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

# --- Choose the GGUF you want from HF and set here (exact names) ---
REPO_ID    = os.getenv("REPO_ID", "TheBloke/airoboros-mistral-7B-GGUF")
MODEL_FILE = os.getenv("MODEL_FILE", "airoboros-mistral-7b.Q4_K_M.gguf")

SYSTEM_PROMPT  = os.getenv("SYSTEM_PROMPT", "You are an uncensored assistant. Answer directly.")
APP_PASSWORD   = os.getenv("APP_PASSWORD", "")  # set in Settings → Variables if you want a password
CTX            = int(os.getenv("CTX", "4096"))
THREADS        = int(os.getenv("THREADS", "4"))
N_GPU_LAYERS   = int(os.getenv("N_GPU_LAYERS", "0"))  # 0 on free CPU Space

# Download model (cached by HF infra)
model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_FILE, local_dir="models")

# Load model
llm = Llama(
    model_path=model_path,
    n_ctx=CTX,
    n_threads=THREADS,
    n_gpu_layers=N_GPU_LAYERS,
    verbose=False
)

def chat_fn(history, user_msg):
    messages = [{"role":"system","content":SYSTEM_PROMPT}]
    for role, content in history:
        messages.append({"role": "user" if role=="user" else "assistant", "content": content})
    messages.append({"role":"user","content":user_msg})

    out = llm.create_chat_completion(
        messages=messages,
        temperature=0.9, top_p=0.92, repeat_penalty=1.05, max_tokens=1024
    )
    reply = out["choices"][0]["message"]["content"]
    history = history + [("user", user_msg), ("assistant", reply)]
    return history, ""

with gr.Blocks(title="Airoboros Mistral 7B (Uncensored)") as demo:
    gr.Markdown("## Airoboros Mistral 7B (Uncensored)\nFree CPU Space is slow. Add a GPU in **Settings → Hardware** for speed.")

    # Simple password gate (optional)
    with gr.Row():
        user = gr.Textbox(label="User (any)")
        pwd  = gr.Textbox(label="Password", type="password")
        enter = gr.Button("Enter")
    gate_info = gr.Markdown(visible=False, value="Access granted. Start chatting below.")
    chat = gr.Chatbot(height=460, visible=False)
    msg  = gr.Textbox(label="Message", visible=False, lines=3, placeholder="Ask anything…")
    send = gr.Button("Send", visible=False)
    state = gr.State([])

    def allow(u, p):
        if not APP_PASSWORD or p == APP_PASSWORD:
            return gr.update(visible=True), gr.update(visible=False), gr.update(visible=True), gr.update(visible=True), gr.update(visible=True)
        return gr.update(value="Wrong password. Try again."), gr.update(), gr.update(), gr.update(), gr.update()

    enter.click(allow, [user, pwd], [gate_info, user, chat, msg, send])

    send.click(chat_fn, [state, msg], [state, msg]).then(lambda s: s, state, chat)

demo.queue().launch()