Make pip nvcc detection robust (glob, not hardcoded path)

Search site-packages for nvidia/cuda_nvcc*/bin/nvcc instead of
assuming the exact path, and only add include/lib dirs that exist.
Also note in the fallback warning that a stale build cache (lingering
cu130 torch + missing nvcc package) needs a Factory Reboot.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -104,21 +104,33 @@ except ImportError:
     # requires, and torch's build check errors on a major CUDA mismatch (nvcc 13 vs
     # torch 12.8). Point CUDA_HOME at the pip nvcc and feed the runtime/cub headers
     # and cudart lib via CPATH/LIBRARY_PATH so nvcc 12.8 matches torch 12.8.
-    import site
-    _site = site.getsitepackages()[0]
-    _nv = os.path.join(_site, "nvidia")
-    _nvcc_home = os.path.join(_nv, "cuda_nvcc")
-    if os.path.isdir(_nvcc_home):
+    import site, glob
+    _search_dirs = list(site.getsitepackages())
+    try:
+        _search_dirs.append(site.getusersitepackages())
+    except Exception:
+        pass
+    _nvcc_bins = []
+    for _d in _search_dirs:
+        _nvcc_bins += glob.glob(os.path.join(_d, "nvidia", "cuda_nvcc*", "bin", "nvcc"))
+    if _nvcc_bins:
+        _nvcc_home = os.path.dirname(os.path.dirname(_nvcc_bins[0]))  # .../nvidia/cuda_nvcc
+        _nv = os.path.dirname(_nvcc_home)                              # .../nvidia
         env["CUDA_HOME"] = _nvcc_home
         env["PATH"] = os.path.join(_nvcc_home, "bin") + os.pathsep + env.get("PATH", "")
-        _incs = [os.path.join(_nv, "cuda_runtime", "include"),
-                 os.path.join(_nv, "cuda_cccl", "include")]
-        env["CPATH"] = os.pathsep.join(_incs + ([env["CPATH"]] if env.get("CPATH") else []))
+        _incs = [p for p in (os.path.join(_nv, "cuda_runtime", "include"),
+                             os.path.join(_nv, "cuda_cccl", "include")) if os.path.isdir(p)]
+        if _incs:
+            env["CPATH"] = os.pathsep.join(_incs + ([env["CPATH"]] if env.get("CPATH") else []))
         _rt_lib = os.path.join(_nv, "cuda_runtime", "lib")
-        env["LIBRARY_PATH"] = _rt_lib + os.pathsep + env.get("LIBRARY_PATH", "")
-        env["LD_LIBRARY_PATH"] = _rt_lib + os.pathsep + env.get("LD_LIBRARY_PATH", "")
+        if os.path.isdir(_rt_lib):
+            env["LIBRARY_PATH"] = _rt_lib + os.pathsep + env.get("LIBRARY_PATH", "")
+            env["LD_LIBRARY_PATH"] = _rt_lib + os.pathsep + env.get("LD_LIBRARY_PATH", "")
+        print(f"[BUILD] using pip CUDA toolkit: CUDA_HOME={_nvcc_home}")
     else:
-        print(f"⚠️  pip CUDA toolkit not found at {_nvcc_home}; falling back to system nvcc")
+        print("⚠️  pip CUDA nvcc (nvidia-cuda-nvcc-cu12) not found under site-packages/nvidia; "
+              "falling back to system nvcc (will mismatch cu128 torch). "
+              "If this persists, Factory Reboot the Space to clear the stale build cache.")
 
     # Build a cross-arch set incl. Blackwell sm_120 (the ZeroGPU GPU is cap 12.0);
     # the build stage may not see a GPU, so don't rely on get_device_capability.