| import os, sys | |
| sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))) | |
| import torch | |
| from bit_transformer import BitTransformerLM | |
| from bit_transformer.training import train_loop as train | |
| def test_train_compression_metrics(): | |
| model = BitTransformerLM(d_model=16, nhead=4, num_layers=1, dim_feedforward=32, max_seq_len=8) | |
| data = torch.zeros((4, 8), dtype=torch.long) | |
| metrics = train(model, data, epochs=1, compress_prob=1.0, log=False) | |
| m = metrics[0] | |
| assert m['compressed_loss'] > 0 | |
| assert m['compression_ratio'] < 1.0 | |
| assert m['raw_loss'] == 0 | |
| def test_train_no_compression(): | |
| model = BitTransformerLM(d_model=16, nhead=4, num_layers=1, dim_feedforward=32, max_seq_len=8) | |
| data = torch.zeros((4, 8), dtype=torch.long) | |
| metrics = train(model, data, epochs=1, compress_prob=0.0, log=False) | |
| m = metrics[0] | |
| assert m['raw_loss'] > 0 | |
| assert m['compressed_loss'] == 0 | |
| def test_train_direct_compression(): | |
| model = BitTransformerLM(d_model=16, nhead=4, num_layers=1, dim_feedforward=32, max_seq_len=8) | |
| data = torch.zeros((4, 8), dtype=torch.long) | |
| metrics = train(model, data, epochs=1, compress_prob=0.0, direct_prob=1.0, log=False) | |
| m = metrics[0] | |
| assert m['direct_loss'] > 0 | |
| def test_diffusion_training_loop(): | |
| model = BitTransformerLM(d_model=16, nhead=4, num_layers=1, dim_feedforward=32, max_seq_len=8) | |
| data = torch.randint(0, 2, (4, 8), dtype=torch.long) | |
| metrics = train(model, data, epochs=1, diffusion=True, log=False) | |
| m = metrics[0] | |
| assert m['raw_loss'] > 0 | |