feat(cugraph): upgrade to RAPIDS 25.12 / CUDA 13.1 with comprehensive e2e tests

.gitignore

@@ -208,4 +208,4 @@ python/mage/link_prediction/random_customer_results.txt
 python/mage/link_prediction/cora_results.txt
 python/mage/link_prediction/random_features_services_results.txt
 python/mage/link_prediction/issue.py
-python/mage/link_prediction/issue2.py
+python/mage/link_prediction/issue2.py.build-staging/

Dockerfile.cugraph

@@ -1,133 +1,109 @@
-ARG CUGRAPH_VERSION=22.02
-ARG CUDA_VERSION=11.5
-ARG CUDA_VERSION_MINOR=11.5.2
-ARG PY_VERSION=3.8
+ARG RAPIDS_VERSION=25.12
+ARG CUDA_VERSION=13
+ARG CUDA_VERSION_MINOR=13.1.0
+ARG PY_VERSION=3.12
 ARG MG_VERSION=3.7.2
 
-FROM rapidsai/rapidsai:${CUGRAPH_VERSION}-cuda${CUDA_VERSION}-runtime-ubuntu20.04-py${PY_VERSION} as cugraph-dev
+FROM nvcr.io/nvidia/rapidsai/base:${RAPIDS_VERSION}-cuda${CUDA_VERSION}-py${PY_VERSION} AS cugraph-dev
 
-FROM nvidia/cuda:${CUDA_VERSION_MINOR}-devel-ubuntu20.04 AS dev
+FROM nvidia/cuda:${CUDA_VERSION_MINOR}-devel-ubuntu24.04 AS dev
 
 USER root
 
 ARG DEBIAN_FRONTEND=noninteractive
 ARG MG_VERSION
 ARG PY_VERSION
-ENV MG_VERSION ${MG_VERSION}
-ENV PY_VERSION ${PY_VERSION}
+ARG CACHE_PRESENT
+ENV MG_VERSION=${MG_VERSION}
+ENV PY_VERSION=${PY_VERSION}
 
-# Copy RAPIDS libraries
-COPY --from=cugraph-dev /opt/conda/envs/rapids/lib/libcugraph.so /opt/conda/envs/rapids/lib/libcugraph.so
-COPY --from=cugraph-dev /opt/conda/envs/rapids/include /opt/conda/envs/rapids/include
+COPY --from=cugraph-dev /opt/conda/lib/libcugraph.so /opt/conda/lib/libcugraph.so
+COPY --from=cugraph-dev /opt/conda/lib/libcugraph_c.so /opt/conda/lib/libcugraph_c.so
+COPY --from=cugraph-dev /opt/conda/lib/librmm.so /opt/conda/lib/librmm.so
+COPY --from=cugraph-dev /opt/conda/lib/librapids_logger.so /opt/conda/lib/librapids_logger.so
+COPY --from=cugraph-dev /opt/conda/include /opt/conda/include
 
-# Prevent from linking the Conda environment
-ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/cuda/lib64:/usr/local/lib
+ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/cuda/lib64:/usr/local/lib:/opt/conda/lib
 
-# NVIDIA key rotation
-RUN rm /etc/apt/sources.list.d/cuda.list
-
-# Essentials for production/dev
 RUN apt-get update && apt-get install -y \
-    libcurl4        `memgraph` \
-    libpython${PY_VERSION}   `memgraph` \
-    libssl-dev       `memgraph` \
-    libssl-dev       `memgraph` \
-    openssl         `memgraph` \
-    build-essential `mage-memgraph` \
-    curl            `mage-memgraph` \
-    g++             `mage-memgraph` \
-    python3         `mage-memgraph` \
-    python3-pip     `mage-memgraph` \
-    python3-setuptools     `mage-memgraph` \
-    python3-dev     `mage-memgraph` \
-    clang           `mage-memgraph` \
-    git             `mage-memgraph` \
-    software-properties-common  `mage-cugraph` \
-    lsb-release `mage-cugraph` \
-    wget `mage-cugraph` \
-    uuid-dev \
-    gdb \
-    procps \
-    linux-perf \
-    libc6-dbg \
-    --no-install-recommends && \
-    rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/* && \
-    # Install newest CMake (cuGraph requires >= 20.01)
-    wget -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc | apt-key add - && \
-    apt-add-repository "deb https://apt.kitware.com/ubuntu/ $(lsb_release -cs) main" && \
-    apt-get install -y \
-    cmake           `mage-memgraph` \
-    --no-install-recommends
-
-ENV PATH=/usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/bin/cmake:/usr/lib/cmake
-
-# Memgraph listens for Bolt Protocol on this port by default.
+    libcurl4t64 libpython${PY_VERSION} libssl-dev openssl build-essential curl g++ \
+    python3 python3-pip python3-setuptools python3-dev clang git \
+    software-properties-common lsb-release wget uuid-dev gdb procps \
+    linux-tools-generic ninja-build libc6-dbg cmake libboost-all-dev \
+    unixodbc-dev --no-install-recommends && \
+    rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+
+ENV PATH=/usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+
 EXPOSE 7687
 
-# Copy and build MAGE
+RUN curl https://download.memgraph.com/memgraph/v${MG_VERSION}/ubuntu-24.04/memgraph_${MG_VERSION}-1_amd64.deb --output memgraph.deb \
+    && dpkg -i memgraph.deb && rm memgraph.deb
+
 WORKDIR /mage
+
 COPY . /mage
 
+ENV CXXFLAGS="-DLIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE"
+ENV CUDAFLAGS="-DLIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE"
+
+# Install PyTorch with CUDA 13.0 support FIRST
+# Then install all dependencies that require matching PyTorch/CUDA versions
 RUN curl https://sh.rustup.rs -sSf | sh -s -- -y && \
-    export PATH="/root/.cargo/bin:${PATH}" && \ 
-    python3 -m pip install -r /mage/python/requirements.txt && \ 
-    python3 -m  pip install -r /mage/python/tests/requirements.txt && \
-    python3 -m pip install dgl -f https://data.dgl.ai/wheels/repo.html && \ 
-    python3 /mage/setup build \ 
-    --gpu \
-    --cpp-build-flags MAGE_CUGRAPH_ROOT=/opt/conda/envs/rapids/ CMAKE_BUILD_TYPE=Release \
-    -p /usr/lib/memgraph/query_modules/ 
-
-#DGL build from source
-RUN git clone --recurse-submodules -b 0.9.x https://github.com/dmlc/dgl.git   \
-    && cd dgl && mkdir build && cd build && cmake -DUSE_CUDA=ON .. \
-    && make -j4 && cd ../python && python3 setup.py install
+    export PATH="/root/.cargo/bin:${PATH}" && \
+    python3 -m pip install --break-system-packages torch torchvision --index-url https://download.pytorch.org/whl/cu130 && \
+    python3 -m pip install --break-system-packages --ignore-installed -r /mage/python/requirements.txt && \
+    python3 -m pip install --break-system-packages --ignore-installed -r /mage/python/tests/requirements.txt && \
+    (if [ "$CACHE_PRESENT" = "true" ]; then \
+        echo "Using cached packages" && \
+        pip install /mage/wheels/*.whl --break-system-packages; \
+    else \
+        python3 -m pip install --break-system-packages dgl -f https://data.dgl.ai/wheels/torch-2.9/cu130/repo.html && \
+        python3 -m pip install --break-system-packages torch_geometric && \
+        python3 -m pip install --break-system-packages ninja wheel && \
+        python3 -m pip install --break-system-packages --no-build-isolation git+https://github.com/pyg-team/pyg-lib.git && \
+        python3 -m pip install --break-system-packages --no-build-isolation git+https://github.com/rusty1s/pytorch_scatter.git && \
+        python3 -m pip install --break-system-packages --no-build-isolation git+https://github.com/rusty1s/pytorch_sparse.git && \
+        python3 -m pip install --break-system-packages --no-build-isolation git+https://github.com/rusty1s/pytorch_cluster.git && \
+        python3 -m pip install --break-system-packages --no-build-isolation git+https://github.com/rusty1s/pytorch_spline_conv.git && \
+        python3 -m pip install --break-system-packages --upgrade numpy gensim; \
+    fi) && \
+    python3 /mage/setup build --gpu \
+    --cpp-build-flags MAGE_CUGRAPH_ROOT=/opt/conda/ CMAKE_BUILD_TYPE=Release \
+    -p /usr/lib/memgraph/query_modules/
 
 USER memgraph
 ENTRYPOINT ["/usr/lib/memgraph/memgraph"]
 
-FROM nvidia/cuda:${CUDA_VERSION_MINOR}-runtime-ubuntu20.04 AS prod
+FROM nvidia/cuda:${CUDA_VERSION_MINOR}-runtime-ubuntu24.04 AS prod
 
 USER root
 
 ARG DEBIAN_FRONTEND=noninteractive
 ARG MG_VERSION
 ARG PY_VERSION
-ENV MG_VERSION ${MG_VERSION}
-ENV PY_VERSION ${PY_VERSION}
+ENV MG_VERSION=${MG_VERSION}
+ENV PY_VERSION=${PY_VERSION}
 
-# Copy modules
 COPY --from=dev /usr/lib/memgraph/query_modules/ /usr/lib/memgraph/query_modules/
-# Copy cugraph library
-COPY --from=dev /opt/conda/envs/rapids/lib/libcugraph.so /opt/conda/envs/rapids/lib/libcugraph.so
-# Copy python build
+COPY --from=dev /opt/conda/lib/libcugraph.so /opt/conda/lib/libcugraph.so
+COPY --from=dev /opt/conda/lib/libcugraph_c.so /opt/conda/lib/libcugraph_c.so
+COPY --from=dev /opt/conda/lib/librmm.so /opt/conda/lib/librmm.so
+COPY --from=dev /opt/conda/lib/librapids_logger.so /opt/conda/lib/librapids_logger.so
 COPY --from=dev /usr/local/lib/python${PY_VERSION}/ /usr/local/lib/python${PY_VERSION}/
 
-# NVIDIA key rotation
-RUN rm /etc/apt/sources.list.d/cuda.list
+ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/cuda/lib64:/usr/local/lib:/opt/conda/lib
 
-# Download and install Memgraph
 RUN apt-get update && apt-get install -y \
-    libcurl4        `memgraph` \
-    libpython${PY_VERSION}   `memgraph` \
-    libssl1.1       `memgraph` \
-    libssl-dev       `memgraph` \
-    openssl         `memgraph` \
-    curl            `mage-memgraph` \
-    libgomp1         `mage-memgraph` \
-    python3         `mage-memgraph` \
-    python3-setuptools     `mage-memgraph` \
-    && curl https://download.memgraph.com/memgraph/v${MG_VERSION}/ubuntu-20.04/memgraph_${MG_VERSION}-1_amd64.deb --output memgraph.deb \
-    && dpkg -i memgraph.deb \
-    && rm memgraph.deb \
+    libcurl4t64 libpython${PY_VERSION} libssl3t64 openssl curl libgomp1 libatomic1 python3 python3-setuptools \
+    unixodbc --no-install-recommends \
+    && curl https://download.memgraph.com/memgraph/v${MG_VERSION}/ubuntu-24.04/memgraph_${MG_VERSION}-1_amd64.deb --output memgraph.deb \
+    && dpkg -i memgraph.deb && rm memgraph.deb \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
 
-RUN export PATH="/usr/local/lib/python${PY_VERSION}:${PATH}"
+ENV PATH="/usr/local/lib/python${PY_VERSION}:${PATH}"
 
-RUN rm -rf /mage \
-    && export PATH="/usr/local/lib/python${PY_VERSION}:${PATH}" \
-    && apt-get -y --purge autoremove curl python3-dev \
-    && apt-get clean
+RUN rm -rf /mage && apt-get -y --purge autoremove curl python3-dev && apt-get clean
 
 USER memgraph
 ENTRYPOINT ["/usr/lib/memgraph/memgraph"]

cpp/cugraph_module/algorithms/balanced_cut_clustering.cu

@@ -12,10 +12,16 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+// NOTE: balancedCutClustering only exists in the legacy cugraph::ext_raft namespace
+// and requires legacy::GraphCSRView. There is no modern API equivalent.
+
+#include <cugraph/legacy/graph.hpp>
+#include <cugraph/algorithms.hpp>
+
 #include "mg_cugraph_utility.hpp"
 
 namespace {
-// TODO: Check Balanced Cut API. Update in new cuGraph API.
+// NOTE: Spectral clustering legacy API only supports int32_t vertex/edge types
 using vertex_t = int32_t;
 using edge_t = int32_t;
 using weight_t = double;
@@ -55,7 +61,6 @@ void InsertBalancedCutResult(mgp_graph *graph, mgp_result *result, mgp_memory *m
 
 void BalancedCutClusteringProc(mgp_list *args, mgp_graph *graph, mgp_result *result, mgp_memory *memory) {
   try {
-    // TODO: Not supporting int64_t
     int num_clusters = mgp::value_get_int(mgp::list_at(args, 0));
     int num_eigenvectors = mgp::value_get_int(mgp::list_at(args, 1));
     double ev_tolerance = mgp::value_get_double(mgp::list_at(args, 2));
@@ -74,20 +79,30 @@ void BalancedCutClusteringProc(mgp_list *args, mgp_graph *graph, mgp_result *res
     raft::handle_t handle{};
     auto stream = handle.get_stream();
 
-    // IMPORTANT: Balanced cut cuGraph algorithm works only on legacy code
+    // IMPORTANT: Balanced cut cuGraph algorithm works only on legacy CSR graph format
     auto cu_graph_ptr =
         mg_cugraph::CreateCugraphLegacyFromMemgraph<vertex_t, edge_t, weight_t>(*mg_graph.get(), handle);
     auto cu_graph_view = cu_graph_ptr->view();
     cu_graph_view.prop.directed = false;
 
     rmm::device_uvector<vertex_t> clustering_result(n_vertices, stream);
-    // Only supported for weighted graphs
-    cugraph::ext_raft::balancedCutClustering(cu_graph_view, num_clusters, num_eigenvectors, ev_tolerance, ev_maxiter,
-                                             kmean_tolerance, kmean_maxiter, clustering_result.data());
 
-    for (vertex_t node_id = 0; node_id < clustering_result.size(); ++node_id) {
-      auto cluster = clustering_result.element(node_id, stream);
-      InsertBalancedCutResult(graph, result, memory, mg_graph->GetMemgraphNodeId(node_id), cluster);
+    // Create RNG state for cuGraph 25.x API
+    raft::random::RngState rng_state(42);
+
+    // Call balancedCutClustering API - cuGraph 25.x requires handle and rng_state
+    cugraph::ext_raft::balancedCutClustering(handle, rng_state, cu_graph_view, num_clusters, num_eigenvectors,
+                                   static_cast<weight_t>(ev_tolerance), ev_maxiter,
+                                   static_cast<weight_t>(kmean_tolerance), kmean_maxiter,
+                                   clustering_result.data());
+
+    // Copy results to host and output
+    std::vector<vertex_t> h_clustering(n_vertices);
+    raft::update_host(h_clustering.data(), clustering_result.data(), n_vertices, stream);
+    handle.sync_stream();
+
+    for (vertex_t node_id = 0; node_id < static_cast<vertex_t>(n_vertices); ++node_id) {
+      InsertBalancedCutResult(graph, result, memory, mg_graph->GetMemgraphNodeId(node_id), h_clustering[node_id]);
     }
   } catch (const std::exception &e) {
     // We must not let any exceptions out of our module.