Update app.py

app.py

@@ -15,17 +15,23 @@ from huggingface_hub import snapshot_download
 import zipfile
 import json
 from pathlib import Path
-import diffusers
-print(f"[DIAG] diffusers version   : {diffusers.__version__}", flush=True)
+
 
 def _get_subprocess_env() -> dict:
-    """Stellt sicher, dass libcudart im Subprocess-Pfad verfügbar ist."""
+    """Baut die Subprocess-Umgebung mit korrekten CUDA- und Library-Pfaden.
+
+    Wichtige Fixes:
+    - LD_LIBRARY_PATH: libcudart.so via nvidia pip-Pakete findbar machen
+    - PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True nutzt NVML, das im
+      Subprocess auf ZeroGPU Blackwell + torch 2.11.0 nicht stabil ist.
+      Im Subprocess max_split_size_mb:512 verwenden (NVML-frei, stabil).
+    """
     import site
     import glob as _glob
 
     env = os.environ.copy()
 
-    # 1. NVIDIA pip-Pakete (nvidia-cuda-runtime-cu12/cu13 etc.)
+    # 1. CUDA-Library-Pfade für libcudart
     nvidia_paths: list[str] = []
     all_site = site.getsitepackages()
     try:
@@ -35,53 +41,44 @@ def _get_subprocess_env() -> dict:
     for sp in all_site:
         matches = _glob.glob(os.path.join(sp, "nvidia", "*", "lib"))
         nvidia_paths.extend(m for m in matches if os.path.isdir(m))
-
-    # 2. PyTorch-eigenes lib/-Verzeichnis
     torch_lib = os.path.join(os.path.dirname(torch.__file__), "lib")
-
-    # 3. System-CUDA-Pfade als Fallback
     system_cuda = [
-        "/usr/local/cuda-13/lib64",
-        "/usr/local/cuda-13.0/lib64",
-        "/usr/local/cuda-12/lib64",
-        "/usr/local/cuda-12.8/lib64",
-        "/usr/local/cuda/lib64",
-        "/usr/local/cuda/targets/x86_64-linux/lib",
+        "/usr/local/cuda-13/lib64", "/usr/local/cuda-13.0/lib64",
+        "/usr/local/cuda-12/lib64", "/usr/local/cuda/lib64",
         "/usr/lib/x86_64-linux-gnu",
     ]
-    system_valid = [p for p in system_cuda if os.path.isdir(p)]
-
-    all_paths = nvidia_paths + [torch_lib] + system_valid
+    extra = [p for p in system_cuda if os.path.isdir(p)]
     existing_ld = env.get("LD_LIBRARY_PATH", "")
     env["LD_LIBRARY_PATH"] = ":".join(
-        all_paths + ([existing_ld] if existing_ld else [])
+        nvidia_paths + [torch_lib] + extra + ([existing_ld] if existing_ld else [])
     )
+
+    # 2. CUDA-Allocator: expandable_segments:True braucht NVML, das im
+    #    ZeroGPU-Subprocess mit torch 2.11.0 zu CUDACachingAllocator-Assertions führt.
+    #    Im Subprocess stabilen Fallback verwenden.
+    env["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:512"
+
     return env
 
 
 def _log_cuda_diagnostics():
-    """Loggt torch-Version, CUDA-Build und libcudart-Pfad beim Start."""
+    """Loggt torch-Version, CUDA, libcudart-Pfad und pytorch3d beim Start."""
     import glob as _glob
+    import diffusers
+    print(f"[DIAG] diffusers version   : {diffusers.__version__}", flush=True)
     print(f"[DIAG] torch version      : {torch.__version__}", flush=True)
     print(f"[DIAG] torch CUDA build   : {torch.version.cuda}", flush=True)
     print(f"[DIAG] CUDA available     : {torch.cuda.is_available()}", flush=True)
     print(f"[DIAG] sys.executable     : {sys.executable}", flush=True)
-
     env = _get_subprocess_env()
     ld_path = env.get("LD_LIBRARY_PATH", "")
     print(f"[DIAG] LD_LIBRARY_PATH    : {ld_path}", flush=True)
-
     found = []
     for d in ld_path.split(":"):
-        if not d:
-            continue
-        hits = _glob.glob(os.path.join(d, "libcudart.so*"))
-        found.extend(hits)
-    if found:
-        print(f"[DIAG] libcudart found    : {found}", flush=True)
-    else:
-        print("[DIAG] libcudart found    : NICHT GEFUNDEN", flush=True)
-
+        if d:
+            found.extend(_glob.glob(os.path.join(d, "libcudart.so*")))
+    print(f"[DIAG] libcudart found    : {found}", flush=True)
+    print(f"[DIAG] subprocess ALLOC   : {env.get('PYTORCH_CUDA_ALLOC_CONF')}", flush=True)
     import importlib.util
     p3d = importlib.util.find_spec("pytorch3d")
     print(f"[DIAG] pytorch3d spec     : {p3d}", flush=True)