#!/usr/bin/env python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Define the directory where your fine-tuned model is saved.
model_dir = "./gpt2-finetuned"

# Load the tokenizer and model from the saved directory.
tokenizer = AutoTokenizer.from_pretrained(model_dir)
model = AutoModelForCausalLM.from_pretrained(model_dir)

# If you are using GPU and it's available, move the model to GPU.
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

print("Chat with the model! Type 'exit' or 'quit' to end the conversation.")

while True:
    # Get user input.
    user_input = input("You: ")
    if user_input.lower() in ["exit", "quit"]:
        print("Exiting chat.")
        break

    # Encode the input text and generate an attention mask.
    inputs = tokenizer(user_input, return_tensors="pt", padding=True, truncation=True)
    input_ids = inputs["input_ids"].to(device)
    attention_mask = inputs["attention_mask"].to(device)  # Explicitly set the attention mask

    # Generate a response. You can tweak the generation parameters as needed.
    output_ids = model.generate(
        input_ids,
        attention_mask=attention_mask,  # Pass the attention mask here
        max_length=100,             # Maximum length of the generated response.
        do_sample=True,             # Use sampling; set to False for greedy decoding.
        top_p=0.95,                 # Top-p (nucleus) sampling.
        top_k=50,                   # Top-k sampling.
        pad_token_id=tokenizer.eos_token_id  # Avoid warnings if no pad token is defined.
    )

    # Decode the generated tokens to a string.
    response = tokenizer.decode(output_ids[0], skip_special_tokens=True)
    print("Bot:", response)