app.py

import gradio as gr
import numpy as np
import random
from diffusers import DiffusionPipeline
import torch
from huggingface_hub import login
import os
import bitsandbytes as bnb
import onnx
import onnxruntime as ort
from onnxruntime.quantization import quantize_dynamic, QuantType

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

# Set your Hugging Face token
HUGGINGFACE_TOKEN = os.getenv("HUGGINGFACE_TOKEN")
login(token=HUGGINGFACE_TOKEN)

# Path to your model repository and safetensors weights
base_model_repo = "stabilityai/stable-diffusion-3-medium-diffusers"
lora_weights_path = "./pytorch_lora_weights.safetensors"

# Load the base model with 8-bit precision
pipeline = DiffusionPipeline.from_pretrained(
    base_model_repo,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    use_auth_token=HUGGINGFACE_TOKEN
)
bnb.optim.load_int8_model(pipeline.model, device=device)

pipeline.load_lora_weights(lora_weights_path)
pipeline.enable_sequential_cpu_offload()  # Efficient memory usage
pipeline.enable_xformers_memory_efficient_attention()  # Enable xformers memory efficient attention
pipeline = pipeline.to(device)

# Export to ONNX
onnx_model_path = "model.onnx"
pipeline.model.eval()
dummy_input = torch.randn(1, 3, 512, 512, device=device)
torch.onnx.export(pipeline.model, dummy_input, onnx_model_path, export_params=True, opset_version=11, do_constant_folding=True, input_names=['input'], output_names=['output'])

# Quantize ONNX model to 8-bit
quantized_model_path = "model_quantized.onnx"
quantize_dynamic(onnx_model_path, quantized_model_path, weight_type=QuantType.QUInt8)

# Load quantized ONNX model
session = ort.InferenceSession(quantized_model_path)

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 768  # Reduce max image size to fit within memory constraints

def infer(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
        
    generator = torch.Generator(device=device).manual_seed(seed)
    
    image = pipeline(
        prompt=prompt, 
        negative_prompt=negative_prompt,
        guidance_scale=guidance_scale, 
        num_inference_steps=num_inference_steps, 
        width=width, 
        height=height,
        generator=generator
    ).images[0] 
    
    return image

examples = [
    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
    "An astronaut riding a green horse",
    "A delicious ceviche cheesecake slice",
]

css = """
#col-container {
    margin: 0 auto;
    max-width: 520px;
}
"""

if torch.cuda.is_available():
    power_device = "GPU"
else:
    power_device = "CPU"

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(f"""
        # Text-to-Image Gradio Template
        Currently running on {power_device}.
        """)
        
        with gr.Row():
            prompt = gr.Textbox(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt",
                container=False,
            )
            
            run_button = gr.Button("Run", scale=0)
        
        result = gr.Image(label="Result", show_label=False)

        with gr.Accordion("Advanced Settings", open=False):
            negative_prompt = gr.Textbox(
                label="Negative prompt",
                max_lines=1,
                placeholder="Enter a negative prompt",
                visible=True,
            )
            
            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )
            
            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
            
            with gr.Row():
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=512,
                )
                
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=512,
                )
            
            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance scale",
                    minimum=0.0,
                    maximum=10.0,
                    step=0.1,
                    value=7.5,
                )
                
                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=30,
                )
        
        gr.Examples(
            examples=examples,
            inputs=[prompt]
        )

    run_button.click(
        fn=infer,
        inputs=[prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
        outputs=[result]
    )

demo.queue().launch()