import torch
import time
from PIL import Image
import numpy as np
from lyrasd_model import LyraSdXLControlnetTxt2ImgPipeline
import GPUtil

# 存放模型文件的路径，应该包含一下结构：
#   1. clip 模型
#   2. 转换好的优化后的 unet 模型
#   3. 转换好的优化后的 controlnet 模型
#   4. vae 模型
#   5. scheduler 配置
lib_path = "./lyrasd_model/lyrasd_lib/libth_lyrasd_cu12_sm80.so"
model_path = "./models/helloworldSDXL20Fp16"
torch.classes.load_library(lib_path)

# 构建 Txt2Img 的 Pipeline
pipe = LyraSdXLControlnetTxt2ImgPipeline()

start = time.perf_counter()
pipe.reload_pipe(model_path)
print(f"pipeline load cost: {time.perf_counter() - start}")

# load Controlnet 模型，最多load 3个
start = time.perf_counter()
pipe.load_controlnet_model_v2("canny", "./models/controlnet-canny-sdxl-1.0")
print(f"controlnet load cost: {time.perf_counter() - start}")

# 可以通过 get_loaded_controlnet 方法获取目前已经load 好的Controlnet list
print(pipe.get_loaded_controlnet())

# 可以通过unload_controlnet_model 方法unload Controlnet
# pipe.unload_controlnet_model("canny")

control_img = Image.open("control_bird_canny.png")

# 准备应用的输入和超参数
prompt = "a bird"
negative_prompt = ""
height, width = 1024, 1024
steps = 20
guidance_scale = 7.5
generator = torch.Generator().manual_seed(123)
num_images = 1
guess_mode = False

# 可以一次性load 3 个 Controlnets，达到multi Controlnet的效果，这里的参数的长度需要对其
# Controlnet 所输入的img list 长度应该和 controlnet scale 与 Controlnet name 一致，而内部的list长度需要和batch size一致
# 对应的index 可以对其
controlnet_images = [[control_img]]
controlnet_scale = [0.5]
controlnet_names = ['canny']

# 推理生成，返回结果都是生成好的 PIL.Image
for batch in [1]:
    print(f"cur batch: {batch}")
    for _ in range(3):
        start = time.perf_counter()
        images = pipe(prompt=prompt, height=height, width=width, num_inference_steps=steps,
                      guidance_scale=guidance_scale, negative_prompt=negative_prompt, num_images_per_prompt=batch,
                      generator=generator, controlnet_images=controlnet_images,
                      controlnet_scale=controlnet_scale, controlnet_names=controlnet_names,
                      guess_mode=guess_mode
                      )
        print("cur cost: ", time.perf_counter() - start)
        GPUtil.showUtilization(all=True)
# 存储生成的图片
for i, image in enumerate(images):
    image.save(f"./outputs/res_controlnet_sdxl_txt2img_{i}.png")