[Feature] Add support for running huggingface models locally (#1287)

docs/components/llms.mdx

@@ -451,7 +451,15 @@ pip install --upgrade 'embedchain[huggingface-hub]'
 
 First, set `HUGGINGFACE_ACCESS_TOKEN` in environment variable which you can obtain from [their platform](https://huggingface.co/settings/tokens).
 
-Once you have the token, load the app using the config yaml file:
+You can load the LLMs from Hugging Face using three ways:
+
+- [Hugging Face Hub](#hugging-face-hub)
+- [Hugging Face Local Pipelines](#hugging-face-local-pipelines)
+- [Hugging Face Inference Endpoint](#hugging-face-inference-endpoint)
+
+### Hugging Face Hub
+
+To load the model from Hugging Face Hub, use the following code:
 
 <CodeGroup>
 
@@ -461,24 +469,49 @@ from embedchain import App
 
 os.environ["HUGGINGFACE_ACCESS_TOKEN"] = "xxx"
 
-# load llm configuration from config.yaml file
-app = App.from_config(config_path="config.yaml")
-```
+config = {
+  "app": {"config": {"id": "my-app"}},
+  "llm": {
+      "provider": "huggingface",
+      "config": {
+          "model": "bigscience/bloom-1b7",
+          "top_p": 0.5,
+          "max_length": 200,
+          "temperature": 0.1,
+      },
+  },
+}
 
-```yaml config.yaml
-llm:
-  provider: huggingface
-  config:
-    model: 'google/flan-t5-xxl'
-    temperature: 0.5
-    max_tokens: 1000
-    top_p: 0.5
-    stream: false
+app = App.from_config(config=config)
 ```
 </CodeGroup>
 
-### Custom Endpoints
+### Hugging Face Local Pipelines
+
+If you want to load the locally downloaded model from Hugging Face, you can do so by following the code provided below:
 
+<CodeGroup>
+```python main.py
+from embedchain import App
+
+config = {
+  "app": {"config": {"id": "my-app"}},
+  "llm": {
+      "provider": "huggingface",
+      "config": {
+          "model": "Trendyol/Trendyol-LLM-7b-chat-v0.1",
+          "local": True,  # Necessary if you want to run model locally
+          "top_p": 0.5,
+          "max_tokens": 1000,
+          "temperature": 0.1,
+      },
+  }
+}
+app = App.from_config(config=config)
+```
+</CodeGroup>
+
+### Hugging Face Inference Endpoint
 
 You can also use [Hugging Face Inference Endpoints](https://huggingface.co/docs/inference-endpoints/index#-inference-endpoints) to access custom endpoints. First, set the `HUGGINGFACE_ACCESS_TOKEN` as above.
 
@@ -487,35 +520,23 @@ Then, load the app using the config yaml file:
 <CodeGroup>
 
 ```python main.py
-import os
 from embedchain import App
 
-os.environ["HUGGINGFACE_ACCESS_TOKEN"] = "xxx"
-
-# load llm configuration from config.yaml file
-app = App.from_config(config_path="config.yaml")
-```
+config = {
+  "app": {"config": {"id": "my-app"}},
+  "llm": {
+      "provider": "huggingface",
+      "config": {
+        "endpoint": "https://api-inference.huggingface.co/models/gpt2",
+        "model_params": {"temprature": 0.1, "max_new_tokens": 100}
+      },
+  },
+}
+app = App.from_config(config=config)
 
-```yaml config.yaml
-llm:
-  provider: huggingface
-  config:
-    endpoint: https://api-inference.huggingface.co/models/gpt2 # replace with your personal endpoint
 ```
 </CodeGroup>
 
-If your endpoint requires additional parameters, you can pass them in the `model_kwargs` field:
-
-```
-llm:
-  provider: huggingface
-  config:
-    endpoint: <YOUR_ENDPOINT_URL_HERE>
-    model_kwargs:
-      max_new_tokens: 100
-      temperature: 0.5
-```
-
 Currently only supports `text-generation` and `text2text-generation` for now [[ref](https://api.python.langchain.com/en/latest/llms/langchain_community.llms.huggingface_endpoint.HuggingFaceEndpoint.html?highlight=huggingfaceendpoint#)].
 
 See langchain's [hugging face endpoint](https://python.langchain.com/docs/integrations/chat/huggingface#huggingfaceendpoint) for more information. 

embedchain/config/llm/base.py

@@ -95,6 +95,7 @@ def __init__(
         api_key: Optional[str] = None,
         endpoint: Optional[str] = None,
         model_kwargs: Optional[dict[str, Any]] = None,
+        local: Optional[bool] = False,
     ):
         """
         Initializes a configuration class instance for the LLM.
@@ -138,6 +139,8 @@ def __init__(
         :type callbacks: Optional[list], optional
         :param query_type: The type of query to use, defaults to None
         :type query_type: Optional[str], optional
+        :param local: If True, the model will be run locally, defaults to False (for huggingface provider)
+        :type local: Optional[bool], optional
         :raises ValueError: If the template is not valid as template should
         contain $context and $query (and optionally $history)
         :raises ValueError: Stream is not boolean
@@ -165,6 +168,7 @@ def __init__(
         self.api_key = api_key
         self.endpoint = endpoint
         self.model_kwargs = model_kwargs
+        self.local = local
 
         if isinstance(prompt, str):
             prompt = Template(prompt)

embedchain/llm/huggingface.py

@@ -5,6 +5,7 @@
 
 from langchain_community.llms.huggingface_endpoint import HuggingFaceEndpoint
 from langchain_community.llms.huggingface_hub import HuggingFaceHub
+from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline
 
 from embedchain.config import BaseLlmConfig
 from embedchain.helpers.json_serializable import register_deserializable
@@ -34,12 +35,15 @@ def get_llm_model_answer(self, prompt):
 
     @staticmethod
     def _get_answer(prompt: str, config: BaseLlmConfig) -> str:
-        if config.model:
+        # If the user wants to run the model locally, they can do so by setting the `local` flag to True
+        if config.model and config.local:
+            return HuggingFaceLlm._from_pipeline(prompt=prompt, config=config)
+        elif config.model:
             return HuggingFaceLlm._from_model(prompt=prompt, config=config)
         elif config.endpoint:
             return HuggingFaceLlm._from_endpoint(prompt=prompt, config=config)
         else:
-            raise ValueError("Either `model` or `endpoint` must be set")
+            raise ValueError("Either `model` or `endpoint` must be set in config")
 
     @staticmethod
     def _from_model(prompt: str, config: BaseLlmConfig) -> str:
@@ -53,15 +57,14 @@ def _from_model(prompt: str, config: BaseLlmConfig) -> str:
         else:
             raise ValueError("`top_p` must be > 0.0 and < 1.0")
 
-        model = config.model or "google/flan-t5-xxl"
+        model = config.model
         logging.info(f"Using HuggingFaceHub with model {model}")
         llm = HuggingFaceHub(
             huggingfacehub_api_token=os.environ["HUGGINGFACE_ACCESS_TOKEN"],
             repo_id=model,
             model_kwargs=model_kwargs,
         )
-
-        return llm(prompt)
+        return llm.invoke(prompt)
 
     @staticmethod
     def _from_endpoint(prompt: str, config: BaseLlmConfig) -> str:
@@ -71,4 +74,23 @@ def _from_endpoint(prompt: str, config: BaseLlmConfig) -> str:
             task="text-generation",
             model_kwargs=config.model_kwargs,
         )
-        return llm(prompt)
+        return llm.invoke(prompt)
+
+    @staticmethod
+    def _from_pipeline(prompt: str, config: BaseLlmConfig) -> str:
+        model_kwargs = {
+            "temperature": config.temperature or 0.1,
+            "max_new_tokens": config.max_tokens,
+        }
+
+        if 0.0 < config.top_p < 1.0:
+            model_kwargs["top_p"] = config.top_p
+        else:
+            raise ValueError("`top_p` must be > 0.0 and < 1.0")
+
+        llm = HuggingFacePipeline.from_model_id(
+            model_id=config.model,
+            task="text-generation",
+            pipeline_kwargs=model_kwargs,
+        )
+        return llm.invoke(prompt)

embedchain/utils/misc.py

@@ -425,6 +425,7 @@ def validate_config(config_data):
                     Optional("api_key"): str,
                     Optional("endpoint"): str,
                     Optional("model_kwargs"): dict,
+                    Optional("local"): bool,
                 },
             },
             Optional("vectordb"): {

tests/llm/test_huggingface.py

@@ -62,18 +62,19 @@ def test_get_llm_model_answer(huggingface_llm_config, mocker):
 
 def test_hugging_face_mock(huggingface_llm_config, mocker):
     mock_llm_instance = mocker.Mock(return_value="Test answer")
-    mocker.patch("embedchain.llm.huggingface.HuggingFaceHub", return_value=mock_llm_instance)
+    mock_hf_hub = mocker.patch("embedchain.llm.huggingface.HuggingFaceHub")
+    mock_hf_hub.return_value.invoke = mock_llm_instance
 
     llm = HuggingFaceLlm(huggingface_llm_config)
     answer = llm.get_llm_model_answer("Test query")
-
     assert answer == "Test answer"
     mock_llm_instance.assert_called_once_with("Test query")
 
 
 def test_custom_endpoint(huggingface_endpoint_config, mocker):
     mock_llm_instance = mocker.Mock(return_value="Test answer")
-    mocker.patch("embedchain.llm.huggingface.HuggingFaceEndpoint", return_value=mock_llm_instance)
+    mock_hf_endpoint = mocker.patch("embedchain.llm.huggingface.HuggingFaceEndpoint")
+    mock_hf_endpoint.return_value.invoke = mock_llm_instance
 
     llm = HuggingFaceLlm(huggingface_endpoint_config)
     answer = llm.get_llm_model_answer("Test query")