From 2ef5579eae206801c4b23dcbfe7171f4a9feabca Mon Sep 17 00:00:00 2001
From: Abdelsalam ElTamawy <40343437+solomspd@users.noreply.github.com>
Date: Fri, 26 May 2023 03:54:52 +0300
Subject: [PATCH] Added pipline args to `HuggingFacePipeline.from_model_id`
 (#5268)

The current `HuggingFacePipeline.from_model_id` does not allow passing
of pipeline arguments to the transformer pipeline.
This PR enables adding important pipeline parameters like setting
`max_new_tokens` for example.
Previous to this PR it would be necessary to manually create the
pipeline through huggingface transformers then handing it to langchain.

For example instead of this
```py
model_id = "gpt2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
pipe = pipeline(
    "text-generation", model=model, tokenizer=tokenizer, max_new_tokens=10
)
hf = HuggingFacePipeline(pipeline=pipe)
```
You can write this
```py
hf = HuggingFacePipeline.from_model_id(
    model_id="gpt2", task="text-generation", pipeline_kwargs={"max_new_tokens": 10}
)
```


Co-authored-by: Dev 2049 <dev.dev2049@gmail.com>
---
 langchain/llms/huggingface_pipeline.py | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/langchain/llms/huggingface_pipeline.py b/langchain/llms/huggingface_pipeline.py
index 9b8d604292..615dcd8e3f 100644
--- a/langchain/llms/huggingface_pipeline.py
+++ b/langchain/llms/huggingface_pipeline.py
@@ -28,7 +28,9 @@ class HuggingFacePipeline(LLM):
 
             from langchain.llms import HuggingFacePipeline
             hf = HuggingFacePipeline.from_model_id(
-                model_id="gpt2", task="text-generation"
+                model_id="gpt2",
+                task="text-generation",
+                pipeline_kwargs={"max_new_tokens": 10},
             )
     Example passing pipeline in directly:
         .. code-block:: python
@@ -49,7 +51,9 @@ class HuggingFacePipeline(LLM):
     model_id: str = DEFAULT_MODEL_ID
     """Model name to use."""
     model_kwargs: Optional[dict] = None
-    """Key word arguments to pass to the model."""
+    """Key word arguments passed to the model."""
+    pipeline_kwargs: Optional[dict] = None
+    """Key word arguments passed to the pipeline."""
 
     class Config:
         """Configuration for this pydantic object."""
@@ -63,6 +67,7 @@ class HuggingFacePipeline(LLM):
         task: str,
         device: int = -1,
         model_kwargs: Optional[dict] = None,
+        pipeline_kwargs: Optional[dict] = None,
         **kwargs: Any,
     ) -> LLM:
         """Construct the pipeline object from model_id and task."""
@@ -119,12 +124,14 @@ class HuggingFacePipeline(LLM):
             _model_kwargs = {
                 k: v for k, v in _model_kwargs.items() if k != "trust_remote_code"
             }
+        _pipeline_kwargs = pipeline_kwargs or {}
         pipeline = hf_pipeline(
             task=task,
             model=model,
             tokenizer=tokenizer,
             device=device,
             model_kwargs=_model_kwargs,
+            **_pipeline_kwargs,
         )
         if pipeline.task not in VALID_TASKS:
             raise ValueError(
@@ -135,6 +142,7 @@ class HuggingFacePipeline(LLM):
             pipeline=pipeline,
             model_id=model_id,
             model_kwargs=_model_kwargs,
+            pipeline_kwargs=_pipeline_kwargs,
             **kwargs,
         )
 
@@ -142,8 +150,9 @@ class HuggingFacePipeline(LLM):
     def _identifying_params(self) -> Mapping[str, Any]:
         """Get the identifying parameters."""
         return {
-            **{"model_id": self.model_id},
-            **{"model_kwargs": self.model_kwargs},
+            "model_id": self.model_id,
+            "model_kwargs": self.model_kwargs,
+            "pipeline_kwargs": self.pipeline_kwargs,
         }
 
     @property