Move hmr2/skel/chumpy to runtime --no-build-isolation; document all ZeroGPU constraints

app.py

@@ -8,29 +8,58 @@ 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")
+# ── ZeroGPU: install packages that can't be built at Docker build time ─────────
+#
+# Two categories of packages must be installed at runtime, not at build time:
+#
+# 1. CUDA-compiled extensions (nvdiffrast, diso, detectron2):
+#    These require nvcc (NVIDIA CUDA compiler). The ZeroGPU Docker build stage
+#    has no GPU/nvcc; only the runtime containers do.
+#
+# 2. Packages with broken build isolation (hmr2, skel → chumpy):
+#    hmr2 and skel declare `chumpy @ git+https://...` as a direct-reference dep.
+#    chumpy's setup.py does `from pip._internal.req import parse_requirements`,
+#    which fails when pip>=21 creates an isolated build environment (pip is not
+#    importable there). Fix: --no-build-isolation skips isolated environments,
+#    making pip importable. This flag can only be passed via subprocess, not
+#    requirements.txt.
+#
+# Packages are installed once on first startup and cached via a marker file.
+# ──────────────────────────────────────────────────────────────────────────────
+_RUNTIME_PKG_MARKER = Path("/tmp/.runtime_pkgs_installed")
+
+# Packages requiring --no-build-isolation (chumpy transitive dep chain)
+_NO_ISOLATION_PACKAGES = [
+    "chumpy @ git+https://github.com/mattloper/chumpy.git@580566eafc9ac68b2614b64d6f7aaa84eebb70da",
+    "hmr2 @ git+https://github.com/shubham-goel/4D-Humans.git@efe18deff163b29dff87ddbd575fa29b716a356c",
+    "skel @ git+https://github.com/MarilynKeller/SKEL.git@c32cf16581295bff19399379efe5b776d707cd95",
+]
+
+# Packages requiring nvcc (compiled at runtime when GPU container is available)
 _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():
+def _install_runtime_packages():
+    if _RUNTIME_PKG_MARKER.exists():
         return
-    print("[startup] Installing CUDA-compiled packages (first run, may take a few minutes)...")
+    print("[startup] Installing runtime packages (first run, ~5-10 min)...")
+    subprocess.run(
+        [sys.executable, "-m", "pip", "install", "--quiet", "--no-build-isolation"]
+        + _NO_ISOLATION_PACKAGES,
+        check=True,
+    )
     subprocess.run(
-        [sys.executable, "-m", "pip", "install", "--quiet"] + _CUDA_PACKAGES,
-        check=True
+        [sys.executable, "-m", "pip", "install", "--quiet"]
+        + _CUDA_PACKAGES,
+        check=True,
     )
-    _CUDA_PKG_MARKER.touch()
-    print("[startup] CUDA packages installed.")
+    _RUNTIME_PKG_MARKER.touch()
+    print("[startup] Runtime packages installed.")
 
-_install_cuda_packages()
+_install_runtime_packages()
 # ──────────────────────────────────────────────────────────────────────────────
 
 import cv2

requirements.txt

@@ -1,16 +1,16 @@
 # HuggingFace ZeroGPU Space — Gradio SDK
-# chumpy vendored locally with patched setup.py (original does `import pip` which breaks isolated builds)
 spaces
 
 # Git-pinned installs
-hmr2 @ git+https://github.com/shubham-goel/4D-Humans.git@efe18deff163b29dff87ddbd575fa29b716a356c
+# NOTE: hmr2, chumpy, skel, nvdiffrast, diso, detectron2 are installed at runtime
+# in app.py because:
+#   - hmr2/skel declare `chumpy @ git+https://...` as a direct-ref dep; chumpy's
+#     setup.py does `from pip._internal.req import parse_requirements` which fails
+#     in pip>=21 isolated builds. Fix: --no-build-isolation at runtime.
+#   - nvdiffrast/diso/detectron2 require nvcc (NVIDIA CUDA compiler) which is only
+#     available in the ZeroGPU runtime container, not the Docker build stage.
 clip @ git+https://github.com/openai/CLIP.git@d05afc436d78f1c48dc0dbf8e5980a9d471f35f6
 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, 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