import os

import gradio as gr
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
import spaces

huggingface_token = os.getenv("HF_TOKEN")

infer_prompt = "فيما يلي تعليمات تصف مهمة. اكتب استجابة تكمل الطلب بشكل مناسب.\n\n### تعليمات:\n{}\n\n### إجابة:\n"
model_id = "nazimali/mistral-7b-v0.3-instruct-arabic"
file_name = "Q4_K_M.gguf"
llm = None


hf_hub_download(
    repo_id=model_id,
    filename=file_name,
    local_dir="./models",
    token=huggingface_token,
)


@spaces.GPU
def respond(
    message,
    history,
):
    global llm
    if llm is None:
        llm = Llama(
            model_path=f"./models/{file_name}",
            flash_attn=True,
            n_gpu_layers=-1,
            n_ctx=2048,
            verbose=True,
        )

    stream = llm.create_chat_completion(
        messages=[{"role": "user", "content": infer_prompt.format(message) }],
        max_tokens=512,
        repeat_penalty=1.2,
        stream=True,
        temperature=0.7,
        top_k=40,
        top_p=0.95,
    )

    outputs = ""
    for output in stream:
        print(output)
        outputs += output["choices"][0]["delta"].get("content", "")
        yield outputs



demo = gr.ChatInterface(respond, examples=["السلام عليكم كيف حالك؟", "hello"], title="Mistral 7B Arabic Fine-tuned")


if __name__ == "__main__":
    demo.launch()