Merge branch 'branch-24.08' of https://github.com/rapidsai/cugraph in…

…to fea_biased_sampling

ci/build_wheel.sh

@@ -56,7 +56,13 @@ fi
 
 cd "${package_dir}"
 
-python -m pip wheel . -w dist -vvv --no-deps --disable-pip-version-check
+python -m pip wheel \
+    -w dist \
+    -vvv \
+    --no-deps \
+    --disable-pip-version-check \
+    --extra-index-url https://pypi.nvidia.com \
+    .
 
 # pure-python packages should be marked as pure, and not have auditwheel run on them.
 if [[ ${package_name} == "nx-cugraph" ]] || \

ci/build_wheel_cugraph.sh

@@ -1,5 +1,5 @@
 #!/bin/bash
-# Copyright (c) 2023, NVIDIA CORPORATION.
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.
 
 set -euo pipefail
 
@@ -12,6 +12,10 @@ RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
 RAPIDS_PY_WHEEL_NAME=pylibcugraph_${RAPIDS_PY_CUDA_SUFFIX} rapids-download-wheels-from-s3 ./local-pylibcugraph
 export PIP_FIND_LINKS=$(pwd)/local-pylibcugraph
 
+PARALLEL_LEVEL=$(python -c \
+  "from math import ceil; from multiprocessing import cpu_count; print(ceil(cpu_count()/4))")
+
 export SKBUILD_CMAKE_ARGS="-DDETECT_CONDA_ENV=OFF;-DFIND_CUGRAPH_CPP=OFF;-DCPM_cugraph-ops_SOURCE=${GITHUB_WORKSPACE}/cugraph-ops/"
+export SKBUILD_BUILD_TOOL_ARGS="-j${PARALLEL_LEVEL};-l${PARALLEL_LEVEL}"
 
 ./ci/build_wheel.sh cugraph python/cugraph

ci/build_wheel_pylibcugraph.sh

@@ -1,8 +1,12 @@
 #!/bin/bash
-# Copyright (c) 2023, NVIDIA CORPORATION.
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.
 
 set -euo pipefail
 
+PARALLEL_LEVEL=$(python -c \
+  "from math import ceil; from multiprocessing import cpu_count; print(ceil(cpu_count()/4))")
+
 export SKBUILD_CMAKE_ARGS="-DDETECT_CONDA_ENV=OFF;-DFIND_CUGRAPH_CPP=OFF;-DCPM_cugraph-ops_SOURCE=${GITHUB_WORKSPACE}/cugraph-ops/"
+export SKBUILD_BUILD_TOOL_ARGS="-j${PARALLEL_LEVEL};-l${PARALLEL_LEVEL}"
 
 ./ci/build_wheel.sh pylibcugraph python/pylibcugraph

ci/test_python.sh

@@ -44,6 +44,8 @@ rapids-mamba-retry install \
 rapids-logger "Check GPU usage"
 nvidia-smi
 
+export LD_PRELOAD="${CONDA_PREFIX}/lib/libgomp.so.1"
+
 # RAPIDS_DATASET_ROOT_DIR is used by test scripts
 export RAPIDS_DATASET_ROOT_DIR="$(realpath datasets)"
 pushd "${RAPIDS_DATASET_ROOT_DIR}"
@@ -191,6 +193,8 @@ if [[ "${RAPIDS_CUDA_VERSION}" == "11.8.0" ]]; then
     conda activate test_cugraph_pyg
     set -u
 
+    rapids-print-env
+
     # TODO re-enable logic once CUDA 12 is testable
     #if [[ "${RAPIDS_CUDA_VERSION}" == "11.8.0" ]]; then
     CONDA_CUDA_VERSION="11.8"
@@ -204,18 +208,9 @@ if [[ "${RAPIDS_CUDA_VERSION}" == "11.8.0" ]]; then
     rapids-mamba-retry install \
       --channel "${CPP_CHANNEL}" \
       --channel "${PYTHON_CHANNEL}" \
-      --channel pytorch \
       --channel pyg \
-      --channel nvidia \
       "cugraph-pyg" \
-      "pytorch=2.1.0" \
-      "pytorch-cuda=${CONDA_CUDA_VERSION}"
-
-    # Install pyg dependencies (which requires pip)
-
-    pip install \
-      ogb \
-      tensordict
+      "ogb"
 
     pip install \
         pyg_lib \

ci/test_wheel_cugraph-pyg.sh

@@ -42,7 +42,6 @@ rapids-retry python -m pip install \
   pyg_lib \
   torch_scatter \
   torch_sparse \
-  tensordict \
   -f ${PYG_URL}
 
 rapids-logger "pytest cugraph-pyg (single GPU)"

ci/test_wheel_nx-cugraph.sh

@@ -1,6 +1,11 @@
 #!/bin/bash
-# Copyright (c) 2023, NVIDIA CORPORATION.
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.
 
 set -eoxu pipefail
 
+# Download wheels built during this job.
+RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
+RAPIDS_PY_WHEEL_NAME="pylibcugraph_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./local-deps
+python -m pip install ./local-deps/*.whl
+
 ./ci/test_wheel.sh nx-cugraph python/nx-cugraph

conda/environments/all_cuda-118_arch-x86_64.yaml

@@ -42,7 +42,7 @@ dependencies:
 - numpy>=1.23,<2.0a0
 - numpydoc
 - nvcc_linux-64=11.8
-- openmpi
+- openmpi<5.0.3
 - packaging>=21
 - pandas
 - pre-commit
@@ -56,6 +56,7 @@ dependencies:
 - pytest-mpl
 - pytest-xdist
 - python-louvain
+- pytorch>=2.0,<2.2.0a0
 - raft-dask==24.8.*
 - rapids-dask-dependency==24.8.*
 - recommonmark

conda/environments/all_cuda-122_arch-x86_64.yaml

@@ -47,7 +47,7 @@ dependencies:
 - numba>=0.57
 - numpy>=1.23,<2.0a0
 - numpydoc
-- openmpi
+- openmpi<5.0.3
 - packaging>=21
 - pandas
 - pre-commit
@@ -61,6 +61,7 @@ dependencies:
 - pytest-mpl
 - pytest-xdist
 - python-louvain
+- pytorch>=2.0,<2.2.0a0
 - raft-dask==24.8.*
 - rapids-dask-dependency==24.8.*
 - recommonmark

conda/recipes/libcugraph/meta.yaml

@@ -42,7 +42,7 @@ requirements:
     - {{ compiler('cxx') }}
     - cmake {{ cmake_version }}
     - ninja
-    - openmpi # Required for building cpp-mgtests (multi-GPU tests)
+    - openmpi<5.0.3 # Required for building cpp-mgtests (multi-GPU tests)
     - {{ stdlib("c") }}
   host:
     {% if cuda_major == "11" %}

cpp/CMakeLists.txt

@@ -193,9 +193,11 @@ set(CUGRAPH_SOURCES
     src/link_prediction/jaccard_sg.cu
     src/link_prediction/sorensen_sg.cu
     src/link_prediction/overlap_sg.cu
+    src/link_prediction/cosine_sg.cu
     src/link_prediction/jaccard_mg.cu
     src/link_prediction/sorensen_mg.cu
     src/link_prediction/overlap_mg.cu
+    src/link_prediction/cosine_mg.cu
     src/layout/legacy/force_atlas2.cu
     src/converters/legacy/COOtoCSR.cu
     src/community/legacy/spectral_clustering.cu

cpp/include/cugraph/algorithms.hpp

@@ -1973,6 +1973,37 @@ rmm::device_uvector<weight_t> jaccard_coefficients(
   std::tuple<raft::device_span<vertex_t const>, raft::device_span<vertex_t const>> vertex_pairs,
   bool do_expensive_check = false);
 
+/**
+ * @brief     Compute Cosine similarity coefficient
+ *
+ * Similarity is computed for every pair of vertices specified. Note that
+ * similarity algorithms expect a symmetric graph.
+ *
+ * @throws                 cugraph::logic_error when an error occurs.
+ *
+ * @tparam vertex_t Type of vertex identifiers. Needs to be an integral type.
+ * @tparam edge_t Type of edge identifiers. Needs to be an integral type.
+ * @tparam weight_t Type of edge weights. Needs to be a floating point type.
+ * @tparam multi_gpu Flag indicating whether template instantiation should target single-GPU (false)
+ * @param handle RAFT handle object to encapsulate resources (e.g. CUDA stream, communicator, and
+ * handles to various CUDA libraries) to run graph algorithms.
+ * @param graph_view Graph view object.
+ * @param edge_weight_view Optional view object holding edge weights for @p graph_view. If @p
+ * edge_weight_view.has_value() == true, use the weights associated with the graph. If false, assume
+ * a weight of 1 for all edges.
+ * @param vertex_pairs tuple of device spans defining the vertex pairs to compute similarity for
+ * In a multi-gpu context each vertex pair should be local to this GPU.
+ * @param do_expensive_check A flag to run expensive checks for input arguments (if set to `true`).
+ * @return similarity coefficient for the corresponding @p vertex_pairs
+ */
+template <typename vertex_t, typename edge_t, typename weight_t, bool multi_gpu>
+rmm::device_uvector<weight_t> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<vertex_t, edge_t, false, multi_gpu> const& graph_view,
+  std::optional<edge_property_view_t<edge_t, weight_t const*>> edge_weight_view,
+  std::tuple<raft::device_span<vertex_t const>, raft::device_span<vertex_t const>> vertex_pairs,
+  bool do_expensive_check = false);
+
 /**
  * @brief     Compute Sorensen similarity coefficient
  *
@@ -2093,6 +2124,62 @@ std::
     std::optional<size_t> topk,
     bool do_expensive_check = false);
 
+/**
+ * @brief     Compute Consine all pairs similarity coefficient
+ *
+ * Similarity is computed for all pairs of vertices.  Note that in a sparse
+ * graph, many of the vertex pairs will have a score of zero.  We actually
+ * compute similarity only for vertices that are two hop neighbors within
+ * the graph, since vertices that are not two hop neighbors will have
+ * a score of 0.
+ *
+ * If @p vertices is specified we will compute similarity on two hop
+ * neighbors the @p vertices.  If @p vertices is not specified it will
+ * compute similarity on all two hop neighbors in the graph.
+ *
+ * If @p topk is specified only the top @p topk scoring vertex pairs
+ * will be returned, if not specified then scores for all computed vertex pairs
+ * will be returned.
+ *
+ * Note the list of two hop neighbors in the entire graph might be a large
+ * number of vertex pairs.  If the graph is dense enough it could be as large
+ * as the the number of vertices squared, which might run out of memory.
+ *
+ * @throws                 cugraph::logic_error when an error occurs.
+ *
+ * @tparam vertex_t Type of vertex identifiers. Needs to be an integral type.
+ * @tparam edge_t Type of edge identifiers. Needs to be an integral type.
+ * @tparam weight_t Type of edge weights. Needs to be a floating point type.
+ * @tparam multi_gpu Flag indicating whether template instantiation should target single-GPU (false)
+ * @param handle RAFT handle object to encapsulate resources (e.g. CUDA stream, communicator, and
+ * handles to various CUDA libraries) to run graph algorithms.
+ * @param graph_view Graph view object.
+ * @param edge_weight_view Optional view object holding edge weights for @p graph_view. If @p
+ * edge_weight_view.has_value() == true, use the weights associated with the graph. If false, assume
+ * a weight of 1 for all edges.
+ * @param vertices optional device span defining the seed vertices. In a multi-gpu context the
+ * vertices should be local to this GPU.
+ * @param topk optional specification of the how many of the top scoring vertex pairs should be
+ * returned
+ * @param do_expensive_check A flag to run expensive checks for input arguments (if set to `true`).
+ * @return tuple containing three device vectors (v1, v2, score) of the same length.  Corresponding
+ * elements in the vectors identify a result, v1 identifying a vertex in the graph, v2 identifying
+ * one of v1's two hop neighors, and the score identifying the similarity score between v1 and v2.
+ * If @p topk was specified then the vectors will be no longer than @p topk elements.  In a
+ * multi-gpu context, if @p topk is specified all results will return on GPU rank 0, otherwise they
+ * will be returned on the local GPU for vertex v1.
+ */
+template <typename vertex_t, typename edge_t, typename weight_t, bool multi_gpu>
+std::
+  tuple<rmm::device_uvector<vertex_t>, rmm::device_uvector<vertex_t>, rmm::device_uvector<weight_t>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<vertex_t, edge_t, false, multi_gpu> const& graph_view,
+    std::optional<edge_property_view_t<edge_t, weight_t const*>> edge_weight_view,
+    std::optional<raft::device_span<vertex_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check = false);
+
 /**
  * @brief     Compute Sorensen similarity coefficient
  *

cpp/src/link_prediction/cosine_mg.cu

@@ -0,0 +1,122 @@
+/*
+ * Copyright (c) 2022-2024, NVIDIA CORPORATION.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "link_prediction/cosine_similarity_impl.cuh"
+
+namespace cugraph {
+
+template rmm::device_uvector<float> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int32_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int32_t, float const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<float> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int64_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<float> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int64_t, int64_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+  std::tuple<raft::device_span<int64_t const>, raft::device_span<int64_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<double> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int32_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int32_t, double const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<double> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int64_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<double> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int64_t, int64_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+  std::tuple<raft::device_span<int64_t const>, raft::device_span<int64_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<float>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int32_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int32_t, float const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<float>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int64_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int64_t>, rmm::device_uvector<int64_t>, rmm::device_uvector<float>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int64_t, int64_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+    std::optional<raft::device_span<int64_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<double>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int32_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int32_t, double const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<double>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int64_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int64_t>, rmm::device_uvector<int64_t>, rmm::device_uvector<double>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int64_t, int64_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+    std::optional<raft::device_span<int64_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+}  // namespace cugraph