[LoRA] allow loras to be loaded with low_cpu_mem_usage. (#9510)

* allow loras to be loaded with low_cpu_mem_usage.

* add flux support but note https://github.com/huggingface/diffusers/pull/9510\#issuecomment-2378316687

* low_cpu_mem_usage.

* fix-copies

* fix-copies again

* tests

* _LOW_CPU_MEM_USAGE_DEFAULT_LORA

* _peft_version default.

* version checks.

* version check.

* version check.

* version check.

* require peft 0.13.1.

* explicitly specify low_cpu_mem_usage=False.

* docs.

* transformers version 4.45.2.

* update

* fix

* empty

* better name initialize_dummy_state_dict.

* doc todos.

* Apply suggestions from code review

Co-authored-by: Steven Liu <[email protected]>

* style

* fix-copies

---------

Co-authored-by: Steven Liu <[email protected]>

Author: sayakpaul

docs/source/en/tutorials/using_peft_for_inference.md

@@ -75,6 +75,12 @@ image
 
 ![pixel-art](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_12_1.png)
 
+<Tip>
+
+By default, if the most up-to-date versions of PEFT and Transformers are detected, `low_cpu_mem_usage` is set to `True` to speed up the loading time of LoRA checkpoints. 
+
+</Tip>
+
 ## Merge adapters
 
 You can also merge different adapter checkpoints for inference to blend their styles together.