fix:multi-gpu support

inference.py

@@ -5,6 +5,14 @@ from threading import RLock, Thread
 from queue import Queue, PriorityQueue
 from typing import Any, Dict, List, Optional, Sequence, Tuple
 
+# Must clear CUDA_VISIBLE_DEVICES before importing torch to ensure all GPUs are seen
+if "CUDA_VISIBLE_DEVICES" in os.environ:
+    # We only clear it if we are sure we want all GPUs. 
+    # In HF Spaces 4xGPU, this var might be set to "0" by default? 
+    # Let's log it first implies we need imports. 
+    # Just Unset it.
+    del os.environ["CUDA_VISIBLE_DEVICES"]
+
 import cv2
 import numpy as np
 import torch