app.py

import subprocess
import sys
import time
from collections import defaultdict, deque

# Otomatik kurulum

def install_and_import(package):
    try:
        __import__(package)
    except ImportError:
        print(f"{package} is not installed, installing...")
        subprocess.check_call([sys.executable, "-m", "pip", "install", package])

install_and_import("gradio")
install_and_import("transformers")
install_and_import("torch")

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

# === RATE LIMIT ===
click_logs = defaultdict(lambda: {"minute": deque(), "hour": deque(), "day": deque()})
LIMITS = {"minute": (13, 60), "hour": (90, 3600), "day": (1350, 86400)}

def check_rate_limit(session_id):
    now = time.time()
    logs = click_logs[session_id]
    remaining, reset_times = {}, {}
    for key, (limit, interval) in LIMITS.items():
        # Geçmiş istekleri temizle
        while logs[key] and now - logs[key][0] > interval:
            logs[key].popleft()
        used = len(logs[key])
        remaining[key] = max(0, limit - used)
        reset_times[key] = int(interval - (now - logs[key][0]) if logs[key] else interval)
        if used >= limit:
            return False, f"⛔ {key.capitalize()} rate limit exceeded ({limit}/{key})", remaining, reset_times
    # Limit aşılmadıysa log'a şimdi ekle
    for key in LIMITS:
        logs[key].append(now)
    return True, None, remaining, reset_times

# === CHAT ÜRETİM FONKSİYONU ===
def extract_response_between_tokens(text: str) -> str:

    start = "<|im_start|>assistant<|im_sep|>"
    end = "<|im_end|>"
    try:
        return text.split(start)[1].split(end)[0]
    except Exception:
        return text

# Model yükleme
model_name = "Bertug1911/BrtGPT-1-Pre"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

# Özel token ID
im_end_id = tokenizer.convert_tokens_to_ids("<|im_end|>")

# Üretim fonksiyonu chat_generate
def chat_generate(prompt, temperature, top_k, max_new_tokens, session_id):
    ok, msg, rem, resets = check_rate_limit(session_id)
    if not ok:
        return msg, format_status(rem, resets)

    # Jinja chat format
    messages = [{"role": "user", "content": prompt}]
    formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(formatted, return_tensors="pt").to(device)
    gen = inputs["input_ids"]

    # Döngüsel üretim
    for _ in range(int(max_new_tokens)):
        out = model(gen)
        logits = out.logits[:, -1, :] / float(temperature)
        if int(top_k) > 0:
            vals, idxs = torch.topk(logits, int(top_k))
            filt = torch.full_like(logits, float('-inf'))
            filt.scatter_(1, idxs, vals)
            logits = filt
        probs = torch.softmax(logits, dim=-1)
        nxt = torch.multinomial(probs, num_samples=1)
        gen = torch.cat([gen, nxt], dim=1)
        if nxt.item() == im_end_id:
            break

    out_text = tokenizer.decode(gen[0], skip_special_tokens=False)
    # Format düzeltme
    no_sp = out_text.replace(" ", "").replace("Ġ", " ")
    formatted_out = no_sp.replace("Ċ", "\n")
    if not formatted_out.strip().endswith("<|im_end|>"):
        formatted_out += "<|im_end|>"
    resp = extract_response_between_tokens(formatted_out)
    return resp, format_status(rem, resets)

# Durum metni formatlama
def format_status(rem, resets):
    return "\n".join([f"🕒 {k.capitalize()}: {rem[k]} left — resets in {resets[k]} sec" for k in ["minute","hour","day"]])

# === UI ===
with gr.Blocks() as app:
    session_id = gr.State(str(time.time()))
    gr.Markdown("""
    # 🤖 BrtGPT-1-Pre
    """ )

    with gr.Row():
        prompt = gr.Textbox(lines=3, placeholder="Enter your message...", label="Prompt")
        output = gr.Textbox(label="Response")

    with gr.Row():
        temperature = gr.Slider(0.01,1.0,value=0.5,step=0.01,label="Temperature")
        top_k = gr.Slider(1,50,value=10,step=1,label="Top-K")
        max_new_tokens = gr.Slider(1,128,value=15,step=1,label="Max New Tokens")

    generate_button = gr.Button("Generate")
    status = gr.Markdown()

    generate_button.click(
        fn=chat_generate,
        inputs=[prompt, temperature, top_k, max_new_tokens, session_id],
        outputs=[output, status]
    )

app.launch()