handler.py

from typing import Dict
import torch
from diffusers import DiffusionPipeline
from compel import Compel
from io import BytesIO
import base64

class EndpointHandler:
    def __init__(self, path: str = ""):
        print(f"Initializing model from: {path}")
        self.pipe = DiffusionPipeline.from_pretrained(
            "black-forest-labs/FLUX.1-dev",
            torch_dtype=torch.float16,
            use_auth_token=True  # Required for gated base model
        )

        # Load LoRA weights from your Hugging Face repo
        print("Loading LoRA weights from: Texttra/Cityscape_Studio")
        self.pipe.load_lora_weights("Texttra/Cityscape_Studio", weight_name="c1t3_v1.safetensors")

        # Send to GPU if available
        if torch.cuda.is_available():
            self.pipe.to("cuda")
        else:
            self.pipe.to("cpu")

        self.pipe.enable_model_cpu_offload()

        # Initialize Compel for prompt conditioning
        self.compel = Compel(
            tokenizer=self.pipe.tokenizer,
            text_encoder=self.pipe.text_encoder
        )
        print("Model initialized successfully.")

    def __call__(self, data: Dict) -> Dict:
        print("Received data:", data)

        inputs = data.get("inputs", {})
        prompt = inputs.get("prompt", "")
        print("Extracted prompt:", prompt)

        if not prompt:
            return {"error": "No prompt provided"}

        # Generate both prompt and pooled embeddings
        conditioning, pooled = self.compel(prompt, return_pooled=True)
        print("Conditioning complete.")

        # Run the model
        image = self.pipe(
            prompt_embeds=conditioning,
            pooled_prompt_embeds=pooled
        ).images[0]
        print("Image generated.")

        # Encode image to base64
        buffer = BytesIO()
        image.save(buffer, format="PNG")
        base64_image = base64.b64encode(buffer.getvalue()).decode("utf-8")
        print("Returning image.")

        return {"image": base64_image}