scorevision: push artifact

chute_config.yml

@@ -3,7 +3,6 @@ Image:
   run_command:
     - pip install --upgrade setuptools wheel
     - "pip install 'numpy>=1.23' 'onnxruntime-gpu>=1.16' 'nvidia-cudnn-cu12' 'nvidia-cublas-cu12' 'opencv-python-headless>=4.7' 'pillow>=9.5' 'huggingface_hub>=0.19.4' 'pydantic>=2.0' 'pyyaml>=6.0' 'aiohttp>=3.9' 'ensemble-boxes>=1.0'"
-    - "python3 -c \"import os,nvidia.cudnn,nvidia.cublas;cudnn=os.path.join(os.path.dirname(nvidia.cudnn.__file__),chr(108)+chr(105)+chr(98));cublas=os.path.join(os.path.dirname(nvidia.cublas.__file__),chr(108)+chr(105)+chr(98));open(chr(47)+chr(101)+chr(116)+chr(99)+chr(47)+chr(108)+chr(100)+chr(46)+chr(115)+chr(111)+chr(46)+chr(99)+chr(111)+chr(110)+chr(102)+chr(46)+chr(100)+chr(47)+chr(110)+chr(118)+chr(105)+chr(100)+chr(105)+chr(97)+chr(45)+chr(111)+chr(114)+chr(116)+chr(46)+chr(99)+chr(111)+chr(110)+chr(102),chr(119)).write(cudnn+chr(10)+cublas+chr(10))\" && ldconfig"
 
 NodeSelector:
   gpu_count: 1

miner.py

@@ -15,6 +15,44 @@ cls_id 0 is shared: "bus" for vehicle eval, "person" for person eval.
 Vehicle eval uses cls_id 0-3. Person eval uses cls_id 0 only.
 """
 
+import os
+import ctypes
+import glob as _glob
+import logging as _logging
+
+_cuda_log = _logging.getLogger(__name__)
+
+def _preload_cuda_libs():
+    """Pre-load CUDA libs from pip nvidia packages so onnxruntime-gpu finds them."""
+    try:
+        lib_dirs = []
+        for mod_name in ['nvidia.cudnn', 'nvidia.cublas']:
+            try:
+                mod = __import__(mod_name, fromlist=['__file__'])
+                lib_dir = os.path.join(os.path.dirname(mod.__file__), 'lib')
+                if os.path.isdir(lib_dir):
+                    lib_dirs.append(lib_dir)
+            except ImportError:
+                pass
+        if not lib_dirs:
+            return
+        # Set LD_LIBRARY_PATH for subprocesses
+        existing = os.environ.get('LD_LIBRARY_PATH', '')
+        os.environ['LD_LIBRARY_PATH'] = ':'.join(lib_dirs + ([existing] if existing else []))
+        # Pre-load .so files with RTLD_GLOBAL so dlopen() finds them
+        for lib_dir in lib_dirs:
+            for so in sorted(_glob.glob(os.path.join(lib_dir, 'lib*.so*'))):
+                try:
+                    ctypes.CDLL(so, mode=ctypes.RTLD_GLOBAL)
+                    _cuda_log.info(f'Preloaded CUDA lib: {os.path.basename(so)}')
+                except OSError:
+                    pass
+    except Exception as e:
+        _cuda_log.warning(f'CUDA preload error: {e}')
+
+_preload_cuda_libs()
+
+
 from pathlib import Path
 import math
 import time