File size: 1,681 Bytes
70546b5
 
 
 
 
6b6c817
 
70546b5
6b6c817
 
 
 
 
70546b5
 
6b6c817
 
70546b5
 
e38f741
6b6c817
 
70546b5
 
6b6c817
 
 
 
 
 
 
 
 
 
 
 
 
 
 
70546b5
 
 
 
6b6c817
 
70546b5
6b6c817
70546b5
 
6b6c817
70546b5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
import gradio as gr
from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
from diffusers.utils import load_image
import torch

# Clear CUDA cache
torch.cuda.empty_cache()

# Set environment variable for memory fragmentation
import os
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:128'

device = "cuda" if torch.cuda.is_available() else "cpu"

pipes = {
    "txt2img": AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16").to(device),
    "img2img": AutoPipelineForImage2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16").to(device)
}

if device == "cpu":
    pipes["txt2img"].enable_model_cpu_offload()
    pipes["img2img"].enable_model_cpu_offload()

def run(prompt, image):
    try:
        print(f"prompt={prompt}, image={image}")
        if image is None:
            return pipes["txt2img"](prompt=prompt, num_inference_steps=1, guidance_scale=0.0).images[0]
        else:
            image = image.resize((512,512))
            print(f"img2img image={image}")
            return pipes["img2img"](prompt, image=image, num_inference_steps=2, strength=0.5, guidance_scale=0.0).images[0]
    except RuntimeError as e:
        if "CUDA out of memory" in str(e):
            print("CUDA out of memory. Trying to clear cache.")
            torch.cuda.empty_cache()
            # Consider additional fallback strategies here
        else:
            raise e

demo = gr.Interface(
    run,
    inputs=[
        gr.Textbox(label="Prompt"),
        gr.Image(type="pil")
    ],
    outputs=gr.Image(width=512, height=512),
    live=True
)

demo.launch()