import os
import spaces
import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

device = "cuda" if torch.cuda.is_available() else "cpu"
model_id = "Qwen/Qwen2.5-Coder-1.5B-Instruct"
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

@spaces.GPU(duration=30)
def infer(message: str, sysprompt: str, tokens: int=30):
    messages = [
        {"role": "system", "content": sysprompt},
        {"role": "user", "content": message}
    ]
    
    input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(text=[input_text], return_tensors="pt").to(model.device)
    generated_ids = model.generate(**inputs, max_new_tokens=tokens)
    generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, generated_ids)]
    output_str = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

    print(message)
    print(output_str)

    return output_str

with gr.Blocks() as demo:
    with gr.Row():
        message = gr.Textbox(label="Message", value="", lines=1)
        sysprompt = gr.Textbox(label="System prompt", value="You are Qwen, created by Alibaba Cloud. You are a helpful assistant.", lines=4)
        tokens = gr.Slider(label="Max tokens", value=30, minimum=1, maximum=2048, step=1)
        #image_url = gr.Textbox(label="Image URL", value=url, lines=1)
    run_button = gr.Button("Run", variant="primary")
    info_md = gr.Markdown("<br><br><br>")

    run_button.click(infer, [message, sysprompt, tokens], [info_md])

demo.launch()