scorevision: push artifact

miner.py

@@ -73,11 +73,14 @@ class Miner:
         offset: int,
         n_keypoints: int,
     ) -> list[TVFrameResult]:
+        # device="cpu": chute GPUs (Blackwell) lack kernels for the image's
+        # torch build, and the ONNX model targets CPU for the latency gate.
         detections = self.model.predict(
             batch_images,
             conf=CONF_THRESHOLD,
             imgsz=IMG_SIZE,
             classes=[PERSON_CLS_ID],
+            device=os.getenv("SV_DEVICE", "cpu"),
             verbose=False,
         )