Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Merge pull request #1497 from basil-k-aji-dev/feat/tool-fallback-olmoe-qwen36
Vincenzo Fornaro739e89912884
diff --cc c/openai_server.pyindex 42fe3635,b43eb7cd..b319b314--- a/c/openai_server.py+++ b/c/openai_server.py@@@ -258,23 -315,16 +315,32 @@@ _PARTIAL_END_RE = re.compile(r"<(?:/(?:# <arg_key>K</arg_key><arg_value> structure. Default OFF (never rewrites well-formed output)._SALVAGE = os.environ.get("COLI_TOOL_SALVAGE", "0") == "1"+ # Families whose chat template has no tool syntax at all (OLMoE, Qwen3.6) refuse+ # tools[] and role:"tool" rather than invent a format. COLI_TOOL_FALLBACK=1 opts+ # into a prompt-injected translation for them: the declaration block, the prior+ # assistant calls and the tool results are written as ordinary turns, in the+ # same wire format parse_tool_calls() already reads back (#1378). Default OFF --+ # these models were never trained on tool syntax, so this trades a clean 400 for+ # output the parser may or may not recognise.+ _TOOL_FALLBACK = os.environ.get("COLI_TOOL_FALLBACK", "0") == "1"++def _tool_choice_name(tool_choice):+ """The tool name a dict `tool_choice` forces, or None.++ `.get` is taken only from a dict. Writing the name where the object goes --+ {"type": "function", "function": "search"} instead of+ {"function": {"name": "search"}} -- raised AttributeError in the five+ renderers that read this and in generation_options() itself, 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+ json_schema fix (#1587): read the member, then check it.+ """+ function = tool_choice.get("function")+ return ((function if isinstance(function, dict) else {}).get("name")+ or tool_choice.get("name"))++def _tool_param_order(tools):"""name -> ordered param names (required first) from the request schema, for de-mangling."""out = {}