Speedup model init on CPU (by 10x+ for llama-3-8B as one example) (#31771)

* 1,100%!

* Clean

* Don't touch DS

* Experiment with dtype allocation

* skip test_load_save_without_tied_weights test

* A little faster

* Include proper upscaling?

* Fixup tests

* Potentially skip?

* Let's see if this fixes git history

* Maintain new dtype

* Fin

* Rm hook idea for now

* New approach, see what breaks

* stage

* Clean

* Stash

* Should be fin now, just need to mark failing models

* Clean up

* Simplify

* Deal with weird models

* Enc/Dec

* Skip w/ reason

* Adjust test

* Fix test

* one more test

* Keep experimenting

* Fix ref

* TO REMOVE: testing feedback CI

* Right push

* Update tests/utils/test_modeling_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* disable

* Add new func

* Test nits from Amy

* Update src/transformers/modeling_utils.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Adjust comment

* Adjust comment on skip

* make private

* Fin

* Should be a not flag

* Clarify and rename test

---------

Co-authored-by: Marc Sun <marc@huggingface.co>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

docs/source/en/main_classes/model.md

@@ -40,6 +40,10 @@ for text generation, [`~generation.GenerationMixin`] (for the PyTorch models),
     - push_to_hub
     - all
 
+Custom models should also include a `_supports_assign_param_buffer`, which determines if superfast init can apply
+on the particular model. Signs that your model needs this are if `test_save_and_load_from_pretrained` fails. If so,
+set this to `False`.
+
 ## ModuleUtilsMixin
 
 [[autodoc]] modeling_utils.ModuleUtilsMixin