Merge pull request #3774 from nv-rliu/branch-23.10-merge-23.08

Forward-merge branch-23.08 to branch-23.10

.github/workflows/build.yaml

@@ -34,6 +34,7 @@ jobs:
       branch: ${{ inputs.branch }}
       date: ${{ inputs.date }}
       sha: ${{ inputs.sha }}
+      node_type: cpu32
   python-build:
     needs: [cpp-build]
     secrets: inherit

.github/workflows/pr.yaml

@@ -36,7 +36,7 @@ jobs:
     uses: rapidsai/shared-action-workflows/.github/workflows/[email protected]
     with:
       build_type: pull-request
-      node_type: cpu16
+      node_type: cpu32
   conda-cpp-tests:
     needs: conda-cpp-build
     secrets: inherit

build.sh

@@ -175,7 +175,7 @@ if hasArg --cmake_default_generator; then
     CMAKE_GENERATOR_OPTION=""
 fi
 if hasArg --pydevelop; then
-    PYTHON_ARGS_FOR_INSTALL="-m pip install --no-build-isolation --no-deps -e ."
+    PYTHON_ARGS_FOR_INSTALL="-m pip install --no-build-isolation --no-deps -e"
 fi
 
 # Append `-DFIND_RAFT_CPP=ON` to EXTRA_CMAKE_ARGS unless a user specified the option.
@@ -302,6 +302,22 @@ if buildAll || hasArg pylibcugraph; then
     if hasArg --clean; then
         cleanPythonDir ${REPODIR}/python/pylibcugraph
     else
+        # FIXME: skbuild with setuptools>=64 has a bug when called from a "pip
+        # install -e" command, resulting in a broken editable wheel. Continue
+        # to use "setup.py bdist_ext --inplace" for a develop build until
+        # https://github.com/scikit-build/scikit-build/issues/981 is closed.
+        if hasArg --pydevelop; then
+            cd ${REPODIR}/python/pylibcugraph
+            python setup.py build_ext \
+                   --inplace \
+                   -- \
+                   -DFIND_CUGRAPH_CPP=ON \
+                   -DUSE_CUGRAPH_OPS=${BUILD_WITH_CUGRAPHOPS} \
+                   -Dcugraph_ROOT=${LIBCUGRAPH_BUILD_DIR} \
+                   -- \
+                   -j${PARALLEL_LEVEL:-1}
+            cd -
+        fi
         SKBUILD_CONFIGURE_OPTIONS="${SKBUILD_EXTRA_CMAKE_ARGS} -DUSE_CUGRAPH_OPS=${BUILD_WITH_CUGRAPHOPS}" \
             SKBUILD_BUILD_OPTIONS="-j${PARALLEL_LEVEL}" \
             python ${PYTHON_ARGS_FOR_INSTALL} ${REPODIR}/python/pylibcugraph
@@ -313,6 +329,22 @@ if buildAll || hasArg cugraph; then
     if hasArg --clean; then
         cleanPythonDir ${REPODIR}/python/cugraph
     else
+        # FIXME: skbuild with setuptools>=64 has a bug when called from a "pip
+        # install -e" command, resulting in a broken editable wheel. Continue
+        # to use "setup.py bdist_ext --inplace" for a develop build until
+        # https://github.com/scikit-build/scikit-build/issues/981 is closed.
+        if hasArg --pydevelop; then
+            cd ${REPODIR}/python/cugraph
+            python setup.py build_ext \
+                   --inplace \
+                   -- \
+                   -DFIND_CUGRAPH_CPP=ON \
+                   -DUSE_CUGRAPH_OPS=${BUILD_WITH_CUGRAPHOPS} \
+                   -Dcugraph_ROOT=${LIBCUGRAPH_BUILD_DIR} \
+                   -- \
+                   -j${PARALLEL_LEVEL:-1}
+            cd -
+        fi
         SKBUILD_CONFIGURE_OPTIONS="${SKBUILD_EXTRA_CMAKE_ARGS} -DUSE_CUGRAPH_OPS=${BUILD_WITH_CUGRAPHOPS}" \
             SKBUILD_BUILD_OPTIONS="-j${PARALLEL_LEVEL}" \
             python ${PYTHON_ARGS_FOR_INSTALL} ${REPODIR}/python/cugraph

ci/test_wheel_cugraph.sh

@@ -9,7 +9,7 @@ RAPIDS_PY_WHEEL_NAME="pylibcugraph_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-whe
 python -m pip install --no-deps ./local-pylibcugraph-dep/pylibcugraph*.whl
 
 # Always install latest dask for testing
-python -m pip install git+https://github.com/dask/dask.git@main git+https://github.com/dask/distributed.git@main git+https://github.com/rapidsai/[email protected]
+python -m pip install git+https://github.com/dask/dask.git@2023.7.1 git+https://github.com/dask/distributed.git@2023.7.1 git+https://github.com/rapidsai/[email protected]
 
 # Only download test data for x86
 arch=$(uname -m)

conda/recipes/cugraph-pyg/meta.yaml

@@ -26,7 +26,7 @@ requirements:
     - python
     - scikit-build >=0.13.1
   run:
-    - distributed >=2023.5.1
+    - distributed ==2023.7.1
     - numba >=0.57
     - numpy >=1.21
     - python

conda/recipes/cugraph-service/meta.yaml

@@ -57,7 +57,7 @@ outputs:
         - cupy >=12.0.0
         - dask-cuda ={{ minor_version }}
         - dask-cudf ={{ minor_version }}
-        - distributed >=2023.5.1
+        - distributed ==2023.7.1
         - numba >=0.57
         - numpy >=1.21
         - python

conda/recipes/cugraph/meta.yaml

@@ -75,9 +75,9 @@ requirements:
     - cupy >=12.0.0
     - dask-cuda ={{ minor_version }}
     - dask-cudf ={{ minor_version }}
-    - dask >=2023.5.1
-    - dask-core >=2023.5.1
-    - distributed >=2023.5.1
+    - dask ==2023.7.1
+    - dask-core ==2023.7.1
+    - distributed ==2023.7.1
     - libcugraph ={{ version }}
     - pylibcugraph ={{ version }}
     - pylibraft ={{ minor_version }}

cpp/include/cugraph/graph_functions.hpp

@@ -922,15 +922,16 @@ rmm::device_uvector<vertex_t> select_random_vertices(
  * This function renumbers sampling function (e.g. uniform_neighbor_sample) outputs satisfying the
  * following requirements.
  *
- * 1. Say @p edgelist_srcs has N unique vertices. These N unique vertices will be mapped to [0, N).
- * 2. Among the N unique vertices, an original vertex with a smaller attached hop number will be
- * renumbered to a smaller vertex ID than any other original vertices with a larger attached hop
- * number (if @p edgelist_hops.has_value() is true). If a single vertex is attached to multiple hop
- * numbers, the minimum hop number is used.
- * 3. Say @p edgelist_dsts has M unique vertices that appear only in @p edgelist_dsts (the set of M
- * unique vertices does not include any vertices that appear in @p edgelist_srcs). Then, these M
- * unique vertices will be mapped to [N, N + M).
- * 4. If label_offsets.has_value() is ture, edge lists for different labels will be renumbered
+ * 1. If @p edgelist_hops is valid, we can consider (vertex ID, flag=src, hop) triplets for each
+ * vertex ID in @p edgelist_srcs and (vertex ID, flag=dst, hop) triplets for each vertex ID in @p
+ * edgelist_dsts. From these triplets, we can find the minimum (hop, flag) pairs for every unique
+ * vertex ID (hop is the primary key and flag is the secondary key, flag=src is considered smaller
+ * than flag=dst if hop numbers are same). Vertex IDs with smaller (hop, flag) pairs precede vertex
+ * IDs with larger (hop, flag) pairs in renumbering. Ordering can be arbitrary among the vertices
+ * with the same (hop, flag) pairs.
+ * 2. If @p edgelist_hops is invalid, unique vertex IDs in @p edgelist_srcs precede vertex IDs that
+ * appear only in @p edgelist_dsts.
+ * 3. If label_offsets.has_value() is ture, edge lists for different labels will be renumbered
  * separately.
  *
  * This function is single-GPU only (we are not aware of any practical multi-GPU use cases).
@@ -940,10 +941,10 @@ rmm::device_uvector<vertex_t> select_random_vertices(
  * @param  handle RAFT handle object to encapsulate resources (e.g. CUDA stream, communicator, and
  * handles to various CUDA libraries) to run graph algorithms.
  * @param edgelist_srcs A vector storing original edgelist source vertices.
- * @param edgelist_hops An optional pointer to the array storing hops for each edge list source
- * vertices (size = @p edgelist_srcs.size()).
  * @param edgelist_dsts A vector storing original edgelist destination vertices (size = @p
  * edgelist_srcs.size()).
+ * @param edgelist_hops An optional pointer to the array storing hops for each edge list (source,
+ * destination) pairs (size = @p edgelist_srcs.size() if valid).
  * @param label_offsets An optional tuple of unique labels and the input edge list (@p
  * edgelist_srcs, @p edgelist_hops, and @p edgelist_dsts) offsets for the labels (siez = # unique
  * labels + 1).
@@ -962,8 +963,8 @@ std::tuple<rmm::device_uvector<vertex_t>,
 renumber_sampled_edgelist(
   raft::handle_t const& handle,
   rmm::device_uvector<vertex_t>&& edgelist_srcs,
-  std::optional<raft::device_span<int32_t const>> edgelist_hops,
   rmm::device_uvector<vertex_t>&& edgelist_dsts,
+  std::optional<raft::device_span<int32_t const>> edgelist_hops,
   std::optional<std::tuple<raft::device_span<label_t const>, raft::device_span<size_t const>>>
     label_offsets,
   bool do_expensive_check = false);

cpp/src/c_api/graph_functions.cpp

@@ -146,6 +146,12 @@ struct two_hop_neighbors_functor : public cugraph::c_api::abstract_functor {
                    start_vertices_->size_,
                    handle_.get_stream());
 
+        if constexpr (multi_gpu) {
+          start_vertices =
+            cugraph::detail::shuffle_ext_vertices_to_local_gpu_by_vertex_partitioning(
+              handle_, std::move(start_vertices));
+        }
+
         cugraph::renumber_ext_vertices<vertex_t, multi_gpu>(
           handle_,
           start_vertices.data(),
@@ -155,11 +161,6 @@ struct two_hop_neighbors_functor : public cugraph::c_api::abstract_functor {
           graph_view.local_vertex_partition_range_last(),
           do_expensive_check_);
 
-        if constexpr (multi_gpu) {
-          start_vertices =
-            cugraph::detail::shuffle_ext_vertices_to_local_gpu_by_vertex_partitioning(
-              handle_, std::move(start_vertices));
-        }
       } else {
         start_vertices.resize(graph_view.local_vertex_partition_range_size(), handle_.get_stream());
         cugraph::detail::sequence_fill(handle_.get_stream(),

cpp/src/c_api/similarity.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022, NVIDIA CORPORATION.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.

cpp/src/c_api/uniform_neighbor_sampling.cpp

@@ -236,9 +236,9 @@ struct uniform_neighbor_sampling_functor : public cugraph::c_api::abstract_funct
         std::tie(src, dst, renumber_map, renumber_map_offsets) = cugraph::renumber_sampled_edgelist(
           handle_,
           std::move(src),
+          std::move(dst),
           hop ? std::make_optional(raft::device_span<int32_t const>{hop->data(), hop->size()})
               : std::nullopt,
-          std::move(dst),
           std::make_optional(std::make_tuple(
             raft::device_span<label_t const>{edge_label->data(), edge_label->size()},
             raft::device_span<size_t const>{offsets->data(), offsets->size()})),

cpp/src/link_prediction/jaccard_impl.cuh

@@ -24,22 +24,15 @@ namespace cugraph {
 namespace detail {
 
 struct jaccard_functor_t {
-  template <typename weight_t>
-  weight_t __device__ compute_score(weight_t cardinality_a,
-                                    weight_t cardinality_b,
-                                    weight_t cardinality_a_intersect_b) const
-  {
-    return cardinality_a_intersect_b / (cardinality_a + cardinality_b - cardinality_a_intersect_b);
-  }
-};
-
-struct weighted_jaccard_functor_t {
   template <typename weight_t>
   weight_t __device__ compute_score(weight_t weight_a,
                                     weight_t weight_b,
-                                    weight_t min_weight_a_intersect_b) const
+                                    weight_t weight_a_intersect_b,
+                                    weight_t weight_a_union_b) const
   {
-    return min_weight_a_intersect_b / (weight_a + weight_b - min_weight_a_intersect_b);
+    return weight_a_union_b <= std::numeric_limits<weight_t>::min()
+             ? weight_t{0}
+             : weight_a_intersect_b / weight_a_union_b;
   }
 };
 
@@ -55,20 +48,12 @@ rmm::device_uvector<weight_t> jaccard_coefficients(
 {
   CUGRAPH_EXPECTS(!graph_view.has_edge_mask(), "unimplemented.");
 
-  if (!edge_weight_view)
-    return detail::similarity(handle,
-                              graph_view,
-                              edge_weight_view,
-                              vertex_pairs,
-                              detail::jaccard_functor_t{},
-                              do_expensive_check);
-  else
-    return detail::similarity(handle,
-                              graph_view,
-                              edge_weight_view,
-                              vertex_pairs,
-                              detail::weighted_jaccard_functor_t{},
-                              do_expensive_check);
+  return detail::similarity(handle,
+                            graph_view,
+                            edge_weight_view,
+                            vertex_pairs,
+                            detail::jaccard_functor_t{},
+                            do_expensive_check);
 }
 
 }  // namespace cugraph

cpp/src/link_prediction/overlap_impl.cuh

@@ -24,22 +24,15 @@ namespace cugraph {
 namespace detail {
 
 struct overlap_functor_t {
-  template <typename weight_t>
-  weight_t __device__ compute_score(weight_t cardinality_a,
-                                    weight_t cardinality_b,
-                                    weight_t cardinality_a_intersect_b) const
-  {
-    return cardinality_a_intersect_b / std::min(cardinality_a, cardinality_b);
-  }
-};
-
-struct weighted_overlap_functor_t {
   template <typename weight_t>
   weight_t __device__ compute_score(weight_t weight_a,
                                     weight_t weight_b,
-                                    weight_t min_weight_a_intersect_b) const
+                                    weight_t weight_a_intersect_b,
+                                    weight_t weight_a_union_b) const
   {
-    return min_weight_a_intersect_b / std::min(weight_a, weight_b);
+    return std::min(weight_a, weight_b) <= std::numeric_limits<weight_t>::min()
+             ? weight_t{0}
+             : weight_a_intersect_b / std::min(weight_a, weight_b);
   }
 };
 
@@ -55,20 +48,12 @@ rmm::device_uvector<weight_t> overlap_coefficients(
 {
   CUGRAPH_EXPECTS(!graph_view.has_edge_mask(), "unimplemented.");
 
-  if (!edge_weight_view)
-    return detail::similarity(handle,
-                              graph_view,
-                              edge_weight_view,
-                              vertex_pairs,
-                              detail::overlap_functor_t{},
-                              do_expensive_check);
-  else
-    return detail::similarity(handle,
-                              graph_view,
-                              edge_weight_view,
-                              vertex_pairs,
-                              detail::weighted_overlap_functor_t{},
-                              do_expensive_check);
+  return detail::similarity(handle,
+                            graph_view,
+                            edge_weight_view,
+                            vertex_pairs,
+                            detail::overlap_functor_t{},
+                            do_expensive_check);
 }
 
 }  // namespace cugraph