generate.py

import torch
from model import MiniGPT
from tokenizer import load_tokenizer

# ------- Paramètres -------
device = 'cuda' if torch.cuda.is_available() else 'cpu'
checkpoint_path = "checkpoints/model_step_best.pt"  # ← remplace par ton fichier
tokenizer_path = "tokenizer_wtw_tinystories.json"
block_size = 128
embed_dim = 128
n_heads = 16
n_layers = 16
max_new_tokens = 500

# ------- Load tokenizer -------
stoi, itos, encode, decode, pad_token_id = load_tokenizer(tokenizer_path)
vocab_size = len(stoi)

# ------- Load model -------
model = MiniGPT(
    vocab_size=vocab_size,
    block_size=block_size,
    embed_dim=embed_dim,
    depth=n_layers,
    heads=n_heads
).to(device)

checkpoint = torch.load(checkpoint_path, map_location=device)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# ------- Contexte initial -------
prompt = "Maman"
context_ids = encode(prompt)
context = torch.tensor([context_ids], dtype=torch.long, device=device)

# ------- Génération -------
with torch.no_grad():
    output_ids = model.generate(context, max_new_tokens=max_new_tokens)[0].tolist()

# ------- Décodage -------
generated_text = decode(output_ids)
print("\n--- Histoire générée ---\n")
print(generated_text)
print("\n------------------------\n")