Fix DEB10 wheel build and verification

- Link amdsminic static library into libamd_smi_python.so to resolve undefined NIC symbols
- Prevent pip-context wrapper from falling back to system libamd_smi.so to avoid dual-loading segfault
- Use RTLD_LOCAL instead of RTLD_GLOBAL when loading .so via ctypes
- Make wheel verification fatal in CI
- Add trailing newline to setup.cfg.in

Author: marifamd

.github/workflows/amdsmi-build.yml

@@ -183,6 +183,11 @@ jobs:
 
               # Verify wheel installation
               echo 'Verifying wheel installation...'
+              # Diagnostics: list installed .so files without importing (import triggers ctypes.CDLL)
+              SITE_DIR=$($PYTHON_CMD -c "import site; print(site.getsitepackages()[0])" 2>/dev/null || echo "/usr/lib/python3/dist-packages")
+              echo "Wheel installed to: $SITE_DIR/amdsmi/"
+              ls -lh "$SITE_DIR/amdsmi/"*.so 2>/dev/null || echo "  No .so files found in wheel install"
+
               cd /tmp && $PYTHON_CMD -c "import amdsmi; print('✓ Import successful'); amdsmi.amdsmi_init(); print('✓ Library loaded'); amdsmi.amdsmi_shut_down(); print('✓ Wheel working!')"
               echo 'Python wheel build and install completed on ${{ matrix.os }}'
               break

projects/amdsmi/py-interface/amdsmi_wrapper.py

@@ -243,11 +243,15 @@ def _build_candidate_paths():
     candidates = []
 
     if context == "pip":
-        # .so lives alongside the wrapper inside the wheel / site-packages
+        # .so is self-contained inside the wheel / site-packages.
+        # Do NOT fall back to the system libamd_smi.so: loading both libraries
+        # in the same process causes segfaults during static initialisation of
+        # std::variant tables on older toolchains (GCC 8 / glibc 2.28).
         candidates.append(base / "libamd_smi_python.so")
-    else:
-        # System package - .so lives under <rocm_root>/lib/
-        candidates.append(base / "lib" / "libamd_smi.so")
+        return candidates
+
+    # System package - .so lives under <rocm_root>/lib/
+    candidates.append(base / "lib" / "libamd_smi.so")
 
     # Fallbacks
     for env_var in ("ROCM_HOME", "ROCM_PATH"):
@@ -276,7 +280,15 @@ def _load_library():
     """
     candidates = _build_candidate_paths()
     last_err = None
-    mode = getattr(ctypes, "RTLD_GLOBAL", 0)
+
+    # Use RTLD_LOCAL (default, mode=0) to keep each library's symbols in its
+    # own scope.  RTLD_GLOBAL is unnecessary because ctypes resolves symbols
+    # via the returned handle (dlsym), and using it leaks template symbols
+    # into the global namespace.  When the pip-context libamd_smi_python.so
+    # is tried first and fails, those leaked symbols collide with the
+    # fallback libamd_smi.so's own copies, causing a segfault during the
+    # latter's static initialisation on older platforms (Debian 10 / glibc 2.28).
+    mode = 0
 
     for candidate in candidates:
         try:

projects/amdsmi/py-interface/setup.cfg.in

@@ -18,4 +18,4 @@ include_package_data = True
 zip_safe = False
 
 [options.package_data]
-* = *.so
+* = *.so

projects/amdsmi/src/CMakeLists.txt

@@ -156,8 +156,12 @@ if(BUILD_PYTHON_LIB)
         rt
         Threads::Threads
         ${CMAKE_DL_LIBS}
+        amdsminic
         ${FILESYSTEM_LIB}
     )
+    target_link_directories(${AMD_SMI}_python PRIVATE
+        ${CMAKE_CURRENT_BINARY_DIR}/nic/ai-nic/amdsmi_unified/build/
+    )
     target_include_directories(${AMD_SMI}_python PRIVATE 
         ${CMAKE_CURRENT_SOURCE_DIR} 
         ${PROJECT_SOURCE_DIR}/rocm_smi/include