fix: Increase queue buffer to absorb GPT latency

inference.py

@@ -972,7 +972,8 @@ def run_inference(
     # queue_out: (frame_idx, processed_frame, detections)
     queue_in = Queue(maxsize=16) 
     # Tuning for A10: buffer at least 32 frames per GPU (batch size)
-    queue_out_max = max(64, (len(detectors) if detectors else 1) * 32)
+    # GPT Latency Buffer: GPT takes ~3s. At 30fps, that's 90 frames. We need to absorb this burst.
+    queue_out_max = max(512, (len(detectors) if detectors else 1) * 64)
     queue_out = Queue(maxsize=queue_out_max)
     
     
@@ -1122,8 +1123,8 @@ def run_inference(
                             # However, if 'buffer' grows too large (because we are missing next_idx),
                             # we are effectively unbounded again if queue_out fills up with future frames.
                             # So we should monitor buffer size.
-                            if len(buffer) > 64:
-                                logging.warning("Writer buffer large (%d items), waiting for frame %d...", len(buffer), next_idx)
+                            if len(buffer) > 200 and len(buffer) % 50 == 0:
+                                logging.warning("Writer buffer large (%d items), waiting for frame %d (GPT Latency?)...", len(buffer), next_idx)
                             
                             item = queue_out.get(timeout=1.0) # wait