v3.30: vehicle TRT background build + disable vehicle TTA + skip plate on vehicle-only hint

sv_gpu.py

@@ -1,5 +1,5 @@
 """
-Score Vision SN44 — Unified miner v3.29 (2026-04-08). R9c vehicle FP16 (mAP50=0.929). Person: TTA consensus + 15% box shrink + NMS 0.35.
+Score Vision SN44 — Unified miner v3.30 (2026-04-08). R9c vehicle FP16 (mAP50=0.929) + TRT background build. Person: TTA consensus + 15% box shrink + NMS 0.35. Vehicle TTA disabled for RTF compliance.
 Dual-model: vehicle (YOLO11m INT8 1280) + person (YOLO12s FP16 960 TRT).
 Pose model: YOLOv8n-pose FP16 640 for false-positive filtering + keypoint box refinement.
 Vehicle weights loaded from secondary HF repo (meaculpitt/ScoreVision-Vehicle).
@@ -368,7 +368,8 @@ TRT_WORKSPACE_GB = 4
 WBF_SKIP_THR = 0.0001
 
 # ── Speed config ────────────────────────────────────────────────────────────
-ENABLE_TTA = True
+ENABLE_TTA = True             # Person TTA (flip consensus merge)
+VEHICLE_TTA = False           # Vehicle flip TTA disabled — saves ~200-400ms, RTF compliance
 ENABLE_PARALLEL = True
 
 # ── Secondary HF repo for vehicle weights ───────────────────────────────────
@@ -608,13 +609,22 @@ class Miner:
         self._trt_ready = False
         logger.info("[init] Person model: CUDA (TRT build starting in background)")
 
-        # Launch background TRT engine build
+        # Launch background TRT engine builds
         os.makedirs(TRT_CACHE_PATH, exist_ok=True)
         threading.Thread(
             target=self._build_trt_engine,
-            args=(per_onnx,),
+            args=(per_onnx, "person"),
             daemon=True,
-            name="trt-builder",
+            name="trt-builder-person",
+        ).start()
+
+        # Vehicle TRT build — same pattern, separate thread
+        self._veh_trt_ready = False
+        threading.Thread(
+            target=self._build_trt_engine,
+            args=(veh_weights, "vehicle"),
+            daemon=True,
+            name="trt-builder-vehicle",
         ).start()
 
         # Pose model — for FP filtering + box refinement
@@ -676,12 +686,12 @@ class Miner:
         logger.info(f"Person ORT providers: {per_prov} (TRT building in background)")
         logger.info(f"TTA={ENABLE_TTA} PARALLEL={ENABLE_PARALLEL}")
 
-    def _build_trt_engine(self, per_onnx):
-        """Build TRT FP16 engine in background, swap person session when ready.
+    def _build_trt_engine(self, onnx_path, model_name="person"):
+        """Build TRT FP16 engine in background, swap session when ready.
 
         On fresh nodes: ~18 min to compile. Cached engine loads in <1s.
-        During build, inference uses CUDAExecutionProvider (passes RTF at ~78ms).
-        After build, atomically swaps to TRT session (~29ms pipeline).
+        During build, inference uses CUDAExecutionProvider.
+        After build, atomically swaps to TRT session.
         """
         try:
             trt_opts = {
@@ -691,9 +701,9 @@ class Miner:
                 "trt_engine_cache_path": TRT_CACHE_PATH,
             }
             t0 = time.monotonic()
-            logger.info("[trt-build] Creating TRT session (may take ~18min on fresh node)...")
+            logger.info("[trt-build-%s] Creating TRT session (may take ~18min on fresh node)...", model_name)
             trt_session = ort.InferenceSession(
-                per_onnx,
+                onnx_path,
                 providers=[
                     ("TensorrtExecutionProvider", trt_opts),
                     "CUDAExecutionProvider",
@@ -703,7 +713,7 @@ class Miner:
 
             provs = trt_session.get_providers()
             if "TensorrtExecutionProvider" not in provs:
-                logger.warning("[trt-build] TRT provider not active (%s), keeping CUDA", provs)
+                logger.warning("[trt-build-%s] TRT provider not active (%s), keeping CUDA", model_name, provs)
                 return
 
             # Run dummy inference to fully materialize the engine
@@ -713,21 +723,24 @@ class Miner:
             trt_session.run(None, {inp_name: dummy})
 
             dt = time.monotonic() - t0
-            logger.info("[trt-build] TRT engine ready in %.1fs — swapping person session", dt)
+            logger.info("[trt-build-%s] TRT engine ready in %.1fs — swapping session", model_name, dt)
 
             # Atomic swap — Python GIL makes single attribute assignment safe.
-            # Any in-flight inference holds a reference to the old session, which
-            # stays alive until that inference completes.
-            self.per_session = trt_session
-            self._trt_ready = True
-
-            logger.info("[trt-build] Person model now using TensorRT FP16")
+            if model_name == "person":
+                self.per_session = trt_session
+                self._trt_ready = True
+                logger.info("[trt-build-person] Person model now using TensorRT FP16")
+            elif model_name == "vehicle":
+                self.veh_session = trt_session
+                self._veh_trt_ready = True
+                logger.info("[trt-build-vehicle] Vehicle model now using TensorRT FP16")
         except Exception as e:
-            logger.warning("[trt-build] TRT build failed (%s), keeping CUDA", e)
+            logger.warning("[trt-build-%s] TRT build failed (%s), keeping CUDA", model_name, e)
 
     def __repr__(self) -> str:
-        trt_status = "TRT" if self._trt_ready else "CUDA (TRT building)"
-        return f"Unified Miner v3.16 — person={trt_status}, background TRT engine build"
+        per_trt = "TRT" if self._trt_ready else "CUDA (TRT building)"
+        veh_trt = "TRT" if self._veh_trt_ready else "CUDA (TRT building)"
+        return f"Unified Miner v3.30 — person={per_trt}, vehicle={veh_trt}"
 
     # ── Vehicle preprocessing (letterbox) ───────────────────────────────────
 
@@ -785,7 +798,7 @@ class Miner:
         boxes, confs, cls_ids = self._veh_run_pass(image_bgr, VEH_CONF_THRES, session)
 
         # Flip TTA pass — horizontal flip, mirror boxes back
-        if ENABLE_TTA:
+        if VEHICLE_TTA:
             flipped = cv2.flip(image_bgr, 1)
             f_boxes, f_confs, f_cls = self._veh_run_pass(flipped, VEH_TTA_CONF, session)
             if len(f_boxes) > 0:
@@ -1161,7 +1174,7 @@ class Miner:
             confs = pred[:, 4] * np.max(pred[:, 5:], axis=1)
         return bool((confs >= VEH_PARTS_PLATE_CONF).any())
 
-    def _vehicle_parts_confirm(self, vehicle_boxes, person_boxes, image_bgr):
+    def _vehicle_parts_confirm(self, vehicle_boxes, person_boxes, image_bgr, skip_plate=False):
         """Parts-based confidence scoring for vehicle detections.
 
         Scoring hierarchy (confidence boosts are additive):
@@ -1182,7 +1195,9 @@ class Miner:
         img_area = float(oh * ow)
         has_plate_model = self.plate_session is not None
         # Skip plate checks on crowded scenes (aerial/drone, plates invisible)
-        skip_plate = len(vehicle_boxes) > 20
+        # Also skip when called from vehicle-only hint (saves ONNX inference latency)
+        if not skip_plate:
+            skip_plate = len(vehicle_boxes) > 20
 
         result = []
         n_driver = 0
@@ -2083,10 +2098,10 @@ class Miner:
 
         if element_hint == 'vehicle':
             # Run vehicle detection + parts confirmation with empty person_boxes.
-            # Plate/headlight/window checks fire normally; driver/rider overlap
-            # check finds no matches (boost=0) but doesn't suppress.
+            # Skip plate ONNX to save latency — no person boxes means driver/rider
+            # checks are no-ops anyway.
             vehicle_boxes = self._infer_vehicle(image_bgr)
-            return self._vehicle_parts_confirm(vehicle_boxes, [], image_bgr)
+            return self._vehicle_parts_confirm(vehicle_boxes, [], image_bgr, skip_plate=True)
 
         # Fallback: run both (original behavior)
         if ENABLE_PARALLEL: