fix(prefill): isolate progress HTTP event loop

Run dashboard telemetry on a dedicated thread so blocking model RPCs cannot hide segmented Prefill progress or stale the Terminal ETA.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: 2 people

scripts/start_grpc_runtime_server.py

@@ -34,6 +34,7 @@
 
 import argparse
 import asyncio
+import threading
 import logging
 import signal
 import sys
@@ -565,7 +566,7 @@ async def _serve(args: argparse.Namespace) -> int:
     _LOG.info("kakeya gRPC RuntimeService listening on %s", args.bind)
 
     http_server = None
-    http_task = None
+    http_thread = None
     if args.network_http_port:
         if registry is None or prefill_store is None:
             raise SystemExit(
@@ -594,7 +595,12 @@ async def _serve(args: argparse.Namespace) -> int:
             port=args.network_http_port,
             log_level=args.log_level.lower(),
         ))
-        http_task = asyncio.create_task(http_server.serve())
+        http_thread = threading.Thread(
+            target=http_server.run,
+            name="kakeya-network-http",
+            daemon=True,
+        )
+        http_thread.start()
         _LOG.info(
             "inference-network dashboard listening on http://%s:%d/network",
             args.network_http_host,
@@ -641,8 +647,8 @@ def _on_signal(sig: int) -> None:
             pass
     if http_server is not None:
         http_server.should_exit = True
-    if http_task is not None:
-        await http_task
+    if http_thread is not None:
+        await asyncio.to_thread(http_thread.join, args.shutdown_grace_s)
     if prefill_hook is not None:
         prefill_hook.close()
     await server.stop(grace=args.shutdown_grace_s)

tests/inference_engine/bridge/test_runtime_http_thread.py

@@ -0,0 +1,14 @@
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[3]
+RUNTIME = ROOT / "scripts" / "start_grpc_runtime_server.py"
+
+
+def test_network_http_runs_outside_blocking_grpc_event_loop():
+    source = RUNTIME.read_text()
+    assert 'name="kakeya-network-http"' in source
+    assert "target=http_server.run" in source
+    assert "http_thread.start()" in source
+    assert "asyncio.create_task(http_server.serve())" not in source
+    assert "asyncio.to_thread(http_thread.join" in source