feat!: Add extra for inference dependencies such as torch (#5147)

* feat!: add extra for inference dependencies such as torch

* add inference extra to 'all' and 'all-gpu' extra

* install inference extra in selected integration tests

* import LazyImport

* review feedback

* add import error messages and update readme

* remove extra dot

.github/workflows/tests.yml

@@ -210,7 +210,7 @@ jobs:
           python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Install Haystack
-        run: pip install .[elasticsearch,dev,preprocessing]
+        run: pip install .[elasticsearch,dev,preprocessing,inference]
 
       - name: Run tests
         run: |
@@ -608,7 +608,7 @@ jobs:
           python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Install Haystack
-        run: pip install .[dev,preprocessing]
+        run: pip install .[dev,preprocessing,inference]
 
       - name: Run tests
         run: |
@@ -662,7 +662,7 @@ jobs:
           python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Install Haystack
-        run: pip install .[dev,preprocessing]
+        run: pip install .[dev,preprocessing,inference]
 
       - name: Run tests
         run: |
@@ -716,7 +716,7 @@ jobs:
           python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Install Haystack
-        run: pip install .[dev,preprocessing]
+        run: pip install .[dev,preprocessing,inference]
 
       - name: Run tests
         run: |

README.md

@@ -76,13 +76,18 @@ This command installs everything needed for basic Pipelines that use an in-memor
 
 **Full Installation**
 
-To use more advanced features, like certain DocumentStores, FileConverters, OCR, or Ray,
+To use more advanced features, like certain DocumentStores, FileConverters, OCR, local inference with pytorch, or Ray,
 you need to install further dependencies. The following command installs the [latest release](https://github.com/deepset-ai/haystack/releases) of Haystack and all its dependencies:
 
 ```sh
 pip install 'farm-haystack[all]' ## or 'all-gpu' for the GPU-enabled dependencies
 ```
 
+If you want to install only the dependencies needed for model inference on your local hardware (not remote API endpoints), such as torch and sentence-transformers, you can use the following command:
+```sh
+pip install 'farm-haystack[inference]' ## installs torch, sentence-transformers, sentencepiece, and huggingface-hub
+```
+
 If you want to try out the newest features that are not in an official release yet, you can install the unstable version from the main branch with the following command:
 
 ```sh

haystack/document_stores/memory.py

@@ -25,7 +25,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_import:
     import torch
     from haystack.modeling.utils import initialize_device_settings  # pylint: disable=ungrouped-imports
 
@@ -92,6 +92,7 @@ def __init__(
                                 You can learn more about these parameters by visiting https://github.com/dorianbrown/rank_bm25
                                 By default, no parameters are set.
         """
+        torch_import.check()
         if bm25_parameters is None:
             bm25_parameters = {}
         super().__init__()

haystack/environment.py

@@ -8,7 +8,7 @@
 from haystack import __version__
 from haystack.lazy_imports import LazyImport
 
-with LazyImport() as torch_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_import:
     import torch
 
 with LazyImport() as transformers_import:

haystack/modeling/model/feature_extraction.py

@@ -32,7 +32,7 @@
 SPECIAL_TOKENIZER_CHARS = r"^(##|Ġ|▁)"
 
 
-with LazyImport(message="Run 'pip install transformers[torch,sentencepiece]'.") as transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as transformers_import:
     import transformers
     from transformers import PreTrainedTokenizer, RobertaTokenizer, AutoConfig, AutoFeatureExtractor, AutoTokenizer
 

haystack/nodes/answer_generator/transformers.py

@@ -13,7 +13,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import (
         RagTokenizer,

haystack/nodes/audio/whisper_transcriber.py

@@ -11,7 +11,7 @@
 from haystack.lazy_imports import LazyImport
 
 
-with LazyImport() as torch_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_import:
     import torch
 
 

haystack/nodes/doc_language_classifier/transformers.py

@@ -12,7 +12,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import pipeline
     from haystack.modeling.utils import initialize_device_settings  # pylint: disable=ungrouped-imports

haystack/nodes/document_classifier/transformers.py

@@ -8,7 +8,7 @@
 from haystack.nodes.document_classifier.base import BaseDocumentClassifier
 from haystack.lazy_imports import LazyImport
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import pipeline
     from haystack.modeling.utils import initialize_device_settings  # pylint: disable=ungrouped-imports

haystack/nodes/extractor/entity.py

@@ -33,7 +33,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from torch.utils.data import Dataset, DataLoader
     from transformers import AutoTokenizer, AutoModelForTokenClassification

haystack/nodes/image_to_text/transformers.py

@@ -13,7 +13,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import pipeline
     from haystack.modeling.utils import initialize_device_settings  # pylint: disable=ungrouped-imports

haystack/nodes/label_generator/pseudo_label_generator.py

@@ -13,7 +13,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from sentence_transformers import CrossEncoder
     from haystack.modeling.utils import initialize_device_settings  # pylint: disable=ungrouped-imports

haystack/nodes/prompt/invocation_layer/hugging_face.py

@@ -10,7 +10,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import (
         pipeline,

haystack/nodes/prompt/prompt_model.py

@@ -7,7 +7,7 @@
 from haystack.schema import Document, MultiLabel
 from haystack.lazy_imports import LazyImport
 
-with LazyImport() as torch_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_import:
     import torch
 
 

haystack/nodes/prompt/prompt_node.py

@@ -10,7 +10,7 @@
 from haystack.nodes.prompt.prompt_template import PromptTemplate
 from haystack.lazy_imports import LazyImport
 
-with LazyImport() as torch_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_import:
     import torch
 
 

haystack/nodes/query_classifier/transformers.py

@@ -11,7 +11,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import pipeline
     from haystack.modeling.utils import initialize_device_settings  # pylint: disable=ungrouped-imports

haystack/nodes/question_generator/question_generator.py

@@ -14,7 +14,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import AutoModelForSeq2SeqLM
     from transformers import AutoTokenizer
@@ -93,6 +93,7 @@ def __init__(
                         parameter is not used and a single CPU device is used for inference.
 
         """
+        torch_and_transformers_import.check()
         super().__init__()
         self.devices, _ = initialize_device_settings(devices=devices, use_cuda=use_gpu, multi_gpu=False)
         if len(self.devices) > 1:

haystack/nodes/ranker/sentence_transformers.py

@@ -13,7 +13,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from torch.nn import DataParallel
     from transformers import AutoModelForSequenceClassification, AutoTokenizer

haystack/nodes/reader/farm.py

@@ -24,7 +24,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from haystack.modeling.data_handler.data_silo import DataSilo, DistillationDataSilo
     from haystack.modeling.data_handler.processor import SquadProcessor, Processor

haystack/nodes/reader/table.py

@@ -18,7 +18,7 @@
 
 TableQuestionAnsweringPipeline = object
 TapasPreTrainedModel = object
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import (  # type: ignore
         TapasTokenizer,

haystack/nodes/reader/transformers.py

@@ -12,7 +12,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import pipeline
     from transformers.data.processors.squad import SquadExample

haystack/nodes/retriever/_embedding_encoder.py

@@ -31,7 +31,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from sentence_transformers import InputExample, SentenceTransformer
     from torch.utils.data import DataLoader

haystack/nodes/retriever/dense.py

@@ -26,7 +26,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from torch.nn import DataParallel
     from torch.utils.data.sampler import SequentialSampler

haystack/nodes/retriever/multimodal/embedder.py

@@ -15,7 +15,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from haystack.utils.torch_utils import get_devices  # pylint: disable=ungrouped-imports
     from haystack.modeling.model.multimodal import get_model  # pylint: disable=ungrouped-imports

haystack/nodes/retriever/multimodal/retriever.py

@@ -15,7 +15,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
 
 

haystack/nodes/sampler/top_p_sampler.py

@@ -11,7 +11,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from sentence_transformers import CrossEncoder
     from haystack.modeling.utils import initialize_device_settings  # pylint: disable=ungrouped-imports

haystack/nodes/summarizer/transformers.py

@@ -13,7 +13,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import pipeline
     from haystack.modeling.utils import initialize_device_settings  # pylint: disable=ungrouped-imports

haystack/nodes/translator/transformers.py

@@ -13,7 +13,7 @@
 logger = logging.getLogger(__name__)
 
 
-with LazyImport() as torch_and_transformers_import:
+with LazyImport(message="Run 'pip install farm-haystack[inference]'") as torch_and_transformers_import:
     import torch
     from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
     from haystack.modeling.utils import initialize_device_settings  # pylint: disable=ungrouped-imports

pyproject.toml

@@ -48,7 +48,7 @@ classifiers = [
 dependencies = [
   "requests",
   "pydantic",
-  "transformers[torch,sentencepiece]==4.30.1",
+  "transformers==4.30.1",
   "pandas",
   "rank_bm25",
   "scikit-learn>=1.0.0", # TF-IDF, SklearnQueryClassifier and metrics
@@ -62,16 +62,15 @@ dependencies = [
   "quantulum3",  # quantities extraction from text
   "posthog",  # telemetry
   # audio's espnet-model-zoo requires huggingface-hub version <0.8 while we need >=0.5 to be able to use create_repo in FARMReader
-  "huggingface-hub>=0.5.0",
   "tenacity",  # retry decorator
   "sseclient-py",  # server side events for OpenAI streaming
   "more_itertools",  # utilities
 
   # Web Retriever
   "boilerpy3",
-
-  # See haystack/nodes/retriever/_embedding_encoder.py, _SentenceTransformersEmbeddingEncoder
-  "sentence-transformers>=2.2.0",
+  
+  # Multimodal Embedder haystack/nodes/retriever/multimodal/embedder.py
+  "Pillow",
 
   # OpenAI tokenizer
   "tiktoken>=0.3.2",
@@ -89,6 +88,11 @@ dependencies = [
 ]
 
 [project.optional-dependencies]
+inference = [
+  "transformers[torch,sentencepiece]==4.30.1",
+  "sentence-transformers>=2.2.0",  # See haystack/nodes/retriever/_embedding_encoder.py, _SentenceTransformersEmbeddingEncoder
+  "huggingface-hub>=0.5.0",
+]
 elasticsearch = [
   "elasticsearch>=7.17,<8",
 ]
@@ -212,11 +216,11 @@ formatting = [
 ]
 
 all = [
-  "farm-haystack[docstores,audio,crawler,preprocessing,file-conversion,pdf,ocr,ray,onnx,beir,metrics]",
+  "farm-haystack[inference,docstores,audio,crawler,preprocessing,file-conversion,pdf,ocr,ray,onnx,beir,metrics]",
 ]
 all-gpu = [
   # beir is incompatible with faiss-gpu: https://github.com/beir-cellar/beir/issues/71
-  "farm-haystack[docstores-gpu,audio,crawler,preprocessing,file-conversion,pdf,ocr,ray,onnx-gpu,metrics]",
+  "farm-haystack[inference,docstores-gpu,audio,crawler,preprocessing,file-conversion,pdf,ocr,ray,onnx-gpu,metrics]",
 ]
 
 [project.scripts]