Add `TrainableLLM` (#11721)

- **Description:** Add `TrainableLLM` for those LLM support fine-tuning
  - **Tag maintainer:** @hwchase17

This PR add training methods to `GradientLLM`

---------

Co-authored-by: Bagatur <baskaryan@gmail.com>

libs/langchain/langchain/llms/base.py

@@ -50,12 +50,7 @@ from langchain.load.dump import dumpd
 from langchain.prompts.base import StringPromptValue
 from langchain.prompts.chat import ChatPromptValue
 from langchain.pydantic_v1 import Field, root_validator, validator
-from langchain.schema import (
-    Generation,
-    LLMResult,
-    PromptValue,
-    RunInfo,
-)
+from langchain.schema import Generation, LLMResult, PromptValue, RunInfo
 from langchain.schema.language_model import BaseLanguageModel, LanguageModelInput
 from langchain.schema.messages import AIMessage, BaseMessage, get_buffer_string
 from langchain.schema.output import GenerationChunk

libs/langchain/langchain/llms/gradient_ai.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, List, Mapping, Optional, Union
+from typing import Any, Dict, List, Mapping, Optional, Sequence, TypedDict, Union
 
 import aiohttp
 import requests
@@ -13,6 +13,10 @@ from langchain.pydantic_v1 import Extra, root_validator
 from langchain.utils import get_from_dict_or_env
 
 
+class TrainResult(TypedDict):
+    loss: float
+
+
 class GradientLLM(LLM):
     """Gradient.ai LLM Endpoints.
 
@@ -125,6 +129,51 @@ class GradientLLM(LLM):
         """Return type of llm."""
         return "gradient"
 
+    def _kwargs_post_fine_tune_request(
+        self, inputs: Sequence[str], kwargs: Mapping[str, Any]
+    ) -> Mapping[str, Any]:
+        """Build the kwargs for the Post request, used by sync
+
+        Args:
+            prompt (str): prompt used in query
+            kwargs (dict): model kwargs in payload
+
+        Returns:
+            Dict[str, Union[str,dict]]: _description_
+        """
+        _model_kwargs = self.model_kwargs or {}
+        _params = {**_model_kwargs, **kwargs}
+
+        multipliers = _params.get("multipliers", None)
+
+        return dict(
+            url=f"{self.gradient_api_url}/models/{self.model_id}/fine-tune",
+            headers={
+                "authorization": f"Bearer {self.gradient_access_token}",
+                "x-gradient-workspace-id": f"{self.gradient_workspace_id}",
+                "accept": "application/json",
+                "content-type": "application/json",
+            },
+            json=dict(
+                samples=tuple(
+                    {
+                        "inputs": input,
+                    }
+                    for input in inputs
+                )
+                if multipliers is None
+                else tuple(
+                    {
+                        "inputs": input,
+                        "fineTuningParameters": {
+                            "multiplier": multiplier,
+                        },
+                    }
+                    for input, multiplier in zip(inputs, multipliers)
+                ),
+            ),
+        )
+
     def _kwargs_post_request(
         self, prompt: str, kwargs: Mapping[str, Any]
     ) -> Mapping[str, Any]:
@@ -234,3 +283,60 @@ class GradientLLM(LLM):
             text = enforce_stop_tokens(text, stop)
 
         return text
+
+    def train_unsupervised(
+        self,
+        inputs: Sequence[str],
+        **kwargs: Any,
+    ) -> TrainResult:
+        try:
+            response = requests.post(
+                **self._kwargs_post_fine_tune_request(inputs, kwargs)
+            )
+            if response.status_code != 200:
+                raise Exception(
+                    f"Gradient returned an unexpected response with status "
+                    f"{response.status_code}: {response.text}"
+                )
+        except requests.exceptions.RequestException as e:
+            raise Exception(f"RequestException while calling Gradient Endpoint: {e}")
+
+        response_json = response.json()
+        loss = response_json["sumLoss"] / response_json["numberOfTrainableTokens"]
+        return TrainResult(loss=loss)
+
+    async def atrain_unsupervised(
+        self,
+        inputs: Sequence[str],
+        **kwargs: Any,
+    ) -> TrainResult:
+        if not self.aiosession:
+            async with aiohttp.ClientSession() as session:
+                async with session.post(
+                    **self._kwargs_post_fine_tune_request(inputs, kwargs)
+                ) as response:
+                    if response.status != 200:
+                        raise Exception(
+                            f"Gradient returned an unexpected response with status "
+                            f"{response.status}: {response.text}"
+                        )
+                    response_json = await response.json()
+                    loss = (
+                        response_json["sumLoss"]
+                        / response_json["numberOfTrainableTokens"]
+                    )
+        else:
+            async with self.aiosession.post(
+                **self._kwargs_post_fine_tune_request(inputs, kwargs)
+            ) as response:
+                if response.status != 200:
+                    raise Exception(
+                        f"Gradient returned an unexpected response with status "
+                        f"{response.status}: {response.text}"
+                    )
+                response_json = await response.json()
+                loss = (
+                    response_json["sumLoss"] / response_json["numberOfTrainableTokens"]
+                )
+
+        return TrainResult(loss=loss)