import argparse
from pathlib import Path

import torch

from train import TinyTransformerLM


@torch.no_grad()
def generate(model, idx, max_new_tokens, temperature, itos):
    model.eval()
    for _ in range(max_new_tokens):
        idx_cond = idx[:, -model.block_size :]
        logits, _ = model(idx_cond)
        logits = logits[:, -1, :] / temperature
        probs = torch.softmax(logits, dim=-1)
        next_id = torch.multinomial(probs, num_samples=1)
        idx = torch.cat((idx, next_id), dim=1)
    return "".join(itos[int(i)] for i in idx[0])


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", default="runs/tiny-char-model.pt")
    parser.add_argument("--prompt", default="hello")
    parser.add_argument("--tokens", type=int, default=400)
    parser.add_argument("--temperature", type=float, default=0.8)
    args = parser.parse_args()

    checkpoint = torch.load(Path(args.model), map_location="cpu")
    config = checkpoint["config"]
    stoi = checkpoint["stoi"]
    itos = {int(k): v for k, v in checkpoint["itos"].items()}

    model = TinyTransformerLM(**config)
    model.load_state_dict(checkpoint["model"])

    fallback = next(iter(stoi.values()))
    encoded = [stoi.get(ch, fallback) for ch in args.prompt]
    idx = torch.tensor([encoded], dtype=torch.long)
    print(generate(model, idx, args.tokens, args.temperature, itos))


if __name__ == "__main__":
    main()