import sys
import time
import warnings
from pathlib import Path


# 配置hugface环境
from huggingface_hub import hf_hub_download
import gradio as gr
import os
import glob
import json

# os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
# torch.set_float32_matmul_precision("high")



def instruct_generate(
    img_path: str = " ",
    prompt: str = "What food do lamas eat?",
    input: str = "",
    max_new_tokens: int = 100,
    top_k: int = 200,
    temperature: float = 0.8,
) -> None:
    """Generates a response based on a given instruction and an optional input.
    This script will only work with checkpoints from the instruction-tuned LLaMA-Adapter model.
    See `finetune_adapter.py`.

    Args:
        prompt: The prompt/instruction (Alpaca style).
        adapter_path: Path to the checkpoint with trained adapter weights, which are the output of
            `finetune_adapter.py`.
        input: Optional input (Alpaca style).
        pretrained_path: The path to the checkpoint with pretrained LLaMA weights.
        tokenizer_path: The tokenizer path to load.
        quantize: Whether to quantize the model and using which method:
            ``"llm.int8"``: LLM.int8() mode,
            ``"gptq.int4"``: GPTQ 4-bit mode.
        max_new_tokens: The number of generation steps to take.
        top_k: The number of top most probable tokens to consider in the sampling process.
        temperature: A value controlling the randomness of the sampling process. Higher values result in more random
    """
    output = [prompt, input, max_new_tokens, top_k, temperature]
    print(output)
    return output

# 配置具体参数

example_path = "example.json"
# 1024如果不够, 调整为512
max_seq_len = 1024
max_batch_size = 1

with open(example_path, 'r') as f:
    content = f.read()
    example_dict = json.loads(content)


def create_instruct_demo():
    with gr.Blocks() as instruct_demo:
        with gr.Row():
            with gr.Column():
                scene_img = gr.Image(label='Scene', type='filepath')
                object_list = gr.Textbox(
                    lines=2, label="Input")

                instruction = gr.Textbox(
                    lines=2, label="Instruction")
                max_len = gr.Slider(minimum=1, maximum=512,
                                    value=128, label="Max length")
                with gr.Accordion(label='Advanced options', open=False):
                    temp = gr.Slider(minimum=0, maximum=1,
                                     value=0.8, label="Temperature")
                    top_k = gr.Slider(minimum=100, maximum=300,
                                      value=200, label="Top k")

                run_botton = gr.Button("Run")

            with gr.Column():
                outputs = gr.Textbox(lines=10, label="Output")

        inputs = [instruction, object_list, max_len, top_k, temp]

        # 接下来设定具体的example格式
        examples_img_list = glob.glob("caption_demo/*.png")
        examples = []
        for example_img_one in examples_img_list:
            scene_name = os.path.basename(example_img_one).split(".")[0]
            example_object_list = example_dict[scene_name]["input"]
            example_instruction = example_dict[scene_name]["instruction"]
            example_one = [example_img_one, example_object_list, example_instruction, 512, 0.8, 200]
            examples.append(example_one)

        gr.Examples(
            examples=examples,
            inputs=inputs,
            outputs=outputs,
            fn=instruct_generate,
            cache_examples=os.getenv('SYSTEM') == 'spaces'
        )
        run_botton.click(fn=instruct_generate, inputs=inputs, outputs=outputs)
    return instruct_demo


# Please refer to our [arXiv paper](https://arxiv.org/abs/2303.16199) and [github](https://github.com/ZrrSkywalker/LLaMA-Adapter) for more details.
description = """
# TaPA
The official demo for **Embodied Task Planning with Large Language Models**.
"""

with gr.Blocks(css='style.css') as demo:
    gr.Markdown(description)
    with gr.TabItem("Instruction-Following"):
        create_instruct_demo()

demo.queue(api_open=True, concurrency_count=1).launch()