import os
import torch
import gradio as gr
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import InferenceClient

# Environment variables
os.environ["TOKENIZERS_PARALLELISM"] = "0"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
# os.environ["GRADIO_CACHE_DIR"] = "/home/jwy4/gradio_cache"

# Initialize Hugging Face Inference Client
client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")

# Load model and tokenizer (if you want to use a local model, uncomment and use the load_model_and_tokenizer function)
model = None
tokenizer = None

def load_model_and_tokenizer(model_name, dtype, kv_bits):
    global model, tokenizer
    if model is None or tokenizer is None:
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        special_tokens = {"pad_token": "<PAD>"}
        tokenizer.add_special_tokens(special_tokens)

        config = AutoConfig.from_pretrained(model_name)
        if kv_bits != "unquantized":
            quantizer_path = f"codebooks/{model_name.split('/')[-1]}_{kv_bits}bit.xmad"
            setattr(config, "quantizer_path", quantizer_path)

        dtype = torch.__dict__.get(dtype, torch.float32)
        model = AutoModelForCausalLM.from_pretrained(model_name, config=config, torch_dtype=dtype, device_map="auto")

        if len(tokenizer) > model.get_input_embeddings().weight.shape[0]:
            model.resize_token_embeddings(len(tokenizer))

        tokenizer.padding_side = "left"
        model.config.pad_token_id = tokenizer.pad_token_id

    return model, tokenizer

def respond(message, history, system_message, max_tokens, temperature, top_p):
    messages = [{"role": "system", "content": system_message}]
    for val in history:
        if val[0]:
            messages.append({"role": "user", "content": val[0]})
        if val[1]:
            messages.append({"role": "assistant", "content": val[1]})
    messages.append({"role": "user", "content": message})

    response = ""
    for message in client.chat_completion(
        messages,
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
    ):
        token = message.choices[0].delta.content
        response += token
        yield response

# Initialize Gradio ChatInterface
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"),
    ],
    theme="default",
    title="1bit llama3 by xMAD.ai",
    description="The first industrial level 1 bit quantization Llama3, we can achieve 800 tokens per second on NVIDIA V100 adn 1200 on NVIDIA A100, 90%% cost down of your cloud hostin cost",
    css=".scrollable { height: 400px; overflow-y: auto; padding: 10px; border: 1px solid #ccc; }"
)

if __name__ == "__main__":
    # Uncomment if using local model loading
    # load_model_and_tokenizer("NousResearch/Meta-Llama-3-8B-Instruct", "fp16", "1")
    demo.launch()