Fix: GPU detection order and distance consistency

inference.py

@@ -455,14 +455,16 @@ def run_inference(
     active_detector = detector_name or "hf_yolov8"
     
     # 3. Parallel Model Loading
-    num_gpus = torch.cuda.device_count()
-    detectors = []
-    depth_estimators = []
     
     # Clear CUDA_VISIBLE_DEVICES to ensure we see all GPUs if not already handled
+    # This must be done BEFORE any torch.cuda calls in this scope if the env was modified externally
     if "CUDA_VISIBLE_DEVICES" in os.environ:
          del os.environ["CUDA_VISIBLE_DEVICES"]
 
+    num_gpus = torch.cuda.device_count()
+    detectors = []
+    depth_estimators = []
+    
     if num_gpus > 0:
         logging.info("Detected %d GPUs. Loading models in parallel...", num_gpus)
         
@@ -671,12 +673,13 @@ def run_segmentation(
     logging.info("Using segmenter: %s with queries: %s", active_segmenter, queries)
 
     # 2. Load Segmenters (Parallel)
-    num_gpus = torch.cuda.device_count()
-    segmenters = []
     
     if "CUDA_VISIBLE_DEVICES" in os.environ:
          del os.environ["CUDA_VISIBLE_DEVICES"]
 
+    num_gpus = torch.cuda.device_count()
+    segmenters = []
+
     if num_gpus > 0:
         logging.info("Detected %d GPUs. Loading segmenters...", num_gpus)
         def load_seg(gpu_id: int):