Skip to content

structured

Logits processors for structured generation.

/ Don't want to self-host? \ Try .json at http://dottxt.co / 
   \   ^__^
    \  (oo)\_______
        (__)\       )\/                ||----w |
            ||     ||

Copyright 2024- the Outlines developers

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

CFGLogitsProcessor

Bases: GuideLogitsProcessor

Bias generation based on a context-free grammar.

Source code in outlines/processors/structured.py 
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
class CFGLogitsProcessor(GuideLogitsProcessor):
    """Bias generation based on a context-free grammar."""
    guide: CFGGuide

    def __init__(
        self, cfg_str: str, tokenizer: "Tokenizer", tensor_library_name: str
    ):
        """
        Parameters
        ----------
        cfg_str
            A string that represents a grammar.
        tokenizer
            The tokenizer used to convert tokens to ids.
        tensor_library_name
            The name of the library to use to manipulate the tensors.

        """
        # Build a guide from the CFG string and then pass it to the
        # GuideLogitsProcessor superclass.
        cfg_guide = CFGGuide(cfg_string=cfg_str, tokenizer=tokenizer)
        super().__init__(
            tokenizer=tokenizer,
            guide=cfg_guide,
            tensor_library_name=tensor_library_name
        )

    def process_logits(
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Same behavior as GuideLogitsProcessor, but uses rejection
        sampling.

        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._seq_start_idx is None:
            self._seq_start_idx = len(input_ids[0]) # type: ignore

        sequence_states: List = []  # vector of states corresponding to `input_ids`

        for seq_ids in input_ids: # type: ignore
            gen_ids = seq_ids[self._seq_start_idx :]
            curr_state_key = hash(tuple(self.tensor_adapter.to_list(gen_ids)))

            if curr_state_key not in self._guide_states: # pragma: no cover
                prev_state = self._guide_states[hash(tuple(self.tensor_adapter.to_list(gen_ids[:-1])))]
                curr_state = self.guide.get_next_state(prev_state, self.tensor_adapter.to_scalar(gen_ids[-1]))
                self._guide_states[curr_state_key] = curr_state

            sequence_states.append(self._guide_states[curr_state_key])

        mask = self.tensor_adapter.full_like(logits, -math.inf)
        for i, guide_state in enumerate(sequence_states):
            first_legal_token = next(
                self.guide.iter_valid_token_ids(
                    guide_state, self.tensor_adapter.argsort_descending(logits[i]) # type: ignore
                )
            )
            mask[i, [first_legal_token]] = logits[i, [first_legal_token]] # type: ignore

        return mask

__init__(cfg_str, tokenizer, tensor_library_name)

Parameters:

Name Type Description Default
cfg_str str

A string that represents a grammar.

 required
tokenizer Tokenizer

The tokenizer used to convert tokens to ids.

 required
tensor_library_name str

The name of the library to use to manipulate the tensors.

 required
Source code in outlines/processors/structured.py 
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
def __init__(
    self, cfg_str: str, tokenizer: "Tokenizer", tensor_library_name: str
):
    """
    Parameters
    ----------
    cfg_str
        A string that represents a grammar.
    tokenizer
        The tokenizer used to convert tokens to ids.
    tensor_library_name
        The name of the library to use to manipulate the tensors.

    """
    # Build a guide from the CFG string and then pass it to the
    # GuideLogitsProcessor superclass.
    cfg_guide = CFGGuide(cfg_string=cfg_str, tokenizer=tokenizer)
    super().__init__(
        tokenizer=tokenizer,
        guide=cfg_guide,
        tensor_library_name=tensor_library_name
    )

process_logits(input_ids, logits)

Same behavior as GuideLogitsProcessor, but uses rejection sampling.

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/processors/structured.py 
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
def process_logits(
    self, input_ids: TensorType, logits: TensorType
) -> TensorType:
    """Same behavior as GuideLogitsProcessor, but uses rejection
    sampling.

    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._seq_start_idx is None:
        self._seq_start_idx = len(input_ids[0]) # type: ignore

    sequence_states: List = []  # vector of states corresponding to `input_ids`

    for seq_ids in input_ids: # type: ignore
        gen_ids = seq_ids[self._seq_start_idx :]
        curr_state_key = hash(tuple(self.tensor_adapter.to_list(gen_ids)))

        if curr_state_key not in self._guide_states: # pragma: no cover
            prev_state = self._guide_states[hash(tuple(self.tensor_adapter.to_list(gen_ids[:-1])))]
            curr_state = self.guide.get_next_state(prev_state, self.tensor_adapter.to_scalar(gen_ids[-1]))
            self._guide_states[curr_state_key] = curr_state

        sequence_states.append(self._guide_states[curr_state_key])

    mask = self.tensor_adapter.full_like(logits, -math.inf)
    for i, guide_state in enumerate(sequence_states):
        first_legal_token = next(
            self.guide.iter_valid_token_ids(
                guide_state, self.tensor_adapter.argsort_descending(logits[i]) # type: ignore
            )
        )
        mask[i, [first_legal_token]] = logits[i, [first_legal_token]] # type: ignore

    return mask

GuideLogitsProcessor

Bases: OutlinesLogitsProcessor

Bias generation using a guide.

Attributes:

Name Type Description
tokenizer Tokenizer

The outlines tokenizer used to convert tokens to ids.
guide Guide

The outlines guide used to bias the logits.
Source code in outlines/processors/structured.py 
 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
class GuideLogitsProcessor(OutlinesLogitsProcessor):
    """Bias generation using a guide.

    Attributes
    ----------
    tokenizer
        The outlines tokenizer used to convert tokens to ids.
    guide
        The outlines guide used to bias the logits.
    """
    tokenizer: "Tokenizer"
    guide: Guide
    _guide_states: Dict[int, Any]
    _seq_start_idx: Optional[int]

    def __init__(
        self, tokenizer: "Tokenizer", guide: Guide, tensor_library_name: str
    ):
        """
        Parameters
        ----------
        tokenizer
            The tokenizer used to convert tokens to ids.
        guide
            The `outlines.processors.guide.Guide` that is used to bias the
            logits.
        tensor_library_name
            The name of the library to use to manipulate the tensors.

        """
        super().__init__(tensor_library_name=tensor_library_name)
        self.tokenizer = tokenizer
        self.guide = guide
        self._guide_states = {hash(tuple([])): self.guide.initial_state}
        self._seq_start_idx = None

    def process_logits(
        self, input_ids: TensorType, logits: TensorType
    ) -> TensorType:
        """Use the Guide to bias the logits before sampling the next token.

        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._seq_start_idx is None:
            self._seq_start_idx = len(input_ids[0]) # type: ignore

        sequence_states: List[int] = []  # vector of states corresponding to `input_ids`

        for seq_ids in input_ids: # type: ignore
            gen_ids = seq_ids[self._seq_start_idx :]
            curr_state_key = hash(tuple(self.tensor_adapter.to_list(gen_ids)))

            if curr_state_key not in self._guide_states:
                prev_state = self._guide_states[hash(tuple(self.tensor_adapter.to_list(gen_ids[:-1])))]
                curr_state = self.guide.get_next_state(prev_state, self.tensor_adapter.to_scalar(gen_ids[-1]))
                self._guide_states[curr_state_key] = curr_state

            sequence_states.append(self._guide_states[curr_state_key])

        allowed_tokens_batch = []
        batch_indices = []
        for i, guide_state in enumerate(sequence_states):
            allowed_tokens = self.guide.get_next_instruction(guide_state).tokens
            allowed_tokens_batch.append(allowed_tokens)
            batch_indices.append(
                self.tensor_adapter.full_like(allowed_tokens, i)
            )  # Store batch index for each allowed token

        device = self.tensor_adapter.get_device(logits)
        allowed_tokens_concat = self.tensor_adapter.to_device(
            self.tensor_adapter.concatenate(allowed_tokens_batch),
            device
        )
        batch_indices_concat = self.tensor_adapter.to_device(
            self.tensor_adapter.concatenate(batch_indices),
            device
        )

        mask = self.tensor_adapter.boolean_ones_like(logits)
        mask[batch_indices_concat, allowed_tokens_concat] = False
        logits = self.tensor_adapter.apply_mask(logits, mask, float("-inf"))

        return logits

    def copy(self) -> "GuideLogitsProcessor":
        """Return a copy of the logits processor."""
        return GuideLogitsProcessor(
            tokenizer=self.tokenizer,
            guide=self.guide.copy(),
            tensor_library_name=self.tensor_adapter.library_name
        )

__init__(tokenizer, guide, tensor_library_name)

Parameters:

Name Type Description Default
tokenizer Tokenizer

The tokenizer used to convert tokens to ids.

 required
guide Guide

The outlines.processors.guide.Guide that is used to bias the logits.

 required
tensor_library_name str

The name of the library to use to manipulate the tensors.

 required
Source code in outlines/processors/structured.py 
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
def __init__(
    self, tokenizer: "Tokenizer", guide: Guide, tensor_library_name: str
):
    """
    Parameters
    ----------
    tokenizer
        The tokenizer used to convert tokens to ids.
    guide
        The `outlines.processors.guide.Guide` that is used to bias the
        logits.
    tensor_library_name
        The name of the library to use to manipulate the tensors.

    """
    super().__init__(tensor_library_name=tensor_library_name)
    self.tokenizer = tokenizer
    self.guide = guide
    self._guide_states = {hash(tuple([])): self.guide.initial_state}
    self._seq_start_idx = None

copy()

Return a copy of the logits processor.

Source code in outlines/processors/structured.py 
143
144
145
146
147
148
149
def copy(self) -> "GuideLogitsProcessor":
    """Return a copy of the logits processor."""
    return GuideLogitsProcessor(
        tokenizer=self.tokenizer,
        guide=self.guide.copy(),
        tensor_library_name=self.tensor_adapter.library_name
    )

process_logits(input_ids, logits)

Use the Guide to bias the logits before sampling the next token.

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/processors/structured.py 
 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
def process_logits(
    self, input_ids: TensorType, logits: TensorType
) -> TensorType:
    """Use the Guide to bias the logits before sampling the next token.

    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._seq_start_idx is None:
        self._seq_start_idx = len(input_ids[0]) # type: ignore

    sequence_states: List[int] = []  # vector of states corresponding to `input_ids`

    for seq_ids in input_ids: # type: ignore
        gen_ids = seq_ids[self._seq_start_idx :]
        curr_state_key = hash(tuple(self.tensor_adapter.to_list(gen_ids)))

        if curr_state_key not in self._guide_states:
            prev_state = self._guide_states[hash(tuple(self.tensor_adapter.to_list(gen_ids[:-1])))]
            curr_state = self.guide.get_next_state(prev_state, self.tensor_adapter.to_scalar(gen_ids[-1]))
            self._guide_states[curr_state_key] = curr_state

        sequence_states.append(self._guide_states[curr_state_key])

    allowed_tokens_batch = []
    batch_indices = []
    for i, guide_state in enumerate(sequence_states):
        allowed_tokens = self.guide.get_next_instruction(guide_state).tokens
        allowed_tokens_batch.append(allowed_tokens)
        batch_indices.append(
            self.tensor_adapter.full_like(allowed_tokens, i)
        )  # Store batch index for each allowed token

    device = self.tensor_adapter.get_device(logits)
    allowed_tokens_concat = self.tensor_adapter.to_device(
        self.tensor_adapter.concatenate(allowed_tokens_batch),
        device
    )
    batch_indices_concat = self.tensor_adapter.to_device(
        self.tensor_adapter.concatenate(batch_indices),
        device
    )

    mask = self.tensor_adapter.boolean_ones_like(logits)
    mask[batch_indices_concat, allowed_tokens_concat] = False
    logits = self.tensor_adapter.apply_mask(logits, mask, float("-inf"))

    return logits

JSONLogitsProcessor

Bases: RegexLogitsProcessor

Bias generation based on a JSON schema.

Source code in outlines/processors/structured.py 
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
class JSONLogitsProcessor(RegexLogitsProcessor):
    """Bias generation based on a JSON schema."""
    def __init__(
        self,
        schema: Union[dict, Type[BaseModel], str],
        tokenizer: "Tokenizer",
        tensor_library_name: str,
        whitespace_pattern: Optional[str] = None,
    ):
        """
        Parameters
        ----------
        schema
            A JSON schema that encodes the structure we want the model to generate.
        tokenizer
            The tokenizer used to convert tokens to ids.
        tensor_library_name
            The name of the library to use to manipulate the tensors.
        whitespace_pattern
            Pattern to use for JSON syntactic whitespace (doesn't impact string
            literals). For example, to allow only a single space or newline with
            `whitespace_pattern=r"[\n ]?"`.

        """
        # Convert the JSON schema into a regex string and then pass it to the
        # RegexLogitsProcessor superclass.
        schema_str = JsonSchema(schema).schema
        regex_string = build_regex_from_schema(schema_str, whitespace_pattern)
        super().__init__(
            regex_string=regex_string,
            tokenizer=tokenizer,
            tensor_library_name=tensor_library_name
        )

__init__(schema, tokenizer, tensor_library_name, whitespace_pattern=None) 

   Parameters

   schema
       A JSON schema that encodes the structure we want the model to generate.
   tokenizer
       The tokenizer used to convert tokens to ids.
   tensor_library_name
       The name of the library to use to manipulate the tensors.
   whitespace_pattern
       Pattern to use for JSON syntactic whitespace (doesn't impact string
       literals). For example, to allow only a single space or newline with
       `whitespace_pattern=r"[

]?"`.

Source code in outlines/processors/structured.py 
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
def __init__(
    self,
    schema: Union[dict, Type[BaseModel], str],
    tokenizer: "Tokenizer",
    tensor_library_name: str,
    whitespace_pattern: Optional[str] = None,
):
    """
    Parameters
    ----------
    schema
        A JSON schema that encodes the structure we want the model to generate.
    tokenizer
        The tokenizer used to convert tokens to ids.
    tensor_library_name
        The name of the library to use to manipulate the tensors.
    whitespace_pattern
        Pattern to use for JSON syntactic whitespace (doesn't impact string
        literals). For example, to allow only a single space or newline with
        `whitespace_pattern=r"[\n ]?"`.

    """
    # Convert the JSON schema into a regex string and then pass it to the
    # RegexLogitsProcessor superclass.
    schema_str = JsonSchema(schema).schema
    regex_string = build_regex_from_schema(schema_str, whitespace_pattern)
    super().__init__(
        regex_string=regex_string,
        tokenizer=tokenizer,
        tensor_library_name=tensor_library_name
    )

RegexLogitsProcessor

Bases: GuideLogitsProcessor

Bias generation based on a regular expression.

Source code in outlines/processors/structured.py 
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
class RegexLogitsProcessor(GuideLogitsProcessor):
    """Bias generation based on a regular expression."""
    guide: RegexGuide

    def __init__(
        self,
        regex_string: str,
        tokenizer: "Tokenizer",
        tensor_library_name: str,
    ):
        """
        Parameters
        ----------
        regex_string
            A string that represents a regular expression.
        tokenizer
            An Outlines tokenizer.
        tensor_library_name
            The name of the library to use to manipulate the tensors.

        """
        # Build a guide from the regex string and then pass it to the
        # GuideLogitsProcessor superclass.
        guide = RegexGuide.from_regex(regex_string, tokenizer)
        super().__init__(tokenizer=tokenizer, guide=guide, tensor_library_name=tensor_library_name)

__init__(regex_string, tokenizer, tensor_library_name)

Parameters:

Name Type Description Default
regex_string str

A string that represents a regular expression.

 required
tokenizer Tokenizer

An Outlines tokenizer.

 required
tensor_library_name str

The name of the library to use to manipulate the tensors.

 required
Source code in outlines/processors/structured.py 
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
def __init__(
    self,
    regex_string: str,
    tokenizer: "Tokenizer",
    tensor_library_name: str,
):
    """
    Parameters
    ----------
    regex_string
        A string that represents a regular expression.
    tokenizer
        An Outlines tokenizer.
    tensor_library_name
        The name of the library to use to manipulate the tensors.

    """
    # Build a guide from the regex string and then pass it to the
    # GuideLogitsProcessor superclass.
    guide = RegexGuide.from_regex(regex_string, tokenizer)
    super().__init__(tokenizer=tokenizer, guide=guide, tensor_library_name=tensor_library_name)