Implement some rethinking about LoRA system

1. Add an option to allow users to use UNet in fp8/gguf but lora in fp16.
2. All FP16 loras do not need patch. Others will only patch again when lora weight change.
3. FP8 unet + fp16 lora are available (somewhat only available) in Forge now. This also solves some “LoRA too subtle” problems.
4. Significantly speed up all gguf models (in Async mode) by using independent thread (CUDA stream) to compute and dequant at the same time, even when low-bit weights are already on GPU.
5. View “online lora” as a module similar to ControlLoRA so that it is moved to GPU together with model when sampling, achieving significant speedup and perfect low VRAM management simultaneously.

backend/sampling/sampling_function.py

@@ -10,6 +10,8 @@ import collections
 from backend import memory_management
 from backend.sampling.condition import Condition, compile_conditions, compile_weighted_conditions
 from backend.operations import cleanup_cache
+from backend.args import dynamic_args
+from backend import utils
 
 
 def get_area_and_mult(conds, x_in, timestep_in):
@@ -353,10 +355,19 @@ def sampling_prepare(unet, x):
         additional_inference_memory += unet.controlnet_linked_list.inference_memory_requirements(unet.model_dtype())
         additional_model_patchers += unet.controlnet_linked_list.get_models()
 
+    if dynamic_args.get('online_lora', False):
+        lora_memory = utils.nested_compute_size(unet.lora_loader.patches)
+        additional_inference_memory += lora_memory
+
     memory_management.load_models_gpu(
         models=[unet] + additional_model_patchers,
         memory_required=unet_inference_memory + additional_inference_memory)
 
+    if dynamic_args.get('online_lora', False):
+        utils.nested_move_to_device(unet.lora_loader.patches, device=unet.current_device)
+
+    unet.lora_loader.patches = {}
+
     real_model = unet.model
 
     percent_to_timestep_function = lambda p: real_model.predictor.percent_to_sigma(p)