further fix for rocWMMA

toolboxes/Dockerfile.rocm-6.4.4-rocwmma

@@ -1,11 +1,11 @@
 # build stage
 FROM registry.fedoraproject.org/fedora:rawhide AS builder
 
-# rocm 6.4.6 repo
+# rocm 6.4.4 repo
 RUN <<'EOF'
 tee /etc/yum.repos.d/rocm.repo <<REPO
-[ROCm-6.4.6]
-name=ROCm6.4.6
+[ROCm-6.4.4]
+name=ROCm6.4.4
 baseurl=https://repo.radeon.com/rocm/el9/6.4.4/main
 enabled=1
 priority=50
@@ -18,9 +18,9 @@ EOF
 RUN dnf -y --nodocs --setopt=install_weak_deps=False \
     --exclude='*sdk*' --exclude='*samples*' --exclude='*-doc*' --exclude='*-docs*' \
     install \
-      make gcc cmake lld clang clang-devel compiler-rt libcurl-devel \
+      make gcc cmake lld clang clang-devel compiler-rt libcurl-devel ninja-build \
       rocm-llvm rocm-device-libs hip-runtime-amd hip-devel \
-      rocblas rocblas-devel hipblas hipblas-devel \
+      rocblas rocblas-devel hipblas hipblas-devel rocm-cmake libomp-devel libomp \
       rocminfo radeontop \
       git-core vim sudo rsync \
  && dnf clean all && rm -rf /var/cache/dnf/*
@@ -32,20 +32,26 @@ ENV ROCM_PATH=/opt/rocm \
     HIP_DEVICE_LIB_PATH=/opt/rocm/amdgcn/bitcode \
     PATH=/opt/rocm/bin:/opt/rocm/llvm/bin:$PATH
 
-# rocWMMA
+# rocWMMA 
 WORKDIR /opt
-RUN git clone -b release/rocm-rel-7.1 https://github.com/ROCm/rocWMMA.git
-RUN sudo mkdir -p /usr/include/rocwmma
-RUN sudo cp -r rocWMMA/library/include/rocwmma /usr/include/
+COPY ./build-rocwmma.sh .
+RUN chmod +x build-rocwmma.sh && ./build-rocwmma.sh
 
 # llama.cpp
 WORKDIR /opt/llama.cpp
 RUN git clone --recursive https://github.com/ggerganov/llama.cpp.git .
 
+# Apply # rocWMMA patch
+COPY ./apply-rocwmma-fix.sh /opt/apply-rocwmma-fix.sh
+RUN chmod +x /opt/apply-rocwmma-fix.sh && /opt/apply-rocwmma-fix.sh /opt/llama.cpp
+
 # build
 RUN git clean -xdf \
  && git pull \
  && git submodule update --recursive \
+ && echo -e '#ifndef HIP_HAS_SHFL_SYNC_FUNCS\n#define HIP_HAS_SHFL_SYNC_FUNCS\n#ifndef __shfl_sync\n#define __shfl_sync(mask,var,srcLane,width) __shfl(var,srcLane,width)\n#endif\n#ifndef __shfl_xor_sync\n#define __shfl_xor_sync(mask,var,laneMask,width) __shfl_xor(var,laneMask,width)\n#endif\n#ifndef __shfl_up_sync\n#define __shfl_up_sync(mask,var,delta,width) __shfl_up(var,delta,width)\n#endif\n#endif\n' \
+    | cat - /opt/llama.cpp/ggml/src/ggml-cuda/mma.cuh > /opt/llama.cpp/ggml/src/ggml-cuda/mma.cuh.tmp \
+ && mv /opt/llama.cpp/ggml/src/ggml-cuda/mma.cuh.tmp /opt/llama.cpp/ggml/src/ggml-cuda/mma.cuh \
  && cmake -S . -B build \
       -DGGML_HIP=ON \
       -DAMDGPU_TARGETS=gfx1151 \

toolboxes/apply-rocwmma-fix.sh

@@ -150,4 +150,6 @@ echo "  cmake -B build -S '$LLAMA_DIR' -DGGML_HIP=ON -DAMDGPU_TARGETS=\"gfx1151\
 echo ""
 echo "Backup files were created with .backup extension in case you need to revert."
 echo ""
-echo "Done! Your llama.cpp checkout now supports rocWMMA builds."
+echo "Done! Your llama.cpp checkout now supports rocWMMA builds."
+
+

toolboxes/build-rocwmma.sh

@@ -2,9 +2,60 @@
 git clone https://github.com/ROCm/rocWMMA
 cd rocWMMA
 
-# Change FP8 check from FAIL to STATUS
+# --- BEGIN: make OpenMP explicit for ROCm toolchains (drop-in) ---
+# find libomp (check ROCM_PATH first, then system)
+CANDIDATES=(
+  "${ROCM_PATH}/llvm/lib/libomp.so"
+  "${ROCM_PATH}/llvm/lib/libomp.a"
+  "/usr/lib64/libomp.so"
+  "/usr/lib64/libomp.a"
+  "/usr/local/lib/libomp.so"
+)
+FOUND_LIBOMP=""
+for p in "${CANDIDATES[@]}"; do
+  if [ -f "$p" ]; then
+    FOUND_LIBOMP="$p"
+    break
+  fi
+done
 
-rm -rf build; mkdir build; CC=$ROCM_PATH/llvm/bin/amdclang CXX=$ROCM_PATH/llvm/bin/amdclang++ cmake -B build  . -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$ROCM_PATH -DROCWMMA_BUILD_TESTS=OFF -DROCWMMA_BUILD_SAMPLES=OFF -DGPU_TARGETS="gfx1151"
+CMAKE_OPTS=""
+if [ -n "$FOUND_LIBOMP" ]; then
+  # directory & basename
+  OMP_LIB_DIR="$(dirname "$FOUND_LIBOMP")"
+  OMP_LIB_BASENAME="$(basename "$FOUND_LIBOMP")"
+  # set cache vars so FindOpenMP will succeed
+  CMAKE_OPTS="${CMAKE_OPTS} -DOpenMP_CXX_FLAGS=-fopenmp=libomp"
+  CMAKE_OPTS="${CMAKE_OPTS} -DOpenMP_C_FLAGS=-fopenmp=libomp"
+  CMAKE_OPTS="${CMAKE_OPTS} -DOpenMP_CXX_LIB_NAMES=omp"
+  CMAKE_OPTS="${CMAKE_OPTS} -DOpenMP_C_LIB_NAMES=omp"
+  CMAKE_OPTS="${CMAKE_OPTS} -DOpenMP_LIBRARY=${FOUND_LIBOMP}"
+  CMAKE_OPTS="${CMAKE_OPTS} -DOpenMP_INCLUDE_DIR=${ROCM_PATH}/llvm/include"
+  export LD_LIBRARY_PATH="${OMP_LIB_DIR}${LD_LIBRARY_PATH:+:}$LD_LIBRARY_PATH"
+  export CXXFLAGS="-fopenmp=libomp ${CXXFLAGS:-}"
+  export LDFLAGS="-L${OMP_LIB_DIR} -lomp ${LDFLAGS:-}"
+else
+  # fallback: force flags so FindOpenMP might at least get flags
+  CMAKE_OPTS="${CMAKE_OPTS} -DOpenMP_CXX_FLAGS=-fopenmp=libomp -DOpenMP_C_FLAGS=-fopenmp=libomp"
+  export CXXFLAGS="-fopenmp=libomp ${CXXFLAGS:-}"
+  export LDFLAGS="${LDFLAGS:-} -lomp"
+fi
+# ---  END: make OpenMP explicit  ---
 
-cmake --build build -j$(nproc)
+CC=$ROCM_PATH/llvm/bin/amdclang \
+CXX=$ROCM_PATH/llvm/bin/amdclang++ \
+cmake -B build -S . -G Ninja \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_INSTALL_PREFIX=$ROCM_PATH \
+  -DROCWMMA_BUILD_TESTS=OFF \
+  -DROCWMMA_BUILD_SAMPLES=OFF \
+  -DGPU_TARGETS="gfx1151" \
+  -DOpenMP_CXX_FLAGS="-fopenmp=libomp" \
+  -DOpenMP_C_FLAGS="-fopenmp=libomp" \
+  -DOpenMP_omp_LIBRARY="/usr/lib64/libomp.so" \
+  -DOpenMP_CXX_LIB_NAMES="omp" \
+  -DOpenMP_C_LIB_NAMES="omp" \
+  -DOpenMP_INCLUDE_DIRS="/usr/lib64/clang/19/include"
+
+cmake --install build
 sudo cmake --install build