multi-turn tool call test

llama_stack/providers/inline/inference/meta_reference/inference.py

@@ -338,6 +338,10 @@ class MetaReferenceInferenceImpl(
             stop_reason = None
 
             for token_result in self.generator.chat_completion(request):
+                from termcolor import cprint
+
+                cprint(token_result.text, "cyan", end="")
+
                 tokens.append(token_result.token)
 
                 if token_result.token == tokenizer.eot_id:
@@ -386,6 +390,10 @@ class MetaReferenceInferenceImpl(
             ipython = False
 
             for token_result in self.generator.chat_completion(request):
+                from termcolor import cprint
+
+                cprint(token_result.text, "cyan", end="")
+
                 tokens.append(token_result.token)
 
                 if not ipython and token_result.text.startswith("<|python_tag|>"):

tests/integration/inference/test_text_inference.py

@@ -491,3 +491,55 @@ def test_text_chat_completion_tool_calling_tools_not_in_request(
     else:
         for tc in response.completion_message.tool_calls:
             assert tc.tool_name == "get_object_namespace_list"
+
+
+@pytest.mark.parametrize(
+    "test_case",
+    [
+        "inference:chat_completion:tool_calling",
+    ],
+)
+def test_text_chat_completion_with_tool_calling_loop_non_streaming(client_with_models, text_model_id, test_case):
+    tc = TestCase(test_case)
+    messages = tc["messages"]
+    messages[0]["content"] += """
+Once you make one or more function calls, try to answer the question using the response if you can.
+NEVER invoke the same function with the same argumennts twice. Use the response of the first call instead."""
+    response = client_with_models.inference.chat_completion(
+        model_id=text_model_id,
+        messages=messages,
+        tools=tc["tools"],
+        stream=False,
+    )
+
+    # some models can return content for the response in addition to the tool call
+    assert response.completion_message.role == "assistant"
+
+    assert len(response.completion_message.tool_calls) == 1
+    assert response.completion_message.tool_calls[0].tool_name == tc["tools"][0]["tool_name"]
+    assert response.completion_message.tool_calls[0].arguments == tc["expected"]
+
+    messages.append(response.completion_message)
+
+    messages.append(
+        # Tool Response Message
+        {
+            "role": "tool",
+            "call_id": response.completion_message.tool_calls[0].call_id,
+            "content": "70 degrees and foggy",
+        }
+    )
+
+    response = client_with_models.inference.chat_completion(
+        model_id=text_model_id,
+        messages=messages,
+        tools=tc["tools"],
+        tool_choice="auto",
+        stream=False,
+    )
+
+    from rich.pretty import pprint
+
+    pprint(response.completion_message)
+    # you would expect no tool call but a text completion message
+    assert len(response.completion_message.tool_calls) == 0