diff --git a/.github/workflows/build-gfx11-rocm.yml b/.github/workflows/build-gfx11-rocm.yml
index 6a811fffbd77..ce91fa77fd87 100644
--- a/.github/workflows/build-gfx11-rocm.yml
+++ b/.github/workflows/build-gfx11-rocm.yml
@@ -15,6 +15,10 @@ on:
         description: 'Publish a dated GitHub Release (b<YYYYMMDD>) with the built binaries. Set true only for the nightly dispatch.'
         required: false
         default: 'false'
+      roofline:
+        description: 'Build with per-op roofline profiling support (GGML_HIP_ROOFLINE) and bundle the profiler libs (librocprofiler-sdk + libdw). Set false for lean builds.'
+        required: false
+        default: 'true'
 
 # Tracks the latest multiarch ROCm nightly (auto-detected). The earlier pin to
 # 7.14.0a20260608 worked around a libhsa-runtime64 regression (build 1b2a555677)
@@ -28,6 +32,9 @@ on:
 
 env:
   ROCM_VERSION: ${{ github.event.inputs.rocm_version || 'latest' }}
+  # Default ON so push/PR/cron (non-dispatch) runs keep roofline, matching prior
+  # always-on behavior; a workflow_dispatch can pass 'false' for a lean build.
+  ROOFLINE: ${{ github.event.inputs.roofline || 'true' }}
 
 jobs:
   build-ubuntu:
@@ -225,7 +232,7 @@ jobs:
           -DGGML_CUDA_FORCE_CUBLAS=OFF \
           -DGGML_RPC=ON \
           -DGGML_HIP_ROCWMMA_FATTN=OFF \
-          -DGGML_HIP_ROOFLINE=ON \
+          -DGGML_HIP_ROOFLINE=${{ env.ROOFLINE == 'true' && 'ON' || 'OFF' }} \
           -DLLAMA_BUILD_BORINGSSL=ON \
           -DGGML_NATIVE=OFF \
           -DGGML_STATIC=OFF \
@@ -263,8 +270,6 @@ jobs:
 
         # Copy required ROCm libraries to build directory
         # If artifacts from ROCm or Llama.cpp change, you may need to update this list.
-        # To regenerate the list, run:
-        #   utils/gather_required_libs.py --rocm-dir /opt/rocm --dest-dir build/bin
         echo "Copying required ROCm libraries to build directory..."
         cp -v /opt/rocm/lib/libhipblas.so* "$build_bin_path/" 2>/dev/null || echo "libhipblas.so* not found"
         cp -v /opt/rocm/lib/librocblas.so* "$build_bin_path/" 2>/dev/null || echo "librocblas.so* not found"
@@ -274,6 +279,14 @@ jobs:
         cp -v /opt/rocm/lib/libhipblaslt.so* "$build_bin_path/" 2>/dev/null || echo "libhipblaslt.so* not found"
         cp -v /opt/rocm/lib/rocm_sysdeps/lib/librocm_sysdeps_liblzma.so* "$build_bin_path/" 2>/dev/null || echo "librocm_sysdeps_liblzma.so* not found"
         cp -v /opt/rocm/lib/librocprofiler-register.so* "$build_bin_path/" 2>/dev/null || echo "librocprofiler-register.so* not found"
+        # Roofline builds dlopen librocprofiler-sdk.so.1 at runtime (ggml-cuda-roofline.cpp);
+        # the release must ship it plus its one bundle-missing dep libdw
+        # (librocm_sysdeps_dw.so.1 — a hard DT_NEEDED of the sdk lib). The $ORIGIN RPATH
+        # step below lets the dlopen resolve both locally with no LD_LIBRARY_PATH.
+        if [ "${{ env.ROOFLINE }}" = "true" ]; then
+          cp -v /opt/rocm/lib/librocprofiler-sdk.so* "$build_bin_path/" 2>/dev/null || echo "librocprofiler-sdk.so* not found"
+          cp -v /opt/rocm/lib/rocm_sysdeps/lib/librocm_sysdeps_dw.so* "$build_bin_path/" 2>/dev/null || echo "librocm_sysdeps_dw.so* not found"
+        fi
         cp -v /opt/rocm/lib/libamd_comgr.so* "$build_bin_path/" 2>/dev/null || echo "libamd_comgr.so* not found"
         cp -v /opt/rocm/lib/libamd_comgr_loader.so* "$build_bin_path/" 2>/dev/null || echo "libamd_comgr_loader.so* not found"
         cp -v /opt/rocm/lib/libhsa-runtime64.so* "$build_bin_path/" 2>/dev/null || echo "libhsa-runtime64.so* not found"