Move CUDA compilation into @spaces.GPU function — nvcc unavailable at APP_STARTING

On ZeroGPU, nvcc is only accessible inside @spaces.GPU decorated calls.
Remove CUDA packages from _install_runtime_packages() (CPU-only startup).
Add _install_cuda_packages() with @spaces.GPU(duration=300) called once
at module level after `import spaces`, so nvcc is available when it runs.

app.py

@@ -126,15 +126,34 @@ def _install_runtime_packages():
         [sys.executable, "-m", "pip", "install", "--quiet", "--no-deps"]
         + _NO_DEPS_PACKAGES, check=True,
     )
-    # A10G = sm_86. Set arch list explicitly because NVML is unavailable at install
-    # time (no GPU allocated yet), so torch can't auto-detect it.
-    # Detect CUDA home dynamically — path varies between ZeroGPU container builds.
+    _RUNTIME_PKG_MARKER.touch()
+    print("[startup] CPU runtime packages installed.")
+
+_install_runtime_packages()
+# ──────────────────────────────────────────────────────────────────────────────
+
+import cv2
+import gradio as gr
+import spaces
+import torch
+import numpy as np
+from PIL import Image
+
+# ── CUDA package installation ─────────────────────────────────────────────────
+# nvcc is only available inside a @spaces.GPU call on ZeroGPU (not at APP_STARTING).
+# Compile nvdiffrast / detectron2 / diso here, on first GPU allocation at startup.
+_CUDA_PKG_MARKER = Path("/tmp/.cuda_pkgs_installed")
+
+@spaces.GPU(duration=300)
+def _install_cuda_packages():
+    if _CUDA_PKG_MARKER.exists():
+        return
+    print("[startup] Installing CUDA packages (nvdiffrast, detectron2, diso)...")
     import shutil as _shutil
     _nvcc = _shutil.which("nvcc")
     if _nvcc:
         _cuda_home = str(Path(_nvcc).parent.parent)
     else:
-        # Scan common ZeroGPU / CUDA container paths
         for _cand in [
             "/usr/local/cuda",
             "/usr/local/cuda-12.9",
@@ -148,10 +167,8 @@ def _install_runtime_packages():
                 _cuda_home = _cand
                 break
         else:
-            raise RuntimeError(
-                "nvcc not found — no CUDA toolkit in PATH or standard locations. "
-                "Cannot build nvdiffrast/detectron2/diso."
-            )
+            print("[startup] WARNING: nvcc not found even with GPU allocated — CUDA extensions unavailable")
+            return
     print(f"[startup] CUDA home: {_cuda_home}")
     _cuda_env = {
         **os.environ,
@@ -165,8 +182,7 @@ def _install_runtime_packages():
         [sys.executable, "-m", "pip", "install", "--quiet", "--no-build-isolation"]
         + _CUDA_PACKAGES, env=_cuda_env, check=True,
     )
-    # diso: pip install git+... does not fetch git submodules, causing undefined
-    # CUDA symbols at import time. Clone with --recurse-submodules first.
+    # diso must be cloned with --recurse-submodules; pip install git+... skips submodules
     _diso_src = Path("/tmp/diso-build")
     if not _diso_src.exists():
         subprocess.run(
@@ -179,19 +195,10 @@ def _install_runtime_packages():
          str(_diso_src)],
         env=_cuda_env, check=True,
     )
-    _RUNTIME_PKG_MARKER.touch()
-    print("[startup] Runtime packages installed.")
-
-_install_runtime_packages()
-# ──────────────────────────────────────────────────────────────────────────────
-
-import cv2
-import gradio as gr
-import spaces
-import torch
-import numpy as np
-from PIL import Image
+    _CUDA_PKG_MARKER.touch()
+    print("[startup] CUDA packages installed.")
 
+_install_cuda_packages()
 # ── Paths ─────────────────────────────────────────────────────────────────────
 HERE        = Path(__file__).parent
 PIPELINE_DIR = HERE / "pipeline"