handler.py

import os
#os.environ["CUDA_VISIBLE_DEVICES"] = "6"
from typing import Dict, List, Any
import torch
from torch import autocast
from diffusers import FluxPipeline, AutoPipelineForText2Image, FluxTransformer2DModel, BitsAndBytesConfig
import base64
from io import BytesIO
from huggingface_hub import login

token = os.getenv("HF_TOKEN")
if token is None:
    raise ValueError("The environment variable HF_TOKEN is not set")

login(token=token)

# set device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

if device.type != 'cuda':
    raise ValueError("need to run on GPU")

class EndpointHandler():
    def __init__(self, path=""):
        # First load and merge LoRA weights
        ckpt_id = "black-forest-labs/FLUX.1-dev"
        fused_transformer_path = os.path.join(path, "fused_transformer")
        if not os.path.exists(fused_transformer_path):
            print("Fusing LoRA weights into transformer...")
            pipeline = FluxPipeline.from_pretrained(
                ckpt_id, 
                text_encoder=None, 
                text_encoder_2=None, 
                torch_dtype=torch.float16
            )
            pipeline.load_lora_weights(path, weight_name="pytorch_lora_weights.safetensors")
            pipeline.fuse_lora()
            pipeline.unload_lora_weights()
            pipeline.transformer.save_pretrained(fused_transformer_path)
            print("Finished fusing LoRA weights")

        # Load quantized model
        bnb_4bit_compute_dtype = torch.float16
        nf4_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=bnb_4bit_compute_dtype,
        )
        transformer = FluxTransformer2DModel.from_pretrained(
            os.path.join(path, "fused_transformer"),
            quantization_config=nf4_config,
            torch_dtype=bnb_4bit_compute_dtype,
        )
        self.pipe = AutoPipelineForText2Image.from_pretrained(
            ckpt_id, 
            transformer=transformer, 
            torch_dtype=bnb_4bit_compute_dtype
        )
        self.pipe.enable_model_cpu_offload()

    def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
        """
        Args:
            data (:obj:):
                includes the input data and the parameters for the inference.
        Return:
            A :obj:`dict`:. base64 encoded image
        """
        inputs = data.pop("inputs", data)
        
        # run inference pipeline
        with autocast(device.type):
            image = self.pipe(
                prompt=inputs,
                num_inference_steps=28,
                generator=torch.Generator(device='cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu'),
                width=1024,
                height=1024,
                guidance_scale=3.5,
            ).images[0]  
        
        #os.makedirs('generated_images', exist_ok=True)
        #image_path = f'generated_images/generated_{hash(inputs)}.jpg'
        #image.save(image_path)
        return image

        # encode image as base 64
        #buffered = BytesIO()
        #image.save(buffered, format="JPEG")
        # img_str = base64.b64encode(buffered.getvalue())

        # postprocess the prediction
        #return {"image": img_str.decode()}