lora_mem_usage_512_768.py

import torch
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler

def create_pipeline():
    pipe = StableDiffusionPipeline.from_pretrained(
        "gsdf/Counterfeit-V2.5", torch_dtype=torch.float16, safety_checker=None
    ).to("cuda")
    pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True)
    pipe.enable_xformers_memory_efficient_attention()
    return pipe

def render(pipe):
    return pipe(
        prompt=prompt,
        width=512,
        height=768,
        num_inference_steps=15,
        num_images_per_prompt=4,
        generator=torch.manual_seed(0),
    ).images

if __name__ == "__main__":

    prompt = "A photo of sks dog in a bucket"

    torch.cuda.reset_peak_memory_stats()
    pipe = create_pipeline()
    render(pipe)
    mem_bytes = torch.cuda.max_memory_allocated()
    print(f"without lora -> {mem_bytes/(10**6)}MB")

    del pipe

    torch.cuda.reset_peak_memory_stats()
    pipe = create_pipeline()
    pipe.load_lora_weights('takuma104/lora_unetonly_rank4')
    render(pipe)
    mem_bytes = torch.cuda.max_memory_allocated()
    print(f"rank4 -> {mem_bytes/(10**6)}MB")

    del pipe

    torch.cuda.reset_peak_memory_stats()
    pipe = create_pipeline()
    pipe.load_lora_weights('takuma104/lora_unetonly_rank128')
    render(pipe)
    mem_bytes = torch.cuda.max_memory_allocated()
    print(f"rank128 -> {mem_bytes/(10**6)}MB")

    del pipe

    torch.cuda.reset_peak_memory_stats()
    pipe = create_pipeline()
    pipe.load_lora_weights("../stable-diffusion-study/models/lora/light_and_shadow.safetensors")
    render(pipe)
    mem_bytes = torch.cuda.max_memory_allocated()
    print(f"light_and_shadow -> {mem_bytes/(10**6)}MB")