Skip to content

backends

Module to define the backends in charge of creating logits processors.

get_cfg_logits_processor(backend_name, model, grammar)

Create a logits processor from a context-free grammar.

Parameters:

Name Type Description Default
backend_name str | None

The name of the backend to use.

 required
model SteerableModel

The Outlines model of the user.

 required
grammar str

The context-free grammar to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessorType

The logits processor.
Source code in outlines/backends/__init__.py 
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
def get_cfg_logits_processor(
    backend_name: str | None,
    model: SteerableModel,
    grammar: str,
) -> LogitsProcessorType:
    """Create a logits processor from a context-free grammar.

    Parameters
    ----------
    backend_name: str | None
        The name of the backend to use.
    model: Model
        The Outlines model of the user.
    grammar: str
        The context-free grammar to create a logits processor from.

    Returns
    -------
    LogitsProcessorType
        The logits processor.

    """
    backend = _get_backend(
        backend_name or CFG_DEFAULT_BACKEND,
        model,
    )
    return backend.get_cfg_logits_processor(grammar)

get_json_schema_logits_processor(backend_name, model, json_schema)

Create a logits processor from a JSON schema.

Parameters:

Name Type Description Default
backend_name str | None

The name of the backend to use.

 required
model SteerableModel

The Outlines model of the user.

 required
json_schema str

The JSON schema to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessorType

The logits processor.
Source code in outlines/backends/__init__.py 
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
def get_json_schema_logits_processor(
    backend_name: str | None,
    model: SteerableModel,
    json_schema: str,
) -> LogitsProcessorType:
    """Create a logits processor from a JSON schema.

    Parameters
    ----------
    backend_name: str | None
        The name of the backend to use.
    model: Model
        The Outlines model of the user.
    json_schema: str
        The JSON schema to create a logits processor from.

    Returns
    -------
    LogitsProcessorType
        The logits processor.

    """
    backend = _get_backend(
        backend_name or JSON_SCHEMA_DEFAULT_BACKEND,
        model,
    )
    return backend.get_json_schema_logits_processor(json_schema)

get_regex_logits_processor(backend_name, model, regex)

Create a logits processor from a regex.

Parameters:

Name Type Description Default
backend_name str | None

The name of the backend to use.

 required
model SteerableModel

The Outlines model of the user.

 required
regex str

The regex to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessorType

The logits processor.
Source code in outlines/backends/__init__.py 
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
def get_regex_logits_processor(
    backend_name: str | None,
    model: SteerableModel,
    regex: str,
) -> LogitsProcessorType:
    """Create a logits processor from a regex.

    Parameters
    ----------
    backend_name: str | None
        The name of the backend to use.
    model: Model
        The Outlines model of the user.
    regex: str
        The regex to create a logits processor from.

    Returns
    -------
    LogitsProcessorType
        The logits processor.

    """
    backend = _get_backend(
        backend_name or REGEX_DEFAULT_BACKEND,
        model,
    )
    return backend.get_regex_logits_processor(regex)

base

Base class for all backends.

BaseBackend

Bases: ABC

Base class for all backends.

The subclasses must implement methods that create a logits processor from a JSON schema, regex or CFG.

Source code in outlines/backends/base.py 
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
class BaseBackend(ABC):
    """Base class for all backends.

    The subclasses must implement methods that create a logits processor
    from a JSON schema, regex or CFG.

    """

    @abstractmethod
    def get_json_schema_logits_processor(
        self, json_schema: str
    ) -> LogitsProcessorType:
        """Create a logits processor from a JSON schema.

        Parameters
        ----------
        json_schema: str
            The JSON schema to create a logits processor from.

        Returns
        -------
        LogitsProcessorType
            The logits processor.

        """
        ...

    @abstractmethod
    def get_regex_logits_processor(self, regex: str) -> LogitsProcessorType:
        """Create a logits processor from a regex.

        Parameters
        ----------
        regex: str
            The regex to create a logits processor from.

        Returns
        -------
        LogitsProcessorType
            The logits processor.

        """
        ...

    @abstractmethod
    def get_cfg_logits_processor(self, grammar: str) -> LogitsProcessorType:
        """Create a logits processor from a context-free grammar.

        Parameters
        ----------
        grammar: str
            The context-free grammar to create a logits processor from.

        Returns
        -------
        LogitsProcessorType
            The logits processor.

        """
        ...

get_cfg_logits_processor(grammar) abstractmethod

Create a logits processor from a context-free grammar.

Parameters:

Name Type Description Default
grammar str

The context-free grammar to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessorType

The logits processor.
Source code in outlines/backends/base.py 
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
@abstractmethod
def get_cfg_logits_processor(self, grammar: str) -> LogitsProcessorType:
    """Create a logits processor from a context-free grammar.

    Parameters
    ----------
    grammar: str
        The context-free grammar to create a logits processor from.

    Returns
    -------
    LogitsProcessorType
        The logits processor.

    """
    ...

get_json_schema_logits_processor(json_schema) abstractmethod

Create a logits processor from a JSON schema.

Parameters:

Name Type Description Default
json_schema str

The JSON schema to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessorType

The logits processor.
Source code in outlines/backends/base.py 
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
@abstractmethod
def get_json_schema_logits_processor(
    self, json_schema: str
) -> LogitsProcessorType:
    """Create a logits processor from a JSON schema.

    Parameters
    ----------
    json_schema: str
        The JSON schema to create a logits processor from.

    Returns
    -------
    LogitsProcessorType
        The logits processor.

    """
    ...

get_regex_logits_processor(regex) abstractmethod

Create a logits processor from a regex.

Parameters:

Name Type Description Default
regex str

The regex to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessorType

The logits processor.
Source code in outlines/backends/base.py 
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
@abstractmethod
def get_regex_logits_processor(self, regex: str) -> LogitsProcessorType:
    """Create a logits processor from a regex.

    Parameters
    ----------
    regex: str
        The regex to create a logits processor from.

    Returns
    -------
    LogitsProcessorType
        The logits processor.

    """
    ...

llguidance

Backend class for LLGuidance.

LLGuidanceBackend

Bases: BaseBackend

Backend for LLGuidance.

Source code in outlines/backends/llguidance.py 
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
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
class LLGuidanceBackend(BaseBackend):
    """Backend for LLGuidance."""

    def __init__(self, model: SteerableModel):
        """
        Parameters
        ----------
        model
            The Outlines model of the user.

        """
        import llguidance as llg

        self.llg = llg
        self.tensor_library_name = model.tensor_library_name
        self.llg_tokenizer = self._create_llg_tokenizer(model)

    def _create_llg_tokenizer(self, model: SteerableModel) -> "LLGTokenizer":
        """Create an llg tokenizer from the Outlines model's tokenizer.

        Parameters
        ----------
        model: Model
            The Outlines model.

        Returns
        -------
        LLGTokenizer
            The llg tokenizer.

        """
        if isinstance(model, Transformers):
            import llguidance.hf

            return llguidance.hf.from_tokenizer(model.hf_tokenizer)

        elif isinstance(model, LlamaCpp):
            import llama_cpp
            import llguidance.llamacpp

            vocab = llama_cpp.llama_model_get_vocab(model.model.model)
            return llguidance.llamacpp.lltokenizer_from_vocab(vocab)

        elif isinstance(model, MLXLM): # pragma: no cover
            import llguidance.hf

            return llguidance.hf.from_tokenizer(
                model.mlx_tokenizer._tokenizer
            )

        else: # pragma: no cover
            raise ValueError(
                f"Unsupported model type: {type(model)}. "
                "Llguidance only supports LlamaCpp, MLXLM "
                "and Transformers models."
            )

    def get_json_schema_logits_processor(
        self, json_schema: str
    ) -> LLGuidanceLogitsProcessor:
        """Create a logits processor from a JSON schema.

        Parameters
        ----------
        json_schema: str
            The JSON schema to create a logits processor from.

        Returns
        -------
        LogitsProcessor
            The logits processor to use to constrain the generation.

        """
        grammar_spec = self.llg.grammar_from("json_schema", json_schema)
        return LLGuidanceLogitsProcessor(
            grammar_spec, self.llg_tokenizer, self.tensor_library_name
        )

    def get_regex_logits_processor(
        self, regex: str
    ) -> LLGuidanceLogitsProcessor:
        """Create a logits processor from a regex.

        Parameters
        ----------
        regex: str
            The regex to create a logits processor from.

        Returns
        -------
        LogitsProcessor
            The logits processor to use to constrain the generation.

        """
        grammar_spec = self.llg.grammar_from("regex", regex)
        return LLGuidanceLogitsProcessor(
            grammar_spec, self.llg_tokenizer, self.tensor_library_name
        )

    def get_cfg_logits_processor(
        self, grammar: str
    ) -> LLGuidanceLogitsProcessor:
        """Create a logits processor from a context-free grammar.

        Parameters
        ----------
        grammar: str
            The context-free grammar to create a logits processor from.

        Returns
        -------
        LogitsProcessor
            The logits processor to use to constrain the generation.

        """
        # We try both lark and ebnf
        try:
            grammar_spec = self.llg.grammar_from("grammar", grammar)
        except ValueError:
            grammar_spec = self.llg.grammar_from("lark", grammar)
        return LLGuidanceLogitsProcessor(
            grammar_spec, self.llg_tokenizer, self.tensor_library_name
        )

__init__(model)

Parameters:

Name Type Description Default
model SteerableModel

The Outlines model of the user.

 required
Source code in outlines/backends/llguidance.py 
186
187
188
189
190
191
192
193
194
195
196
197
198
def __init__(self, model: SteerableModel):
    """
    Parameters
    ----------
    model
        The Outlines model of the user.

    """
    import llguidance as llg

    self.llg = llg
    self.tensor_library_name = model.tensor_library_name
    self.llg_tokenizer = self._create_llg_tokenizer(model)

get_cfg_logits_processor(grammar)

Create a logits processor from a context-free grammar.

Parameters:

Name Type Description Default
grammar str

The context-free grammar to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessor

The logits processor to use to constrain the generation.
Source code in outlines/backends/llguidance.py 
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
def get_cfg_logits_processor(
    self, grammar: str
) -> LLGuidanceLogitsProcessor:
    """Create a logits processor from a context-free grammar.

    Parameters
    ----------
    grammar: str
        The context-free grammar to create a logits processor from.

    Returns
    -------
    LogitsProcessor
        The logits processor to use to constrain the generation.

    """
    # We try both lark and ebnf
    try:
        grammar_spec = self.llg.grammar_from("grammar", grammar)
    except ValueError:
        grammar_spec = self.llg.grammar_from("lark", grammar)
    return LLGuidanceLogitsProcessor(
        grammar_spec, self.llg_tokenizer, self.tensor_library_name
    )

get_json_schema_logits_processor(json_schema)

Create a logits processor from a JSON schema.

Parameters:

Name Type Description Default
json_schema str

The JSON schema to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessor

The logits processor to use to constrain the generation.
Source code in outlines/backends/llguidance.py 
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
def get_json_schema_logits_processor(
    self, json_schema: str
) -> LLGuidanceLogitsProcessor:
    """Create a logits processor from a JSON schema.

    Parameters
    ----------
    json_schema: str
        The JSON schema to create a logits processor from.

    Returns
    -------
    LogitsProcessor
        The logits processor to use to constrain the generation.

    """
    grammar_spec = self.llg.grammar_from("json_schema", json_schema)
    return LLGuidanceLogitsProcessor(
        grammar_spec, self.llg_tokenizer, self.tensor_library_name
    )

get_regex_logits_processor(regex)

Create a logits processor from a regex.

Parameters:

Name Type Description Default
regex str

The regex to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessor

The logits processor to use to constrain the generation.
Source code in outlines/backends/llguidance.py 
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
def get_regex_logits_processor(
    self, regex: str
) -> LLGuidanceLogitsProcessor:
    """Create a logits processor from a regex.

    Parameters
    ----------
    regex: str
        The regex to create a logits processor from.

    Returns
    -------
    LogitsProcessor
        The logits processor to use to constrain the generation.

    """
    grammar_spec = self.llg.grammar_from("regex", regex)
    return LLGuidanceLogitsProcessor(
        grammar_spec, self.llg_tokenizer, self.tensor_library_name
    )

LLGuidanceLogitsProcessor

Bases: OutlinesLogitsProcessor

Logits Processor for the LLGuidance backend.

Source code in outlines/backends/llguidance.py 
 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
class LLGuidanceLogitsProcessor(OutlinesLogitsProcessor):
    """Logits Processor for the LLGuidance backend."""

    def __init__(
        self,
        grammar: str,
        llg_tokenizer,
        tensor_library_name: str,
    ) -> None:
        """
        Parameters
        ----------
        grammar: str
            The grammar spec to use to create the LLMatcher
        llg_tokenizer: LLTokenizer
            The LLGuidance tokenizer
        tensor_library_name: str
            The name of the tensor library used by the model

        """
        if tensor_library_name not in SUPPORTED_TENSOR_LIBRARIES:
            raise TypeError(f"Unsupported tensor library: {tensor_library_name}")

        self.is_first_token = True
        self.grammar = grammar
        self.llg_tokenizer = llg_tokenizer
        self.tensor_library_name = tensor_library_name
        super().__init__(tensor_library_name)

    def reset(self):
        """Ensure self._setup is called again for the next generation."""
        self.is_first_token = True

    def _setup(self, batch_size: int) -> None:
        """Setup the LLMatchers, the bitmask and some functions used in the
        `process_logits` method.

        This method is called when the first token is generated instead of
        at initialization because we need to know the batch size.

        Parameters
        ----------
        batch_size: int
            The batch size of the input

        """
        from llguidance import LLMatcher

        self.ll_matchers = [
            LLMatcher(self.llg_tokenizer, self.grammar)
            for _ in range(batch_size)
        ]

        # we must adapt the bitmask creation and the bias function to the
        # tensor library used by the model
        if self.tensor_library_name == "torch":
            import llguidance.torch

            self.bitmask = llguidance.torch.allocate_token_bitmask(batch_size, self.llg_tokenizer.vocab_size)
            self._bias_logits = self._bias_logits_torch
        elif self.tensor_library_name == "numpy":
            import llguidance.numpy

            self.bitmask = llguidance.numpy.allocate_token_bitmask(batch_size, self.llg_tokenizer.vocab_size)
            self._bias_logits = self._bias_logits_numpy
        elif self.tensor_library_name == "mlx": # pragma: no cover
            import llguidance.numpy

            self.bitmask = llguidance.numpy.allocate_token_bitmask(batch_size, self.llg_tokenizer.vocab_size)
            self._bias_logits = self._bias_logits_mlx
        else: # pragma: no cover
            raise ValueError(f"Unsupported tensor library: {self.tensor_library_name}")

    def _bias_logits_mlx( # pragma: no cover
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Bias the logits for the MLX backend."""
        import llguidance.mlx
        import llguidance.numpy

        biased_logits_array = []
        for i in range(self.tensor_adapter.shape(input_ids)[0]):
            llguidance.numpy.fill_next_token_bitmask(self.ll_matchers[i], self.bitmask, i)
            biased_logits = llguidance.mlx.apply_token_bitmask(
                logits[i], self.bitmask[i] # type: ignore
            )
            biased_logits_array.append(biased_logits)

        return self.tensor_adapter.concatenate(biased_logits_array)

    def _bias_logits_torch(
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Bias the logits for the Torch backend."""
        import llguidance.torch

        for i in range(self.tensor_adapter.shape(input_ids)[0]):
            llguidance.torch.fill_next_token_bitmask(self.ll_matchers[i], self.bitmask, i)
            self.bitmask = self.tensor_adapter.to_device(
                self.bitmask,
                self.tensor_adapter.get_device(logits)
            )
            llguidance.torch.apply_token_bitmask_inplace(
                logits[i], # type: ignore
                self.bitmask[i]
            )
            self.bitmask = self.tensor_adapter.to_device(
                self.bitmask,
                "cpu"
            )

        return logits

    def _bias_logits_numpy(
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Bias the logits for the Numpy backend."""
        import llguidance.numpy

        for i in range(self.tensor_adapter.shape(input_ids)[0]):
            llguidance.numpy.fill_next_token_bitmask(self.ll_matchers[i], self.bitmask, i)
            llguidance.numpy.apply_token_bitmask_inplace(
                logits[i], self.bitmask[i] # type: ignore
            )

        return logits

    def process_logits(
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Use the instances of LLMatcher to bias the logits.

        Parameters
        ----------
        input_ids
            The ids of the tokens of the existing sequences.
        logits
            The logits for the current generation step.

        Returns
        -------
        TensorType
            The biased logits.

        """
        if self.is_first_token:
            self._setup(self.tensor_adapter.shape(input_ids)[0])
            self.is_first_token = False

        # we do not make the matchers consume the last token during the first
        # generation step because no tokens have been generated yet
        else:
            for i in range(self.tensor_adapter.shape(input_ids)[0]):
                sequence = input_ids[i] # type: ignore
                last_token = sequence[-1].item()
                self.ll_matchers[i].consume_token(last_token)
                error = self.ll_matchers[i].get_error()
                if error:
                    warnings.warn(f"Error in LLMatcher: {error}")

        return self._bias_logits(input_ids, logits)

__init__(grammar, llg_tokenizer, tensor_library_name)

Parameters:

Name Type Description Default
grammar str

The grammar spec to use to create the LLMatcher

 required
llg_tokenizer

The LLGuidance tokenizer

 required
tensor_library_name str

The name of the tensor library used by the model

 required
Source code in outlines/backends/llguidance.py 
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
def __init__(
    self,
    grammar: str,
    llg_tokenizer,
    tensor_library_name: str,
) -> None:
    """
    Parameters
    ----------
    grammar: str
        The grammar spec to use to create the LLMatcher
    llg_tokenizer: LLTokenizer
        The LLGuidance tokenizer
    tensor_library_name: str
        The name of the tensor library used by the model

    """
    if tensor_library_name not in SUPPORTED_TENSOR_LIBRARIES:
        raise TypeError(f"Unsupported tensor library: {tensor_library_name}")

    self.is_first_token = True
    self.grammar = grammar
    self.llg_tokenizer = llg_tokenizer
    self.tensor_library_name = tensor_library_name
    super().__init__(tensor_library_name)

process_logits(input_ids, logits)

Use the instances of LLMatcher to bias the logits.

Parameters:

Name Type Description Default
input_ids TensorType

The ids of the tokens of the existing sequences.

 required
logits TensorType

The logits for the current generation step.

 required

Returns:

Type Description
TensorType

The biased logits.
Source code in outlines/backends/llguidance.py 
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
def process_logits(
    self, input_ids: TensorType, logits: TensorType
) -> TensorType:
    """Use the instances of LLMatcher to bias the logits.

    Parameters
    ----------
    input_ids
        The ids of the tokens of the existing sequences.
    logits
        The logits for the current generation step.

    Returns
    -------
    TensorType
        The biased logits.

    """
    if self.is_first_token:
        self._setup(self.tensor_adapter.shape(input_ids)[0])
        self.is_first_token = False

    # we do not make the matchers consume the last token during the first
    # generation step because no tokens have been generated yet
    else:
        for i in range(self.tensor_adapter.shape(input_ids)[0]):
            sequence = input_ids[i] # type: ignore
            last_token = sequence[-1].item()
            self.ll_matchers[i].consume_token(last_token)
            error = self.ll_matchers[i].get_error()
            if error:
                warnings.warn(f"Error in LLMatcher: {error}")

    return self._bias_logits(input_ids, logits)

reset()

Ensure self._setup is called again for the next generation.

Source code in outlines/backends/llguidance.py 
49
50
51
def reset(self):
    """Ensure self._setup is called again for the next generation."""
    self.is_first_token = True

outlines_core

Backend class for Outlines Core.

OutlinesCoreBackend

Bases: BaseBackend

Backend for Outlines Core.

Source code in outlines/backends/outlines_core.py 
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
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
class OutlinesCoreBackend(BaseBackend):
    """Backend for Outlines Core."""

    def __init__(self, model: SteerableModel):
        """
        Parameters
        ----------
        model
            The Outlines model of the user.

        """
        if isinstance(model, Transformers):
            tokenizer = model.tokenizer
            vocabulary = tokenizer.get_vocab()
            eos_token_id = tokenizer.eos_token_id
            eos_token = tokenizer.eos_token
            token_to_str = tokenizer.convert_token_to_string
        elif isinstance(model, LlamaCpp):
            tokenizer = model.tokenizer # type: ignore
            vocabulary = tokenizer.vocabulary
            eos_token_id = tokenizer.eos_token_id
            eos_token = tokenizer.eos_token
            token_to_str = tokenizer.convert_token_to_string
        elif isinstance(model, MLXLM):
            tokenizer = model.mlx_tokenizer # type: ignore
            vocabulary = tokenizer.get_vocab()
            eos_token_id = tokenizer.eos_token_id
            eos_token = tokenizer.eos_token
            token_to_str = lambda token: tokenizer.convert_tokens_to_string([token]) # type: ignore
        else:
            raise ValueError(f"Unsupported model type: {type(model)}")

        self.eos_token_id = eos_token_id
        self.vocabulary = self.create_outlines_core_vocabulary(
            vocabulary, eos_token_id, eos_token, token_to_str
        )
        self.tensor_library_name = model.tensor_library_name

    def get_json_schema_logits_processor(
        self, json_schema: str
    ):
        """Create a logits processor from a JSON schema.

        Parameters
        ----------
        json_schema: str
            The JSON schema to create a logits processor from.

        Returns
        -------
        LogitsProcessor
            The logits processor to use to constrain the generation.

        """
        regex = outlines_core.json_schema.build_regex_from_schema(json_schema)
        return self.get_regex_logits_processor(regex)

    def get_regex_logits_processor(self, regex: str):
        """Create a logits processor from a regex.

        Parameters
        ----------
        regex: str
            The regex to create a logits processor from.

        Returns
        -------
        LogitsProcessor
            The logits processor to use to constrain the generation.

        """
        index = Index(regex, self.vocabulary)
        return OutlinesCoreLogitsProcessor(index, self.tensor_library_name)

    def get_cfg_logits_processor(self, grammar):
        raise NotImplementedError(
            "Outlines Core does not support context-free grammar."
        )

    @staticmethod
    def create_outlines_core_vocabulary(
        vocab: Dict[str, int],
        eos_token_id: int,
        eos_token: str,
        token_to_str: Callable[[str], str]
    ) -> Vocabulary:
        """Create an Outlines Core Vocabulary instance.

        Parameters
        ----------
        vocab: Dict[str, int]
            The vocabulary to create an Outlines Core vocabulary from.
        eos_token_id: int
            The EOS token ID.
        eos_token: str
            The EOS token.
        token_to_str: Callable[[str], str]
            The function to convert a token to a string.

        Returns
        -------
        Vocabulary
            The Outlines Core Vocabulary instance.

        """
        formatted_vocab = {}
        for token, token_id in vocab.items():
            # This step is necessary to transform special tokens into their
            # string representation, in particular for spacing. We need those
            # string representations as outlines core first builds an FSM from
            # the regex provided that only contains regular strings.
            token_as_str = token_to_str(token)
            formatted_vocab[token_as_str] = [token_id]
        formatted_vocab.pop(eos_token)
        return Vocabulary(eos_token_id, formatted_vocab)

__init__(model)

Parameters:

Name Type Description Default
model SteerableModel

The Outlines model of the user.

 required
Source code in outlines/backends/outlines_core.py 
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
def __init__(self, model: SteerableModel):
    """
    Parameters
    ----------
    model
        The Outlines model of the user.

    """
    if isinstance(model, Transformers):
        tokenizer = model.tokenizer
        vocabulary = tokenizer.get_vocab()
        eos_token_id = tokenizer.eos_token_id
        eos_token = tokenizer.eos_token
        token_to_str = tokenizer.convert_token_to_string
    elif isinstance(model, LlamaCpp):
        tokenizer = model.tokenizer # type: ignore
        vocabulary = tokenizer.vocabulary
        eos_token_id = tokenizer.eos_token_id
        eos_token = tokenizer.eos_token
        token_to_str = tokenizer.convert_token_to_string
    elif isinstance(model, MLXLM):
        tokenizer = model.mlx_tokenizer # type: ignore
        vocabulary = tokenizer.get_vocab()
        eos_token_id = tokenizer.eos_token_id
        eos_token = tokenizer.eos_token
        token_to_str = lambda token: tokenizer.convert_tokens_to_string([token]) # type: ignore
    else:
        raise ValueError(f"Unsupported model type: {type(model)}")

    self.eos_token_id = eos_token_id
    self.vocabulary = self.create_outlines_core_vocabulary(
        vocabulary, eos_token_id, eos_token, token_to_str
    )
    self.tensor_library_name = model.tensor_library_name

create_outlines_core_vocabulary(vocab, eos_token_id, eos_token, token_to_str) staticmethod

Create an Outlines Core Vocabulary instance.

Parameters:

Name Type Description Default
vocab Dict[str, int]

The vocabulary to create an Outlines Core vocabulary from.

 required
eos_token_id int

The EOS token ID.

 required
eos_token str

The EOS token.

 required
token_to_str Callable[[str], str]

The function to convert a token to a string.

 required

Returns:

Type Description
Vocabulary

The Outlines Core Vocabulary instance.
Source code in outlines/backends/outlines_core.py 
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
@staticmethod
def create_outlines_core_vocabulary(
    vocab: Dict[str, int],
    eos_token_id: int,
    eos_token: str,
    token_to_str: Callable[[str], str]
) -> Vocabulary:
    """Create an Outlines Core Vocabulary instance.

    Parameters
    ----------
    vocab: Dict[str, int]
        The vocabulary to create an Outlines Core vocabulary from.
    eos_token_id: int
        The EOS token ID.
    eos_token: str
        The EOS token.
    token_to_str: Callable[[str], str]
        The function to convert a token to a string.

    Returns
    -------
    Vocabulary
        The Outlines Core Vocabulary instance.

    """
    formatted_vocab = {}
    for token, token_id in vocab.items():
        # This step is necessary to transform special tokens into their
        # string representation, in particular for spacing. We need those
        # string representations as outlines core first builds an FSM from
        # the regex provided that only contains regular strings.
        token_as_str = token_to_str(token)
        formatted_vocab[token_as_str] = [token_id]
    formatted_vocab.pop(eos_token)
    return Vocabulary(eos_token_id, formatted_vocab)

get_json_schema_logits_processor(json_schema)

Create a logits processor from a JSON schema.

Parameters:

Name Type Description Default
json_schema str

The JSON schema to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessor

The logits processor to use to constrain the generation.
Source code in outlines/backends/outlines_core.py 
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
def get_json_schema_logits_processor(
    self, json_schema: str
):
    """Create a logits processor from a JSON schema.

    Parameters
    ----------
    json_schema: str
        The JSON schema to create a logits processor from.

    Returns
    -------
    LogitsProcessor
        The logits processor to use to constrain the generation.

    """
    regex = outlines_core.json_schema.build_regex_from_schema(json_schema)
    return self.get_regex_logits_processor(regex)

get_regex_logits_processor(regex)

Create a logits processor from a regex.

Parameters:

Name Type Description Default
regex str

The regex to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessor

The logits processor to use to constrain the generation.
Source code in outlines/backends/outlines_core.py 
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
def get_regex_logits_processor(self, regex: str):
    """Create a logits processor from a regex.

    Parameters
    ----------
    regex: str
        The regex to create a logits processor from.

    Returns
    -------
    LogitsProcessor
        The logits processor to use to constrain the generation.

    """
    index = Index(regex, self.vocabulary)
    return OutlinesCoreLogitsProcessor(index, self.tensor_library_name)

OutlinesCoreLogitsProcessor

Bases: OutlinesLogitsProcessor

Logits processor for Outlines Core.

Source code in outlines/backends/outlines_core.py 
 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
class OutlinesCoreLogitsProcessor(OutlinesLogitsProcessor):
    """Logits processor for Outlines Core."""

    def __init__(
        self, index: Index, tensor_library_name: str
    ):
        """
        Parameters
        ----------
        index: Index
            The Outlines Core `Index` instance to use to create the Outlines
            Core `Guide` instances that will be used to bias the logits
        tensor_library_name: str
            The tensor library name to use for the logits processor.

        """
        self.index = index
        self.tensor_library_name = tensor_library_name
        self.is_first_token = True
        super().__init__(tensor_library_name)

    def reset(self) -> None:
        """Reset the logits processor."""
        self.is_first_token = True

    def _setup(self, batch_size: int, vocab_size: int) -> None:
        """Set the guides, bitmasks and some functions used in the
        `process_logits` method.

        This method is called when the first token is generated instead of
        at initialization because we need to know the batch size and the device
        of the logits.

        Parameters
        ----------
        batch_size: int
            The batch size.
        vocab_size: int
            The vocabulary size.

        """
        if self.tensor_library_name == "torch":
            from outlines_core.kernels.torch import allocate_token_bitmask

            self.allocate_token_bitmask = allocate_token_bitmask
            self.bias_logits = self._bias_logits_torch

        elif self.tensor_library_name == "numpy":
            from outlines_core.kernels.numpy import allocate_token_bitmask

            self.allocate_token_bitmask = allocate_token_bitmask
            self.bias_logits = self._bias_logits_numpy

        elif self.tensor_library_name == "mlx":
            from outlines_core.kernels.mlx import (
                allocate_token_bitmask
            )

            self.allocate_token_bitmask = allocate_token_bitmask
            self.bias_logits = self._bias_logits_mlx

        else:
            raise ValueError(
                f"Unsupported tensor library: {self.tensor_library_name}"
            )

        self._guides = [Guide(self.index) for _ in range(batch_size)]
        self._bitmasks = [
            self.allocate_token_bitmask(vocab_size)
            for _ in range(batch_size)
        ]

    def _bias_logits_mlx( # pragma: no cover
        self, batch_size: int, logits: TensorType
    ) -> TensorType:
        """Bias the logits for MLX tensors."""
        from outlines_core.kernels.mlx import (
            apply_token_bitmask,
            fill_next_token_bitmask
        )

        biased_logits_array = []
        for i in range(batch_size):
            fill_next_token_bitmask(self._guides[i], self._bitmasks[i])
            biased_logits = apply_token_bitmask(
                self.tensor_adapter.unsqueeze(logits[i]), self._bitmasks[i] # type: ignore
            )
            biased_logits_array.append(biased_logits)

        return self.tensor_adapter.concatenate(biased_logits_array)

    def _bias_logits_torch(
        self, batch_size: int, logits: TensorType
    ) -> TensorType:
        """Bias the logits for Torch tensors."""
        from outlines_core.kernels.torch import (
            apply_token_bitmask_inplace,
            fill_next_token_bitmask
        )

        for i in range(batch_size):
            fill_next_token_bitmask(self._guides[i], self._bitmasks[i])
            self._bitmasks[i] = self.tensor_adapter.to_device(
                self._bitmasks[i],
                self.tensor_adapter.get_device(logits)
            )
            apply_token_bitmask_inplace(
                self.tensor_adapter.unsqueeze(logits[i]), # type: ignore
                self._bitmasks[i]
            )
            self._bitmasks[i] = self.tensor_adapter.to_device(
                self._bitmasks[i],
                "cpu"
            )

        return logits

    def _bias_logits_numpy(
        self, batch_size: int, logits: TensorType
    ) -> TensorType:
        """Bias the logits for Numpy tensors."""
        from outlines_core.kernels.numpy import (
            apply_token_bitmask_inplace,
            fill_next_token_bitmask
        )

        for i in range(batch_size):
            fill_next_token_bitmask(self._guides[i], self._bitmasks[i])
            apply_token_bitmask_inplace(
                self.tensor_adapter.unsqueeze(logits[i]), # type: ignore
                self._bitmasks[i]
            )

        return logits

    def process_logits(
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Use the guides to bias the logits.

        Parameters
        ----------
        input_ids
            The ids of the tokens of the existing sequences.
        logits
            The logits for the current generation step.

        Returns
        -------
        TensorType
            The biased logits.

        """
        batch_size = self.tensor_adapter.shape(input_ids)[0]
        vocab_size = self.tensor_adapter.shape(logits)[1]

        if self.is_first_token:
            self._setup(batch_size, vocab_size)
            self.is_first_token = False
        else:
            for i in range(batch_size):
                last_token_id = self.tensor_adapter.to_scalar(input_ids[i][-1]) # type: ignore
                if not self._guides[i].is_finished():
                    self._guides[i].advance(
                        token_id=last_token_id,
                        return_tokens=False
                    )

        return self.bias_logits(batch_size, logits)

__init__(index, tensor_library_name)

Parameters:

Name Type Description Default
index Index

The Outlines Core Index instance to use to create the Outlines Core Guide instances that will be used to bias the logits

 required
tensor_library_name str

The tensor library name to use for the logits processor.

 required
Source code in outlines/backends/outlines_core.py 
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
def __init__(
    self, index: Index, tensor_library_name: str
):
    """
    Parameters
    ----------
    index: Index
        The Outlines Core `Index` instance to use to create the Outlines
        Core `Guide` instances that will be used to bias the logits
    tensor_library_name: str
        The tensor library name to use for the logits processor.

    """
    self.index = index
    self.tensor_library_name = tensor_library_name
    self.is_first_token = True
    super().__init__(tensor_library_name)

process_logits(input_ids, logits)

Use the guides to bias the logits.

Parameters:

Name Type Description Default
input_ids TensorType

The ids of the tokens of the existing sequences.

 required
logits TensorType

The logits for the current generation step.

 required

Returns:

Type Description
TensorType

The biased logits.
Source code in outlines/backends/outlines_core.py 
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
def process_logits(
    self, input_ids: TensorType, logits: TensorType
) -> TensorType:
    """Use the guides to bias the logits.

    Parameters
    ----------
    input_ids
        The ids of the tokens of the existing sequences.
    logits
        The logits for the current generation step.

    Returns
    -------
    TensorType
        The biased logits.

    """
    batch_size = self.tensor_adapter.shape(input_ids)[0]
    vocab_size = self.tensor_adapter.shape(logits)[1]

    if self.is_first_token:
        self._setup(batch_size, vocab_size)
        self.is_first_token = False
    else:
        for i in range(batch_size):
            last_token_id = self.tensor_adapter.to_scalar(input_ids[i][-1]) # type: ignore
            if not self._guides[i].is_finished():
                self._guides[i].advance(
                    token_id=last_token_id,
                    return_tokens=False
                )

    return self.bias_logits(batch_size, logits)

reset()

Reset the logits processor.

Source code in outlines/backends/outlines_core.py 
41
42
43
def reset(self) -> None:
    """Reset the logits processor."""
    self.is_first_token = True

xgrammar

Backend class for XGrammar.

XGrammarBackend

Bases: BaseBackend

Backend for XGrammar.

Source code in outlines/backends/xgrammar.py 
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
class XGrammarBackend(BaseBackend):
    """Backend for XGrammar."""

    def __init__(self, model: SteerableModel):
        """
        Parameters
        ----------
        model
            The Outlines model of the user.

        """
        import xgrammar as xgr

        if isinstance(model, Transformers):
            tokenizer = model.hf_tokenizer
        elif isinstance(model, MLXLM):
            tokenizer = model.mlx_tokenizer._tokenizer
        else: # pragma: no cover
            raise ValueError(
                "The xgrammar backend only supports Transformers and "
                + "MLXLM models"
            )

        tokenizer_info = xgr.TokenizerInfo.from_huggingface(
            tokenizer,
            vocab_size=len(tokenizer.get_vocab())
        )
        self.grammar_compiler = xgr.GrammarCompiler(tokenizer_info)
        self.tensor_library_name = model.tensor_library_name

    def get_json_schema_logits_processor(
        self, json_schema: str
    ) -> XGrammarLogitsProcessor:
        """Create a logits processor from a JSON schema.

        Parameters
        ----------
        json_schema: str
            The JSON schema to create a logits processor from.

        Returns
        -------
        LogitsProcessor
            The logits processor to use to constrain the generation.

        """
        compiled_grammar = self.grammar_compiler.compile_json_schema(
            json_schema
        )
        return XGrammarLogitsProcessor(
            compiled_grammar,
            self.tensor_library_name
        )

    def get_regex_logits_processor(
        self, regex: str
    ) -> XGrammarLogitsProcessor:
        """Create a logits processor from a regex.

        Parameters
        ----------
        regex: str
            The regex to create a logits processor from.

        Returns
        -------
        LogitsProcessor
            The logits processor to use to constrain the generation.

        """
        compiled_grammar = self.grammar_compiler.compile_regex(regex)
        return XGrammarLogitsProcessor(
            compiled_grammar,
            self.tensor_library_name
        )

    def get_cfg_logits_processor(
        self, grammar: str
    ) -> XGrammarLogitsProcessor:
        """Create a logits processor from a context-free grammar.

        Parameters
        ----------
        grammar: str
            The context-free grammar to create a logits processor from.

        Returns
        -------
        LogitsProcessor
            The logits processor to use to constrain the generation.

        """
        compiled_grammar = self.grammar_compiler.compile_grammar(grammar)
        return XGrammarLogitsProcessor(
            compiled_grammar,
            self.tensor_library_name
        )

__init__(model)

Parameters:

Name Type Description Default
model SteerableModel

The Outlines model of the user.

 required
Source code in outlines/backends/xgrammar.py 
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
def __init__(self, model: SteerableModel):
    """
    Parameters
    ----------
    model
        The Outlines model of the user.

    """
    import xgrammar as xgr

    if isinstance(model, Transformers):
        tokenizer = model.hf_tokenizer
    elif isinstance(model, MLXLM):
        tokenizer = model.mlx_tokenizer._tokenizer
    else: # pragma: no cover
        raise ValueError(
            "The xgrammar backend only supports Transformers and "
            + "MLXLM models"
        )

    tokenizer_info = xgr.TokenizerInfo.from_huggingface(
        tokenizer,
        vocab_size=len(tokenizer.get_vocab())
    )
    self.grammar_compiler = xgr.GrammarCompiler(tokenizer_info)
    self.tensor_library_name = model.tensor_library_name

get_cfg_logits_processor(grammar)

Create a logits processor from a context-free grammar.

Parameters:

Name Type Description Default
grammar str

The context-free grammar to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessor

The logits processor to use to constrain the generation.
Source code in outlines/backends/xgrammar.py 
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
def get_cfg_logits_processor(
    self, grammar: str
) -> XGrammarLogitsProcessor:
    """Create a logits processor from a context-free grammar.

    Parameters
    ----------
    grammar: str
        The context-free grammar to create a logits processor from.

    Returns
    -------
    LogitsProcessor
        The logits processor to use to constrain the generation.

    """
    compiled_grammar = self.grammar_compiler.compile_grammar(grammar)
    return XGrammarLogitsProcessor(
        compiled_grammar,
        self.tensor_library_name
    )

get_json_schema_logits_processor(json_schema)

Create a logits processor from a JSON schema.

Parameters:

Name Type Description Default
json_schema str

The JSON schema to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessor

The logits processor to use to constrain the generation.
Source code in outlines/backends/xgrammar.py 
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
def get_json_schema_logits_processor(
    self, json_schema: str
) -> XGrammarLogitsProcessor:
    """Create a logits processor from a JSON schema.

    Parameters
    ----------
    json_schema: str
        The JSON schema to create a logits processor from.

    Returns
    -------
    LogitsProcessor
        The logits processor to use to constrain the generation.

    """
    compiled_grammar = self.grammar_compiler.compile_json_schema(
        json_schema
    )
    return XGrammarLogitsProcessor(
        compiled_grammar,
        self.tensor_library_name
    )

get_regex_logits_processor(regex)

Create a logits processor from a regex.

Parameters:

Name Type Description Default
regex str

The regex to create a logits processor from.

 required

Returns:

Type Description
LogitsProcessor

The logits processor to use to constrain the generation.
Source code in outlines/backends/xgrammar.py 
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
def get_regex_logits_processor(
    self, regex: str
) -> XGrammarLogitsProcessor:
    """Create a logits processor from a regex.

    Parameters
    ----------
    regex: str
        The regex to create a logits processor from.

    Returns
    -------
    LogitsProcessor
        The logits processor to use to constrain the generation.

    """
    compiled_grammar = self.grammar_compiler.compile_regex(regex)
    return XGrammarLogitsProcessor(
        compiled_grammar,
        self.tensor_library_name
    )

XGrammarLogitsProcessor

Bases: OutlinesLogitsProcessor

Logits processor for XGrammar.

Source code in outlines/backends/xgrammar.py 
 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
class XGrammarLogitsProcessor(OutlinesLogitsProcessor):
    """Logits processor for XGrammar."""

    def __init__(self, compiled_grammar: str, tensor_library_name: str,):
        """
        Parameters
        ----------
        compiled_grammar: str
            The compiled grammar to use to create the logits processor.
        tensor_library_name: str
            The name of the tensor library used by the model

        """
        import xgrammar as xgr

        self.xgr = xgr
        self.is_first_token = True
        self.compiled_grammar = compiled_grammar
        self.tensor_library_name = tensor_library_name
        super().__init__(tensor_library_name)

    def reset(self):
        """Ensure self._setup is called again for the next generation."""
        self.is_first_token = True

    def _setup(self, batch_size: int, vocab_size: int) -> None:
        """Setup the logits processor for a new generation."""
        if self.tensor_library_name == "torch":
            self._bias_logits = self._bias_logits_torch
        elif self.tensor_library_name == "mlx":
            self._bias_logits = self._bias_logits_mlx
        else: # pragma: no cover
            raise ValueError(
                f"Unsupported tensor library: {self.tensor_library_name}"
            )

        self._matchers = [
            self.xgr.GrammarMatcher(self.compiled_grammar)
            for _ in range(batch_size)
        ]
        self._bitmask = self.xgr.allocate_token_bitmask(batch_size, vocab_size)

    def _bias_logits_torch(
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Bias the logits for Torch tensors."""
        for i in range(self.tensor_adapter.shape(input_ids)[0]):
            if not self._matchers[i].is_terminated():
                self._matchers[i].fill_next_token_bitmask(self._bitmask, i)

        self._bitmask = self.tensor_adapter.to_device(
            self._bitmask,
            self.tensor_adapter.get_device(logits)
        )
        self.xgr.apply_token_bitmask_inplace(logits, self._bitmask)
        self._bitmask = self.tensor_adapter.to_device(
            self._bitmask,
            "cpu"
        )

        return logits

    def _bias_logits_mlx( # pragma: no cover
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Bias the logits for MLX tensors."""
        import mlx.core as mx
        from xgrammar.kernels.apply_token_bitmask_mlx import apply_token_bitmask_mlx

        for i in range(self.tensor_adapter.shape(input_ids)[0]):
            if not self._matchers[i].is_terminated():
                self._matchers[i].fill_next_token_bitmask(self._bitmask, i)

        biased_logits = apply_token_bitmask_mlx(
            mx.array(self._bitmask.numpy()), logits, self.tensor_adapter.shape(logits)[1]
        )

        return biased_logits

    def process_logits(
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Use the XGrammar matchers to bias the logits."""
        batch_size = self.tensor_adapter.shape(input_ids)[0]
        vocab_size = self.tensor_adapter.shape(logits)[1]

        if self.is_first_token:
            self._setup(batch_size, vocab_size)
            self.is_first_token = False
        else:
            for i in range(batch_size):
                if not self._matchers[i].is_terminated():
                    last_token_id = self.tensor_adapter.to_scalar(
                        input_ids[i][-1] # type: ignore
                    )
                    assert self._matchers[i].accept_token(last_token_id)

        return self._bias_logits(input_ids, logits)

__init__(compiled_grammar, tensor_library_name)

Parameters:

Name Type Description Default
compiled_grammar str

The compiled grammar to use to create the logits processor.

 required
tensor_library_name str

The name of the tensor library used by the model

 required
Source code in outlines/backends/xgrammar.py 
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
def __init__(self, compiled_grammar: str, tensor_library_name: str,):
    """
    Parameters
    ----------
    compiled_grammar: str
        The compiled grammar to use to create the logits processor.
    tensor_library_name: str
        The name of the tensor library used by the model

    """
    import xgrammar as xgr

    self.xgr = xgr
    self.is_first_token = True
    self.compiled_grammar = compiled_grammar
    self.tensor_library_name = tensor_library_name
    super().__init__(tensor_library_name)

process_logits(input_ids, logits)

Use the XGrammar matchers to bias the logits.

Source code in outlines/backends/xgrammar.py 
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
def process_logits(
    self, input_ids: TensorType, logits: TensorType
) -> TensorType:
    """Use the XGrammar matchers to bias the logits."""
    batch_size = self.tensor_adapter.shape(input_ids)[0]
    vocab_size = self.tensor_adapter.shape(logits)[1]

    if self.is_first_token:
        self._setup(batch_size, vocab_size)
        self.is_first_token = False
    else:
        for i in range(batch_size):
            if not self._matchers[i].is_terminated():
                last_token_id = self.tensor_adapter.to_scalar(
                    input_ids[i][-1] # type: ignore
                )
                assert self._matchers[i].accept_token(last_token_id)

    return self._bias_logits(input_ids, logits)

reset()

Ensure self._setup is called again for the next generation.

Source code in outlines/backends/xgrammar.py 
34
35
36
def reset(self):
    """Ensure self._setup is called again for the next generation."""
    self.is_first_token = True