Preload libcudart.so.13 via ctypes.CDLL before nvdiffrast import

app.py

@@ -25,14 +25,29 @@ import sys
 import time
 from pathlib import Path
 
-# nvdiffrast is compiled against the machine's CUDA 13.0 toolkit, so its .so
-# links to libcudart.so.13.  Add the CUDA 13.0 lib dir before any import so
-# the dynamic linker can find it.  Must happen before `import nvdiffrast`.
-_cuda_lib = "/cuda-image/usr/local/cuda-13.0/lib64"
-if os.path.isdir(_cuda_lib):
-    os.environ["LD_LIBRARY_PATH"] = (
-        _cuda_lib + ":" + os.environ.get("LD_LIBRARY_PATH", "")
-    )
+# nvdiffrast is compiled against CUDA 13.0 and its .so links to libcudart.so.13.
+# os.environ["LD_LIBRARY_PATH"] only affects child processes, not the running
+# Python interpreter.  Use ctypes.CDLL with RTLD_GLOBAL to load libcudart.so.13
+# into the current process before nvdiffrast is imported.
+import ctypes, glob as _glob
+
+def _preload_cudart() -> None:
+    cuda_lib_dir = "/cuda-image/usr/local/cuda-13.0/lib64"
+    # Try versioned name first, then unversioned fallback.
+    for pattern in (
+        f"{cuda_lib_dir}/libcudart.so.13*",
+        f"{cuda_lib_dir}/libcudart.so",
+    ):
+        matches = _glob.glob(pattern)
+        if matches:
+            try:
+                ctypes.CDLL(matches[0], mode=ctypes.RTLD_GLOBAL)
+                print(f"Preloaded {matches[0]}", flush=True)
+            except OSError as e:
+                print(f"Preload failed for {matches[0]}: {e}", flush=True)
+            return
+
+_preload_cudart()
 
 import gradio as gr
 import spaces  # HF ZeroGPU