app.py

import gradio as gr
from unsloth import FastLanguageModel
import torch

# Load the model and tokenizer locally
max_seq_length = 2048
model_name_or_path = "michailroussos/model_llama_8d"

# Load model and tokenizer using unsloth
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_name_or_path,
    max_seq_length=max_seq_length,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)  # Enable optimized inference

# Define the response function
def respond(message, history, system_message, max_tokens, temperature, top_p):
    print("\n" + "="*50)
    print("===== RESPOND FUNCTION CALLED =====")
    print("="*50)
    
    # Print input parameters
    print(f"Input Message: {message}")
    print(f"System Message: {system_message}")
    print(f"Max Tokens: {max_tokens}")
    print(f"Temperature: {temperature}")
    print(f"Top-p: {top_p}")
    
    # Debug history
    print("\n--- Current History ---")
    print(f"History Type: {type(history)}")
    print(f"History Content: {history}")
    
    # Prepare the messages for the model
    messages = []
    try:
        if history:
            print("\n--- Processing Existing History ---")
            for entry in history:
                messages.append({"role": "user", "content": entry[0]})
                messages.append({"role": "assistant", "content": entry[1]})
        
        # Add the current user message
        print("\n--- Adding Current Message ---")
        messages.append({"role": "user", "content": message})
        
        # Debug messages before tokenization
        print("\n--- Messages Before Tokenization ---")
        for msg in messages:
            print(f"Role: {msg['role']}, Content: {msg['content']}")
        
        # Tokenize the input
        print("\n--- Tokenizing Input ---")
        inputs = tokenizer.apply_chat_template(
            messages,
            tokenize=True,
            add_generation_prompt=True,
            return_tensors="pt",
        ).to("cuda" if torch.cuda.is_available() else "cpu")
        
        print(f"Tokenized Inputs Shape: {inputs.shape}")
        print(f"Tokenized Inputs Device: {inputs.device}")
        
        # Generate response
        attention_mask = inputs.ne(tokenizer.pad_token_id).long()
        
        try:
            generated_tokens = model.generate(
                input_ids=inputs,
                attention_mask=attention_mask,
                max_new_tokens=max_tokens,
                use_cache=True,
                temperature=temperature,
                top_p=top_p,
            )
            
            # Decode the generated response
            response = tokenizer.decode(generated_tokens[0], skip_special_tokens=True)
            print("\n--- Generated Response ---")
            print(f"Raw Response: {response}")
            
            # Check and filter response
            #if "system" in response.lower():
            #    print("WARNING: System message detected in response")
            #    response = "Hello! How can I assist you today?"
            
            # Prepare return history in OpenAI messages format
            return_messages = []
            for entry in (history or []):
                return_messages.append({"role": "user", "content": entry[0]})
                return_messages.append({"role": "assistant", "content": entry[1]})
            
            # Add current conversation turn
            return_messages.append({"role": "user", "content": message})
            return_messages.append({"role": "assistant", "content": response})
            
            print("\n--- Return Messages ---")
            for msg in return_messages:
                print(f"Role: {msg['role']}, Content: {msg['content'][:100]}...")
            
            return return_messages
        
        except Exception as gen_error:
            print("\n--- GENERATION ERROR ---")
            print(f"Error during model generation: {gen_error}")
            return []
    
    except Exception as prep_error:
        print("\n--- PREPARATION ERROR ---")
        print(f"Error during message preparation: {prep_error}")
        return []

# Define the Gradio interface
demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Textbox(value="You are a helpful assistant.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p"),
    ],
    type="messages"  # Explicitly set to messages type
)

if __name__ == "__main__":
    demo.launch(share=False)  # Use share=False for local testing