diff --git a/mlir/lib/Target/Target.cpp b/mlir/lib/Target/Target.cpp
index 9f4bd666e0e4..290cd13d7009 100644
--- a/mlir/lib/Target/Target.cpp
+++ b/mlir/lib/Target/Target.cpp
@@ -128,6 +128,11 @@ AMDGPUSerializer::loadBitcodeFiles(llvm::Module &module) {
                                  ", error message:" + error.getMessage());
         return std::nullopt;
       }
+      // Packaged libraries can come from a newer LLVM than ours. Give them the
+      // same handling as libraries loaded from disk so they adopt our triple
+      // and data layout instead of warning on every kernel that links them.
+      if (failed(handleBitcodeFile(*library)))
+        return std::nullopt;
       // Unset the lib so we don't add it with `appendStandardLibs`.
       libs = libs & ~lib;
       bcFiles.push_back(std::move(library));
diff --git a/mlir/test/Target/packaged-device-libs.mlir b/mlir/test/Target/packaged-device-libs.mlir
new file mode 100644
index 000000000000..9b41f8276fb0
--- /dev/null
+++ b/mlir/test/Target/packaged-device-libs.mlir
@@ -0,0 +1,17 @@
+// RUN: rocmlir-opt %s --rocdl-attach-target=chip=gfx942 --gpu-module-to-binary 2>&1 \
+// RUN: | FileCheck %s --implicit-check-not="Linking two modules"
+
+// The packaged device libraries can be built by a newer LLVM than ours, with a
+// different triple spelling and data layout. Linking them must not warn.
+
+// CHECK: gpu.binary @kernels
+module attributes {gpu.container_module} {
+  gpu.module @kernels {
+    llvm.func @__ocml_exp_f32(f32) -> f32
+    llvm.func @kernel(%arg0: f32, %arg1: !llvm.ptr<1>) attributes {rocdl.kernel} {
+      %0 = llvm.call @__ocml_exp_f32(%arg0) : (f32) -> f32
+      llvm.store %0, %arg1 : f32, !llvm.ptr<1>
+      llvm.return
+    }
+  }
+}