app.py

import spaces
import torch
import gradio as gr
from diffusers import StableDiffusionPipeline
from torchao.quantization import quantize_, Float8DynamicActivationFloat8WeightConfig
import os

# --- 1. Model Loading and Optimization (AoT Compilation) ---

# Choose a stable diffusion model
MODEL_ID = "runwayml/stable-diffusion-v1-5"

# Initialize pipeline, disable safety checker for faster compilation and inference
# Use torch.float16 for efficiency on CUDA hardware
pipe = StableDiffusionPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    safety_checker=None,
    requires_safety_checker=False
)
pipe.to('cuda')
pipe.scheduler.set_timesteps(50) # Set max steps for consistent performance testing

print("Starting AoT Compilation...")

@spaces.GPU(duration=1500)  # Reserve maximum time for startup compilation
def compile_optimized_unet():
    # 1. Apply FP8 quantization (optional, requires H200/H100 for maximum benefit)
    try:
        quantize_(pipe.unet, Float8DynamicActivationFloat8WeightConfig())
        print("✅ Applied FP8 quantization to UNet.")
    except Exception as e:
        print(f"⚠️ FP8 Quantization failed (may require specific hardware/libraries): {e}")

    # 2. Define and capture example inputs for the UNet (the core engine)
    # Standard Stable Diffusion UNet inputs (batch_size=2 for classifier-free guidance)
    bsz = 2
    latent_model_input = torch.randn(bsz, 4, 64, 64, device="cuda", dtype=torch.float16)
    t = torch.randint(0, 1000, (bsz,), device="cuda')
    encoder_hidden_states = torch.randn(bsz, 77, 768, device="cuda", dtype=torch.float16)
    
    with spaces.aoti_capture(pipe.unet) as call:
        pipe.unet(latent_model_input, t, encoder_hidden_states)
    
    # 3. Export the model
    exported = torch.export.export(
        pipe.unet,
        args=call.args,
        kwargs=call.kwargs,
    )
    
    # 4. Compile the exported model using AoT
    return spaces.aoti_compile(exported)

# Execute compilation during startup
compiled_unet = compile_optimized_unet()
# 5. Apply compiled model to the pipeline's UNet component
spaces.aoti_apply(compiled_unet, pipe.unet)

print("✅ AoT Compilation completed successfully.")

# --- 2. Inference Function (Running on GPU) ---

@spaces.GPU(duration=60) # Standard duration for image generation
def generate_image(
    prompt: str, 
    negative_prompt: str, 
    steps: int, 
    seed: int
):
    if not prompt:
        raise gr.Error("Prompt cannot be empty.")
    
    generator = torch.Generator(device="cuda").manual_seed(seed) if seed != -1 else None
    
    steps = int(steps)
    
    # Run inference using the optimized pipeline
    result = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=steps,
        guidance_scale=7.5,
        generator=generator
    ).images
    
    return result

# --- 3. Gradio Interface ---

with gr.Blocks(title="Optimized Vision Model (AoT Powered)") as demo:
    gr.HTML(
        """
        <div style="text-align: center; max-width: 800px; margin: 0 auto;">
            <h1><a href="https://huggingface.co/spaces/akhaliq/anycoder" target="_blank">Built with anycoder</a></h1>
            <h2>High-Performance Creative VLM Simulator (AoT Optimized)</h2>
            <p>This demo simulates a creative Vision Language Model using AoT-compiled Stable Diffusion for lightning-fast image generation.</p>
        </div>
        """
    )
    
    with gr.Row():
        with gr.Column(scale=1):
            prompt = gr.Textbox(
                label="Prompt (Input to VLM)",
                placeholder="A futuristic city painted by Van Gogh, highly detailed.",
                lines=3
            )
            negative_prompt = gr.Textbox(
                label="Negative Prompt (What to avoid)",
                placeholder="Blurry, bad quality, low resolution",
                lines=2
            )
            
            with gr.Accordion("Generation Settings", open=True):
                steps = gr.Slider(
                    minimum=10, 
                    maximum=50, 
                    step=1, 
                    value=30, 
                    label="Inference Steps (Higher = Slower/Better)"
                )
                seed = gr.Number(
                    value=-1, 
                    label="Seed (-1 for random)"
                )
            
            generate_btn = gr.Button("Generate Image (AoT Fast!)", variant="primary")
            
        with gr.Column(scale=2):
            output_gallery = gr.Gallery(
                label="Creative VLM Output",
                show_label=True,
                height=512,
                columns=2,
                object_fit="contain"
            )

    generate_btn.click(
        fn=generate_image,
        inputs=[prompt, negative_prompt, steps, seed],
        outputs=output_gallery
    )
    
    gr.Examples(
        examples=[
            ["A majestic wolf standing on a snowy mountain peak, cinematic lighting", "ugly, deformed, low detail", 30],
            ["Cyberpunk cat sitting in a neon-lit alley, 8k, digital art", "human, blurry, messy background", 40],
            ["A vintage photograph of a space shuttle launching from a tropical island", "modern, cartoon, painting", 25]
        ],
        inputs=[prompt, negative_prompt, steps],
        outputs=output_gallery,
        fn=generate_image,
        cache_examples=False,
    )

demo.queue()
demo.launch()