Do not check trace for diffusers, saving memory and time for FLUX

[mvafin]

What does this PR do?

This is further optimization of memory consumption of diffusers conversion, continuation of this PR: #1033
When check_trace=True the TorchScript graph is generated second time and is compared with the graph generated first time. It is useful to catch incorrect traced graph sometimes, but in optimum we control which models are supported and such issues shouldn't happen.
Currently only introduce this for diffusers, but can be done for all the models.
The most impact on memory is demonstrated for FLUX, for other diffusers it significantly reduces conversion time.

Model	Before	After
FLUX	~58Gb/~240s	~33Gb/~130s
SD-3.5-medium	~18Gb/~250s	~18Gb/~100s

|

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[AlexKoff88]

Awesome! Thanks @mvafin.

Please check and fix the quality:

black --check .
ruff check .

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[mvafin]

ruff check .

Done

[AlexKoff88]

• @alexsu52 . Shall we have an option that disables this behavior during model export? Such an option can be enabled by default.

[IlyasMoutawwakil]

Not a fan of patches, long term solution would be to decouple the export process by doing torch.jit.trace with the arguments we want then ov.convert_model with the resulting model.

[AlexKoff88]

Not a fan of patches, long term solution would be decouple the export process by doing torch.jit.trace with the arguments we want then ov.convert_model with the resulting model.

I agree. I see that torch.jit.trace in one place in the Optimum-Intel code. @mvafin, why did you just not pass the argument there?

[mvafin]

Not a fan of patches, long term solution would be decouple the export process by doing torch.jit.trace with the arguments we want then ov.convert_model with the resulting model.

I agree. I see that torch.jit.trace in one place in the Optimum-Intel code. @mvafin, why did you just not pass the argument there?

torch.jit.trace is called inside the ov.convert_model, there is no control over it from outside. We could do torch.jit.trace before calling convert_model but that would require to also do the necessary preparations we do inside the convert_model which we will not do for TorchScript module. This approach is not good.
Ideal solution would be to allow optimum-intel to set the flag when calling convert_model and this is what we want to do for 2025.0, but for 2024.6 unfortunately we only can use this patch.

[IlyasMoutawwakil]

okay, we can go with a patch but let's at least make it a clean one, the naming in the PR is a bit confusing and I don't think the patching should be in __main__.py but rather in convert where lower level logic is defined. What about this:

• add an argument check_trace=None to the export function(s) so that users can control this behavior.
• if None, set this value to False in the case of diffusers.
• do the patching around the convert_model function with patch_torch_jit_trace(check_trace=check_trace).