Add RTMLib CUDA backends and staged timing

servers/unraid/tail2-pose-server/app/tail2/core.py

@@ -41,6 +41,8 @@ def inverse_letterbox(points: np.ndarray, scale: float, pad_x: float, pad_y: flo
 class LatestQueue:
     def __init__(self) -> None:
         self._q: queue.Queue[Any] = queue.Queue(maxsize=1)
+        self._dropped = 0
+        self._lock = threading.Lock()
 
     def put(self, item: Any) -> None:
         try:
@@ -48,12 +50,17 @@ class LatestQueue:
         except queue.Full:
             try:
                 self._q.get_nowait()
+                with self._lock:
+                    self._dropped += 1
             except queue.Empty:
                 pass
             self._q.put_nowait(item)
 
-    def get(self, timeout: float | None = None) -> Any:
-        return self._q.get(timeout=timeout)
+    def get(self, timeout: float | None = None) -> tuple[Any, int]:
+        item = self._q.get(timeout=timeout)
+        with self._lock:
+            dropped, self._dropped = self._dropped, 0
+        return item, dropped
 
     def qsize(self) -> int:
         return self._q.qsize()
@@ -74,7 +81,7 @@ class BroadcastHub:
         with self._lock:
             self._clients.discard(q)
 
-    def publish_on_loop(self, loop: asyncio.AbstractEventLoop, message: dict) -> None:
+    def publish_on_loop(self, loop: asyncio.AbstractEventLoop, message: str) -> None:
         def publish() -> None:
             with self._lock:
                 clients = list(self._clients)