outlines-core 0.2.14

Structured Generation
Documentation 
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
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271

import importlib

import numpy as np
import pytest
import torch

from outlines_core import Guide, Index, Vocabulary

VOCAB = Vocabulary.from_pretrained("gpt2", None, None)
VOCAB_LEN = len(VOCAB)


@pytest.fixture(scope="session")
def guide() -> Guide:
    return Guide(Index("\\+?[1-9][0-9]{7,14}", VOCAB))


@pytest.mark.no_cover
def test_interface_torch():
    from outlines_core.kernels.torch import (
        allocate_token_bitmask,
        apply_token_bitmask_inplace,
        fill_next_token_bitmask,
    )

    mask = allocate_token_bitmask(VOCAB_LEN)
    assert mask.shape[1] == 1571, "Mask sized incorrectly."
    assert mask.dim() == 2, "Mask should be 2D"
    assert mask.is_contiguous(), "Mask must be contiguous"
    assert mask.dtype == torch.int32, "Mask must be dtype torch.int32"

    # test apply_token_bitmask_inplace
    logits = torch.randn(1, 100)

    wrong_dtype_mask = mask.to(torch.int64)
    with pytest.raises(ValueError, match="Invalid mask dtype"):
        apply_token_bitmask_inplace(logits, wrong_dtype_mask)

    mask_1d = torch.full((mask.shape[1],), -1, dtype=torch.int32)
    with pytest.raises(
        ValueError, match="Invalid mask dimensions: Expected a 2D array"
    ):
        apply_token_bitmask_inplace(logits, mask_1d)

    logits_1d = torch.randn(100)
    with pytest.raises(
        ValueError, match="Invalid logits dimensions: Expected a 2D array, but got 1D."
    ):
        apply_token_bitmask_inplace(logits_1d, mask)

    mask_batch2 = torch.full((2, mask.shape[1]), -1, dtype=torch.int32)
    with pytest.raises(ValueError, match="Invalid batch size"):
        apply_token_bitmask_inplace(logits, mask_batch2)

    # test fill_next_token_bitmask
    wrong_dtype_mask = mask.to(torch.int64)
    with pytest.raises(ValueError, match="Invalid mask dtype"):
        fill_next_token_bitmask(None, wrong_dtype_mask)

    mask_1d = torch.full((mask.shape[1],), -1, dtype=torch.int32)
    with pytest.raises(ValueError, match="Invalid mask dimensions"):
        fill_next_token_bitmask(None, mask_1d)

    mask_batch2 = torch.full((2, mask.shape[1]), -1, dtype=torch.int32)
    with pytest.raises(
        ValueError, match="Invalid batch size: Batch mask writes are not supported"
    ):
        fill_next_token_bitmask(None, mask_batch2)


@pytest.mark.no_cover
def test_interface_numpy():
    from outlines_core.kernels.numpy import (
        allocate_token_bitmask,
        apply_token_bitmask_inplace,
        fill_next_token_bitmask,
    )

    mask = allocate_token_bitmask(VOCAB_LEN)
    assert mask.shape[1] == 1571, "Mask sized incorrectly."
    assert mask.ndim == 2, "Mask should be 2D"
    assert mask.flags["C_CONTIGUOUS"], "Mask must be contiguous"
    assert mask.dtype == np.int32, "Mask must be dtype np.int32"

    logits = np.random.randn(1, 100).astype(np.float32)

    wrong_dtype_mask = mask.astype(np.int64)
    with pytest.raises(ValueError, match="Invalid mask dtype"):
        apply_token_bitmask_inplace(logits, wrong_dtype_mask)

    # 3d because 1d will be fixed automatically
    mask_3d = np.full((mask.shape[1], 1, mask.shape[1]), -1, dtype=np.int32)
    with pytest.raises(
        ValueError, match="Invalid mask dimensions: Expected a 2D array"
    ):
        apply_token_bitmask_inplace(logits, mask_3d)

    with pytest.raises(
        ValueError, match="Invalid logits dimensions: Expected a 2D array"
    ):
        apply_token_bitmask_inplace(mask_3d, mask)

    mask_batch2 = np.full((2, mask.shape[1]), -1, dtype=np.int32)
    with pytest.raises(ValueError, match="Invalid batch size"):
        apply_token_bitmask_inplace(logits, mask_batch2)

    wrong_dtype_mask = mask.astype(np.int64)
    with pytest.raises(ValueError, match="Invalid mask dtype"):
        fill_next_token_bitmask(None, wrong_dtype_mask)

    mask_1d = np.full((mask.shape[1],), -1, dtype=np.int32)
    with pytest.raises(ValueError, match="Invalid mask dimensions"):
        fill_next_token_bitmask(None, mask_1d)

    mask_batch2 = np.full((2, mask.shape[1]), -1, dtype=np.int32)
    with pytest.raises(
        ValueError, match="Invalid batch size: Batch mask writes are not supported"
    ):
        fill_next_token_bitmask(None, mask_batch2)


@pytest.mark.no_cover
@pytest.mark.skipif(
    not importlib.util.find_spec("mlx"), reason="mlx is required to test mlx kernels"  # type: ignore
)
def test_interface_mlx():
    import mlx.core as mx
    import numpy as np
    import pytest

    from outlines_core.kernels.mlx import (
        allocate_token_bitmask,
        apply_token_bitmask,
        fill_next_token_bitmask,
    )

    mask = allocate_token_bitmask(VOCAB_LEN)
    assert mask.shape[1] == (VOCAB_LEN + 31) // 32, "Mask sized incorrectly."
    assert mask.ndim == 2, "Mask should be 2D"
    assert mask.flags["C_CONTIGUOUS"], "Mask must be contiguous"
    assert mask.dtype == np.int32, "Mask must be dtype np.int32"

    logits = mx.array(np.random.randn(1, 100).astype(np.float32))
    wrong_dtype_mask = mask.astype(np.int64)
    with pytest.raises(ValueError, match="Invalid mask dtype"):
        apply_token_bitmask(logits, wrong_dtype_mask)

    # 3d because 1d will be fixed automatically
    mask_3d = np.full((mask.shape[1], 1, mask.shape[1]), -1, dtype=np.int32)
    with pytest.raises(
        ValueError, match="Invalid mask dimensions: Expected a 2D array"
    ):
        apply_token_bitmask(logits, mask_3d)

    with pytest.raises(
        ValueError, match="Invalid logits dimensions: Expected a 2D array"
    ):
        apply_token_bitmask(mask_3d, mask)

    mask_batch2 = np.full((2, mask.shape[1]), -1, dtype=np.int32)
    with pytest.raises(ValueError, match="Invalid batch size"):
        apply_token_bitmask(logits, mask_batch2)

    wrong_dtype_mask = mask.astype(np.int64)
    with pytest.raises(ValueError, match="Invalid mask dtype"):
        fill_next_token_bitmask(None, wrong_dtype_mask)

    mask_1d = np.full((mask.shape[1],), -1, dtype=np.int32)
    with pytest.raises(ValueError, match="Invalid mask dimensions"):
        fill_next_token_bitmask(None, mask_1d)

    mask_batch2 = np.full((2, mask.shape[1]), -1, dtype=np.int32)
    with pytest.raises(
        ValueError, match="Invalid batch size: Batch mask writes are not supported"
    ):
        fill_next_token_bitmask(None, mask_batch2)


@pytest.mark.no_cover
def test_torch_correctness(guide):
    from outlines_core.kernels.torch import _apply_token_bitmask_inplace_kernel

    allowed_tokens = set(guide.get_tokens())

    logits = torch.tensor(torch.randn(1, VOCAB_LEN))

    orig_logits = logits.clone()

    mask = torch.tensor(
        torch.full((1, ((VOCAB_LEN + 31) // 32)), -1, dtype=torch.int32)
    )

    guide.write_mask_into(mask.data_ptr(), mask.numel(), mask.element_size())

    _apply_token_bitmask_inplace_kernel(logits, mask)

    for j in range(VOCAB_LEN):
        if j in allowed_tokens:
            assert torch.isclose(
                logits[0, j], orig_logits[0, j], equal_nan=True
            ), f"Token {j} should be allowed but was masked."
        else:
            assert logits[0, j] == -float(
                "inf"
            ), f"Token {j} should be masked but was {logits[0, j].item()}."


@pytest.mark.no_cover
def test_numpy_correctness(guide):
    from outlines_core.kernels.numpy import _apply_token_bitmask_inplace_kernel

    allowed_tokens = set(guide.get_tokens())

    logits = np.random.randn(1, VOCAB_LEN).astype(np.float32)
    orig_logits = logits.copy()

    mask = np.full((1, ((VOCAB_LEN + 31) // 32)), -1, dtype=np.int32)

    guide.write_mask_into(mask.ctypes.data, mask.size, mask.itemsize)

    _apply_token_bitmask_inplace_kernel(logits, mask)

    for j in range(VOCAB_LEN):
        if j in allowed_tokens:
            np.testing.assert_allclose(
                logits[0, j],
                orig_logits[0, j],
                err_msg=f"Token {j} should be allowed but was masked.",
            )
        else:
            assert (
                logits[0, j] == -np.inf
            ), f"Token {j} should be masked was got {logits[0, j]}."


@pytest.mark.no_cover
@pytest.mark.skipif(
    not importlib.util.find_spec("mlx"), reason="mlx is required to test mlx kernels"  # type: ignore
)
def test_mlx_correctness(guide):
    import mlx.core as mx

    from outlines_core.kernels.mlx import _apply_token_bitmask_kernel

    allowed_tokens = set(guide.get_tokens())

    np_logits = np.random.randn(1, VOCAB_LEN).astype(np.float32)

    orig_logits = np_logits.copy()

    logits_mlx = mx.array(np_logits)

    mask = np.full((1, (VOCAB_LEN + 31) // 32), -1, dtype=np.int32)

    guide.write_mask_into(mask.ctypes.data, mask.size, mask.itemsize)

    logits_mlx_out = _apply_token_bitmask_kernel(logits_mlx, mx.array(mask))

    logits_out = np.array(logits_mlx_out)

    for j in range(VOCAB_LEN):
        if j in allowed_tokens:
            np.testing.assert_allclose(
                logits_out[0, j],
                orig_logits[0, j],
                err_msg=f"Token {j} should be allowed but was masked.",
            )
        else:
            assert (
                logits_out[0, j] == -np.inf
            ), f"Token {j} should be masked was got {logits_out[0, j]}."