File size: 2,306 Bytes
86e3bb9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 | import numpy as np
import torch
from kawk50m.binary_dataset import PackedTokenDataset
def test_dataset_builds_causal_lm_batches_and_epoch_permutation(tmp_path) -> None:
token_path = tmp_path / "tokens.bin"
np.arange(21, dtype=np.uint16).tofile(token_path)
dataset = PackedTokenDataset(token_path, sequence_length=4)
assert len(dataset) == 5
inputs, labels = dataset.evaluation_batch(
batch_index=0,
batch_size=2,
rank=0,
world_size=1,
device=torch.device("cpu"),
)
assert inputs.tolist() == [[0, 1, 2, 3], [4, 5, 6, 7]]
# Hugging Face causal-LM heads shift labels internally. Returning an
# already shifted target here would train the model to predict x[t+2].
assert labels.tolist() == inputs.tolist()
assert labels.data_ptr() != inputs.data_ptr()
visited = []
for micro_step in range(len(dataset)):
train_inputs, _ = dataset.training_batch(
micro_step=micro_step,
batch_size=1,
rank=0,
world_size=1,
seed=42,
device=torch.device("cpu"),
)
visited.append(train_inputs[0, 0].item() // 4)
assert sorted(visited) == list(range(len(dataset)))
offset_inputs, _ = dataset.training_batch(
micro_step=0,
batch_size=1,
rank=0,
world_size=1,
seed=42,
sample_offset=1,
device=torch.device("cpu"),
)
next_inputs, _ = dataset.training_batch(
micro_step=1,
batch_size=1,
rank=0,
world_size=1,
seed=42,
device=torch.device("cpu"),
)
assert offset_inputs.tolist() == next_inputs.tolist()
def test_dataset_reads_multiple_binary_shards_from_glob(tmp_path) -> None:
np.arange(9, dtype=np.uint16).tofile(tmp_path / "part-000.bin")
np.arange(100, 109, dtype=np.uint16).tofile(tmp_path / "part-001.bin")
dataset = PackedTokenDataset(
str(tmp_path / "part-*.bin"),
sequence_length=4,
)
assert len(dataset) == 4
assert dataset.sequence(0).tolist() == [0, 1, 2, 3, 4]
assert dataset.sequence(1).tolist() == [4, 5, 6, 7, 8]
assert dataset.sequence(2).tolist() == [100, 101, 102, 103, 104]
assert dataset.sequence(3).tolist() == [104, 105, 106, 107, 108]
|