Respect CUDA_VISIBLE_DEVICES env var (do not delete it)

inference.py

@@ -1,8 +1,9 @@
 # CRITICAL: Clear CUDA_VISIBLE_DEVICES BEFORE any imports
 # HF Spaces may set this to "0" dynamically, locking us to a single GPU
+# import os
+# if "CUDA_VISIBLE_DEVICES" in os.environ:
+#     del os.environ["CUDA_VISIBLE_DEVICES"]
 import os
-if "CUDA_VISIBLE_DEVICES" in os.environ:
-    del os.environ["CUDA_VISIBLE_DEVICES"]
 
 import logging
 import time
@@ -1453,12 +1454,13 @@ def run_segmentation(
     logging.info(f"[DEBUG] Segmentation PID: {os.getpid()}")
     logging.info(f"[DEBUG] CUDA_VISIBLE_DEVICES before clear: {os.environ.get('CUDA_VISIBLE_DEVICES')}")
     
-    if "CUDA_VISIBLE_DEVICES" in os.environ:
-         logging.info("[DEBUG] Deleting CUDA_VISIBLE_DEVICES from env (segmentation)")
-         del os.environ["CUDA_VISIBLE_DEVICES"]
+    
+    # if "CUDA_VISIBLE_DEVICES" in os.environ:
+    #      logging.info("[DEBUG] Deleting CUDA_VISIBLE_DEVICES from env (segmentation)")
+    #      del os.environ["CUDA_VISIBLE_DEVICES"]
 
     num_gpus = torch.cuda.device_count()
-    logging.info(f"[DEBUG] num_gpus after clear: {num_gpus}")
+    logging.info(f"[DEBUG] num_gpus: {num_gpus}")
     segmenters = []
 
     if num_gpus > 0:
@@ -1680,8 +1682,8 @@ def run_depth_inference(
     num_gpus = torch.cuda.device_count()
     estimators = []
     
-    if "CUDA_VISIBLE_DEVICES" in os.environ:
-         del os.environ["CUDA_VISIBLE_DEVICES"]
+    # if "CUDA_VISIBLE_DEVICES" in os.environ:
+    #      del os.environ["CUDA_VISIBLE_DEVICES"]
 
     if num_gpus > 0:
         logging.info("Detected %d GPUs. Loading depth estimators...", num_gpus)