Reasons about inference speed difference with diffusers on single GPU without parrallelism?

[xyyan0123]

Hi, I'm trying to run Stable Diffusion 3 with xDiT on single 4090 via following command. I set all parrallel related parameter to 1 to disable parrallelism. And I got epoch time: 7.96 sec, parameter memory: 17.58 GB, peak memory: 20.16 GB.

torchrun --nproc_per_node=1 examples/sd3_example.py \
--model "stabilityai/stable-diffusion-3-medium-diffusers" \
--pipefusion_parallel_degree 1 --ulysses_degree 1 \
--data_parallel_degree 1 --ring_degree 1 --tensor_parallel_degree 1 \
--num_inference_steps 50 --warmup_steps 0 \
--prompt "A sign that reads Raining Cats and Dogs with a dog smiling and wagging its tail."

However, I found that the inference time only with diffusers is slower than xDiT. I coded as below and got a result of 9.28s. It makes me confused about whether some optimization used in the situation of single gpu and non parrallelism. And I tried some difference inference steps , but there is always a gap around 1.4 seconds. So my question is that, what's reason of this difference?

import torch
from diffusers import StableDiffusion3Pipeline
import time

pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16)
pipe = pipe.to("cuda")

start_time = time.time()
image = pipe(
    "A sign that reads Raining Cats and Dogs with a dog smiling and wagging its tail.",
    negative_prompt="",
    num_inference_steps=50,
    guidance_scale=7.0,
).images[0]

end_time = time.time()
print(f"inference time: {end_time - start_time:.2f} s")

image.save("sign.png")

[feifeibear]

They should be the same, no quatization and other optimazation is applied. When parallel degree is 1, it use the diffusers code.

[xyyan0123]

Yeah, I know it "should be the same". But i tested on multiple machines with different GPUs. Tha gap is still exists. I don't know why this phenomena happens. Is there something I did wrong? I refactor my scirpt format above. Could you help me?