Optional get_example_inputs method of model

[Xia-Weiwen]

This PR allows users to write an optional get_example_inputs method of the wrapper model so that it generates desired input tensors. It is useful when the inputs are complex and cannot be simply generated by random numbers and a common dtype.
If this method is not provided, Xe-Forge still generates random inputs as before.
Users should remove this method if it is not needed.

[danielfleischer]

Thank you for the PR! It is a good idea to add. Two things we'd like to see:

• Make sure it's callable

method = getattr(obj, "greet", None)
if callable(method):
      result = method()

• Please add a sentence in the writing-the-model-class section in the README about custom initialization and mention this method.

[Xia-Weiwen]

Thank you for the PR! It is a good idea to add. Two things we'd like to see:

• Make sure it's callable

method = getattr(obj, "greet", None)
if callable(method):
      result = method()

• Please add a sentence in the writing-the-model-class section in the README about custom initialization and mention this method.

Thanks. I have updated per your comments. Please review again.

[Xia-Weiwen]

I also fixed an issue in spec_loader.get_flop. It substitutes dimension values into formula without checking if the name of one variable is the prefix of another, like B (batch) and BK (block K). Now it traverse the dict from the longest key string to shortest.

[danielfleischer]

LGTM, thank you!

[mzweilin]

Shall we adopt the same formatting style used in the KernelBench dataset?

https://github.com/ScalingIntelligence/KernelBench/blob/423217d9fda91e0c2d67e4a43bf62f96f6d104f1/KernelBench/level1/100_HingeLoss.py#L21

[danielfleischer]

Shall we adopt the same formatting style used in the KernelBench dataset?

Are you saying this PR functionality is the same as what get_inputs does or you're talking about some other formatting issue?

[mzweilin]

Shall we adopt the same formatting style used in the KernelBench dataset?

Are you saying this PR functionality is the same as what get_inputs does or you're talking about some other formatting issue?

It looks like get_example_inputs() serves the same purpose as get_inputs() does in the KernelBench.

BTW, can we make the spec file optional when the KernelBench problem format is given? I think every spec we need is already in the .py file.