fix: Move CUDA_VISIBLE_DEVICES clearing to module import time

- Moved env var clearing to very top of inference.py before torch import
- Removed redundant debug logging and clearing logic from run_inference
- Ensures all 4 GPUs are visible before PyTorch CUDA context initializes
- Fixes HF Spaces issue where CUDA_VISIBLE_DEVICES=0 locks to single GPU

inference.py

@@ -1,18 +1,15 @@
-import logging
+# 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 logging
 import time
 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
@@ -612,21 +609,8 @@ def run_inference(
     logging.info("Detection queries: %s", queries)
     active_detector = detector_name or "hf_yolov8"
     
-    # 3. Parallel Model Loading
-    
-    # DEBUG: Log current state
-    logging.info(f"[DEBUG] PID: {os.getpid()}")
-    logging.info(f"[DEBUG] CUDA_VISIBLE_DEVICES before clear: {os.environ.get('CUDA_VISIBLE_DEVICES')}")
-    logging.info(f"[DEBUG] torch.cuda.device_count() before clear: {torch.cuda.device_count()}")
-
-    # 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:
-         logging.info("[DEBUG] Deleting CUDA_VISIBLE_DEVICES from env")
-         del os.environ["CUDA_VISIBLE_DEVICES"]
-
+    # Parallel Model Loading
     num_gpus = torch.cuda.device_count()
-    logging.info(f"[DEBUG] num_gpus after clear: {num_gpus}")
     detectors = []
     depth_estimators = []