app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import time
import os
from huggingface_hub import login

# Hugging Face login
login(token=os.environ["HF_TOKEN"])

print(f"CUDA is available: {torch.cuda.is_available()}")
print(f"CUDA device count: {torch.cuda.device_count()}")
if torch.cuda.is_available():
    print(f"Current CUDA device: {torch.cuda.current_device()}")
    print(f"CUDA device name: {torch.cuda.get_device_name(0)}")

class ConversationManager:
    def __init__(self):
        self.models = {}
        self.conversation = []
        self.delay = 3
        self.is_paused = False
        self.current_model = None
        self.initial_prompt = ""
        self.task_complete = False

    def load_model(self, model_name):
        if not model_name:
            print("Error: Empty model name provided")
            return None
    
        if model_name in self.models:
            return self.models[model_name]

        try:
            print(f"Attempting to load model: {model_name}")
            tokenizer = AutoTokenizer.from_pretrained(model_name)
            # Try to load the model with 8-bit quantization
            model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_8bit=True)
        except RuntimeError as e:
            print(f"8-bit quantization not available, falling back to full precision: {e}")
            if torch.cuda.is_available():
                model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
            else:
                model = AutoModelForCausalLM.from_pretrained(model_name)
        except Exception as e:
            print(f"Failed to load model {model_name}: {e}")
            print(f"Error type: {type(e).__name__}")
            print(f"Error details: {str(e)}")
            return None

        self.models[model_name] = (model, tokenizer)
        print(f"Successfully loaded model: {model_name}")
        return self.models[model_name]

    def generate_response(self, model_name, prompt):
        model, tokenizer = self.load_model(model_name)
        
        formatted_prompt = f"Human: {prompt.strip()}\n\nAssistant:"
        
        inputs = tokenizer(formatted_prompt, return_tensors="pt", max_length=1024, truncation=True)
        with torch.no_grad():
            outputs = model.generate(**inputs, max_length=200, num_return_sequences=1, do_sample=True)
        return tokenizer.decode(outputs[0], skip_special_tokens=True)

    def add_to_conversation(self, model_name, response):
        self.conversation.append((model_name, response))
        if "task complete?" in response.lower():
            self.task_complete = True

    def get_conversation_history(self):
        return "\n".join([f"{model}: {msg}" for model, msg in self.conversation])

    def clear_conversation(self):
        self.conversation = []
        self.initial_prompt = ""
        self.models = {}
        self.current_model = None
        self.task_complete = False

    def rewind_conversation(self, steps):
        self.conversation = self.conversation[:-steps]
        self.task_complete = False

    def rewind_and_insert(self, steps, inserted_response):
        if steps > 0:
            self.conversation = self.conversation[:-steps]
        if inserted_response.strip():
            last_model = self.conversation[-1][0] if self.conversation else "User"
            next_model = "Model 1" if last_model == "Model 2" or last_model == "User" else "Model 2"
            self.conversation.append((next_model, inserted_response))
        self.current_model = last_model
        self.task_complete = False

manager = ConversationManager()

def get_model(dropdown, custom):
    return custom if custom and custom.strip() else dropdown

def chat(model1, model2, user_input, history, inserted_response=""):
    try:
        print(f"Starting chat with models: {model1}, {model2}")
        print(f"User input: {user_input}")
        
        model1 = get_model(model1, model1_custom.value)
        model2 = get_model(model2, model2_custom.value)
        
        print(f"Selected models: {model1}, {model2}")
        
        if not manager.load_model(model1) or not manager.load_model(model2):
            return "Error: Failed to load one or both models. Please check the model names and try again.", ""
        
        if not manager.conversation:
            manager.initial_prompt = user_input
            manager.clear_conversation()
            manager.add_to_conversation("User", user_input)
        
        models = [model1, model2]
        current_model_index = 0 if manager.current_model in ["User", "Model 2"] else 1
        
        while not manager.task_complete:
            if manager.is_paused:
                yield history, "Conversation paused."
                return

            model = models[current_model_index]
            manager.current_model = model
            
            if inserted_response and current_model_index == 0:
                response = inserted_response
                inserted_response = ""
            else:
                conversation_history = manager.get_conversation_history()
                prompt = f"{conversation_history}\n\nPlease continue the conversation. If you believe the task is complete, end your response with 'Task complete?'"
                response = manager.generate_response(model, prompt)
            
            manager.add_to_conversation(model, response)
            history = manager.get_conversation_history()
            
            for i in range(manager.delay, 0, -1):
                yield history, f"{model} is writing... {i}"
                time.sleep(1)
            
            yield history, ""
            
            if manager.task_complete:
                yield history, "Models believe the task is complete. Are you satisfied with the result? (Yes/No)"
                return
            
            current_model_index = (current_model_index + 1) % 2
        
        return history, "Conversation completed."
    except Exception as e:
        print(f"Error in chat function: {str(e)}")
        print(f"Error type: {type(e).__name__}")
        print(f"Error details: {str(e)}")
        return f"An error occurred: {str(e)}", ""

def user_satisfaction(satisfied, history):
    if satisfied.lower() == 'yes':
        return history, "Task completed successfully."
    else:
        manager.task_complete = False
        return history, "Continuing the conversation..."

def pause_conversation():
    manager.is_paused = True
    return "Conversation paused. Press Resume to continue."

def resume_conversation():
    manager.is_paused = False
    return "Conversation resumed."

def edit_response(edited_text):
    if manager.conversation:
        manager.conversation[-1] = (manager.current_model, edited_text)
    manager.task_complete = False
    return manager.get_conversation_history()

def restart_conversation(model1, model2, user_input):
    manager.clear_conversation()
    return chat(model1, model2, user_input, "")

def rewind_and_insert(steps, inserted_response, history):
    manager.rewind_and_insert(int(steps), inserted_response)
    return manager.get_conversation_history(), ""

open_source_models = [
    "mistralai/Mixtral-8x7B-Instruct-v0.1",
    "bigcode/starcoder2-15b",
    "bigcode/starcoder2-3b",
    "tiiuae/falcon-7b",
    "EleutherAI/gpt-neox-20b",
    "google/flan-ul2",
    "stabilityai/stablelm-zephyr-3b",
    "HuggingFaceH4/zephyr-7b-beta",
    "microsoft/phi-2",
    "google/gemma-7b-it",
    "OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5",
    "mosaicml/mpt-7b-chat",
    "databricks/dolly-v2-12b",
    "thebloke/Wizard-Vicuna-13B-Uncensored-HF",
    "bigscience/bloom-560m"
]

with gr.Blocks() as demo:
    gr.Markdown("# ConversAI Playground")
    
    with gr.Row():
        with gr.Column(scale=1):
            model1_dropdown = gr.Dropdown(choices=open_source_models, label="Model 1")
            model1_custom = gr.Textbox(label="Custom Model 1")
        with gr.Column(scale=1):
            model2_dropdown = gr.Dropdown(choices=open_source_models, label="Model 2")
            model2_custom = gr.Textbox(label="Custom Model 2")
    
    user_input = gr.Textbox(label="Initial prompt", lines=2)
    chat_history = gr.Textbox(label="Conversation", lines=20)
    current_response = gr.Textbox(label="Current model response", lines=3)
    
    with gr.Row():
        pause_btn = gr.Button("Pause")
        edit_btn = gr.Button("Edit")
        rewind_btn = gr.Button("Rewind")
        resume_btn = gr.Button("Resume")
        restart_btn = gr.Button("Restart")
        clear_btn = gr.Button("Clear")
    
    with gr.Row():
        rewind_steps = gr.Slider(0, 10, 1, label="Steps to rewind")
        inserted_response = gr.Textbox(label="Insert response after rewind", lines=2)
    
    delay_slider = gr.Slider(0, 10, 3, label="Response Delay (seconds)")

    user_satisfaction_input = gr.Textbox(label="Are you satisfied with the result? (Yes/No)", visible=False)

    gr.Markdown("""
    ## Button Descriptions
    - **Pause**: Temporarily stops the conversation. The current model will finish its response.
    - **Edit**: Allows you to modify the last response in the conversation.
    - **Rewind**: Removes the specified number of last responses from the conversation.
    - **Resume**: Continues the conversation from where it was paused.
    - **Restart**: Begins a new conversation with the same or different models, keeping the initial prompt.
    - **Clear**: Resets everything, including loaded models, conversation history, and initial prompt.
    """)

    def on_chat_update(history, response):
        if response and "Models believe the task is complete" in response:
            return gr.update(visible=True), gr.update(visible=False)
        return gr.update(visible=False), gr.update(visible=True)

    start_btn = gr.Button("Start Conversation")
    chat_output = start_btn.click(
        chat,
        inputs=[
            model1_dropdown,
            model2_dropdown,
            user_input,
            chat_history
        ],
        outputs=[chat_history, current_response]
    )
    
    chat_output.then(
        on_chat_update,
        inputs=[chat_history, current_response],
        outputs=[user_satisfaction_input, start_btn]
    )
    
    user_satisfaction_input.submit(
        user_satisfaction,
        inputs=[user_satisfaction_input, chat_history],
        outputs=[chat_history, current_response]
    ).then(
        chat,
        inputs=[
            model1_dropdown,
            model2_dropdown,
            user_input,
            chat_history
        ],
        outputs=[chat_history, current_response]
    )
    
    pause_btn.click(pause_conversation, outputs=[current_response])
    resume_btn.click(
        chat,
        inputs=[
            model1_dropdown,
            model2_dropdown,
            user_input,
            chat_history,
            inserted_response
        ],
        outputs=[chat_history, current_response]
    )
    edit_btn.click(edit_response, inputs=[current_response], outputs=[chat_history])
    rewind_btn.click(rewind_and_insert, inputs=[rewind_steps, inserted_response, chat_history], outputs=[chat_history, current_response])
    restart_btn.click(
        restart_conversation,
        inputs=[
            model1_dropdown,
            model2_dropdown,
            user_input
        ],
        outputs=[chat_history, current_response]
    )
    clear_btn.click(manager.clear_conversation, outputs=[chat_history, current_response, user_input])
    delay_slider.change(lambda x: setattr(manager, 'delay', x), inputs=[delay_slider])

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