app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
import spaces  # Import the spaces library

# Model IDs from Hugging Face Hub (Fixed to 7B and 32B Unsloth)
model_ids = {
    "7B": "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
    "32B-Unsloth": "unsloth/DeepSeek-R1-Distill-Qwen-32B-bnb-4bit", # Unsloth 32B model
}

models = {}  # Keep models as a dictionary, but initially empty
tokenizers = {} # Keep tokenizers as a dictionary, initially empty

# BitsAndBytesConfig for 4-bit quantization (for the 32B model)
bnb_config_4bit = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16, # Or torch.float16 if needed
)


def get_model_and_tokenizer(size): # Function to load model on demand
    if size not in models: # Load only if not already loaded
        model_id = model_ids[size]
        print(f"Loading {size} model: {model_id} on demand")
        tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
        if size == "32B-Unsloth": # Apply 4-bit config for 32B model
            model = AutoModelForCausalLM.from_pretrained(
                model_id,
                quantization_config=bnb_config_4bit,
                torch_dtype=torch.bfloat16, # Or torch.float16 if needed
                device_map='auto',
                trust_remote_code=True
            )
        else: # 7B model - standard loading
            model = AutoModelForCausalLM.from_pretrained(
                model_id,
                torch_dtype=torch.bfloat16, # Or torch.float16 if needed
                device_map='auto',
                trust_remote_code=True
            )
        models[size] = model
        tokenizers[size] = tokenizer
        print(f"Loaded {size} model on demand.")
    return models[size], tokenizers[size]


# Revised Default Prompts (as defined previously - these are still good)
default_prompt_1_5b = """**Code Analysis Task**
As a Senior Code Analyst, analyze this programming problem:

**User Request:**
{user_prompt}

**Relevant Context:**
{context_1_5b}

**Analysis Required:**
1. Briefly break down the problem, including key constraints and edge cases.
2. Suggest 2-3 potential approach options (algorithms/data structures).
3. Recommend ONE primary strategy and briefly justify your choice.
4. Provide a very brief initial pseudocode sketch of the core logic."""


default_prompt_7b = """**Code Implementation Task**
As a Principal Software Engineer, provide production-ready Streamlit/Python code based on this analysis:

**Initial Analysis:**
{response_1_5b}

**Relevant Context:**
{context_7b}

**Code Requirements:**
1.  Generate concise, production-grade Python code for a Streamlit app.
2.  Include necessary imports, UI elements, and basic functionality.
3.  Add comments for clarity.
    """


# --- Shared Memory Implementation --- (Same)
shared_memory = []

def store_in_memory(memory_item):
    shared_memory.append(memory_item)
    print(f"\n[Memory Stored]: {memory_item[:50]}...")

def retrieve_from_memory(query, top_k=2):
    relevant_memories = []
    query_lower = query.lower()
    for memory_item in shared_memory:
        if query_lower in memory_item.lower():
            relevant_memories.append(memory_item)

    if not relevant_memories:
        print("\n[Memory Retrieval]: No relevant memories found.")
        return []

    print(f"\n[Memory Retrieval]: Found {len(relevant_memories)} relevant memories.")
    return relevant_memories[:top_k]


# --- Swarm Agent Function - Fixed Models (7B and 32B Unsloth) ---
@spaces.GPU  # <----  GPU DECORATOR ADDED HERE!
def swarm_agent_sequential_rag(user_prompt, prompt_1_5b_template, prompt_7b_template, temperature=0.5, top_p=0.9, max_new_tokens=300): # Removed final_model_size
    global shared_memory
    shared_memory = [] # Clear memory for each new request

    print(f"\n--- Swarm Agent Processing with Shared Memory (RAG) - GPU ACCELERATED - Final Model: 32B Unsloth ---") # Updated message

    # 7B Model - Brainstorming/Initial Draft (Lazy Load and get model)
    print("\n[7B Model - Brainstorming] - GPU Accelerated") # Now 7B is brainstorming
    model_7b, tokenizer_7b = get_model_and_tokenizer("7B") # Lazy load 7B
    retrieved_memory_7b = retrieve_from_memory(user_prompt)
    context_7b = "\n".join([f"- {mem}" for mem in retrieved_memory_7b]) if retrieved_memory_7b else "No relevant context found in memory."

    # Use user-provided prompt template for 7B model (as brainstorming model now)
    prompt_7b_brainstorm = prompt_1_5b_template.format(user_prompt=user_prompt, context_1_5b=context_7b) # Reusing 1.5B template - adjust if needed

    input_ids_7b = tokenizer_7b.encode(prompt_7b_brainstorm, return_tensors="pt").to(model_7b.device)
    output_7b = model_7b.generate(
        input_ids_7b,
        max_new_tokens=max_new_tokens, # Use user-defined max_new_tokens
        temperature=temperature, # Use user-defined temperature
        top_p=top_p,           # Use user-defined top_p
        do_sample=True
    )
    response_7b = tokenizer_7b.decode(output_7b[0], skip_special_tokens=True)
    print(f"7B Response (Brainstorming):\n{response_7b}") # Updated message
    store_in_memory(f"7B Model Initial Response: {response_7b[:200]}...")

    # 32B Unsloth Model - Final Code Generation (Lazy Load and get model)
    final_model, final_tokenizer = get_model_and_tokenizer("32B-Unsloth") # Lazy load 32B Unsloth
    print("\n[32B Unsloth Model - Final Code Generation] - GPU Accelerated") # Model-specific message
    model_stage_name = "32B Unsloth Model - Final Code"
    final_max_new_tokens = max_new_tokens + 200 # More tokens for 32B model

    retrieved_memory_final = retrieve_from_memory(response_7b) # Memory from 7B brainstorm
    context_final = "\n".join([f"- {mem}" for mem in retrieved_memory_final]) if retrieved_memory_final else "No relevant context found in memory."

    # Use user-provided prompt template for final model (using 7B template)
    prompt_final = prompt_7b_template.format(response_1_5b=response_7b, context_7b=context_final) # Using prompt_7b_template for final stage


    input_ids_final = final_tokenizer.encode(prompt_final, return_tensors="pt").to(final_model.device)
    output_final = final_model.generate(
        input_ids_final,
        max_new_tokens=final_max_new_tokens,
        temperature=temperature,
        top_p=top_p,
        do_sample=True
    )
    response_final = final_tokenizer.decode(output_final[0], skip_special_tokens=True)
    print(f"{model_stage_name} Response:\n{response_final}")
    store_in_memory(f"{model_stage_name} Response: {response_final[:200]}...")

    return response_final # Returns final model's response


# --- Gradio ChatInterface --- (No Model Selection Dropdown anymore)
def gradio_interface(message, history, temp, top_p, max_tokens, prompt_1_5b_text, prompt_7b_text): # Removed final_model_selector
    # history is automatically managed by ChatInterface
    response = swarm_agent_sequential_rag(
        message,
        prompt_1_5b_template=prompt_1_5b_text, # Pass prompt templates
        prompt_7b_template=prompt_7b_text,
        temperature=temp,
        top_p=top_p,
        max_new_tokens=int(max_tokens) # Ensure max_tokens is an integer
    )
    return response

iface = gr.ChatInterface( # Using ChatInterface now
    fn=gradio_interface,
    # Define additional inputs for settings and prompts (NO model dropdown)
    additional_inputs=[
        gr.Slider(minimum=0.1, maximum=1.0, step=0.1, value=0.5, label="Temperature"), # Lowered default temp to 0.5
        gr.Slider(minimum=0.01, maximum=1.0, step=0.05, value=0.9, label="Top P"),
        gr.Number(value=300, label="Max Tokens", precision=0), # Use Number for integer tokens
        gr.Textbox(value=default_prompt_1_5b, lines=10, label="Brainstorming Model Prompt Template (7B Model)"), # Updated label - 7B now brainstormer
        gr.Textbox(value=default_prompt_7b, lines=10, label="Code Generation Prompt Template (32B Unsloth Model)"),   # Updated label - 32B is code generator
    ],
    title="DeepSeek Agent Swarm Chat (ZeroGPU Demo - Fixed Models: 7B + 32B Unsloth)", # Updated title
    description="Chat with a DeepSeek agent swarm (7B + 32B Unsloth) with shared memory, adjustable settings, **and customizable prompts!** **GPU accelerated using ZeroGPU!** (Requires Pro Space)", # Updated description
)

if __name__ == "__main__":
    iface.launch() # Only launch locally if running this script directly