app.py

import os
import time
import gc
import threading
from itertools import islice
from datetime import datetime
import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from duckduckgo_search import DDGS
import spaces  # Import spaces early to enable ZeroGPU support

# Disable GPU visibility if you wish to force CPU usage outside of GPU functions
# (Not strictly needed for ZeroGPU as the decorator handles allocation)
# os.environ["CUDA_VISIBLE_DEVICES"] = ""

# ------------------------------
# Global Cancellation Event
# ------------------------------
cancel_event = threading.Event()

# ------------------------------
# Torch-Compatible Model Definitions with Adjusted Descriptions
# ------------------------------
MODELS = {
    "Taiwan-tinyllama-v1.0-chat (Q8_0)": {
        "repo_id": "DavidLanz/Taiwan-tinyllama-v1.0-chat",
        "description": "Taiwan-tinyllama-v1.0-chat (Q8_0) – Torch-compatible version converted from GGUF."
    },
    "Llama-3.2-Taiwan-3B-Instruct (Q4_K_M)": {
        "repo_id": "https://huggingface.co/lianghsun/Llama-3.2-Taiwan-3B-Instruct",
        "description": "Llama-3.2-Taiwan-3B-Instruct (Q4_K_M) – Torch-compatible version converted from GGUF."
    },
    "MiniCPM3-4B (Q4_K_M)": {
        "repo_id": "openbmb/MiniCPM3-4B",
        "description": "MiniCPM3-4B (Q4_K_M) – Torch-compatible version converted from GGUF."
    },
    "Qwen2.5-3B-Instruct (Q4_K_M)": {
        "repo_id": "Qwen/Qwen2.5-3B-Instruct",
        "description": "Qwen2.5-3B-Instruct (Q4_K_M) – Torch-compatible version converted from GGUF."
    },
    "Qwen2.5-7B-Instruct (Q2_K)": {
        "repo_id": "Qwen/Qwen2.5-7B-Instruct",
        "description": "Qwen2.5-7B-Instruct (Q2_K) – Torch-compatible version converted from GGUF."
    },
    "Gemma-3-4B-IT (Q4_K_M)": {
        "repo_id": "unsloth/gemma-3-4b-it",
        "description": "Gemma-3-4B-IT (Q4_K_M) – Torch-compatible version converted from GGUF."
    },
    "Phi-4-mini-Instruct (Q4_K_M)": {
        "repo_id": "unsloth/Phi-4-mini-instruct",
        "description": "Phi-4-mini-Instruct (Q4_K_M) – Torch-compatible version converted from GGUF."
    },
    "Meta-Llama-3.1-8B-Instruct (Q2_K)": {
        "repo_id": "MaziyarPanahi/Meta-Llama-3.1-8B-Instruct",
        "description": "Meta-Llama-3.1-8B-Instruct (Q2_K) – Torch-compatible version converted from GGUF."
    },
    "DeepSeek-R1-Distill-Llama-8B (Q2_K)": {
        "repo_id": "unsloth/DeepSeek-R1-Distill-Llama-8B",
        "description": "DeepSeek-R1-Distill-Llama-8B (Q2_K) – Torch-compatible version converted from GGUF."
    },
    "Mistral-7B-Instruct-v0.3 (IQ3_XS)": {
        "repo_id": "MaziyarPanahi/Mistral-7B-Instruct-v0.3",
        "description": "Mistral-7B-Instruct-v0.3 (IQ3_XS) – Torch-compatible version converted from GGUF."
    },
    "Qwen2.5-Coder-7B-Instruct (Q2_K)": {
        "repo_id": "Qwen/Qwen2.5-Coder-7B-Instruct",
        "description": "Qwen2.5-Coder-7B-Instruct (Q2_K) – Torch-compatible version converted from GGUF."
    },
}

LOADED_MODELS = {}
CURRENT_MODEL_NAME = None

# ------------------------------
# Model Loading Helper Function (PyTorch/Transformers)
# ------------------------------
def load_model(model_name):
    global LOADED_MODELS, CURRENT_MODEL_NAME
    if model_name in LOADED_MODELS:
        return LOADED_MODELS[model_name]
    selected_model = MODELS[model_name]
    # Load the model and tokenizer using Transformers.
    model = AutoModelForCausalLM.from_pretrained(selected_model["repo_id"], trust_remote_code=True)
    tokenizer = AutoTokenizer.from_pretrained(selected_model["repo_id"], trust_remote_code=True)
    LOADED_MODELS[model_name] = (model, tokenizer)
    CURRENT_MODEL_NAME = model_name
    return model, tokenizer

# ------------------------------
# Web Search Context Retrieval Function
# ------------------------------
def retrieve_context(query, max_results=6, max_chars_per_result=600):
    try:
        with DDGS() as ddgs:
            results = list(islice(ddgs.text(query, region="wt-wt", safesearch="off", timelimit="y"), max_results))
            context = ""
            for i, result in enumerate(results, start=1):
                title = result.get("title", "No Title")
                snippet = result.get("body", "")[:max_chars_per_result]
                context += f"Result {i}:\nTitle: {title}\nSnippet: {snippet}\n\n"
            return context.strip()
    except Exception:
        return ""

# ------------------------------
# Chat Response Generation with ZeroGPU
# ------------------------------
@spaces.GPU(duration=60)  # This decorator triggers GPU allocation for up to 60 seconds.
def chat_response(user_message, chat_history, system_prompt, enable_search,
                  max_results, max_chars, model_name, max_tokens, temperature, top_k, top_p, repeat_penalty):
    # Reset the cancellation event.
    cancel_event.clear()
    
    # Prepare internal chat history.
    internal_history = list(chat_history) if chat_history else []
    internal_history.append({"role": "user", "content": user_message})
    
    # Retrieve web search context (with debug feedback).
    debug_message = ""
    if enable_search:
        debug_message = "Initiating web search..."
        yield internal_history, debug_message
        search_result = [""]
        def do_search():
            search_result[0] = retrieve_context(user_message, max_results, max_chars)
        search_thread = threading.Thread(target=do_search)
        search_thread.start()
        search_thread.join(timeout=2)
        retrieved_context = search_result[0]
        if retrieved_context:
            debug_message = f"Web search results:\n\n{retrieved_context}"
        else:
            debug_message = "Web search returned no results or timed out."
    else:
        retrieved_context = ""
        debug_message = "Web search disabled."
    
    # Augment the prompt with search context if available.
    if enable_search and retrieved_context:
        augmented_user_input = (
            f"{system_prompt.strip()}\n\n"
            "Use the following recent web search context to help answer the query:\n\n"
            f"{retrieved_context}\n\n"
            f"User Query: {user_message}"
        )
    else:
        augmented_user_input = f"{system_prompt.strip()}\n\nUser Query: {user_message}"
    
    # Append a placeholder for the assistant's response.
    internal_history.append({"role": "assistant", "content": ""})
    
    try:
        # Load the model and tokenizer.
        model, tokenizer = load_model(model_name)
        # Move the model to GPU (using .to('cuda')) inside the GPU-decorated function.
        model = model.to('cuda')
        # Tokenize the augmented prompt and move input tensors to GPU.
        input_ids = tokenizer(augmented_user_input, return_tensors="pt").input_ids.to('cuda')
        
        with torch.no_grad():
            output_ids = model.generate(
                input_ids,
                max_new_tokens=max_tokens,
                temperature=temperature,
                top_k=top_k,
                top_p=top_p,
                repetition_penalty=repeat_penalty,
                do_sample=True
            )
        # Decode the generated tokens.
        generated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
        # Remove the original prompt to isolate the assistant's reply.
        assistant_text = generated_text[len(augmented_user_input):].strip()
        
        # Simulate streaming output by yielding word-by-word.
        words = assistant_text.split()
        assistant_message = ""
        for word in words:
            if cancel_event.is_set():
                assistant_message += "\n\n[Response generation cancelled by user]"
                internal_history[-1]["content"] = assistant_message
                yield internal_history, debug_message
                return
            assistant_message += word + " "
            internal_history[-1]["content"] = assistant_message
            yield internal_history, debug_message
            time.sleep(0.05)  # Short delay to simulate streaming
    except Exception as e:
        internal_history[-1]["content"] = f"Error: {e}"
        yield internal_history, debug_message
    gc.collect()

# ------------------------------
# Cancel Function
# ------------------------------
def cancel_generation():
    cancel_event.set()
    return "Cancellation requested."

# ------------------------------
# Gradio UI Definition
# ------------------------------
with gr.Blocks(title="LLM Inference with ZeroGPU") as demo:
    gr.Markdown("## 🧠 ZeroGPU LLM Inference with Web Search")
    gr.Markdown("Interact with the model. Select your model, set your system prompt, and adjust parameters on the left.")
    
    with gr.Row():
        with gr.Column(scale=3):
            default_model = list(MODELS.keys())[0] if MODELS else "No models available"
            model_dropdown = gr.Dropdown(
                label="Select Model", 
                choices=list(MODELS.keys()) if MODELS else [], 
                value=default_model,
                info="Choose from available models."
            )
            today = datetime.now().strftime('%Y-%m-%d')
            default_prompt = f"You are a helpful assistant. Today is {today}. Please leverage the latest web data when responding to queries."
            system_prompt_text = gr.Textbox(label="System Prompt",
                                            value=default_prompt,
                                            lines=3,
                                            info="Define the base context for the AI's responses.")
            gr.Markdown("### Generation Parameters")
            max_tokens_slider = gr.Slider(label="Max Tokens", minimum=64, maximum=1024, value=1024, step=32,
                                          info="Maximum tokens for the response.")
            temperature_slider = gr.Slider(label="Temperature", minimum=0.1, maximum=2.0, value=0.7, step=0.1,
                                           info="Controls the randomness of the output.")
            top_k_slider = gr.Slider(label="Top-K", minimum=1, maximum=100, value=40, step=1,
                                     info="Limits token candidates to the top-k tokens.")
            top_p_slider = gr.Slider(label="Top-P (Nucleus Sampling)", minimum=0.1, maximum=1.0, value=0.95, step=0.05,
                                     info="Limits token candidates to a cumulative probability threshold.")
            repeat_penalty_slider = gr.Slider(label="Repetition Penalty", minimum=1.0, maximum=2.0, value=1.1, step=0.1,
                                              info="Penalizes token repetition to improve diversity.")
            gr.Markdown("### Web Search Settings")
            enable_search_checkbox = gr.Checkbox(label="Enable Web Search", value=False,
                                                 info="Include recent search context to improve answers.")
            max_results_number = gr.Number(label="Max Search Results", value=6, precision=0,
                                           info="Maximum number of search results to retrieve.")
            max_chars_number = gr.Number(label="Max Chars per Result", value=600, precision=0,
                                         info="Maximum characters to retrieve per search result.")
            clear_button = gr.Button("Clear Chat")
            cancel_button = gr.Button("Cancel Generation")
        with gr.Column(scale=7):
            chatbot = gr.Chatbot(label="Chat", type="messages")
            msg_input = gr.Textbox(label="Your Message", placeholder="Enter your message and press Enter")
            search_debug = gr.Markdown(label="Web Search Debug")
    
    def clear_chat():
        return [], "", ""
    
    clear_button.click(fn=clear_chat, outputs=[chatbot, msg_input, search_debug])
    cancel_button.click(fn=cancel_generation, outputs=search_debug)
    
    # Submission: the chat_response function is now decorated with @spaces.GPU.
    msg_input.submit(
        fn=chat_response,
        inputs=[msg_input, chatbot, system_prompt_text, enable_search_checkbox,
                max_results_number, max_chars_number, model_dropdown,
                max_tokens_slider, temperature_slider, top_k_slider, top_p_slider, repeat_penalty_slider],
        outputs=[chatbot, search_debug],
    )
    
demo.launch()