app.py

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

REPO_ID = "bartowski/SmolLM2-135M-Instruct-GGUF"
FILENAME = "SmolLM2-135M-Instruct-Q4_K_M.gguf"

@lru_cache()
def load_llm():
    model_path = hf_hub_download(
        repo_id=REPO_ID,
        filename=FILENAME,
        local_dir=".",
        local_dir_use_symlinks=False,
    )
    llm = Llama(
        model_path=model_path,
        n_ctx=512,
        n_threads=max(2, os.cpu_count() or 2),
        n_gpu_layers=0,
        n_batch=32,
        verbose=False,
    )
    return llm

# پاسخ‌های کوتاه‌تر و سریع‌تر
SYSTEM_PROMPT = "به فارسی، واضح و خیلی کوتاه جواب بده (حداکثر ۲ جمله)."

def build_prompt(message, history):
    prompt = f"<s>[SYSTEM]\n{SYSTEM_PROMPT}\n[/SYSTEM]\n"
    for user, assistant in history:
        prompt += f"[USER]\n{user}\n[/USER]\n[ASSISTANT]\n{assistant}\n[/ASSISTANT]\n"
    prompt += f"[USER]\n{message}\n[/USER]\n[ASSISTANT]\n"
    return prompt

def respond(message, history):
    llm = load_llm()
    prompt = build_prompt(message, history)
    # تغییرات سرعت: max_tokens و temperature
    stream = llm.create_completion(
        prompt=prompt,
        max_tokens=80,                      
        temperature=0.5,                
        top_p=0.9,
        stop=["[/ASSISTANT]", "[USER]", "\n[USER]"],
        stream=True,
    )
    partial = ""
    for out in stream:
        token = out["choices"][0]["text"]
        partial += token
        yield partial

demo = gr.ChatInterface(
    fn=respond,
    title="چت‌بات خیلی ساده (CPU رایگان)",
    description="TinyLlama 1.1B (GGUF) روی llama.cpp برای یادگیری راه‌اندازی LLM.",
)

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