Per step time and end to end time gap

[BugFreeee]

Hi,
I saw the reference perf you provided is based on time/step. But I'm wondering about the e2e time.
I ran the same workload on H100-sxm5, at bs4, fp32, 4steps, scale8 (Same as the recomanned setting.)
Per step dur is 0.1647. So 0.1647*4steps = 0.6588s. But if I insert time tag in python, that gives e2e duration as 1.147s. I assume the trailing decoder and clip text encoder would takes up ~50% of entire pipline? Does that time consumpation match your experiment? If so, do you have plan to optmize those non unet component?

The actual code I used:

from diffusers import DiffusionPipeline
import torch
import argparse
import time
import os
pipe = DiffusionPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7", torch_dtype=torch.float32)
parser = argparse.ArgumentParser(description="Generate images using a diffusion model.")
parser.add_argument('--bs', type=int, default=1, help='Batch size for image generation')
args = parser.parse_args()
pipe.to("cuda")

prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
prompt_lst = [prompt] * args.bs

num_inference_steps = 4
output_dir = "generated_images"
os.makedirs(output_dir, exist_ok=True)

dur = 0
warmup = 5
iter = 50

for i in range(warmup):
    images = pipe(prompt=prompt_lst, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images

s_time = time.time()
for i in range(iter):
    images = pipe(prompt=prompt_lst, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
e_time = time.time()
print((e_time-s_time)/iter)
print(args.bs/((e_time-s_time)/iter))```