app.py

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

# Model configuration
model_name = "HelpingAI/Dhanishtha-2.0-preview"

# Global variables for model and tokenizer
model = None
tokenizer = None

def load_model():
    """Load the model and tokenizer"""
    global model, tokenizer

    print("Loading tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(model_name)

    # Ensure pad token is set
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    print("Loading model...")
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype="auto",
        device_map="auto",
        trust_remote_code=True
    )

    print("Model loaded successfully!")

def format_thinking_text(text):
    """Format text to properly display <think> tags in Gradio with blue border styling like HelpingAI"""
    if not text:
        return text

    # More sophisticated formatting for thinking blocks with blue styling
    formatted_text = text

    # Handle thinking blocks with proper HTML-like styling for Gradio
    thinking_pattern = r'<think>(.*?)</think>'

    def replace_thinking_block(match):
        thinking_content = match.group(1).strip()
        # Use HTML div with inline CSS for blue border styling like HelpingAI
        return f'''

<div style="border-left: 4px solid #4a90e2; background: linear-gradient(135deg, #f0f8ff 0%, #e6f3ff 100%); padding: 16px 20px; margin: 16px 0; border-radius: 12px; font-family: 'Segoe UI', sans-serif; box-shadow: 0 2px 8px rgba(74, 144, 226, 0.15); border: 1px solid rgba(74, 144, 226, 0.2);">
<div style="color: #4a90e2; font-weight: 600; margin-bottom: 10px; display: flex; align-items: center; font-size: 14px;">
<span style="margin-right: 8px;">🧠</span> Think
</div>
<div style="color: #2c3e50; line-height: 1.6; font-size: 14px;">
{thinking_content}
</div>
</div>

'''

    formatted_text = re.sub(thinking_pattern, replace_thinking_block, formatted_text, flags=re.DOTALL)

    # Clean up any remaining raw tags that might not have been caught
    formatted_text = re.sub(r'</?think>', '', formatted_text)

    return formatted_text.strip()

@spaces.GPU()
def generate_response(message, history, max_tokens, temperature, top_p):
    """Generate streaming response without threading"""
    global model, tokenizer

    if model is None or tokenizer is None:
        yield "Model is still loading. Please wait..."
        return

    # Prepare conversation history
    messages = []
    for user_msg, assistant_msg in history:
        messages.append({"role": "user", "content": user_msg})
        if assistant_msg:
            messages.append({"role": "assistant", "content": assistant_msg})

    # Add current message
    messages.append({"role": "user", "content": message})

    # Apply chat template
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

    # Tokenize input
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

    try:
        with torch.no_grad():
            # Use transformers streaming with custom approach
            generated_text = ""
            current_input_ids = model_inputs["input_ids"]
            current_attention_mask = model_inputs["attention_mask"]

            for _ in range(max_tokens):
                # Generate next token
                outputs = model(
                    input_ids=current_input_ids,
                    attention_mask=current_attention_mask,
                    use_cache=True
                )

                # Get logits for the last token
                logits = outputs.logits[0, -1, :]

                # Apply temperature
                if temperature != 1.0:
                    logits = logits / temperature

                # Apply top-p sampling
                if top_p < 1.0:
                    sorted_logits, sorted_indices = torch.sort(logits, descending=True)
                    cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
                    sorted_indices_to_remove = cumulative_probs > top_p
                    sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].clone()
                    sorted_indices_to_remove[0] = 0
                    indices_to_remove = sorted_indices[sorted_indices_to_remove]
                    logits[indices_to_remove] = float('-inf')

                # Sample next token
                probs = torch.softmax(logits, dim=-1)
                next_token = torch.multinomial(probs, num_samples=1)

                # Check for EOS token
                if next_token.item() == tokenizer.eos_token_id:
                    break

                # Decode the new token (preserve special tokens like <think>)
                new_token_text = tokenizer.decode(next_token, skip_special_tokens=False)
                generated_text += new_token_text

                # Format and yield the current text
                formatted_text = format_thinking_text(generated_text)
                yield formatted_text

                # Update inputs for next iteration
                current_input_ids = torch.cat([current_input_ids, next_token.unsqueeze(0)], dim=-1)
                current_attention_mask = torch.cat([current_attention_mask, torch.ones((1, 1), device=model.device)], dim=-1)

    except Exception as e:
        yield f"Error generating response: {str(e)}"
        return

    # Final yield with complete formatted text
    final_text = format_thinking_text(generated_text) if generated_text else "No response generated."
    yield final_text

def chat_interface(message, history, max_tokens, temperature, top_p):
    """Main chat interface with improved streaming"""
    if not message.strip():
        return history, ""

    # Add user message to history
    history.append([message, ""])

    # Generate response with streaming
    for partial_response in generate_response(message, history[:-1], max_tokens, temperature, top_p):
        history[-1][1] = partial_response
        yield history, ""

    return history, ""

# Load model on startup
print("Initializing model...")
load_model()

# Custom CSS for better styling and thinking blocks
custom_css = """
/* Main chatbot styling */
.chatbot {
    font-size: 14px;
    font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
}

/* Enhanced thinking block styling - now handled via inline HTML */
.thinking-block {
    background: linear-gradient(135deg, #f0f8ff 0%, #e6f3ff 100%);
    border-left: 4px solid #4a90e2;
    border-radius: 8px;
    padding: 12px 16px;
    margin: 12px 0;
    font-family: 'Segoe UI', sans-serif;
    box-shadow: 0 2px 4px rgba(0,0,0,0.1);
    position: relative;
}

/* Support for HTML content in chatbot */
.chatbot .message {
    overflow: visible;
}

.chatbot .message div {
    max-width: none;
}

/* Message styling */
.message {
    padding: 10px 14px;
    margin: 6px 0;
    border-radius: 12px;
    line-height: 1.5;
}

.user-message {
    background: linear-gradient(135deg, #e3f2fd 0%, #bbdefb 100%);
    margin-left: 15%;
    border-bottom-right-radius: 4px;
}

.assistant-message {
    background: linear-gradient(135deg, #f5f5f5 0%, #eeeeee 100%);
    margin-right: 15%;
    border-bottom-left-radius: 4px;
}

/* Code block styling */
pre {
    background-color: #f8f9fa;
    border: 1px solid #e9ecef;
    border-radius: 6px;
    padding: 12px;
    overflow-x: auto;
    font-family: 'Consolas', 'Monaco', 'Courier New', monospace;
    font-size: 13px;
    line-height: 1.4;
}

/* Button styling */
.gradio-button {
    border-radius: 8px;
    font-weight: 500;
    transition: all 0.2s ease;
}

.gradio-button:hover {
    transform: translateY(-1px);
    box-shadow: 0 4px 8px rgba(0,0,0,0.15);
}

/* Input styling */
.gradio-textbox {
    border-radius: 8px;
    border: 2px solid #e0e0e0;
    transition: border-color 0.2s ease;
}

.gradio-textbox:focus {
    border-color: #4a90e2;
    box-shadow: 0 0 0 3px rgba(74, 144, 226, 0.1);
}

/* Slider styling */
.gradio-slider {
    margin: 8px 0;
}

/* Examples styling */
.gradio-examples {
    margin-top: 16px;
}

.gradio-examples .gradio-button {
    background: linear-gradient(135deg, #f8f9fa 0%, #e9ecef 100%);
    border: 1px solid #dee2e6;
    color: #495057;
    font-size: 13px;
    padding: 8px 12px;
}

.gradio-examples .gradio-button:hover {
    background: linear-gradient(135deg, #e9ecef 0%, #dee2e6 100%);
    color: #212529;
}
"""

# Create Gradio interface
with gr.Blocks(
    title="🤖 Dhanishtha-2.0-preview Chat",
    theme=gr.themes.Soft(),
    css=custom_css
) as demo:
    gr.Markdown(
        """
        # 🤖 Dhanishtha-2.0-preview Chat

        Chat with the **HelpingAI/Dhanishtha-2.0-preview** model - The world's first LLM designed to think between responses!

        ### ✨ Key Features:
        - 🧠 **Multi-step Reasoning**: Unlike other LLMs that think once, Dhanishtha can think, rethink, self-evaluate, and refine using multiple `<think>` blocks
        - 🔄 **Iterative Thinking**: Watch the model's thought process unfold in real-time
        - 💡 **Enhanced Problem Solving**: Better reasoning capabilities through structured thinking

        **Note**: The `<think>` blocks show the model's internal reasoning process and will be displayed in a formatted way below.
        """
    )
    
    with gr.Row():
        with gr.Column(scale=4):
            chatbot = gr.Chatbot(
                [],
                elem_id="chatbot",
                bubble_full_width=False,
                height=600,
                show_copy_button=True,
                show_share_button=True,
                avatar_images=("👤", "🤖"),
                render_markdown=True,
                sanitize_html=False,  # Allow HTML for thinking blocks
                latex_delimiters=[
                    {"left": "$$", "right": "$$", "display": True},
                    {"left": "$", "right": "$", "display": False}
                ]
            )

            with gr.Row():
                msg = gr.Textbox(
                    container=False,
                    placeholder="Ask me anything! I'll show you my thinking process...",
                    label="Message",
                    autofocus=True,
                    scale=8,
                    lines=1,
                    max_lines=5
                )
                send_btn = gr.Button("🚀 Send", variant="primary", scale=1, size="lg")

        with gr.Column(scale=1, min_width=300):
            gr.Markdown("### ⚙️ Generation Parameters")

            max_tokens = gr.Slider(
                minimum=50,
                maximum=8192,
                value=2048,
                step=50,
                label="🎯 Max Tokens",
                info="Maximum number of tokens to generate"
            )

            temperature = gr.Slider(
                minimum=0.1,
                maximum=2.0,
                value=0.7,
                step=0.1,
                label="🌡️ Temperature",
                info="Higher = more creative, Lower = more focused"
            )

            top_p = gr.Slider(
                minimum=0.1,
                maximum=1.0,
                value=0.9,
                step=0.05,
                label="🎲 Top-p",
                info="Nucleus sampling threshold"
            )

            with gr.Row():
                clear_btn = gr.Button("🗑️ Clear Chat", variant="secondary", scale=1)
                stop_btn = gr.Button("⏹️ Stop", variant="stop", scale=1)

            gr.Markdown("### 📊 Model Info")
            gr.Markdown(
                """
                **Model**: HelpingAI/Dhanishtha-2.0-preview
                **Type**: Reasoning LLM with thinking blocks
                **Features**: Multi-step reasoning, self-evaluation
                """
            )
    
    # Event handlers
    def clear_chat():
        """Clear the chat history"""
        return [], ""

    # Message submission events
    msg.submit(
        chat_interface,
        inputs=[msg, chatbot, max_tokens, temperature, top_p],
        outputs=[chatbot, msg],
        concurrency_limit=1,
        show_progress="minimal"
    )

    send_btn.click(
        chat_interface,
        inputs=[msg, chatbot, max_tokens, temperature, top_p],
        outputs=[chatbot, msg],
        concurrency_limit=1,
        show_progress="minimal"
    )

    # Clear chat event
    clear_btn.click(
        clear_chat,
        outputs=[chatbot, msg],
        show_progress=False
    )

    # Example prompts section
    with gr.Row():
        gr.Examples(
            examples=[
                ["Hello! Can you introduce yourself and show me how you think?"],
                ["Solve this step by step: What is 15% of 240?"],
                ["Explain quantum entanglement in simple terms"],
                ["Write a short Python function to find the factorial of a number"],
                ["What are the pros and cons of renewable energy?"],
                ["Help me understand the difference between AI and machine learning"],
                ["Create a haiku about artificial intelligence"],
                ["Explain why the sky is blue using physics principles"]
            ],
            inputs=msg,
            label="💡 Example Prompts - Try these to see the thinking process!",
            examples_per_page=4
        )

    # Footer with information
    gr.Markdown(
        """
        ---
        ### 🔧 Technical Details
        - **Model**: HelpingAI/Dhanishtha-2.0-preview
        - **Framework**: Transformers + Gradio
        - **Features**: Real-time streaming, thinking process visualization, custom sampling
        - **Reasoning**: Multi-step thinking with `<think>` blocks for transparent AI reasoning

        **Note**: This interface streams responses token by token and formats thinking blocks for better readability.
        The model's internal reasoning process is displayed in formatted code blocks.

        ---
        *Built with ❤️ using Gradio and Transformers*
        """
    )

if __name__ == "__main__":
    demo.queue(
        max_size=20,
        default_concurrency_limit=1
    ).launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False,
        show_error=True,
        quiet=False
    )