15276913207

Committed 27 May 2025 01:41PM UTC coverage: 90.41% (+0.02%) from 90.388%

Build # 15276913207

Build Type

Pull #9449

github

Committed by

web-flow

Commit Message

Merge f55ee47c3 into 3deaa20cb

Pull Request Pull Request #9449: refactor: Refactor hf api chat generator

Run Details

11464 of 12680 relevant lines covered (90.41%)

0.9 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

96.3

haystack/components/generators/hugging_face_api.py

# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
#
# SPDX-License-Identifier: Apache-2.0

from dataclasses import asdict
from datetime import datetime
from typing import Any, Dict, Iterable, List, Optional, Union, cast

from haystack import component, default_from_dict, default_to_dict
from haystack.dataclasses import ComponentInfo, StreamingCallbackT, StreamingChunk, select_streaming_callback
from haystack.lazy_imports import LazyImport
from haystack.utils import Secret, deserialize_callable, deserialize_secrets_inplace, serialize_callable
from haystack.utils.hf import HFGenerationAPIType, HFModelType, check_valid_model
from haystack.utils.url_validation import is_valid_http_url

with LazyImport(message="Run 'pip install \"huggingface_hub>=0.27.0\"'") as huggingface_hub_import:
    from huggingface_hub import (
        InferenceClient,
        TextGenerationOutput,
        TextGenerationStreamOutput,
        TextGenerationStreamOutputToken,
    )


@component
class HuggingFaceAPIGenerator:
    """
    Generates text using Hugging Face APIs.

    Use it with the following Hugging Face APIs:
    - [Free Serverless Inference API]((https://huggingface.co/inference-api)
    - [Paid Inference Endpoints](https://huggingface.co/inference-endpoints)
    - [Self-hosted Text Generation Inference](https://github.com/huggingface/text-generation-inference)

    ### Usage examples

    #### With the free serverless inference API

    ```python
    from haystack.components.generators import HuggingFaceAPIGenerator
    from haystack.utils import Secret

    generator = HuggingFaceAPIGenerator(api_type="serverless_inference_api",
                                        api_params={"model": "HuggingFaceH4/zephyr-7b-beta"},
                                        token=Secret.from_token("<your-api-key>"))

    result = generator.run(prompt="What's Natural Language Processing?")
    print(result)
    ```

    #### With paid inference endpoints

    ```python
    from haystack.components.generators import HuggingFaceAPIGenerator
    from haystack.utils import Secret

    generator = HuggingFaceAPIGenerator(api_type="inference_endpoints",
                                        api_params={"url": "<your-inference-endpoint-url>"},
                                        token=Secret.from_token("<your-api-key>"))

    result = generator.run(prompt="What's Natural Language Processing?")
    print(result)

    #### With self-hosted text generation inference
    ```python
    from haystack.components.generators import HuggingFaceAPIGenerator

    generator = HuggingFaceAPIGenerator(api_type="text_generation_inference",
                                        api_params={"url": "http://localhost:8080"})

    result = generator.run(prompt="What's Natural Language Processing?")
    print(result)
    ```
    """

    def __init__(  # pylint: disable=too-many-positional-arguments
        self,
        api_type: Union[HFGenerationAPIType, str],
        api_params: Dict[str, str],
        token: Optional[Secret] = Secret.from_env_var(["HF_API_TOKEN", "HF_TOKEN"], strict=False),
        generation_kwargs: Optional[Dict[str, Any]] = None,
        stop_words: Optional[List[str]] = None,
        streaming_callback: Optional[StreamingCallbackT] = None,
    ):
        """
        Initialize the HuggingFaceAPIGenerator instance.

        :param api_type:
            The type of Hugging Face API to use. Available types:
            - `text_generation_inference`: See [TGI](https://github.com/huggingface/text-generation-inference).
            - `inference_endpoints`: See [Inference Endpoints](https://huggingface.co/inference-endpoints).
            - `serverless_inference_api`: See [Serverless Inference API](https://huggingface.co/inference-api).
        :param api_params:
            A dictionary with the following keys:
            - `model`: Hugging Face model ID. Required when `api_type` is `SERVERLESS_INFERENCE_API`.
            - `url`: URL of the inference endpoint. Required when `api_type` is `INFERENCE_ENDPOINTS` or
            `TEXT_GENERATION_INFERENCE`.
        :param token: The Hugging Face token to use as HTTP bearer authorization.
            Check your HF token in your [account settings](https://huggingface.co/settings/tokens).
        :param generation_kwargs:
            A dictionary with keyword arguments to customize text generation. Some examples: `max_new_tokens`,
            `temperature`, `top_k`, `top_p`.
            For details, see [Hugging Face documentation](https://huggingface.co/docs/huggingface_hub/en/package_reference/inference_client#huggingface_hub.InferenceClient.text_generation)
            for more information.
        :param stop_words: An optional list of strings representing the stop words.
        :param streaming_callback: An optional callable for handling streaming responses.
        """

        huggingface_hub_import.check()

        if isinstance(api_type, str):
            api_type = HFGenerationAPIType.from_str(api_type)

        if api_type == HFGenerationAPIType.SERVERLESS_INFERENCE_API:
            model = api_params.get("model")
            if model is None:
                raise ValueError(
                    "To use the Serverless Inference API, you need to specify the `model` parameter in `api_params`."
                )
            check_valid_model(model, HFModelType.GENERATION, token)
            model_or_url = model
        elif api_type in [HFGenerationAPIType.INFERENCE_ENDPOINTS, HFGenerationAPIType.TEXT_GENERATION_INFERENCE]:
            url = api_params.get("url")
            if url is None:
                msg = (
                    "To use Text Generation Inference or Inference Endpoints, you need to specify the `url` "
                    "parameter in `api_params`."
                )
                raise ValueError(msg)
            if not is_valid_http_url(url):
                raise ValueError(f"Invalid URL: {url}")
            model_or_url = url
        else:
            msg = f"Unknown api_type {api_type}"
            raise ValueError(msg)

        # handle generation kwargs setup
        generation_kwargs = generation_kwargs.copy() if generation_kwargs else {}
        generation_kwargs["stop_sequences"] = generation_kwargs.get("stop_sequences", [])
        generation_kwargs["stop_sequences"].extend(stop_words or [])
        generation_kwargs.setdefault("max_new_tokens", 512)

        self.api_type = api_type
        self.api_params = api_params
        self.token = token
        self.generation_kwargs = generation_kwargs
        self.streaming_callback = streaming_callback
        self._client = InferenceClient(model_or_url, token=token.resolve_value() if token else None)

    def to_dict(self) -> Dict[str, Any]:
        """
        Serialize this component to a dictionary.

        :returns:
            A dictionary containing the serialized component.
        """
        callback_name = serialize_callable(self.streaming_callback) if self.streaming_callback else None
        return default_to_dict(
            self,
            api_type=str(self.api_type),
            api_params=self.api_params,
            token=self.token.to_dict() if self.token else None,
            generation_kwargs=self.generation_kwargs,
            streaming_callback=callback_name,
        )

    @classmethod
    def from_dict(cls, data: Dict[str, Any]) -> "HuggingFaceAPIGenerator":
        """
        Deserialize this component from a dictionary.
        """
        deserialize_secrets_inplace(data["init_parameters"], keys=["token"])
        init_params = data["init_parameters"]
        serialized_callback_handler = init_params.get("streaming_callback")
        if serialized_callback_handler:
            init_params["streaming_callback"] = deserialize_callable(serialized_callback_handler)
        return default_from_dict(cls, data)

    @component.output_types(replies=List[str], meta=List[Dict[str, Any]])
    def run(
        self,
        prompt: str,
        streaming_callback: Optional[StreamingCallbackT] = None,
        generation_kwargs: Optional[Dict[str, Any]] = None,
    ):
        """
        Invoke the text generation inference for the given prompt and generation parameters.

        :param prompt:
            A string representing the prompt.
        :param streaming_callback:
            A callback function that is called when a new token is received from the stream.
        :param generation_kwargs:
            Additional keyword arguments for text generation.
        :returns:
            A dictionary with the generated replies and metadata. Both are lists of length n.
            - replies: A list of strings representing the generated replies.
        """
        # update generation kwargs by merging with the default ones
        generation_kwargs = {**self.generation_kwargs, **(generation_kwargs or {})}

        # check if streaming_callback is passed
        streaming_callback = select_streaming_callback(
            init_callback=self.streaming_callback, runtime_callback=streaming_callback, requires_async=False
        )

        hf_output = self._client.text_generation(
            prompt, details=True, stream=streaming_callback is not None, **generation_kwargs
        )

        if streaming_callback is not None:
            return self._stream_and_build_response(hf_output, streaming_callback)

        # mypy doesn't know that hf_output is a TextGenerationOutput, so we cast it
        return self._build_non_streaming_response(cast(TextGenerationOutput, hf_output))

    def _stream_and_build_response(
        self, hf_output: Iterable["TextGenerationStreamOutput"], streaming_callback: StreamingCallbackT
    ):
        chunks: List[StreamingChunk] = []
        first_chunk_time = None

        component_info = ComponentInfo.from_component(self)
        for chunk in hf_output:
            token: TextGenerationStreamOutputToken = chunk.token
            if token.special:
                continue

            chunk_metadata = {**asdict(token), **(asdict(chunk.details) if chunk.details else {})}
            if first_chunk_time is None:
                first_chunk_time = datetime.now().isoformat()

            stream_chunk = StreamingChunk(content=token.text, meta=chunk_metadata, component_info=component_info)
            chunks.append(stream_chunk)
            streaming_callback(stream_chunk)

        metadata = {
            "finish_reason": chunks[-1].meta.get("finish_reason", None),
            "model": self._client.model,
            "usage": {"completion_tokens": chunks[-1].meta.get("generated_tokens", 0)},
            "completion_start_time": first_chunk_time,
        }
        return {"replies": ["".join([chunk.content for chunk in chunks])], "meta": [metadata]}

    def _build_non_streaming_response(self, hf_output: "TextGenerationOutput"):
        meta = [
            {
                "model": self._client.model,
                "finish_reason": hf_output.details.finish_reason if hf_output.details else None,
                "usage": {"completion_tokens": len(hf_output.details.tokens) if hf_output.details else 0},
            }
        ]
        return {"replies": [hf_output.generated_text], "meta": meta}

1	# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
2	#
3	# SPDX-License-Identifier: Apache-2.0
4
5	from dataclasses import asdict	1✔
6	from datetime import datetime	1✔
7	from typing import Any, Dict, Iterable, List, Optional, Union, cast	1✔
8
9	from haystack import component, default_from_dict, default_to_dict	1✔
10	from haystack.dataclasses import ComponentInfo, StreamingCallbackT, StreamingChunk, select_streaming_callback	1✔
11	from haystack.lazy_imports import LazyImport	1✔
12	from haystack.utils import Secret, deserialize_callable, deserialize_secrets_inplace, serialize_callable	1✔
13	from haystack.utils.hf import HFGenerationAPIType, HFModelType, check_valid_model	1✔
14	from haystack.utils.url_validation import is_valid_http_url	1✔
15
16	with LazyImport(message="Run 'pip install \"huggingface_hub>=0.27.0\"'") as huggingface_hub_import:	1✔
17	from huggingface_hub import (	1✔
18	InferenceClient,
19	TextGenerationOutput,
20	TextGenerationStreamOutput,
21	TextGenerationStreamOutputToken,
22	)
23
24
25	@component	1✔
26	class HuggingFaceAPIGenerator:	1✔
27	"""
28	Generates text using Hugging Face APIs.
29
30	Use it with the following Hugging Face APIs:
31	- [Free Serverless Inference API]((https://huggingface.co/inference-api)
32	- [Paid Inference Endpoints](https://huggingface.co/inference-endpoints)
33	- [Self-hosted Text Generation Inference](https://github.com/huggingface/text-generation-inference)
34
35	### Usage examples
36
37	#### With the free serverless inference API
38
39	```python
40	from haystack.components.generators import HuggingFaceAPIGenerator
41	from haystack.utils import Secret
42
43	generator = HuggingFaceAPIGenerator(api_type="serverless_inference_api",
44	api_params={"model": "HuggingFaceH4/zephyr-7b-beta"},
45	token=Secret.from_token("<your-api-key>"))
46
47	result = generator.run(prompt="What's Natural Language Processing?")
48	print(result)
49	```
50
51	#### With paid inference endpoints
52
53	```python
54	from haystack.components.generators import HuggingFaceAPIGenerator
55	from haystack.utils import Secret
56
57	generator = HuggingFaceAPIGenerator(api_type="inference_endpoints",
58	api_params={"url": "<your-inference-endpoint-url>"},
59	token=Secret.from_token("<your-api-key>"))
60
61	result = generator.run(prompt="What's Natural Language Processing?")
62	print(result)
63
64	#### With self-hosted text generation inference
65	```python
66	from haystack.components.generators import HuggingFaceAPIGenerator
67
68	generator = HuggingFaceAPIGenerator(api_type="text_generation_inference",
69	api_params={"url": "http://localhost:8080"})
70
71	result = generator.run(prompt="What's Natural Language Processing?")
72	print(result)
73	```
74	"""
75
76	def __init__( # pylint: disable=too-many-positional-arguments	1✔
77	self,
78	api_type: Union[HFGenerationAPIType, str],
79	api_params: Dict[str, str],
80	token: Optional[Secret] = Secret.from_env_var(["HF_API_TOKEN", "HF_TOKEN"], strict=False),
81	generation_kwargs: Optional[Dict[str, Any]] = None,
82	stop_words: Optional[List[str]] = None,
83	streaming_callback: Optional[StreamingCallbackT] = None,
84	):
85	"""
86	Initialize the HuggingFaceAPIGenerator instance.
87
88	:param api_type:
89	The type of Hugging Face API to use. Available types:
90	- `text_generation_inference`: See [TGI](https://github.com/huggingface/text-generation-inference).
91	- `inference_endpoints`: See [Inference Endpoints](https://huggingface.co/inference-endpoints).
92	- `serverless_inference_api`: See [Serverless Inference API](https://huggingface.co/inference-api).
93	:param api_params:
94	A dictionary with the following keys:
95	- `model`: Hugging Face model ID. Required when `api_type` is `SERVERLESS_INFERENCE_API`.
96	- `url`: URL of the inference endpoint. Required when `api_type` is `INFERENCE_ENDPOINTS` or
97	`TEXT_GENERATION_INFERENCE`.
98	:param token: The Hugging Face token to use as HTTP bearer authorization.
99	Check your HF token in your [account settings](https://huggingface.co/settings/tokens).
100	:param generation_kwargs:
101	A dictionary with keyword arguments to customize text generation. Some examples: `max_new_tokens`,
102	`temperature`, `top_k`, `top_p`.
103	For details, see [Hugging Face documentation](https://huggingface.co/docs/huggingface_hub/en/package_reference/inference_client#huggingface_hub.InferenceClient.text_generation)
104	for more information.
105	:param stop_words: An optional list of strings representing the stop words.
106	:param streaming_callback: An optional callable for handling streaming responses.
107	"""
108
109	huggingface_hub_import.check()	1✔
110
111	if isinstance(api_type, str):	1✔
112	api_type = HFGenerationAPIType.from_str(api_type)	1✔
113
114	if api_type == HFGenerationAPIType.SERVERLESS_INFERENCE_API:	1✔
115	model = api_params.get("model")	1✔
116	if model is None:	1✔
117	raise ValueError(	1✔
118	"To use the Serverless Inference API, you need to specify the `model` parameter in `api_params`."
119	)
120	check_valid_model(model, HFModelType.GENERATION, token)	1✔
121	model_or_url = model	1✔
122	elif api_type in [HFGenerationAPIType.INFERENCE_ENDPOINTS, HFGenerationAPIType.TEXT_GENERATION_INFERENCE]:	1✔
123	url = api_params.get("url")	1✔
124	if url is None:	1✔
125	msg = (	1✔
126	"To use Text Generation Inference or Inference Endpoints, you need to specify the `url` "
127	"parameter in `api_params`."
128	)
129	raise ValueError(msg)	1✔
130	if not is_valid_http_url(url):	1✔
131	raise ValueError(f"Invalid URL: {url}")	1✔
132	model_or_url = url	1✔
133	else:
134	msg = f"Unknown api_type {api_type}"	×
135	raise ValueError(msg)	×
136
137	# handle generation kwargs setup
138	generation_kwargs = generation_kwargs.copy() if generation_kwargs else {}	1✔
139	generation_kwargs["stop_sequences"] = generation_kwargs.get("stop_sequences", [])	1✔
140	generation_kwargs["stop_sequences"].extend(stop_words or [])	1✔
141	generation_kwargs.setdefault("max_new_tokens", 512)	1✔
142
143	self.api_type = api_type	1✔
144	self.api_params = api_params	1✔
145	self.token = token	1✔
146	self.generation_kwargs = generation_kwargs	1✔
147	self.streaming_callback = streaming_callback	1✔
148	self._client = InferenceClient(model_or_url, token=token.resolve_value() if token else None)	1✔
149
150	def to_dict(self) -> Dict[str, Any]:	1✔
151	"""
152	Serialize this component to a dictionary.
153
154	:returns:
155	A dictionary containing the serialized component.
156	"""
157	callback_name = serialize_callable(self.streaming_callback) if self.streaming_callback else None	1✔
158	return default_to_dict(	1✔
159	self,
160	api_type=str(self.api_type),
161	api_params=self.api_params,
162	token=self.token.to_dict() if self.token else None,
163	generation_kwargs=self.generation_kwargs,
164	streaming_callback=callback_name,
165	)
166
167	@classmethod	1✔
168	def from_dict(cls, data: Dict[str, Any]) -> "HuggingFaceAPIGenerator":	1✔
169	"""
170	Deserialize this component from a dictionary.
171	"""
172	deserialize_secrets_inplace(data["init_parameters"], keys=["token"])	1✔
173	init_params = data["init_parameters"]	1✔
174	serialized_callback_handler = init_params.get("streaming_callback")	1✔
175	if serialized_callback_handler:	1✔
176	init_params["streaming_callback"] = deserialize_callable(serialized_callback_handler)	1✔
177	return default_from_dict(cls, data)	1✔
178
179	@component.output_types(replies=List[str], meta=List[Dict[str, Any]])	1✔
180	def run(	1✔
181	self,
182	prompt: str,
183	streaming_callback: Optional[StreamingCallbackT] = None,
184	generation_kwargs: Optional[Dict[str, Any]] = None,
185	):
186	"""
187	Invoke the text generation inference for the given prompt and generation parameters.
188
189	:param prompt:
190	A string representing the prompt.
191	:param streaming_callback:
192	A callback function that is called when a new token is received from the stream.
193	:param generation_kwargs:
194	Additional keyword arguments for text generation.
195	:returns:
196	A dictionary with the generated replies and metadata. Both are lists of length n.
197	- replies: A list of strings representing the generated replies.
198	"""
199	# update generation kwargs by merging with the default ones
200	generation_kwargs = {self.generation_kwargs, (generation_kwargs or {})}	1✔
201
202	# check if streaming_callback is passed
203	streaming_callback = select_streaming_callback(	1✔
204	init_callback=self.streaming_callback, runtime_callback=streaming_callback, requires_async=False
205	)
206
207	hf_output = self._client.text_generation(	1✔
208	prompt, details=True, stream=streaming_callback is not None, **generation_kwargs
209	)
210
211	if streaming_callback is not None:	1✔
212	return self._stream_and_build_response(hf_output, streaming_callback)	1✔
213
214	# mypy doesn't know that hf_output is a TextGenerationOutput, so we cast it
215	return self._build_non_streaming_response(cast(TextGenerationOutput, hf_output))	1✔
216
217	def _stream_and_build_response(	1✔
218	self, hf_output: Iterable["TextGenerationStreamOutput"], streaming_callback: StreamingCallbackT
219	):
220	chunks: List[StreamingChunk] = []	1✔
221	first_chunk_time = None	1✔
222
223	component_info = ComponentInfo.from_component(self)	1✔
224	for chunk in hf_output:	1✔
225	token: TextGenerationStreamOutputToken = chunk.token	1✔
226	if token.special:	1✔
227	continue	×
228
229	chunk_metadata = {asdict(token), (asdict(chunk.details) if chunk.details else {})}	1✔
230	if first_chunk_time is None:	1✔
231	first_chunk_time = datetime.now().isoformat()	1✔
232
233	stream_chunk = StreamingChunk(content=token.text, meta=chunk_metadata, component_info=component_info)	1✔
234	chunks.append(stream_chunk)	1✔
235	streaming_callback(stream_chunk)	1✔
236
237	metadata = {	1✔
238	"finish_reason": chunks[-1].meta.get("finish_reason", None),
239	"model": self._client.model,
240	"usage": {"completion_tokens": chunks[-1].meta.get("generated_tokens", 0)},
241	"completion_start_time": first_chunk_time,
242	}
243	return {"replies": ["".join([chunk.content for chunk in chunks])], "meta": [metadata]}	1✔
244
245	def _build_non_streaming_response(self, hf_output: "TextGenerationOutput"):	1✔
246	meta = [	1✔
247	{
248	"model": self._client.model,
249	"finish_reason": hf_output.details.finish_reason if hf_output.details else None,
250	"usage": {"completion_tokens": len(hf_output.details.tokens) if hf_output.details else 0},
251	}
252	]
253	return {"replies": [hf_output.generated_text], "meta": meta}	1✔

deepset-ai / haystack / 15276913207

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous