import torch
from diffusers import FluxPipeline

# Usamos la variante en precisión FP8, que reduce la huella de memoria a la mitad
pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", 
    torch_dtype=torch.float8_e4m3fn, 
    device_map="balanced"
).to("cuda")

# Cargar el LoRA
pipe.load_lora_weights("joyfox/Kontext-Cosplay-Lora", weight_name="kontext_cosplay_lora_20000.safetensors")