Fixes streaming for llama3 models. (#116)

Author: 3coins

libs/aws/langchain_aws/chat_models/bedrock.py

@@ -565,20 +565,16 @@ def _generate(
             )
         else:
             usage_metadata = None
+
         llm_output["model_id"] = self.model_id
-        if len(tool_calls) > 0:
-            msg = AIMessage(
-                content=completion,
-                additional_kwargs=llm_output,
-                tool_calls=cast(List[ToolCall], tool_calls),
-                usage_metadata=usage_metadata,
-            )
-        else:
-            msg = AIMessage(
-                content=completion,
-                additional_kwargs=llm_output,
-                usage_metadata=usage_metadata,
-            )
+
+        msg = AIMessage(
+            content=completion,
+            additional_kwargs=llm_output,
+            tool_calls=cast(List[ToolCall], tool_calls),
+            usage_metadata=usage_metadata,
+        )
+
         return ChatResult(
             generations=[
                 ChatGeneration(

libs/aws/langchain_aws/llms/bedrock.py

@@ -117,7 +117,11 @@ def _stream_response_to_generation_chunk(
             return None
     else:
         # chunk obj format varies with provider
-        generation_info = {k: v for k, v in stream_response.items() if k != output_key}
+        generation_info = {
+            k: v
+            for k, v in stream_response.items()
+            if k not in [output_key, "prompt_token_count", "generation_token_count"]
+        }
         return GenerationChunk(
             text=(
                 stream_response[output_key]
@@ -347,6 +351,10 @@ def prepare_output_stream(
                 yield _get_invocation_metrics_chunk(chunk_obj)
                 return
 
+            elif provider == "meta" and chunk_obj.get("stop_reason", "") == "stop":
+                yield _get_invocation_metrics_chunk(chunk_obj)
+                return
+
             elif messages_api and (chunk_obj.get("type") == "message_stop"):
                 yield _get_invocation_metrics_chunk(chunk_obj)
                 return

libs/aws/tests/integration_tests/chat_models/test_bedrock.py

@@ -84,17 +84,18 @@ def test_chat_bedrock_streaming() -> None:
 @pytest.mark.scheduled
 def test_chat_bedrock_streaming_llama3() -> None:
     """Test that streaming correctly invokes on_llm_new_token callback."""
-    callback_handler = FakeCallbackHandler()
     chat = ChatBedrock(  # type: ignore[call-arg]
-        model_id="meta.llama3-8b-instruct-v1:0",
-        streaming=True,
-        callbacks=[callback_handler],
-        verbose=True,
+        model_id="meta.llama3-8b-instruct-v1:0"
     )
     message = HumanMessage(content="Hello")
-    response = chat([message])
-    assert callback_handler.llm_streams > 0
-    assert isinstance(response, BaseMessage)
+
+    response = AIMessageChunk(content="")
+    for chunk in chat.stream([message]):
+        response += chunk  # type: ignore[assignment]
+
+    assert response.content
+    assert response.response_metadata
+    assert response.usage_metadata
 
 
 @pytest.mark.scheduled