Fix(critical): Add rsmi_init shim to satisfy PyTorch linker dependencies when SMI is disabled

projects/rccl/CMakeLists.txt

@@ -793,6 +793,12 @@ if(ENABLE_AMDSMI)
   else()
      list(APPEND SRC_FILES src/misc/rocm_smi_wrap.cc)
   endif()
+else()
+  # When SMI is disabled, compile the shim to provide dummy symbols (rsmi_init)
+  # This satisfies external dependencies (like PyTorch) that expect SMI symbols
+  # to be present, preventing them from failing to load or trying to load
+  # the broken system library.
+  list(APPEND SRC_FILES src/misc/smi_shim.cc)
 endif()
 list(APPEND SRC_FILES ${SMI_SOURCES})
 

projects/rccl/src/misc/smi_shim.cc

@@ -0,0 +1,10 @@
+#include <cstdint>
+
+extern "C" {
+    // Dummy implementation of rsmi_init to satisfy linker dependencies
+    // when the real ROCm SMI library is broken or causes Bus Errors (gfx1151).
+    // Returns 0 (RSMI_STATUS_SUCCESS).
+    int rsmi_init(uint64_t flags) {
+        return 0;
+    }
+}