Skip to content

vllm

Integration with a vLLM server.

AsyncVLLM

Bases: AsyncModel

Thin async wrapper around the openai.OpenAI client used to communicate with a vllm server.

This wrapper is used to convert the input and output types specified by the users at a higher level to arguments to the openai.OpenAI client for the vllm server.

Source code in outlines/models/vllm.py 
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
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
class AsyncVLLM(AsyncModel):
    """Thin async wrapper around the `openai.OpenAI` client used to communicate
    with a `vllm` server.

    This wrapper is used to convert the input and output types specified by the
    users at a higher level to arguments to the `openai.OpenAI` client for the
    `vllm` server.
    """

    def __init__(
        self,
        client: "AsyncOpenAI",
        model_name: Optional[str] = None,
    ):
        """
        Parameters
        ----------
        client
            An `openai.AsyncOpenAI` client instance.

        """
        self.client = client
        self.model_name = model_name
        self.type_adapter = VLLMTypeAdapter()

    async def generate(
        self,
        model_input: Union[str, Vision],
        output_type: Optional[Any] = None,
        **inference_kwargs: Any,
    ) -> Union[str, list[str]]:
        """Generate text using vLLM.

        Parameters
        ----------
        model_input
            The prompt based on which the model will generate a response.
        output_type
            The desired format of the response generated by the model. All
            output types available in Outlines are supported provided your
            server uses a structured generation backend that supports them.
        inference_kwargs
            Additional keyword arguments to pass to the client.

        Returns
        -------
        Union[str, list[str]]
            The text generated by the model.

        """
        client_args = self._build_client_args(
            model_input, output_type, **inference_kwargs,
        )

        response = await self.client.chat.completions.create(**client_args)

        messages = [choice.message for choice in response.choices]
        for message in messages:
            if message.refusal is not None:  # pragma: no cover
                raise ValueError(
                    f"The vLLM server refused to answer the request: "
                    f"{message.refusal}"
                )

        if len(messages) == 1:
            return messages[0].content
        else:
            return [message.content for message in messages]

    async def generate_stream( # type: ignore
        self,
        model_input: Union[str, Vision],
        output_type: Optional[Any] = None,
        **inference_kwargs: Any,
    ) -> AsyncIterator[str]:
        """Stream text using vLLM.

        Parameters
        ----------
        model_input
            The prompt based on which the model will generate a response.
        output_type
            The desired format of the response generated by the model. All
            output types available in Outlines are supported provided your
            server uses a structured generation backend that supports them.
        inference_kwargs
            Additional keyword arguments to pass to the client.

        Returns
        -------
        AsyncIterator[str]
            An async iterator that yields the text generated by the model.
        """
        client_args = self._build_client_args(
            model_input, output_type, **inference_kwargs,
        )

        stream = await self.client.chat.completions.create(
            **client_args,
            stream=True,
        )

        async for chunk in stream:  # pragma: no cover
            if chunk.choices and chunk.choices[0].delta.content is not None:
                yield chunk.choices[0].delta.content

    def _build_client_args(
        self,
        model_input: Union[str, Vision],
        output_type: Optional[Any] = None,
        **inference_kwargs: Any,
    ) -> dict:
        """Build the arguments to pass to the OpenAI client."""
        messages = self.type_adapter.format_input(model_input)
        output_type_args = self.type_adapter.format_output_type(output_type)
        extra_body = inference_kwargs.pop("extra_body", {})
        extra_body.update(output_type_args)

        if "model" not in inference_kwargs and self.model_name is not None:
            inference_kwargs["model"] = self.model_name

        client_args = {
            **messages,
            **inference_kwargs,
        }
        if extra_body:
            client_args["extra_body"] = extra_body

        return client_args

__init__(client, model_name=None)

Parameters:

Name Type Description Default
client AsyncOpenAI

An openai.AsyncOpenAI client instance.

 required
Source code in outlines/models/vllm.py 
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
def __init__(
    self,
    client: "AsyncOpenAI",
    model_name: Optional[str] = None,
):
    """
    Parameters
    ----------
    client
        An `openai.AsyncOpenAI` client instance.

    """
    self.client = client
    self.model_name = model_name
    self.type_adapter = VLLMTypeAdapter()

generate(model_input, output_type=None, **inference_kwargs) async

Generate text using vLLM.

Parameters:

Name Type Description Default
model_input Union[str, Vision]

The prompt based on which the model will generate a response.

 required
output_type Optional[Any]

The desired format of the response generated by the model. All output types available in Outlines are supported provided your server uses a structured generation backend that supports them.

 None
inference_kwargs Any

Additional keyword arguments to pass to the client.

 {}

Returns:

Type Description
Union[str, list[str]]

The text generated by the model.
Source code in outlines/models/vllm.py 
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
async def generate(
    self,
    model_input: Union[str, Vision],
    output_type: Optional[Any] = None,
    **inference_kwargs: Any,
) -> Union[str, list[str]]:
    """Generate text using vLLM.

    Parameters
    ----------
    model_input
        The prompt based on which the model will generate a response.
    output_type
        The desired format of the response generated by the model. All
        output types available in Outlines are supported provided your
        server uses a structured generation backend that supports them.
    inference_kwargs
        Additional keyword arguments to pass to the client.

    Returns
    -------
    Union[str, list[str]]
        The text generated by the model.

    """
    client_args = self._build_client_args(
        model_input, output_type, **inference_kwargs,
    )

    response = await self.client.chat.completions.create(**client_args)

    messages = [choice.message for choice in response.choices]
    for message in messages:
        if message.refusal is not None:  # pragma: no cover
            raise ValueError(
                f"The vLLM server refused to answer the request: "
                f"{message.refusal}"
            )

    if len(messages) == 1:
        return messages[0].content
    else:
        return [message.content for message in messages]

generate_stream(model_input, output_type=None, **inference_kwargs) async

Stream text using vLLM.

Parameters:

Name Type Description Default
model_input Union[str, Vision]

The prompt based on which the model will generate a response.

 required
output_type Optional[Any]

The desired format of the response generated by the model. All output types available in Outlines are supported provided your server uses a structured generation backend that supports them.

 None
inference_kwargs Any

Additional keyword arguments to pass to the client.

 {}

Returns:

Type Description
AsyncIterator[str]

An async iterator that yields the text generated by the model.
Source code in outlines/models/vllm.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
async def generate_stream( # type: ignore
    self,
    model_input: Union[str, Vision],
    output_type: Optional[Any] = None,
    **inference_kwargs: Any,
) -> AsyncIterator[str]:
    """Stream text using vLLM.

    Parameters
    ----------
    model_input
        The prompt based on which the model will generate a response.
    output_type
        The desired format of the response generated by the model. All
        output types available in Outlines are supported provided your
        server uses a structured generation backend that supports them.
    inference_kwargs
        Additional keyword arguments to pass to the client.

    Returns
    -------
    AsyncIterator[str]
        An async iterator that yields the text generated by the model.
    """
    client_args = self._build_client_args(
        model_input, output_type, **inference_kwargs,
    )

    stream = await self.client.chat.completions.create(
        **client_args,
        stream=True,
    )

    async for chunk in stream:  # pragma: no cover
        if chunk.choices and chunk.choices[0].delta.content is not None:
            yield chunk.choices[0].delta.content

VLLM

Bases: Model

Thin wrapper around the openai.OpenAI client used to communicate with a vllm server.

This wrapper is used to convert the input and output types specified by the users at a higher level to arguments to the openai.OpenAI client for the vllm server.

Source code in outlines/models/vllm.py 
 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
class VLLM(Model):
    """Thin wrapper around the `openai.OpenAI` client used to communicate with
    a `vllm` server.

    This wrapper is used to convert the input and output types specified by the
    users at a higher level to arguments to the `openai.OpenAI` client for the
    `vllm` server.
    """

    def __init__(
        self,
        client: "OpenAI",
        model_name: Optional[str] = None,
    ):
        """
        Parameters
        ----------
        client
            An `openai.OpenAI` client instance.

        """
        self.client = client
        self.model_name = model_name
        self.type_adapter = VLLMTypeAdapter()

    def generate(
        self,
        model_input: Union[str, Vision],
        output_type: Optional[Any] = None,
        **inference_kwargs: Any,
    ) -> Union[str, list[str]]:
        """Generate text using vLLM.

        Parameters
        ----------
        model_input
            The prompt based on which the model will generate a response.
        output_type
            The desired format of the response generated by the model. All
            output types available in Outlines are supported provided your
            server uses a structured generation backend that supports them.
        inference_kwargs
            Additional keyword arguments to pass to the client.

        Returns
        -------
        Union[str, list[str]]
            The text generated by the model.

        """
        client_args = self._build_client_args(
            model_input,
            output_type,
            **inference_kwargs,
        )

        response = self.client.chat.completions.create(**client_args)

        messages = [choice.message for choice in response.choices]
        for message in messages:
            if message.refusal is not None:  # pragma: no cover
                raise ValueError(
                    f"The vLLM server refused to answer the request: "
                    f"{message.refusal}"
                )

        if len(messages) == 1:
            return messages[0].content
        else:
            return [message.content for message in messages]

    def generate_stream(
        self,
        model_input: Union[str, Vision],
        output_type: Optional[Any] = None,
        **inference_kwargs: Any,
    ) -> Iterator[str]:
        """Stream text using vLLM.

        Parameters
        ----------
        model_input
            The prompt based on which the model will generate a response.
        output_type
            The desired format of the response generated by the model. All
            output types available in Outlines are supported provided your
            server uses a structured generation backend that supports them.
        inference_kwargs
            Additional keyword arguments to pass to the client.

        Returns
        -------
        Iterator[str]
            An iterator that yields the text generated by the model.

        """
        client_args = self._build_client_args(
            model_input, output_type, **inference_kwargs,
        )

        stream = self.client.chat.completions.create(
            **client_args, stream=True,
        )

        for chunk in stream:  # pragma: no cover
            if chunk.choices and chunk.choices[0].delta.content is not None:
                yield chunk.choices[0].delta.content

    def _build_client_args(
        self,
        model_input: Union[str, Vision],
        output_type: Optional[Any] = None,
        **inference_kwargs: Any,
    ) -> dict:
        """Build the arguments to pass to the OpenAI client."""
        messages = self.type_adapter.format_input(model_input)
        output_type_args = self.type_adapter.format_output_type(output_type)
        extra_body = inference_kwargs.pop("extra_body", {})
        extra_body.update(output_type_args)

        if "model" not in inference_kwargs and self.model_name is not None:
            inference_kwargs["model"] = self.model_name

        client_args = {
            **messages,
            **inference_kwargs,
        }
        if extra_body:
            client_args["extra_body"] = extra_body

        return client_args

__init__(client, model_name=None)

Parameters:

Name Type Description Default
client OpenAI

An openai.OpenAI client instance.

 required
Source code in outlines/models/vllm.py 
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
def __init__(
    self,
    client: "OpenAI",
    model_name: Optional[str] = None,
):
    """
    Parameters
    ----------
    client
        An `openai.OpenAI` client instance.

    """
    self.client = client
    self.model_name = model_name
    self.type_adapter = VLLMTypeAdapter()

generate(model_input, output_type=None, **inference_kwargs)

Generate text using vLLM.

Parameters:

Name Type Description Default
model_input Union[str, Vision]

The prompt based on which the model will generate a response.

 required
output_type Optional[Any]

The desired format of the response generated by the model. All output types available in Outlines are supported provided your server uses a structured generation backend that supports them.

 None
inference_kwargs Any

Additional keyword arguments to pass to the client.

 {}

Returns:

Type Description
Union[str, list[str]]

The text generated by the model.
Source code in outlines/models/vllm.py 
 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
def generate(
    self,
    model_input: Union[str, Vision],
    output_type: Optional[Any] = None,
    **inference_kwargs: Any,
) -> Union[str, list[str]]:
    """Generate text using vLLM.

    Parameters
    ----------
    model_input
        The prompt based on which the model will generate a response.
    output_type
        The desired format of the response generated by the model. All
        output types available in Outlines are supported provided your
        server uses a structured generation backend that supports them.
    inference_kwargs
        Additional keyword arguments to pass to the client.

    Returns
    -------
    Union[str, list[str]]
        The text generated by the model.

    """
    client_args = self._build_client_args(
        model_input,
        output_type,
        **inference_kwargs,
    )

    response = self.client.chat.completions.create(**client_args)

    messages = [choice.message for choice in response.choices]
    for message in messages:
        if message.refusal is not None:  # pragma: no cover
            raise ValueError(
                f"The vLLM server refused to answer the request: "
                f"{message.refusal}"
            )

    if len(messages) == 1:
        return messages[0].content
    else:
        return [message.content for message in messages]

generate_stream(model_input, output_type=None, **inference_kwargs)

Stream text using vLLM.

Parameters:

Name Type Description Default
model_input Union[str, Vision]

The prompt based on which the model will generate a response.

 required
output_type Optional[Any]

The desired format of the response generated by the model. All output types available in Outlines are supported provided your server uses a structured generation backend that supports them.

 None
inference_kwargs Any

Additional keyword arguments to pass to the client.

 {}

Returns:

Type Description
Iterator[str]

An iterator that yields the text generated by the model.
Source code in outlines/models/vllm.py 
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
def generate_stream(
    self,
    model_input: Union[str, Vision],
    output_type: Optional[Any] = None,
    **inference_kwargs: Any,
) -> Iterator[str]:
    """Stream text using vLLM.

    Parameters
    ----------
    model_input
        The prompt based on which the model will generate a response.
    output_type
        The desired format of the response generated by the model. All
        output types available in Outlines are supported provided your
        server uses a structured generation backend that supports them.
    inference_kwargs
        Additional keyword arguments to pass to the client.

    Returns
    -------
    Iterator[str]
        An iterator that yields the text generated by the model.

    """
    client_args = self._build_client_args(
        model_input, output_type, **inference_kwargs,
    )

    stream = self.client.chat.completions.create(
        **client_args, stream=True,
    )

    for chunk in stream:  # pragma: no cover
        if chunk.choices and chunk.choices[0].delta.content is not None:
            yield chunk.choices[0].delta.content

VLLMTypeAdapter

Bases: ModelTypeAdapter

Type adapter for the VLLM and AsyncVLLM models.

Source code in outlines/models/vllm.py 
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
class VLLMTypeAdapter(ModelTypeAdapter):
    """Type adapter for the `VLLM` and `AsyncVLLM` models."""

    def format_input(self, model_input: Union[str, Vision]) -> dict:
        """Generate the prompt argument to pass to the client.

        We rely on the OpenAITypeAdapter to format the input as the vLLM server
        expects input in the same format as OpenAI.

        Parameters
        ----------
        model_input
            The input passed by the user.

        Returns
        -------
        dict
            The formatted input to be passed to the model.

        """
        return OpenAITypeAdapter().format_input(model_input)

    def format_output_type(self, output_type: Optional[Any] = None) -> dict:
        """Generate the structured output argument to pass to the client.

        Parameters
        ----------
        output_type
            The structured output type provided.

        Returns
        -------
        dict
            The structured output argument to pass to the model.

        """
        if output_type is None:
            return {}

        term = python_types_to_terms(output_type)
        if isinstance(term, CFG):
            return {"guided_grammar": term.definition}
        elif isinstance(term, JsonSchema):
            extra_body = {"guided_json": json.loads(term.schema)}
            if term.whitespace_pattern:
                extra_body["whitespace_pattern"] = term.whitespace_pattern
            return extra_body
        else:
            return {"guided_regex": to_regex(term)}

format_input(model_input)

Generate the prompt argument to pass to the client.

We rely on the OpenAITypeAdapter to format the input as the vLLM server expects input in the same format as OpenAI.

Parameters:

Name Type Description Default
model_input Union[str, Vision]

The input passed by the user.

 required

Returns:

Type Description
dict

The formatted input to be passed to the model.
Source code in outlines/models/vllm.py 
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
def format_input(self, model_input: Union[str, Vision]) -> dict:
    """Generate the prompt argument to pass to the client.

    We rely on the OpenAITypeAdapter to format the input as the vLLM server
    expects input in the same format as OpenAI.

    Parameters
    ----------
    model_input
        The input passed by the user.

    Returns
    -------
    dict
        The formatted input to be passed to the model.

    """
    return OpenAITypeAdapter().format_input(model_input)

format_output_type(output_type=None)

Generate the structured output argument to pass to the client.

Parameters:

Name Type Description Default
output_type Optional[Any]

The structured output type provided.

 None

Returns:

Type Description
dict

The structured output argument to pass to the model.
Source code in outlines/models/vllm.py 
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
def format_output_type(self, output_type: Optional[Any] = None) -> dict:
    """Generate the structured output argument to pass to the client.

    Parameters
    ----------
    output_type
        The structured output type provided.

    Returns
    -------
    dict
        The structured output argument to pass to the model.

    """
    if output_type is None:
        return {}

    term = python_types_to_terms(output_type)
    if isinstance(term, CFG):
        return {"guided_grammar": term.definition}
    elif isinstance(term, JsonSchema):
        extra_body = {"guided_json": json.loads(term.schema)}
        if term.whitespace_pattern:
            extra_body["whitespace_pattern"] = term.whitespace_pattern
        return extra_body
    else:
        return {"guided_regex": to_regex(term)}

from_vllm(client, model_name=None)

Create an Outlines VLLM or AsyncVLLM model instance from an openai.OpenAI or openai.AsyncOpenAI instance.

Parameters:

Name Type Description Default
client Union[OpenAI, AsyncOpenAI]

An openai.OpenAI or openai.AsyncOpenAI instance.

 required
model_name Optional[str]

The name of the model to use.

 None

Returns:

Type Description
Union[VLLM, AsyncVLLM]

An Outlines VLLM or AsyncVLLM model instance.
Source code in outlines/models/vllm.py 
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
def from_vllm(
    client: Union["OpenAI", "AsyncOpenAI"],
    model_name: Optional[str] = None,
) -> Union[VLLM, AsyncVLLM]:
    """Create an Outlines `VLLM` or `AsyncVLLM` model instance from an
    `openai.OpenAI` or `openai.AsyncOpenAI` instance.

    Parameters
    ----------
    client
        An `openai.OpenAI` or `openai.AsyncOpenAI` instance.
    model_name
        The name of the model to use.

    Returns
    -------
    Union[VLLM, AsyncVLLM]
        An Outlines `VLLM` or `AsyncVLLM` model instance.

    """
    from openai import AsyncOpenAI, OpenAI

    if isinstance(client, OpenAI):
        return VLLM(client, model_name)
    elif isinstance(client, AsyncOpenAI):
        return AsyncVLLM(client, model_name)
    else:
        raise ValueError(
            f"Unsupported client type: {type(client)}.\n"
            "Please provide an OpenAI or AsyncOpenAI instance."
        )