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 past tool call whose function is not an object answered 500
kevin93277c991ce79222
diff --git a/c/openai_server.py b/c/openai_server.pyindex 9d595ac9..a3b66a56 100644--- a/c/openai_server.py+++ b/c/openai_server.py@@ -326,6 +326,28 @@ def _fallback_tool_preamble(tools):return "".join(out)+def _history_tool_calls(tool_calls, index):+ """A past assistant message's tool_calls, checked as _fallback_tool_calls and+ _qwen38_tool_calls check them, for the renderers that read `function` without+ looking: GLM, GLM-5.3 and DeepSeek V4/V4.1 raised AttributeError on a+ `function` that is not an object, which do_POST answers with HTTP 500."""+ if tool_calls is None:+ return []+ if not isinstance(tool_calls, list):+ raise APIError(400, "`tool_calls` must be an array.", f"messages.{index}.tool_calls")+ for position, call in enumerate(tool_calls):+ where = f"messages.{index}.tool_calls.{position}"+ if not isinstance(call, dict):+ raise APIError(400, "Each tool call must be an object.", where)+ fn = call.get("function", call)+ if not isinstance(fn, dict):+ raise APIError(400, "`function` must be an object.", f"{where}.function")+ name = fn.get("name")+ if name is not None and not isinstance(name, str):+ raise APIError(400, "`function.name` must be a string.", f"{where}.function.name")+ return tool_calls++def _fallback_tool_calls(tool_calls, index):"""Render assistant tool_calls in the format parse_tool_calls() reads."""out = []@@ -1301,7 +1323,7 @@ def render_chat_v4(messages, enable_thinking=False, reasoning_effort=None, toolscontent = content_text(raw, f"messages.{index}.content") if raw is not None else ""merged.append({"role": role, "content": content,"reasoning_content": message.get("reasoning_content"),- "tool_calls": message.get("tool_calls")})+ "tool_calls": _history_tool_calls(message.get("tool_calls"), index)})continueraw = message.get("content")text = content_text(raw, f"messages.{index}.content") if raw is not None else ""@@ -1905,8 +1927,8 @@ def render_chat(messages, enable_thinking=False, reasoning_effort=None, tools=Noraise APIError(400, "`reasoning_content` must be a string.",f"messages.{index}.reasoning_content")prompt.append(f"<|assistant|><think>{reasoning}</think>{text.strip()}")- for tc in (message.get("tool_calls") or []):- fn = tc.get("function", tc) if isinstance(tc, dict) else {}+ for tc in _history_tool_calls(message.get("tool_calls"), index):+ fn = tc.get("function", tc)args = fn.get("arguments", "{}")if isinstance(args, str):try:@@ -2302,7 +2324,7 @@ def render_chat_glm53(messages, enable_thinking=False, reasoning_effort=None, toif tools:prompt.append(_glm53_tool_block(tools))- for message in messages:+ for index, message in enumerate(messages):if not isinstance(message, dict):raise APIError(400, "each message must be an object.", "messages")role = message.get("role")@@ -2324,7 +2346,7 @@ def render_chat_glm53(messages, enable_thinking=False, reasoning_effort=None, tocontent = content.split("</think>")[-1]opened = f"<think>{reasoning}</think>" if isinstance(reasoning, str) else "<think></think>"body = content.strip()- calls = _glm53_tool_calls(message.get("tool_calls"))+ calls = _glm53_tool_calls(_history_tool_calls(message.get("tool_calls"), index))# The template writes "\n<tool_call>". The model, on a turn that is# nothing but a tool call, writes "</think><tool_call>" with no# newline between them, and that one token is enough to throw away@@ -2429,7 +2451,7 @@ def _dsv41_merge_turns(messages):f"messages.{index}.reasoning_content")turns.append({"role": "assistant", "content": text,"reasoning_content": reasoning,- "tool_calls": message.get("tool_calls")})+ "tool_calls": _history_tool_calls(message.get("tool_calls"), index)})elif role in ("user", "tool"):block = ({"kind": "tool_result", "id": message.get("tool_call_id") or "","text": v41_dsml.render_tool_result(text)} if role == "tool"diff --git a/c/tests/test_openai_server.py b/c/tests/test_openai_server.pyindex e6af5a13..47c77e6c 100644--- a/c/tests/test_openai_server.py+++ b/c/tests/test_openai_server.py@@ -2315,6 +2315,41 @@ class HTTPTest(unittest.TestCase):self.assertEqual(response.status, 200)self.assertEqual(json.load(response)["object"], "chat.completion")+ def test_a_past_tool_call_whose_function_is_not_an_object_is_a_client_error(self):+ """A replayed tool call with `function: "search"` answered HTTP 500.++ The fallback, Kimi and Qwen3.8 renderers already answer 400. GLM, GLM-5.3+ and DeepSeek V4/V4.1 called .get() on the value, and the AttributeError+ became do_POST's 500 "The colibri engine failed to process the request."+ """+ def history(function):+ return {"model": "test-model", "messages": [+ {"role": "user", "content": "run it"},+ {"role": "assistant", "content": "", "tool_calls": [+ {"id": "x", "type": "function", "function": function}]},+ {"role": "tool", "tool_call_id": "x", "content": "done"},+ {"role": "user", "content": "and now?"},+ ]}+ cases = [("search", "messages.1.tool_calls.0.function"),+ (["search"], "messages.1.tool_calls.0.function"),+ (5, "messages.1.tool_calls.0.function"),+ (None, "messages.1.tool_calls.0.function"),+ ({"name": 5, "arguments": "{}"}, "messages.1.tool_calls.0.function.name")]+ for arch in ("glm", "glm53", "deepseek_v4", "deepseek_v41"):+ for function, param in cases:+ with self.subTest(arch=arch, function=function):+ with patch("openai_server.ARCH", arch):+ with self.assertRaises(HTTPError) as caught:+ self.request("/v1/chat/completions", history(function))+ self.addCleanup(caught.exception.close)+ self.assertEqual(caught.exception.code, 400)+ self.assertEqual(json.loads(caught.exception.read())["error"]["param"], param)+ with self.subTest(arch=arch, function="well formed"):+ with patch("openai_server.ARCH", arch):+ with self.request("/v1/chat/completions",+ history({"name": "fn", "arguments": "{}"})) as response:+ self.assertEqual(response.status, 200)+class ClientHangupTest(unittest.TestCase):"""A client that disconnects mid-response must not print a traceback.