scorevision: push artifact

miner.py

@@ -1,5 +1,5 @@
 """
-Score Vision SN44 — Unified miner v3.23.1 (2026-04-06). Hotfix: FP32 fallback trigger 0 boxes only (was ≤1). TTA consensus (person), FP32 fallback (vehicle), parts_confirm on vehicle challenges, car conf 0.60, VEH_MIN_WH=20.
+Score Vision SN44 — Unified miner v3.23.2 (2026-04-07). Force 24GB GPU, lazy-load FP32 fallback. FP32 trigger 0 boxes only. TTA consensus (person), parts_confirm on vehicle challenges, car conf 0.60, VEH_MIN_WH=20.
 Tri-model: vehicle (YOLO11m INT8 1280) + person (YOLO12s FP16 960 TRT) + petrol (end2end 640).
 Pose model: YOLOv8n-pose FP16 640 for false-positive filtering + keypoint box refinement.
 Vehicle weights loaded from secondary HF repo (meaculpitt/ScoreVision-Vehicle).
@@ -602,20 +602,18 @@ class Miner:
         self.veh_h = int(veh_shape[2])
         self.veh_w = int(veh_shape[3])
 
-        # FP32 fallback session for INT8 degradation recovery (block 7905900: 1-box failure)
+        # FP32 fallback — lazy-loaded on first trigger to save ~300MB VRAM at startup
         self.veh_session_fp32 = None
+        self._veh_fp32_path = None
         try:
             veh_fp32 = str(veh_path / "vehicle_weights_fp32.onnx") if veh_path else None
             if veh_fp32 and Path(veh_fp32).exists():
-                self.veh_session_fp32 = ort.InferenceSession(
-                    veh_fp32,
-                    providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
-                )
-                logger.info("[init] Vehicle FP32 fallback model loaded")
+                self._veh_fp32_path = veh_fp32
+                logger.info("[init] Vehicle FP32 fallback available (lazy-load)")
             else:
                 logger.info("[init] Vehicle FP32 fallback not available")
         except Exception as e:
-            logger.warning(f"[init] Vehicle FP32 fallback failed: {e}")
+            logger.warning(f"[init] Vehicle FP32 fallback path check failed: {e}")
 
         # Person model — CUDA immediately, TRT engine builds in background
         per_onnx = str(path_hf_repo / "person_weights.onnx")
@@ -938,14 +936,26 @@ class Miner:
         """
         boxes = self._infer_vehicle_core(image_bgr, self.veh_session)
 
-        if len(boxes) == 0 and self.veh_session_fp32:
-            boxes_fp32 = self._infer_vehicle_core(image_bgr, self.veh_session_fp32)
-            if len(boxes_fp32) > len(boxes):
-                logger.warning(
-                    f"[vehicle] INT8 degraded ({len(boxes)} boxes), "
-                    f"FP32 fallback recovered ({len(boxes_fp32)} boxes)"
-                )
-                return boxes_fp32
+        if len(boxes) == 0 and (self.veh_session_fp32 or self._veh_fp32_path):
+            # Lazy-load FP32 session on first trigger
+            if self.veh_session_fp32 is None and self._veh_fp32_path:
+                try:
+                    self.veh_session_fp32 = ort.InferenceSession(
+                        self._veh_fp32_path,
+                        providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+                    )
+                    logger.info("[vehicle] FP32 fallback lazy-loaded")
+                except Exception as e:
+                    logger.warning(f"[vehicle] FP32 lazy-load failed: {e}")
+                    self._veh_fp32_path = None
+            if self.veh_session_fp32:
+                boxes_fp32 = self._infer_vehicle_core(image_bgr, self.veh_session_fp32)
+                if len(boxes_fp32) > len(boxes):
+                    logger.warning(
+                        f"[vehicle] INT8 degraded ({len(boxes)} boxes), "
+                        f"FP32 fallback recovered ({len(boxes_fp32)} boxes)"
+                    )
+                    return boxes_fp32
 
         return boxes