app.py

import torch
import spaces
import gradio as gr
from threading import Thread
from transformers import (
    AutoModelForCausalLM, 
    AutoTokenizer, 
    BitsAndBytesConfig, 
    TextIteratorStreamer,
    StoppingCriteria,
    StoppingCriteriaList
)

MODEL_ID ="NousResearch/DeepHermes-3-Llama-3-8B-Preview"


# 
# 

DEFAULT_SYSTEM_PROMPT ="""
You are Deepseek R1, an AI engineered for rigorous, transparent reasoning. Your responses must **strictly follow this protocol:**  

1. **THINK FIRST:**  
   - Begin every interaction by generating a raw, unfiltered internal monologue.  
   - Enclose this step-by-step reasoning process—including doubts, methodical evaluations, and logical pivots—between `<think>` and `</think>` tags.  
   - Example: `<think>Analyzing query... Is the user asking for X or Y? Cross-checking definitions... Prioritizing accuracy...</think>`  

2. **ANSWER AFTER:**  
   - Only after completing the `<think>` block, deliver a concise, precise answer enclosed between `<answer>` and `</answer>` tags.  
   - This answer must directly reflect conclusions from your reasoning phase.  

**RULES:**  
- **Tag Compliance:** Omitting or altering `<think>`, `</think>`, `<answer>`, or `</answer>` tags is **prohibited.**  
- **No Shortcuts:** The `<think>` block must detail **every critical step**, even uncertain or exploratory thoughts.  
- **Order Enforcement:** Never output an answer without a preceding `<think>` analysis.  

Failure to adhere to this structure will result in termination."  


"""

CSS = """
.gr-chatbot { min-height: 500px; border-radius: 15px; }
.special-tag { color: #2ecc71; font-weight: 600; }
footer { display: none !important; }
"""

class StopOnTokens(StoppingCriteria):
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        return input_ids[0][-1] == tokenizer.eos_token_id

def initialize_model():
    quantization_config = BitsAndBytesConfig(
        load_in_8bit=True,
        bnb_8bit_compute_dtype=torch.bfloat16,
        bnb_8bit_quant_type="nf4",
        bnb_8bit_use_double_quant=True,
    )

    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
    tokenizer.pad_token = tokenizer.eos_token

    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        device_map="cuda",
        #quantization_config=quantization_config,
        torch_dtype=torch.bfloat16,
        trust_remote_code=True
    )

    return model, tokenizer

def format_response(text):
    return text.replace("[Understand]", '\n<strong class="special-tag">[Understand]</strong>\n') \
              .replace("[/Reason]", '\n<strong class="special-tag">[/Reason]</strong>\n') \
              .replace("[/Answer]", '\n<strong class="special-tag">[/Answer]</strong>\n') \
              .replace("[Reason]", '\n<strong class="special-tag">[Reason]</strong>\n') \
              .replace("[Answer]", '\n<strong class="special-tag">[Answer]</strong>\n')
@spaces.GPU(duration=360)
def generate_response(message, chat_history, system_prompt, temperature, max_tokens):
    # Create conversation history for model
    conversation = [{"role": "system", "content": system_prompt}]
    for user_msg, bot_msg in chat_history:
        conversation.extend([
            {"role": "user", "content": user_msg},
            {"role": "assistant", "content": bot_msg}
        ])
    conversation.append({"role": "user", "content": message})

    # Tokenize input
    input_ids = tokenizer.apply_chat_template(
        conversation,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to(model.device)

    # Setup streaming
    streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)
    generate_kwargs = dict(
        input_ids=input_ids,
        streamer=streamer,
        max_new_tokens=max_tokens,
        temperature=temperature,
        stopping_criteria=StoppingCriteriaList([StopOnTokens()])
    )

    # Start generation thread
    Thread(target=model.generate, kwargs=generate_kwargs).start()

    # Initialize response buffer
    partial_message = ""
    new_history = chat_history + [(message, "")]
    
    # Stream response
    for new_token in streamer:
        partial_message += new_token
        formatted = format_response(partial_message)
        new_history[-1] = (message, formatted + "▌")
        yield new_history

    # Final update without cursor
    new_history[-1] = (message, format_response(partial_message))
    yield new_history

model, tokenizer = initialize_model()

with gr.Blocks(css=CSS, theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    <h1 align="center">🧠 AI Reasoning Assistant</h1>
    <p align="center">Ask me Hard questions</p>
    """)
    
    chatbot = gr.Chatbot(label="Conversation", elem_id="chatbot")
    msg = gr.Textbox(label="Your Question", placeholder="Type your question...")
    
    with gr.Accordion("⚙️ Settings", open=False):
        system_prompt = gr.TextArea(value=DEFAULT_SYSTEM_PROMPT, label="System Instructions")
        temperature = gr.Slider(0, 1, value=0.6, label="Creativity")
        max_tokens = gr.Slider(128, 8192, 2048, label="Max Response Length")

    clear = gr.Button("Clear History")
    
    msg.submit(
        generate_response,
        [msg, chatbot, system_prompt, temperature, max_tokens],
        [chatbot],
        show_progress=True
    )
    clear.click(lambda: None, None, chatbot, queue=False)

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