app.py

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

# ---------------------------------------------------------
# Updated Model ID for the 60M version
MODEL_ID = "Yangyang1205/MobileLLM-60M" 
# ---------------------------------------------------------

model_loaded = False
load_error_msg = ""

def load_model():
    global model_loaded, load_error_msg
    print(f"🚀 Starting... Loading model: {MODEL_ID}")
    try:
        # 1. Force Config Fix (Tie Weights)
        # Essential for MobileLLM architecture to prevent output gibberish
        config = AutoConfig.from_pretrained(MODEL_ID)
        config.tie_word_embeddings = True 
        
        # 2. Load Model
        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=False)
        model = AutoModelForCausalLM.from_pretrained(
            MODEL_ID, 
            config=config,
            use_safetensors=False, # Kept as requested
            trust_remote_code=True
        )
        
        model = model.to("cpu")
        model.eval()
        model_loaded = True
        return tokenizer, model
    except Exception as e:
        model_loaded = False
        load_error_msg = str(e)
        print(f"Error loading model: {e}")
        return None, None

tokenizer, model = load_model()

# --- Core Generation Function ---
def generate_text(prompt, max_len, temp):
    if not model_loaded:
        return f"Model not loaded: {load_error_msg}"

    try:
        inputs = tokenizer(prompt, return_tensors="pt")

        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=max_len,
                do_sample=False,        # 必须设为 False 才能完全利用 Beam Search 的确定性
                num_beams=3,            # ✅ 加上这行：开启束搜索，寻找 3 条路径
                repetition_penalty=1.0,
                pad_token_id=tokenizer.eos_token_id 
                # 注意：我删掉了 temperature，因为 do_sample=False 时它不起作用
            )

        # Decode full response
        full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
        
        # Clean result: remove the input prompt from the output
        new_text = full_response[len(prompt):]
        
        # Heuristic: Cut off incomplete sentences (stop at the first newline if present)
        if "\n" in new_text.strip():
             lines = [line for line in new_text.split('\n') if line.strip()]
             if lines:
                 return lines[0] 
        
        return new_text
        
    except Exception as e:
        return str(e)

# --- UI Layout (Blocks) ---
# Note: No 'theme' or 'title' arguments in Blocks() to prevent Gradio version errors
with gr.Blocks() as demo:
    
    # Header Section
    gr.Markdown(
        """
        # 📱 MobileLLM 60M Demo
        This is a **60M parameter** base model trained on <1% of FineWeb data.
        It is a base model (not a chat model) and excels at **In-Context Learning**.
        """
    )
    
    # Split Layout
    with gr.Row():
        # Left Column: Input
        with gr.Column():
            input_box = gr.Textbox(
                label="Input Prompt",
                lines=10, 
                placeholder="Enter text patterns here...",
                value="The capital of China is Beijing.\nThe capital of Japan is Tokyo.\nThe capital of Germany is Berlin.\nThe capital of France is"
            )
            
            # Advanced Settings
            with gr.Accordion("⚙️ Advanced Settings", open=False):
                slider_len = gr.Slider(minimum=1, maximum=100, value=20, label="Max Length", step=1)
                slider_temp = gr.Slider(minimum=0.1, maximum=1.0, value=0.6, label="Temperature", step=0.1)
            
            submit_btn = gr.Button("🚀 Generate", variant="primary")

        # Right Column: Output
        with gr.Column():
            output_box = gr.Textbox(
                label="Model Output",
                lines=10, 
                interactive=False 
            )

    # Event Binding
    submit_btn.click(
        fn=generate_text, 
        inputs=[input_box, slider_len, slider_temp], 
        outputs=output_box
    )

    # Examples
    gr.Examples(
        examples=[
            ["The capital of China is Beijing.\nThe capital of Japan is Tokyo.\nThe capital of Germany is Berlin.\nThe capital of France is"],
            ["Artificial Intelligence is a field of computer science that"],
            ["def add(a, b):\n    return a + b\n\ndef multiply(a, b):"],
        ],
        inputs=input_box
    )

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