example/compare_checkpoints.py

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

MODEL_ID = "ibm-granite/granite-4.0-micro"

CHECKPOINTS = {
    "Base model": None,
    "LoRA checkpoint-30": "./lora-out/checkpoint-30",
    "LoRA checkpoint-60": "./lora-out/checkpoint-60",
    "LoRA checkpoint-90": "./lora-out/checkpoint-90",
    "LoRA checkpoint-120": "./lora-out/checkpoint-120",
}

MAX_NEW_TOKENS = 300


def load_model(checkpoint_path=None):
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        dtype=torch.float16,
        device_map="cuda"
    )

    if checkpoint_path is not None:
        model = PeftModel.from_pretrained(model, checkpoint_path)

    model.eval()
    return model


def generate_answer(model, tokenizer, question):
    prompt = f"Frage:\n{question}\n\nAntwort:\n"
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        output = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,   # deterministisch
        )

    text = tokenizer.decode(output[0], skip_special_tokens=True)
    return text[len(prompt):].strip()


def main():
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

    print("=" * 80)
    question = input("Bitte eine Frage eingeben:\n> ").strip()
    print("=" * 80)

    for label, checkpoint in CHECKPOINTS.items():
        print(f"\n=== {label} ===\n")

        model = load_model(checkpoint)
        answer = generate_answer(model, tokenizer, question)

        print(answer)
        print("\n" + "-" * 80)

        # Speicher sauber freigeben (optional, aber sauber)
        del model
        torch.cuda.empty_cache()


if __name__ == "__main__":
    main()