Implement load_into for mmap data loader

[JacobSzwejbka]

🚀 The feature, motivation and pitch

load_into is used if mutable state has a meaningful initial value. This primarily happens with on-device training but could also happen if some model had read before write state that we serialized an initial value for.

For this task we need to implement this function https://github.com/pytorch/executorch/blob/main/runtime/core/data_loader.h#L111 in https://github.com/pytorch/executorch/blob/main/extension/data_loader/mmap_data_loader.h

example of implementation in filedataloader https://github.com/pytorch/executorch/blob/main/extension/data_loader/file_data_loader.h#L75

Then add a corresponding test to https://github.com/pytorch/executorch/blob/main/extension/data_loader/test/mmap_data_loader_test.cpp

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

cc @larryliu0820 @lucylq

[keyprocedure]

I’d be happy to work on this!

[keyprocedure]

Hi @JacobSzwejbka, I created a PR for this issue. I mirrored the logic for page mapping and mlock from load(). Would you recommend refactoring that code, along with moving the Range struct to the header file?

[JacobSzwejbka]

Taking a look

[JacobSzwejbka]

Would you recommend refactoring that code, along with moving the Range struct to the header file?

I dont think we need to move it to the header. I think it pretty much looks fine. There is not much duplication if we remove the mlock code from load_into.

[keyprocedure]

Okay, awesome. Yeah, mlock was a big portion of the mirrored code.