feat: support ainvoke, astream

Author: samuelint

langchain_llamacpp_chat_model/llama_chat_model.py

@@ -1,52 +1,21 @@
 from llama_cpp import Llama
 
 from langchain_openai.chat_models.base import BaseChatOpenAI
-from pydantic import Field
 
-
-class LLamaOpenAIClientProxy:
-    def __init__(self, llama: Llama):
-        self.llama = llama
-
-    def create(self, **kwargs):
-        proxy = LlamaCreateContextManager(llama=self.llama, **kwargs)
-        if "stream" in kwargs and kwargs["stream"] is True:
-            return proxy
-        else:
-            return proxy()
-
-
-class LlamaCreateContextManager:
-
-    def __init__(self, llama: Llama, **kwargs):
-        self.llama = llama
-        self.kwargs = kwargs
-        self.response = None
-
-    def __call__(self):
-        self.kwargs.pop("n", None)
-        self.kwargs.pop(
-            "parallel_tool_calls", None
-        )  # LLamaCPP does not support parallel tool calls for now
-
-        self.response = self.llama.create_chat_completion(**self.kwargs)
-        return self.response
-
-    def __enter__(self):
-        return self()
-
-    def __exit__(self, exception_type, exception_value, exception_traceback):
-        if hasattr(self.response, "close"):
-            self.response.close()
-        return False
+from .llama_client_proxy import LLamaOpenAIClientProxy
+from .llama_client_async_proxy import LLamaOpenAIClientAsyncProxy
 
 
 class LlamaChatModel(BaseChatOpenAI):
-    model_name: str = Field(default="", alias="model")
+    model_name: str = "unknown"
 
     def __init__(
         self,
         llama: Llama,
         **kwargs,
     ):
-        super().__init__(**kwargs, client=LLamaOpenAIClientProxy(llama=llama))
+        super().__init__(
+            **kwargs,
+            client=LLamaOpenAIClientProxy(llama=llama),
+            async_client=LLamaOpenAIClientAsyncProxy(llama=llama),
+        )

langchain_llamacpp_chat_model/llama_client_async_proxy.py

@@ -0,0 +1,52 @@
+from llama_cpp import Llama
+
+
+async def to_async_iterator(iterator):
+    for item in iterator:
+        yield item
+
+
+class LlamaCreateAsyncContextManager:
+
+    def __init__(self, llama: Llama, **kwargs):
+        self.llama = llama
+        self.kwargs = kwargs
+        self.response = None
+
+    def __aiter__(self):
+        return self
+
+    async def __anext__(self):
+        try:
+            return next(self.response)
+        except Exception:
+            raise StopAsyncIteration()
+
+    def __call__(self):
+        self.kwargs.pop("n", None)
+        self.kwargs.pop(
+            "parallel_tool_calls", None
+        )  # LLamaCPP does not support parallel tool calls for now
+
+        self.response = self.llama.create_chat_completion(**self.kwargs)
+        return self.response
+
+    async def __aenter__(self):
+        return self()
+
+    async def __aexit__(self, exception_type, exception_value, exception_traceback):
+        if hasattr(self.response, "close"):
+            self.response.close()
+        return False
+
+
+class LLamaOpenAIClientAsyncProxy:
+    def __init__(self, llama: Llama):
+        self.llama = llama
+
+    async def create(self, **kwargs):
+        proxy = LlamaCreateAsyncContextManager(llama=self.llama, **kwargs)
+        if "stream" in kwargs and kwargs["stream"] is True:
+            return proxy
+        else:
+            return proxy()

langchain_llamacpp_chat_model/llama_client_proxy.py

@@ -0,0 +1,38 @@
+from llama_cpp import Llama
+
+
+class LlamaCreateContextManager:
+
+    def __init__(self, llama: Llama, **kwargs):
+        self.llama = llama
+        self.kwargs = kwargs
+        self.response = None
+
+    def __call__(self):
+        self.kwargs.pop("n", None)
+        self.kwargs.pop(
+            "parallel_tool_calls", None
+        )  # LLamaCPP does not support parallel tool calls for now
+
+        self.response = self.llama.create_chat_completion(**self.kwargs)
+        return self.response
+
+    def __enter__(self):
+        return self()
+
+    def __exit__(self, exception_type, exception_value, exception_traceback):
+        if hasattr(self.response, "close"):
+            self.response.close()
+        return False
+
+
+class LLamaOpenAIClientProxy:
+    def __init__(self, llama: Llama):
+        self.llama = llama
+
+    def create(self, **kwargs):
+        proxy = LlamaCreateContextManager(llama=self.llama, **kwargs)
+        if "stream" in kwargs and kwargs["stream"] is True:
+            return proxy
+        else:
+            return proxy()

tests/test_functional/test_ainvoke.py

@@ -0,0 +1,44 @@
+from llama_cpp import Llama
+import pytest
+from langchain_core.messages import AIMessage, HumanMessage
+
+from langchain_llamacpp_chat_model import LlamaChatModel
+
+from langchain_core.pydantic_v1 import BaseModel, Field
+from tests.test_functional.models_configuration import create_llama, models_to_test
+
+
+class Joke(BaseModel):
+    setup: str = Field(description="The setup of the joke")
+    punchline: str = Field(description="The punchline to the joke")
+
+
+class TestAInvoke:
+
+    @pytest.fixture(
+        params=models_to_test, ids=[config["repo_id"] for config in models_to_test]
+    )
+    def llama(self, request) -> Llama:
+        return create_llama(request)
+
+    @pytest.fixture
+    def instance(self, llama):
+        return LlamaChatModel(llama=llama)
+
+    @pytest.mark.asyncio
+    async def test_ainvoke(self, instance: LlamaChatModel):
+        result = await instance.ainvoke("Say Hi!")
+
+        assert len(result.content) > 0
+
+    @pytest.mark.asyncio
+    async def test_conversation_memory(self, instance: LlamaChatModel):
+        result = await instance.ainvoke(
+            input=[
+                HumanMessage(content="Remember that I like bananas"),
+                AIMessage(content="Okay"),
+                HumanMessage(content="What do I like?"),
+            ]
+        )
+
+        assert "banana" in result.content

tests/test_functional/test_astream.py

@@ -0,0 +1,47 @@
+from llama_cpp import Llama
+import pytest
+from langchain_core.messages import AIMessage, HumanMessage
+
+from langchain_llamacpp_chat_model import LlamaChatModel
+from tests.test_functional.models_configuration import create_llama, models_to_test
+
+
+class TestAStream:
+
+    @pytest.fixture(
+        params=models_to_test, ids=[config["repo_id"] for config in models_to_test]
+    )
+    def llama(self, request) -> Llama:
+        return create_llama(request)
+
+    @pytest.fixture
+    def instance(self, llama):
+        return LlamaChatModel(llama=llama)
+
+    @pytest.mark.asyncio
+    async def test_astream(self, instance: LlamaChatModel):
+
+        chunks = []
+        async for chunk in instance.astream("Say Hi!"):
+            chunks.append(chunk)
+
+        final_content = "".join(chunk.content for chunk in chunks)
+
+        assert len(final_content) > 0
+
+    @pytest.mark.asyncio
+    async def test_conversation_memory(self, instance: LlamaChatModel):
+        stream = instance.astream(
+            input=[
+                HumanMessage(content="Remember that I like bananas"),
+                AIMessage(content="Okay"),
+                HumanMessage(content="What do I like?"),
+            ]
+        )
+
+        final_content = ""
+        async for token in stream:
+            final_content += token.content
+
+        assert len(final_content) > 0
+        assert "banana" in final_content

tests/test_functional/test_invoke.py

@@ -67,3 +67,22 @@ def magic_number_tool(input: int) -> int:
         result = llm_with_tool.invoke("What is the magic mumber of 2?")
 
         assert result.tool_calls[0]["name"] == "magic_number_tool"
+
+
+class TestAInvoke:
+
+    @pytest.fixture(
+        params=models_to_test, ids=[config["repo_id"] for config in models_to_test]
+    )
+    def llama(self, request) -> Llama:
+        return create_llama(request)
+
+    @pytest.fixture
+    def instance(self, llama):
+        return LlamaChatModel(llama=llama)
+
+    @pytest.mark.asyncio
+    async def test_ainvoke(self, instance: LlamaChatModel):
+        result = await instance.ainvoke("Say Hi!")
+
+        assert len(result.content) > 0