Fix CUDA compile: set TORCH_CUDA_ARCH_LIST=8.6 and CUDA headers path for A10G

app.py

@@ -63,9 +63,21 @@ 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.
+    # CUDA headers live at a non-standard path on ZeroGPU runtime containers.
+    _cuda_home = "/cuda-image/usr/local/cuda-12.9"
+    _cuda_env = {
+        **os.environ,
+        "TORCH_CUDA_ARCH_LIST": "8.6",
+        "CUDA_HOME": _cuda_home,
+        "CPATH": f"{_cuda_home}/include:{os.environ.get('CPATH', '')}",
+        "C_INCLUDE_PATH": f"{_cuda_home}/include:{os.environ.get('C_INCLUDE_PATH', '')}",
+        "CPLUS_INCLUDE_PATH": f"{_cuda_home}/include:{os.environ.get('CPLUS_INCLUDE_PATH', '')}",
+    }
     subprocess.run(
         [sys.executable, "-m", "pip", "install", "--quiet", "--no-build-isolation"]
-        + _CUDA_PACKAGES, check=True,
+        + _CUDA_PACKAGES, env=_cuda_env, check=True,
     )
     _RUNTIME_PKG_MARKER.touch()
     print("[startup] Runtime packages installed.")