[TorchFX] Do not propagate quantizers through __get_item__ when input/output has different numel

[daniil-lyakhov]

🚀 Feature request

Some TorchFX operations are returning a tuple instead of a torch.Tesnor. In case an algorithm requests a statistic from such node nncf will raise an error like tuple has not is_empty method. This affects the yolon11 model, current workaround is the ignored scope with the type of operation which returns a tuple:

nncf.IgnoredScope(types=["__getitem__"])

Solver debug illustration:

The task is to:

• Find a way to automatically detect __getitem__ which have a tuple as an input (or __getitem__ which gets a piece of the input tenor, not the whole tensor) and do not propagate quantizers up through such operations. __getitem__ which input numel is the same as output numel should be marked as quantize-agnostic

Feature Use Case

No response

Are you going to submit a PR?

• Yes I'd like to help by submitting a PR!