import torch
from transformers import PreTrainedTokenizerFast, LlamaForCausalLM

MODEL_DIR = "sapbot/toyllama-13m"
# --- Generation Settings ---
MAX_NEW_TOKENS = 150  # Increased slightly so it can finish thoughts
TEMPERATURE = 0.7
TOP_P = 0.9

def main():
    device = "cuda" if torch.cuda.is_available() else "cpu"
    print(f"Running inference on: {device.upper()}")

    try:
        print(f"Loading model from {MODEL_DIR}...")
        tokenizer = PreTrainedTokenizerFast.from_pretrained(MODEL_DIR)
        model = LlamaForCausalLM.from_pretrained(MODEL_DIR)
        model.to(device)
        model.eval()
        print("Model loaded successfully!\n")
    except Exception as e:
        print(f"Failed to load. Error: {e}")
        return

    print("=" * 60)
    print("INTERACTIVE MODE: Ready! (Type 'quit' or 'exit' to stop)")
    print("=" * 60)

    # Get special tokens to ignore during decoding
    special_tokens = {tokenizer.bos_token, tokenizer.eos_token, tokenizer.pad_token, None}

    while True:
        try:
            prompt = input(f"\n>>> Enter prompt (temp: {TEMPERATURE}): ")
        except (KeyboardInterrupt, EOFError):
            print("\nExiting...")
            break
            
        if prompt.strip().lower() in ["quit", "exit"]:
            print("Goodbye!")
            break
            
        if not prompt.strip():
            continue

        inputs = tokenizer(prompt, return_tensors="pt").to(device)
        inputs.pop("token_type_ids", None)
        
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=MAX_NEW_TOKENS,
                temperature=TEMPERATURE,
                top_p=TOP_P,
                do_sample=True,
                pad_token_id=tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
            )

        # -------------------------------------------------------------------
        # THE NEW FIX: Manual Subword Gluing
        # -------------------------------------------------------------------
        # 1. Get the raw list of token string pieces from the model output
        token_ids = outputs[0].tolist()
        raw_tokens = tokenizer.convert_ids_to_tokens(token_ids)
        
        # 2. Filter out special tokens like <s> or </s>
        clean_tokens = [tok for tok in raw_tokens if tok not in special_tokens]
        
        # 3. Glue them together tightly with NO spaces (fixes "M iddle" -> "Middle")
        raw_text = "".join(clean_tokens)
        
        # 4. NOW convert the ByteLevel characters into real spaces and newlines
        generated_text = raw_text.replace("Ġ", " ").replace("Ċ", "\n")
        # -------------------------------------------------------------------

        print("-" * 60)
        print(generated_text.strip())
        print("-" * 60)

if __name__ == "__main__":
    main()