ZeroGPU: compile CUDA packages at runtime startup (no nvcc in build container)

app.py

@@ -1,5 +1,6 @@
 import sys
 import os
+import subprocess
 import tempfile
 import shutil
 import traceback
@@ -7,6 +8,31 @@ import json
 import random
 from pathlib import Path
 
+# ── ZeroGPU: compile CUDA extensions on first startup ─────────────────────────
+# nvdiffrast, diso, and detectron2 require nvcc to compile their CUDA kernels.
+# HuggingFace ZeroGPU build containers have no GPU/nvcc; only the runtime
+# containers do. We install these once on startup and cache via a marker file.
+_CUDA_PKG_MARKER = Path("/tmp/.cuda_pkgs_installed")
+_CUDA_PACKAGES = [
+    "nvdiffrast @ git+https://github.com/NVlabs/nvdiffrast.git@253ac4fcea7de5f396371124af597e6cc957bfae",
+    "diso @ git+https://github.com/SarahWeiii/diso.git@9792ad928ccb09bdec938779651ee03e395758a6",
+    "detectron2 @ git+https://github.com/facebookresearch/detectron2.git@8a9d885b3d4dcf1bef015f0593b872ed8d32b4ab",
+]
+
+def _install_cuda_packages():
+    if _CUDA_PKG_MARKER.exists():
+        return
+    print("[startup] Installing CUDA-compiled packages (first run, may take a few minutes)...")
+    subprocess.run(
+        [sys.executable, "-m", "pip", "install", "--quiet"] + _CUDA_PACKAGES,
+        check=True
+    )
+    _CUDA_PKG_MARKER.touch()
+    print("[startup] CUDA packages installed.")
+
+_install_cuda_packages()
+# ──────────────────────────────────────────────────────────────────────────────
+
 import cv2
 import gradio as gr
 import spaces

packages.txt

@@ -8,4 +8,3 @@ cmake
 ninja-build
 build-essential
 pkg-config
-nvidia-cuda-toolkit

requirements.txt

@@ -8,9 +8,9 @@ clip @ git+https://github.com/openai/CLIP.git@d05afc436d78f1c48dc0dbf8e5980a9d47
 mvadapter @ git+https://github.com/huanngzh/MV-Adapter.git@4277e0018232bac82bb2c103caf0893cedb711be
 chumpy @ https://huggingface.co/datasets/Daankular/wheels/resolve/main/chumpy-0.71-py3-none-any.whl
 skel @ git+https://github.com/MarilynKeller/SKEL.git@c32cf16581295bff19399379efe5b776d707cd95
-nvdiffrast @ git+https://github.com/NVlabs/nvdiffrast.git@253ac4fcea7de5f396371124af597e6cc957bfae
-diso @ git+https://github.com/SarahWeiii/diso.git@9792ad928ccb09bdec938779651ee03e395758a6
-detectron2 @ git+https://github.com/facebookresearch/detectron2.git@8a9d885b3d4dcf1bef015f0593b872ed8d32b4ab
+# nvdiffrast, diso, detectron2 require nvcc (CUDA compiler) which is only available
+# in the ZeroGPU runtime container, not the Docker build container.
+# These are compiled and installed on first startup — see app.py _install_cuda_packages().
 
 # Core ML
 accelerate