Fix OLMoE support and silent model-init failures

moe_infinity/common/constants.py

@@ -39,7 +39,7 @@
     "gptoss": 4,
     "qwen3": 5,
     "dbrx": 4,
-    "olmoe": 4,
+    "olmoe": 5,
     "jamba": 4,
 }
 

moe_infinity/entrypoints/openai/api_server_v2.py

@@ -142,6 +142,7 @@ def _decorator(func: Any) -> Any:
 _contextpilot_fallback_count: int = 0
 _contextpilot_last_fallback_count: int = 0
 
+logger = logging.getLogger(__name__)
 _cp_logger = logging.getLogger("moe_infinity.contextpilot")
 _cp_middleware: Optional[Any] = None
 _eviction_sync: Optional[Any] = None
@@ -1093,12 +1094,19 @@ async def _initialize_model() -> None:
             )
 
         _ensure_engine_loop_running()
+        _health_state.set_healthy()
+    except Exception as exc:
+        # _model_init_task is fire-and-forget (asyncio.create_task, never
+        # awaited) -- without this, an exception here is silently dropped:
+        # the process keeps running, /health stays stuck on "starting"
+        # forever, and every request gets a 503 indistinguishable from a
+        # genuine hang. Log it and surface it through /health instead.
+        logger.exception("Model initialization failed")
+        _health_state.set_unhealthy(f"{type(exc).__name__}: {exc}")
     finally:
         if _startup_watchdog is not None:
             _startup_watchdog.cancel()
 
-    _health_state.set_healthy()
-
 
 @app.on_event("startup")
 async def startup_event() -> None:

moe_infinity/utils/hf_config.py

@@ -78,7 +78,7 @@ def parse_moe_param(config: PretrainedConfig) -> Tuple[int, int, int]:
         num_decoder_layers = config.num_hidden_layers
         num_layers = config.num_hidden_layers
         num_experts = config.num_local_experts
-    elif "qwen3" in arch:
+    elif "qwen3" in arch or "olmoe" in arch:
         num_encoder_layers = 0
         num_decoder_layers = config.num_hidden_layers
         num_layers = config.num_hidden_layers
@@ -147,7 +147,7 @@ def parse_expert_id(
             layer_id, expert_id = result[0]
             layer_id = int(layer_id)
             expert_id = int(expert_id)
-    elif "deepseek" in arch or "qwen3" in arch:
+    elif "deepseek" in arch or "qwen3" in arch or "olmoe" in arch:
         decoder_sparse_step = 1
         layer_type = "decoder"
 

pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools==75.3.2", "wheel", "torch"]
+requires = ["setuptools>=75.3.2", "wheel", "torch"]
 build-backend = "setuptools.build_meta"