fix(sglang): make prefill warmup failure fatal (#7944)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

components/src/dynamo/sglang/init_llm.py

@@ -25,9 +25,12 @@ from dynamo.sglang.request_handlers import DecodeWorkerHandler, PrefillWorkerHan
 
 
 async def _warmup_prefill_engine(engine: sgl.Engine, server_args) -> None:
-    """Perform warmup request for prefill engine to reduce initial TTFT."""
+    """Perform warmup request for prefill engine to reduce initial TTFT.
+
+    Raises on failure so the caller can prevent the worker from registering
+    with a broken engine (silent request drops).
+    """
     logging.info("Start of prefill disaggregation warmup ...")
-    try:
     from sglang.srt.disaggregation.utils import FAKE_BOOTSTRAP_HOST
 
     sampling_params = {
@@ -50,10 +53,6 @@ async def _warmup_prefill_engine(engine: sgl.Engine, server_args) -> None:
 
     await asyncio.wait_for(_do_warmup(), timeout=1800)
     logging.info("Prefill warmup completed")
-    except asyncio.TimeoutError:
-        logging.warning("Prefill warmup timed out after 1800s")
-    except Exception as e:
-        logging.warning(f"Prefill warmup failed: {e}")
 
 
 async def init_decode(
@@ -184,7 +183,16 @@ async def init_prefill(
         await handle_non_leader_node(engine, publisher, metrics_task)
         return
 
+    try:
         await _warmup_prefill_engine(engine, server_args)
+    except asyncio.TimeoutError as e:
+        logging.error("Prefill warmup timed out after 1800s — aborting worker startup")
+        raise RuntimeError(
+            "Prefill warmup timed out; worker cannot serve requests"
+        ) from e
+    except Exception as e:
+        logging.error(f"Prefill warmup failed: {e} — aborting worker startup")
+        raise RuntimeError(f"Prefill warmup failed: {e}") from e
 
     handler = PrefillWorkerHandler(
         engine, config, publisher, generate_endpoint, shutdown_event