11591697173

Committed 30 Oct 2024 10:51AM UTC coverage: 90.104%. Remained the same

Build # 11591697173

Build Type

push

github

Committed by

web-flow

Commit Message

fix: `HuggingFaceAPIGenerator` - use forward references (#8502)

* hf API generator: forward references + refactor

* release note

Run Details

7712 of 8559 relevant lines covered (90.1%)

0.9 hits per line

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

96.1

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 typing import Any, Callable, Dict, Iterable, List, Optional, Union

from haystack import component, default_from_dict, default_to_dict, logging
from haystack.dataclasses import StreamingChunk
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>=0.23.0\"'") as huggingface_hub_import:
    from huggingface_hub import (
        InferenceClient,
        TextGenerationOutput,
        TextGenerationOutputToken,
        TextGenerationStreamOutput,
    )


logger = logging.getLogger(__name__)


@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[Callable[[StreamingChunk], None]] = 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(api_type)

        # 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[Callable[[StreamingChunk], None]] = 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 = streaming_callback or self.streaming_callback

        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)

        return self._build_non_streaming_response(hf_output)

    def _stream_and_build_response(
        self, hf_output: Iterable["TextGenerationStreamOutput"], streaming_callback: Callable[[StreamingChunk], None]
    ):
        chunks: List[StreamingChunk] = []
        for chunk in hf_output:
            token: TextGenerationOutputToken = chunk.token
            if token.special:
                continue
            chunk_metadata = {**asdict(token), **(asdict(chunk.details) if chunk.details else {})}
            stream_chunk = StreamingChunk(token.text, chunk_metadata)
            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)},
        }
        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 typing import Any, Callable, Dict, Iterable, List, Optional, Union	1✔
7
8	from haystack import component, default_from_dict, default_to_dict, logging	1✔
9	from haystack.dataclasses import StreamingChunk	1✔
10	from haystack.lazy_imports import LazyImport	1✔
11	from haystack.utils import Secret, deserialize_callable, deserialize_secrets_inplace, serialize_callable	1✔
12	from haystack.utils.hf import HFGenerationAPIType, HFModelType, check_valid_model	1✔
13	from haystack.utils.url_validation import is_valid_http_url	1✔
14
15	with LazyImport(message="Run 'pip install \"huggingface_hub>=0.23.0\"'") as huggingface_hub_import:	1✔
16	from huggingface_hub import (	1✔
17	InferenceClient,
18	TextGenerationOutput,
19	TextGenerationOutputToken,
20	TextGenerationStreamOutput,
21	)
22
23
24	logger = logging.getLogger(__name__)	1✔
25
26
27	@component	1✔
28	class HuggingFaceAPIGenerator:	1✔
29	"""
30	Generates text using Hugging Face APIs.
31
32	Use it with the following Hugging Face APIs:
33	- [Free Serverless Inference API]((https://huggingface.co/inference-api)
34	- [Paid Inference Endpoints](https://huggingface.co/inference-endpoints)
35	- [Self-hosted Text Generation Inference](https://github.com/huggingface/text-generation-inference)
36
37	### Usage examples
38
39	#### With the free serverless inference API
40
41	```python
42	from haystack.components.generators import HuggingFaceAPIGenerator
43	from haystack.utils import Secret
44
45	generator = HuggingFaceAPIGenerator(api_type="serverless_inference_api",
46	api_params={"model": "HuggingFaceH4/zephyr-7b-beta"},
47	token=Secret.from_token("<your-api-key>"))
48
49	result = generator.run(prompt="What's Natural Language Processing?")
50	print(result)
51	```
52
53	#### With paid inference endpoints
54
55	```python
56	from haystack.components.generators import HuggingFaceAPIGenerator
57	from haystack.utils import Secret
58
59	generator = HuggingFaceAPIGenerator(api_type="inference_endpoints",
60	api_params={"url": "<your-inference-endpoint-url>"},
61	token=Secret.from_token("<your-api-key>"))
62
63	result = generator.run(prompt="What's Natural Language Processing?")
64	print(result)
65
66	#### With self-hosted text generation inference
67	```python
68	from haystack.components.generators import HuggingFaceAPIGenerator
69
70	generator = HuggingFaceAPIGenerator(api_type="text_generation_inference",
71	api_params={"url": "http://localhost:8080"})
72
73	result = generator.run(prompt="What's Natural Language Processing?")
74	print(result)
75	```
76	"""
77
78	def __init__( # pylint: disable=too-many-positional-arguments	1✔
79	self,
80	api_type: Union[HFGenerationAPIType, str],
81	api_params: Dict[str, str],
82	token: Optional[Secret] = Secret.from_env_var(["HF_API_TOKEN", "HF_TOKEN"], strict=False),
83	generation_kwargs: Optional[Dict[str, Any]] = None,
84	stop_words: Optional[List[str]] = None,
85	streaming_callback: Optional[Callable[[StreamingChunk], None]] = None,
86	):
87	"""
88	Initialize the HuggingFaceAPIGenerator instance.
89
90	:param api_type:
91	The type of Hugging Face API to use. Available types:
92	- `text_generation_inference`: See [TGI](https://github.com/huggingface/text-generation-inference).
93	- `inference_endpoints`: See [Inference Endpoints](https://huggingface.co/inference-endpoints).
94	- `serverless_inference_api`: See [Serverless Inference API](https://huggingface.co/inference-api).
95	:param api_params:
96	A dictionary with the following keys:
97	- `model`: Hugging Face model ID. Required when `api_type` is `SERVERLESS_INFERENCE_API`.
98	- `url`: URL of the inference endpoint. Required when `api_type` is `INFERENCE_ENDPOINTS` or
99	`TEXT_GENERATION_INFERENCE`.
100	:param token: The Hugging Face token to use as HTTP bearer authorization.
101	Check your HF token in your [account settings](https://huggingface.co/settings/tokens).
102	:param generation_kwargs:
103	A dictionary with keyword arguments to customize text generation. Some examples: `max_new_tokens`,
104	`temperature`, `top_k`, `top_p`.
105	For details, see [Hugging Face documentation](https://huggingface.co/docs/huggingface_hub/en/package_reference/inference_client#huggingface_hub.InferenceClient.text_generation)
106	for more information.
107	:param stop_words: An optional list of strings representing the stop words.
108	:param streaming_callback: An optional callable for handling streaming responses.
109	"""
110
111	huggingface_hub_import.check()	1✔
112
113	if isinstance(api_type, str):	1✔
114	api_type = HFGenerationAPIType.from_str(api_type)	1✔
115
116	if api_type == HFGenerationAPIType.SERVERLESS_INFERENCE_API:	1✔
117	model = api_params.get("model")	1✔
118	if model is None:	1✔
119	raise ValueError(	1✔
120	"To use the Serverless Inference API, you need to specify the `model` parameter in `api_params`."
121	)
122	check_valid_model(model, HFModelType.GENERATION, token)	1✔
123	model_or_url = model	1✔
124	elif api_type in [HFGenerationAPIType.INFERENCE_ENDPOINTS, HFGenerationAPIType.TEXT_GENERATION_INFERENCE]:	1✔
125	url = api_params.get("url")	1✔
126	if url is None:	1✔
127	msg = (	1✔
128	"To use Text Generation Inference or Inference Endpoints, you need to specify the `url` "
129	"parameter in `api_params`."
130	)
131	raise ValueError(msg)	1✔
132	if not is_valid_http_url(url):	1✔
133	raise ValueError(f"Invalid URL: {url}")	1✔
134	model_or_url = url	1✔
135	else:
136	msg = f"Unknown api_type {api_type}"	×
137	raise ValueError(api_type)	×
138
139	# handle generation kwargs setup
140	generation_kwargs = generation_kwargs.copy() if generation_kwargs else {}	1✔
141	generation_kwargs["stop_sequences"] = generation_kwargs.get("stop_sequences", [])	1✔
142	generation_kwargs["stop_sequences"].extend(stop_words or [])	1✔
143	generation_kwargs.setdefault("max_new_tokens", 512)	1✔
144
145	self.api_type = api_type	1✔
146	self.api_params = api_params	1✔
147	self.token = token	1✔
148	self.generation_kwargs = generation_kwargs	1✔
149	self.streaming_callback = streaming_callback	1✔
150	self._client = InferenceClient(model_or_url, token=token.resolve_value() if token else None)	1✔
151
152	def to_dict(self) -> Dict[str, Any]:	1✔
153	"""
154	Serialize this component to a dictionary.
155
156	:returns:
157	A dictionary containing the serialized component.
158	"""
159	callback_name = serialize_callable(self.streaming_callback) if self.streaming_callback else None	1✔
160	return default_to_dict(	1✔
161	self,
162	api_type=str(self.api_type),
163	api_params=self.api_params,
164	token=self.token.to_dict() if self.token else None,
165	generation_kwargs=self.generation_kwargs,
166	streaming_callback=callback_name,
167	)
168
169	@classmethod	1✔
170	def from_dict(cls, data: Dict[str, Any]) -> "HuggingFaceAPIGenerator":	1✔
171	"""
172	Deserialize this component from a dictionary.
173	"""
174	deserialize_secrets_inplace(data["init_parameters"], keys=["token"])	1✔
175	init_params = data["init_parameters"]	1✔
176	serialized_callback_handler = init_params.get("streaming_callback")	1✔
177	if serialized_callback_handler:	1✔
178	init_params["streaming_callback"] = deserialize_callable(serialized_callback_handler)	1✔
179	return default_from_dict(cls, data)	1✔
180
181	@component.output_types(replies=List[str], meta=List[Dict[str, Any]])	1✔
182	def run(	1✔
183	self,
184	prompt: str,
185	streaming_callback: Optional[Callable[[StreamingChunk], None]] = None,
186	generation_kwargs: Optional[Dict[str, Any]] = None,
187	):
188	"""
189	Invoke the text generation inference for the given prompt and generation parameters.
190
191	:param prompt:
192	A string representing the prompt.
193	:param streaming_callback:
194	A callback function that is called when a new token is received from the stream.
195	:param generation_kwargs:
196	Additional keyword arguments for text generation.
197	:returns:
198	A dictionary with the generated replies and metadata. Both are lists of length n.
199	- replies: A list of strings representing the generated replies.
200	"""
201	# update generation kwargs by merging with the default ones
202	generation_kwargs = {self.generation_kwargs, (generation_kwargs or {})}	1✔
203
204	# check if streaming_callback is passed
205	streaming_callback = streaming_callback or self.streaming_callback	1✔
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	return self._build_non_streaming_response(hf_output)	1✔
215
216	def _stream_and_build_response(	1✔
217	self, hf_output: Iterable["TextGenerationStreamOutput"], streaming_callback: Callable[[StreamingChunk], None]
218	):
219	chunks: List[StreamingChunk] = []	1✔
220	for chunk in hf_output:	1✔
221	token: TextGenerationOutputToken = chunk.token	1✔
222	if token.special:	1✔
223	continue	×
224	chunk_metadata = {asdict(token), (asdict(chunk.details) if chunk.details else {})}	1✔
225	stream_chunk = StreamingChunk(token.text, chunk_metadata)	1✔
226	chunks.append(stream_chunk)	1✔
227	streaming_callback(stream_chunk)	1✔
228	metadata = {	1✔
229	"finish_reason": chunks[-1].meta.get("finish_reason", None),
230	"model": self._client.model,
231	"usage": {"completion_tokens": chunks[-1].meta.get("generated_tokens", 0)},
232	}
233	return {"replies": ["".join([chunk.content for chunk in chunks])], "meta": [metadata]}	1✔
234
235	def _build_non_streaming_response(self, hf_output: "TextGenerationOutput"):	1✔
236	meta = [	1✔
237	{
238	"model": self._client.model,
239	"finish_reason": hf_output.details.finish_reason if hf_output.details else None,
240	"usage": {"completion_tokens": len(hf_output.details.tokens) if hf_output.details else 0},
241	}
242	]
243	return {"replies": [hf_output.generated_text], "meta": meta}	1✔

deepset-ai / haystack / 11591697173

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