Drop the pip-installed cmake from the pytorch base image

base_images/CHANGELOG.md

@@ -1,3 +1,17 @@
+# OpenProblems Base Images v1.2.0
+
+## BUG FIXES
+
+* `base_pytorch_nvidia`: drop the pip-installed `cmake` in favour of apt's (PR #48). The pip
+  shims in `/usr/local/bin` shadow `/usr/bin/cmake`, and they fail with `ModuleNotFoundError:
+  No module named 'cmake'` when a build calls them from inside a pip build isolation
+  environment -- which broke, among others, building `louvain` for `cellplm`.
+
+## TESTING
+
+* Check that `cmake`, `cpack` and `ctest` resolve to the apt-provided binaries in
+  `base_pytorch_nvidia` (PR #48).
+
 # OpenProblems Base Images v1.1.0
 
 ## MAJOR CHANGES

base_images/_viash.yaml

@@ -1,5 +1,5 @@
 organization: openproblems
-version: 1.1.0
+version: 1.2.0
 viash_version: 0.9.7
 links:
   repository: https://github.com/openproblems-bio/base_images

base_images/src/pytorch_nvidia/config.vsh.yaml

@@ -3,6 +3,8 @@ description: An nvcr.io pytorch with anndata preinstalled.
 test_resources:
   - type: python_script
     path: test_pytorch.py
+  - type: python_script
+    path: test_cmake.py
   - type: python_script
     path: ../python/test_anndata.py
 engines:
@@ -13,6 +15,12 @@ engines:
         packages:
           - procps
           - git
+          - cmake
+      # The nvcr image pip-installs cmake, whose shims in /usr/local/bin shadow
+      # apt's cmake because /usr/local/bin comes first on PATH. Drop the package
+      # so builds pick up the working /usr/bin/cmake instead.
+      - type: docker
+        run: pip uninstall -y cmake
       - type: python
         packages:
           - anndata~=0.12.0

base_images/src/pytorch_nvidia/test_cmake.py

@@ -0,0 +1,40 @@
+import shutil
+import subprocess
+import tempfile
+
+## VIASH START
+meta = {
+    "temp_dir": tempfile.gettempdir(),
+}
+## VIASH END
+
+# The nvcr pytorch image ships a pip-installed cmake whose shims land in
+# /usr/local/bin and shadow apt's cmake. If a base image rebuild ever
+# reintroduces them, components that compile from source break in confusing
+# ways, so check that cmake resolves to the apt-provided binary and runs.
+print("--- Checking cmake ---", flush=True)
+
+cmake = shutil.which("cmake")
+print(f"cmake resolves to: {cmake}", flush=True)
+
+assert cmake is not None, "cmake not found on PATH"
+assert cmake == "/usr/bin/cmake", (
+    f"expected apt's /usr/bin/cmake on PATH, got {cmake} -- a pip-installed "
+    "cmake is probably shadowing it"
+)
+
+out = subprocess.run([cmake, "--version"], capture_output=True, text=True)
+print(out.stdout.strip(), flush=True)
+
+assert out.returncode == 0, f"`cmake --version` exited {out.returncode}: {out.stderr.strip()}"
+
+# The pip package ships cpack and ctest alongside cmake; removing only the
+# cmake shim would leave those two broken.
+for tool in ["cpack", "ctest"]:
+    path = shutil.which(tool)
+    print(f"{tool} resolves to: {path}", flush=True)
+    assert path is None or path.startswith("/usr/bin/"), (
+        f"{tool} resolves to {path}, expected /usr/bin/{tool} or nothing"
+    )
+
+print("\ncmake test passed!", flush=True)