feat: fix multi-gpu visibility and improve logging

- Unset CUDA_VISIBLE_DEVICES at start of app.py to force visibility of all GPUs
- Change GPU parallelism and distance logs to INFO level for visibility

app.py

@@ -1,6 +1,15 @@
-import asyncio
-import logging
 import os
+import logging
+# Debug/Fix: Unset CUDA_VISIBLE_DEVICES to ensure all GPUs are visible
+# Some environments (like HF Spaces) might set this to "0" by default.
+if "CUDA_VISIBLE_DEVICES" in os.environ:
+    # Use print because logging config might not be set yet
+    print(f"Found CUDA_VISIBLE_DEVICES={os.environ['CUDA_VISIBLE_DEVICES']}. Unsetting it to enable all GPUs.")
+    del os.environ["CUDA_VISIBLE_DEVICES"]
+else:
+    print("CUDA_VISIBLE_DEVICES not set. All GPUs should be visible.")
+
+import asyncio
 import shutil
 import tempfile
 import uuid

inference.py

@@ -301,7 +301,7 @@ def infer_frame(
                 # Add depth to label, e.g. "car 12m"
                 depth_str = f"{int(det['depth_est_m'])}m"
                 label = f"{label} {depth_str}"
-                logging.debug("Object '%s' at %s (bbox: %s)", label, depth_str, det['bbox'])
+                logging.info("Object '%s' at %s (bbox: %s)", label, depth_str, det['bbox'])
             display_labels.append(label)
 
     except Exception:
@@ -468,7 +468,7 @@ def run_inference(
             detector_instance = detectors[gpu_idx]
             depth_instance = depth_estimators[gpu_idx] if depth_estimators else None
             
-            logging.debug("Processing frame %d on GPU %d (cuda:%d)", frame_idx, gpu_idx, gpu_idx)
+            logging.info("Processing frame %d on GPU %d (cuda:%d)", frame_idx, gpu_idx, gpu_idx)
             
             # Run depth estimation every 3 frames if configured
             active_depth_name = depth_estimator_name if (frame_idx % 3 == 0) else None
@@ -573,7 +573,7 @@ def run_segmentation(
             gpu_idx = frame_idx % len(segmenters)
             segmenter_instance = segmenters[gpu_idx]
             
-            logging.debug("Segmenting frame %d on GPU %d (cuda:%d)", frame_idx, gpu_idx, gpu_idx)
+            logging.info("Segmenting frame %d on GPU %d (cuda:%d)", frame_idx, gpu_idx, gpu_idx)
             
             processed, _ = infer_segmentation_frame(
                 frame_data, 
@@ -716,7 +716,7 @@ def process_frames_depth(
             gpu_idx = frame_idx % len(estimators)
             estimator_instance = estimators[gpu_idx]
             
-            logging.debug("Estimating depth for frame %d on GPU %d (cuda:%d)", frame_idx, gpu_idx, gpu_idx)
+            logging.info("Estimating depth for frame %d on GPU %d (cuda:%d)", frame_idx, gpu_idx, gpu_idx)
             
             # Use instance lock
             if hasattr(estimator_instance, "lock"):