward(inp)
    return exec

class M(nn.Module):
    def __init__(self):
        super().__init__()
        self.l1 = nn.Linear(64, 64)
    def forward(self, inp):
        return self.l1(inp)

c = torch.compile(M(), backend=backend)
c(torch.zeros(64, 64))

For the sake of brevity, this example returns gm.forward and does not use torch-mlir in the backend implementation. However, the bug is also present when torch-mlir is used in the backend.

Possible solution

In the file torch_mlir/dynamo.py the dynamo_callable function takes in *inputs. This should only be the case if len(example_inputs) > 1.

[stellaraccident]

I'm not aware of anyone who is actively working on/maintaining these paths. Started a thread on discord as the conversation about what to mark as deprecated has looped a couple of times: https://discord.com/channels/636084430946959380/742573221882364009/1232091021450547221

[stellaraccident]

(but if you see a fix, patches welcome)