Skip to content

Text generation inference

TextGenerationInference #

Bases: FunctionCallingLLM

Source code in llama-index-integrations/llms/llama-index-llms-text-generation-inference/llama_index/llms/text_generation_inference/base.py
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
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
330
331
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
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
class TextGenerationInference(FunctionCallingLLM):
    model_name: Optional[str] = Field(
        default=None,
        description=("The name of the model served at the TGI endpoint"),
    )
    temperature: float = Field(
        default=DEFAULT_TEMPERATURE,
        description=("The temperature to use for sampling."),
        gte=0.0,
        lte=1.0,
    )
    max_tokens: int = Field(
        default=DEFAULT_NUM_OUTPUTS,
        description=("The maximum number of tokens to generate."),
        gt=0,
    )
    token: Union[str, bool, None] = Field(
        default=None,
        description=(
            "Hugging Face token. Will default to the locally saved token. Pass "
            "token=False if you don’t want to send your token to the server."
        ),
    )
    timeout: float = Field(
        default=120, description=("The timeout to use in seconds."), gte=0
    )
    max_retries: int = Field(
        default=5, description=("The maximum number of API retries."), gte=0
    )
    headers: Optional[Dict[str, str]] = Field(
        default=None,
        description=(
            "Additional headers to send to the server. By default only the"
            " authorization headers are sent. Values in this dictionary"
            " will override the default values."
        ),
    )
    cookies: Optional[Dict[str, str]] = Field(
        default=None, description=("Additional cookies to send to the server.")
    )
    seed: Optional[str] = Field(
        default=None, description=("The random seed to use for sampling.")
    )
    additional_kwargs: Dict[str, Any] = Field(
        default_factory=dict, description=("Additional kwargs for the TGI API.")
    )

    _sync_client: "TGIClient" = PrivateAttr()
    _async_client: "TGIAsyncClient" = PrivateAttr()

    context_window: int = Field(
        default=DEFAULT_CONTEXT_WINDOW,
        description=("Maximum input length in tokens returned from TGI endpoint"),
    )
    is_chat_model: bool = Field(
        default=True,
        description=(
            LLMMetadata.__fields__["is_chat_model"].field_info.description
            + " TGI makes use of chat templating,"
            " function call is available only for '/v1/chat/completions' route"
            " of TGI endpoint"
        ),
    )
    is_function_calling_model: bool = Field(
        default=False,
        description=(
            LLMMetadata.__fields__["is_function_calling_model"].field_info.description
            + " 'text-generation-inference' supports function call"
            " starting from v1.4.3"
        ),
    )

    def __init__(
        self,
        model_url,
        model_name: Optional[str] = None,
        cookies: Optional[dict] = None,
        temperature: float = DEFAULT_TEMPERATURE,
        max_tokens: int = DEFAULT_NUM_OUTPUTS,
        timeout: int = 120,
        max_retries: int = 5,
        seed: Optional[int] = None,
        token: Optional[str] = None,
        additional_kwargs: Optional[Dict[str, Any]] = None,
        callback_manager: Optional[CallbackManager] = None,
        system_prompt: Optional[str] = None,
        messages_to_prompt: Optional[Callable[[Sequence[ChatMessage]], str]] = None,
        completion_to_prompt: Optional[Callable[[str], str]] = None,
        pydantic_program_mode: PydanticProgramMode = PydanticProgramMode.DEFAULT,
        output_parser: Optional[BaseOutputParser] = None,
    ) -> None:
        additional_kwargs = additional_kwargs or {}
        callback_manager = callback_manager or CallbackManager([])

        token = get_from_param_or_env("token", token, "HF_TOKEN", "")

        headers = {}
        if token:
            headers.update({"Authorization": f"Bearer {token}"})

        self._sync_client = TGIClient(
            base_url=model_url,
            headers=headers,
            cookies=cookies,
            timeout=timeout,
        )
        self._async_client = TGIAsyncClient(
            base_url=model_url,
            headers=headers,
            cookies=cookies,
            timeout=timeout,
        )

        try:
            is_function_calling_model = resolve_tgi_function_call(model_url)
        except Exception as e:
            logger.warning(f"TGI client has no function call support: {e}")
            is_function_calling_model = False

        context_window = get_max_input_length(model_url) or DEFAULT_CONTEXT_WINDOW

        super().__init__(
            context_window=context_window,
            temperature=temperature,
            max_tokens=max_tokens,
            additional_kwargs=additional_kwargs,
            timeout=timeout,
            max_retries=max_retries,
            seed=seed,
            model_name=model_name,
            is_function_calling_model=is_function_calling_model,
            callback_manager=callback_manager,
            system_prompt=system_prompt,
            messages_to_prompt=messages_to_prompt,
            completion_to_prompt=completion_to_prompt,
            pydantic_program_mode=pydantic_program_mode,
            output_parser=output_parser,
        )

    @classmethod
    def class_name(cls) -> str:
        return "TextGenerationInference"

    @property
    def metadata(self) -> LLMMetadata:
        return LLMMetadata(
            context_window=self.context_window,
            num_output=self.max_tokens,
            is_chat_model=True,
            model_name=self.model_name,
            random_seed=self.seed,
            is_function_calling_model=self.is_function_calling_model,
        )

    @property
    def _model_kwargs(self) -> Dict[str, Any]:
        base_kwargs = {
            "temperature": self.temperature,
            "max_tokens": self.max_tokens,
            "seed": self.seed,
        }
        return {
            **base_kwargs,
            **self.additional_kwargs,
        }

    def _get_all_kwargs(self, **kwargs: Any) -> Dict[str, Any]:
        return {
            **self._model_kwargs,
            **kwargs,
        }

    @llm_chat_callback()
    def chat(self, messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponse:
        # convert to TGI Message
        messages = to_tgi_messages(messages)
        all_kwargs = self._get_all_kwargs(**kwargs)
        response = self._sync_client.chat(messages=messages, **all_kwargs)
        tool_calls = response.choices[0].message.tool_calls

        return ChatResponse(
            message=ChatMessage(
                role=MessageRole.ASSISTANT,
                content=response.choices[0].message.content,
                additional_kwargs=(
                    {"tool_calls": tool_calls} if tool_calls is not None else {}
                ),
            ),
            raw=dict(response),
        )

    @llm_completion_callback()
    def complete(
        self, prompt: str, formatted: bool = False, **kwargs: Any
    ) -> CompletionResponse:
        complete_fn = chat_to_completion_decorator(self.chat)
        return complete_fn(prompt, **kwargs)

    @llm_chat_callback()
    def stream_chat(
        self, messages: Sequence[ChatMessage], **kwargs: Any
    ) -> ChatResponseGen:
        # convert to TGI Message
        messages = to_tgi_messages(messages)
        all_kwargs = self._get_all_kwargs(**kwargs)
        response = self._sync_client.chat(messages=messages, stream=True, **all_kwargs)

        def generator() -> ChatResponseGen:
            content = ""
            role = MessageRole.ASSISTANT
            for chunk in response:
                content_delta = chunk.choices[0].delta.content
                if content_delta is None:
                    continue
                content += content_delta
                yield ChatResponse(
                    message=ChatMessage(role=role, content=content),
                    delta=content_delta,
                    raw=chunk,
                )

        return generator()

    @llm_completion_callback()
    def stream_complete(
        self, prompt: str, formatted: bool = False, **kwargs: Any
    ) -> CompletionResponseGen:
        stream_complete_fn = stream_chat_to_completion_decorator(self.stream_chat)
        return stream_complete_fn(prompt, **kwargs)

    @llm_chat_callback()
    async def achat(
        self, messages: Sequence[ChatMessage], **kwargs: Any
    ) -> ChatResponse:
        # convert to TGI Message
        messages = to_tgi_messages(messages)
        all_kwargs = self._get_all_kwargs(**kwargs)
        response = await self._async_client.chat(messages=messages, **all_kwargs)
        tool_calls = response.choices[0].message.tool_calls

        return ChatResponse(
            message=ChatMessage(
                role=MessageRole.ASSISTANT,
                content=response.choices[0].message.content,
                additional_kwargs=(
                    {"tool_calls": tool_calls} if tool_calls is not None else {}
                ),
            ),
            raw=dict(response),
        )

    @llm_completion_callback()
    async def acomplete(
        self, prompt: str, formatted: bool = False, **kwargs: Any
    ) -> CompletionResponse:
        acomplete_fn = achat_to_completion_decorator(self.achat)
        return await acomplete_fn(prompt, **kwargs)

    @llm_chat_callback()
    async def astream_chat(
        self, messages: Sequence[ChatMessage], **kwargs: Any
    ) -> ChatResponseAsyncGen:
        # convert to TGI Message
        messages = to_tgi_messages(messages)
        all_kwargs = self._get_all_kwargs(**kwargs)
        response = await self._async_client.chat(
            messages=messages, stream=True, **all_kwargs
        )

        async def generator() -> ChatResponseAsyncGen:
            content = ""
            role = MessageRole.ASSISTANT
            async for chunk in response:
                content_delta = chunk.choices[0].delta.content
                if content_delta is None:
                    continue
                content += content_delta
                yield ChatResponse(
                    message=ChatMessage(role=role, content=content),
                    delta=content_delta,
                    raw=chunk,
                )

        return generator()

    @llm_completion_callback()
    async def astream_complete(
        self, prompt: str, formatted: bool = False, **kwargs: Any
    ) -> CompletionResponseAsyncGen:
        astream_complete_fn = astream_chat_to_completion_decorator(self.astream_chat)
        return await astream_complete_fn(prompt, **kwargs)

    def chat_with_tools(
        self,
        tools: List["BaseTool"],
        user_msg: Optional[Union[str, ChatMessage]] = None,
        chat_history: Optional[List[ChatMessage]] = None,
        verbose: bool = False,
        allow_parallel_tool_calls: bool = False,
        tool_choice: str = "auto",
        **kwargs: Any,
    ) -> ChatResponse:
        """Predict and call the tool."""
        # use openai tool format
        tool_specs = [
            tool.metadata.to_openai_tool(skip_length_check=True) for tool in tools
        ]

        if isinstance(user_msg, str):
            user_msg = ChatMessage(role=MessageRole.USER, content=user_msg)

        messages = chat_history or []
        if user_msg:
            messages.append(user_msg)

        response = self.chat(
            messages=messages,
            tools=tool_specs or None,
            tool_choice=resolve_tool_choice(tool_specs, tool_choice),
            **kwargs,
        )
        if not allow_parallel_tool_calls:
            force_single_tool_call(response)
        return response

    async def achat_with_tools(
        self,
        tools: List["BaseTool"],
        user_msg: Optional[Union[str, ChatMessage]] = None,
        chat_history: Optional[List[ChatMessage]] = None,
        verbose: bool = False,
        allow_parallel_tool_calls: bool = False,
        tool_choice: str = "auto",
        **kwargs: Any,
    ) -> ChatResponse:
        # use openai tool format
        tool_specs = [
            tool.metadata.to_openai_tool(skip_length_check=True) for tool in tools
        ]

        if isinstance(user_msg, str):
            user_msg = ChatMessage(role=MessageRole.USER, content=user_msg)

        messages = chat_history or []
        if user_msg:
            messages.append(user_msg)

        response = self.achat(
            messages=messages,
            tools=tool_specs or None,
            tool_choice=resolve_tool_choice(tool_specs, tool_choice),
            **kwargs,
        )
        if not allow_parallel_tool_calls:
            force_single_tool_call(response)
        return response

    def get_tool_calls_from_response(
        self,
        response: "AgentChatResponse",
        error_on_no_tool_call: bool = True,
    ) -> List[ToolSelection]:
        """Predict and call the tool."""
        tool_calls = response.message.additional_kwargs.get("tool_calls", [])

        if len(tool_calls) < 1:
            if error_on_no_tool_call:
                raise ValueError(
                    f"Expected at least one tool call, but got {len(tool_calls)} tool calls."
                )
            else:
                return []

        tool_selections = []
        for tool_call in tool_calls:
            # TODO Add typecheck with ToolCall from TGI once the client is updated
            if tool_call and (tc_type := tool_call["type"]) != "function":
                raise ValueError(
                    f"Invalid tool type: got {tc_type}, expect 'function'."
                )
            argument_dict = tool_call["function"]["parameters"]

            tool_selections.append(
                ToolSelection(
                    tool_id=tool_call["id"],
                    tool_name=tool_call["function"][
                        "name"
                    ],  # NOTE for now the tool_name is hardcoded 'tools' in TGI
                    tool_kwargs=argument_dict,
                )
            )

        return tool_selections

chat_with_tools #

chat_with_tools(tools: List[BaseTool], user_msg: Optional[Union[str, ChatMessage]] = None, chat_history: Optional[List[ChatMessage]] = None, verbose: bool = False, allow_parallel_tool_calls: bool = False, tool_choice: str = 'auto', **kwargs: Any) -> ChatResponse

Predict and call the tool.

Source code in llama-index-integrations/llms/llama-index-llms-text-generation-inference/llama_index/llms/text_generation_inference/base.py
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
def chat_with_tools(
    self,
    tools: List["BaseTool"],
    user_msg: Optional[Union[str, ChatMessage]] = None,
    chat_history: Optional[List[ChatMessage]] = None,
    verbose: bool = False,
    allow_parallel_tool_calls: bool = False,
    tool_choice: str = "auto",
    **kwargs: Any,
) -> ChatResponse:
    """Predict and call the tool."""
    # use openai tool format
    tool_specs = [
        tool.metadata.to_openai_tool(skip_length_check=True) for tool in tools
    ]

    if isinstance(user_msg, str):
        user_msg = ChatMessage(role=MessageRole.USER, content=user_msg)

    messages = chat_history or []
    if user_msg:
        messages.append(user_msg)

    response = self.chat(
        messages=messages,
        tools=tool_specs or None,
        tool_choice=resolve_tool_choice(tool_specs, tool_choice),
        **kwargs,
    )
    if not allow_parallel_tool_calls:
        force_single_tool_call(response)
    return response

get_tool_calls_from_response #

get_tool_calls_from_response(response: AgentChatResponse, error_on_no_tool_call: bool = True) -> List[ToolSelection]

Predict and call the tool.

Source code in llama-index-integrations/llms/llama-index-llms-text-generation-inference/llama_index/llms/text_generation_inference/base.py
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
def get_tool_calls_from_response(
    self,
    response: "AgentChatResponse",
    error_on_no_tool_call: bool = True,
) -> List[ToolSelection]:
    """Predict and call the tool."""
    tool_calls = response.message.additional_kwargs.get("tool_calls", [])

    if len(tool_calls) < 1:
        if error_on_no_tool_call:
            raise ValueError(
                f"Expected at least one tool call, but got {len(tool_calls)} tool calls."
            )
        else:
            return []

    tool_selections = []
    for tool_call in tool_calls:
        # TODO Add typecheck with ToolCall from TGI once the client is updated
        if tool_call and (tc_type := tool_call["type"]) != "function":
            raise ValueError(
                f"Invalid tool type: got {tc_type}, expect 'function'."
            )
        argument_dict = tool_call["function"]["parameters"]

        tool_selections.append(
            ToolSelection(
                tool_id=tool_call["id"],
                tool_name=tool_call["function"][
                    "name"
                ],  # NOTE for now the tool_name is hardcoded 'tools' in TGI
                tool_kwargs=argument_dict,
            )
        )

    return tool_selections