Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
fix(serve): a tool whose function is not an object answered 500
kevin9327c64c8d6f01ac
diff --git a/c/openai_server.py b/c/openai_server.pyindex 76b4a904..94fd6967 100644--- a/c/openai_server.py+++ b/c/openai_server.py@@ -392,6 +392,21 @@ def _tool_choice_name(tool_choice):or tool_choice.get("name"))+def _tool_function(tool):+ """The function object on a tools[] entry, or {} if it is missing or not an object.++ OpenAI dual spelling: {"function": {"name": ...}} or a bare function object.+ .items() is taken only from a dict. Writing the name where the object goes+ ({"type": "function", "function": "search"}) raised AttributeError in the GLM+ and DeepSeek declaration blocks, and do_POST answered HTTP 500 "The colibri+ engine failed to process the request." for a payload generation_options()+ already has a 400 for. Same shape as the tool_choice fix (#1598): read the+ member, then check it.+ """+ fn = tool.get("function", tool) if isinstance(tool, dict) else {}+ return fn if isinstance(fn, dict) else {}++def _tool_param_order(tools):"""name -> ordered param names (required first) from the request schema, for de-mangling."""out = {}@@ -564,7 +579,7 @@ def _dsv4_tools_block(tools):"""V4 tool-declaration block, rendered by the vendored reference template."""schemas = []for tool in (tools or []):- fn = tool.get("function", tool) if isinstance(tool, dict) else {}+ fn = _tool_function(tool)# Gateway-side scrub: OpenAI clients attach routing hints the model# schema must not carry.schemas.append({k: v for k, v in fn.items() if k not in ("defer_loading", "strict")})@@ -1798,7 +1813,7 @@ def render_chat(messages, enable_thinking=False, reasoning_effort=None, tools=No"user query.\n\nYou are provided with function signatures within <tools></tools> ""XML tags:\n<tools>\n")for tool in tools:- fn = tool.get("function", tool) if isinstance(tool, dict) else {}+ fn = _tool_function(tool)clean = {k: v for k, v in fn.items() if k not in ("defer_loading", "strict")}prompt.append(json.dumps(clean, ensure_ascii=False) + "\n")prompt.append("</tools>\n\nFor each function call, output the function name and arguments "@@ -2124,8 +2139,7 @@ def _glm53_tool_block(tools):somiglia a quello dell'addestramento non e' quello dell'addestramento."""body = "".join(f"\n{_glm53_tool_json(tool)}\n\n"for tool in tools- if not (isinstance(tool, dict)- and (tool.get("function", tool) or {}).get("defer_loading")))+ if not _tool_function(tool).get("defer_loading"))return GLM53_TOOL_PREAMBLE + body + GLM53_TOOL_EPILOGUE@@ -2305,7 +2319,7 @@ def _dsv41_tools_block(tools):"""V4.1 tool-declaration block, rendered by the vendored reference template."""schemas = []for tool in (tools or []):- fn = tool.get("function", tool) if isinstance(tool, dict) else {}+ fn = _tool_function(tool)# Gateway-side scrub: OpenAI clients attach routing hints the model# schema must not carry.clean = {k: v for k, v in fn.items() if k not in ("defer_loading", "strict")}diff --git a/c/tests/test_openai_server.py b/c/tests/test_openai_server.pyindex afe7b80b..c6ce0e2c 100644--- a/c/tests/test_openai_server.py+++ b/c/tests/test_openai_server.py@@ -458,6 +458,33 @@ class TemplateTest(unittest.TestCase):render_chat(list(messages), tools=list(tools),tool_choice=choice))+ def test_tool_function_that_is_not_an_object_is_a_400(self):+ """The same slip as tool_choice, on tools[] this time.++ generation_options already has 400 "Tool function must be an object".+ The GLM and DeepSeek declaration blocks then did fn.items() on+ {"type": "function", "function": "search"} and raised AttributeError+ before that 400 ran, so do_POST answered 500 "engine failed".+ """+ import openai_server as srv+ messages = [{"role": "user", "content": "hi"}]+ for function in ("search", ["search"], 5, True):+ with self.subTest(function=function):+ tools = [{"type": "function", "function": function}]+ with self.assertRaises(APIError) as caught:+ generation_options({"tools": tools}, 8)+ self.assertEqual(caught.exception.status, 400)+ self.assertEqual(caught.exception.param, "tools.0.function")+ for render in (render_chat, srv.render_chat_glm53, render_chat_kimi,+ render_chat_v4, srv.render_chat_dsv41):+ try:+ render(list(messages), tools=list(tools))+ except APIError as error:+ self.assertEqual(error.status, 400)+ well = [{"type": "function", "function": {"name": "search"}}]+ generation_options({"tools": well}, 8)+ self.assertIn('"name": "search"', render_chat(list(messages), tools=well))+def test_coli_temp_is_the_default_for_requests_that_omit_temperature(self):with patch.dict("openai_server.os.environ", {"COLI_TEMP": "0.25"}):self.assertEqual(generation_options({}, 8)[1], 0.25)@@ -1980,6 +2007,44 @@ class HTTPTest(unittest.TestCase):self.assertEqual(response.status, 200)self.assertIn("You must call the function `search`", self.engine.calls[-1][0])+ def test_tool_with_a_non_object_function_is_a_client_error(self):+ """{"type": "function", "function": "search"} on tools[] answered HTTP 500.++ generation_options already has the 400, but /v1/chat/completions renders+ first. GLM, GLM-5.3, DeepSeek V4 and V4.1 did fn.items() on a string+ and the AttributeError became do_POST's 500 "The colibri engine failed+ to process the request." for a request the engine never saw.+ """+ for arch in ("glm", "glm53", "kimi", "deepseek_v4", "deepseek_v41",+ "olmoe", "inkling", "qwen36", "qwen38"):+ for function in ("search", ["search"], 5):+ with self.subTest(arch=arch, function=function):+ with patch("openai_server.ARCH", arch):+ with self.assertRaises(HTTPError) as caught:+ self.request("/v1/chat/completions", {+ "model": "test-model",+ "messages": [{"role": "user", "content": "hi"}],+ "tools": [{"type": "function",+ "function": function}]})+ self.addCleanup(caught.exception.close)+ self.assertEqual(caught.exception.code, 400)+ with self.assertRaises(HTTPError) as caught:+ self.request("/v1/completions", {+ "model": "test-model", "prompt": "hi",+ "tools": [{"type": "function", "function": "search"}]})+ self.addCleanup(caught.exception.close)+ self.assertEqual(caught.exception.code, 400)++ def test_a_well_formed_tool_function_still_runs(self):+ """The read above must not change the shape clients actually send."""+ with self.request("/v1/chat/completions", {+ "model": "test-model",+ "messages": [{"role": "user", "content": "hi"}],+ "tools": [{"type": "function",+ "function": {"name": "search"}}]}) as response:+ self.assertEqual(response.status, 200)+ self.assertIn('"name": "search"', self.engine.calls[-1][0])+def test_tool_call_arguments_that_are_not_an_object_do_not_fail_the_request(self):"""A replayed tool call with `arguments: "[1, 2]"` answered HTTP 500.