Merge remote-tracking branch 'upstream/branch-23.06' into branch-23.08

.github/workflows/pr.yaml

@@ -105,5 +105,5 @@ jobs:
       # On arm also need to install CMake because treelite needs to be compiled (no wheels available for arm).
       test-before-arm64: "pip install cmake && pip install git+https://github.com/dask/[email protected] git+https://github.com/dask/[email protected] git+https://github.com/rapidsai/[email protected]"
       # parallelization is based on current test memory usage
-      test-unittest: "pytest ./python/cuml/tests -k 'not test_sparse_pca_inputs' -n 8 --ignore=python/cuml/tests/dask && pytest ./python/cuml/tests -k 'test_sparse_pca_inputs' && pytest ./python/cuml/tests/dask"
+      test-unittest: "python -m pytest ./python/cuml/tests -k 'not test_sparse_pca_inputs' -n 8 --ignore=python/cuml/tests/dask && python -m pytest ./python/cuml/tests -k 'test_sparse_pca_inputs' && python -m pytest ./python/cuml/tests/dask"
       test-smoketest: "python ci/wheel_smoke_test_cuml.py"

ci/release/apply_wheel_modifications.sh

@@ -19,5 +19,4 @@ sed -i "s/rmm/rmm${CUDA_SUFFIX}/g" python/pyproject.toml
 if [[ $CUDA_SUFFIX == "-cu12" ]]; then
     sed -i "s/cuda-python[<=>\.,0-9]*/cuda-python>=12.0,<13.0/g" python/pyproject.toml
     sed -i "s/cupy-cuda11x/cupy-cuda12x/g" python/pyproject.toml
-    sed -i "s/numba[<=>\.,0-9]*/numba>=0.57/g" python/pyproject.toml
 fi

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

@@ -46,7 +46,7 @@ dependencies:
 - nbsphinx
 - ninja
 - nltk
-- numba>=0.56.4,<0.57
+- numba>=0.57
 - numpydoc
 - pip
 - pydata-sphinx-theme

conda/recipes/cuml/meta.yaml

@@ -80,7 +80,6 @@ requirements:
     - python x.x
     - raft-dask ={{ minor_version }}
     - treelite {{ treelite_version }}
-    - seaborn
 
 tests:
   requirements:

cpp/bench/sg/svc.cu

@@ -102,7 +102,7 @@ std::vector<SvcParams<D>> getInputs()
 
   // SvmParameter{C, cache_size, max_iter, nochange_steps, tol, verbosity})
   p.svm_param = ML::SVM::SvmParameter{1, 200, 100, 100, 1e-3, CUML_LEVEL_INFO, 0, ML::SVM::C_SVC};
-  p.model     = ML::SVM::SvmModel<D>{0, 0, 0, nullptr, nullptr, nullptr, 0, nullptr};
+  p.model     = ML::SVM::SvmModel<D>{0, 0, 0, nullptr, {}, nullptr, 0, nullptr};
 
   std::vector<Triplets> rowcols = {{50000, 2, 2}, {2048, 100000, 2}, {50000, 1000, 2}};
 

cpp/include/cuml/experimental/fil/decision_forest.hpp

@@ -105,6 +105,7 @@ struct decision_forest {
   decision_forest()
     : nodes_{},
       root_node_indexes_{},
+      node_id_mapping_{},
       vector_output_{},
       categorical_storage_{},
       num_features_{},
@@ -125,6 +126,8 @@ struct decision_forest {
    * @param nodes A buffer containing all nodes within the forest
    * @param root_node_indexes A buffer containing the index of the root node
    * of every tree in the forest
+   * @param node_id_mapping Mapping to use to convert FIL's internal node ID into Treelite's node
+   * ID. Only relevant when predict_type == infer_kind::leaf_id
    * @param num_features The number of features per input sample for this model
    * @param num_outputs The number of outputs per row from this model
    * @param has_categorical_nodes Whether this forest contains any
@@ -155,6 +158,7 @@ struct decision_forest {
    */
   decision_forest(raft_proto::buffer<node_type>&& nodes,
                   raft_proto::buffer<index_type>&& root_node_indexes,
+                  raft_proto::buffer<index_type>&& node_id_mapping,
                   index_type num_features,
                   index_type num_outputs                                     = index_type{2},
                   bool has_categorical_nodes                                 = false,
@@ -169,6 +173,7 @@ struct decision_forest {
                   io_type postproc_constant = io_type{1})
     : nodes_{nodes},
       root_node_indexes_{root_node_indexes},
+      node_id_mapping_{node_id_mapping},
       vector_output_{vector_output},
       categorical_storage_{categorical_storage},
       num_features_{num_features},
@@ -207,6 +212,8 @@ struct decision_forest {
     if (inference_kind == infer_kind::per_tree) {
       result = num_trees();
       if (has_vector_leaves()) { result *= num_outputs_; }
+    } else if (inference_kind == infer_kind::leaf_id) {
+      result = num_trees();
     }
     return result;
   }
@@ -233,6 +240,8 @@ struct decision_forest {
    * @param[in] predict_type Type of inference to perform. Defaults to summing
    * the outputs of all trees and produce an output per row. If set to
    * "per_tree", we will instead output all outputs of individual trees.
+   * If set to "leaf_id", we will output the integer ID of the leaf node
+   * for each tree.
    * @param[in] specified_rows_per_block_iter If non-nullopt, this value is
    * used to determine how many rows are evaluated for each inference
    * iteration within a CUDA block. Runtime performance is quite sensitive
@@ -301,6 +310,8 @@ struct decision_forest {
   raft_proto::buffer<node_type> nodes_;
   /** The index of the root node for each tree in the forest */
   raft_proto::buffer<index_type> root_node_indexes_;
+  /** Mapping to apply to node IDs. Only relevant when predict_type == infer_kind::leaf_id */
+  raft_proto::buffer<index_type> node_id_mapping_;
   /** Buffer of outputs for all leaves in vector-leaf models */
   std::optional<raft_proto::buffer<io_type>> vector_output_;
   /** Buffer of elements used as backing data for bitsets which specify
@@ -323,6 +334,7 @@ struct decision_forest {
   {
     return forest_type{nodes_.data(),
                        root_node_indexes_.data(),
+                       node_id_mapping_.data(),
                        static_cast<index_type>(root_node_indexes_.size()),
                        num_outputs_};
   }

cpp/include/cuml/experimental/fil/detail/decision_forest_builder.hpp

@@ -67,7 +67,7 @@ struct decision_forest_builder {
         if (cur_tree_size_ % alignment_ != index_type{}) {
           auto padding = (alignment_ - cur_tree_size_ % alignment_);
           for (auto i = index_type{}; i < padding; ++i) {
-            add_node(typename node_type::threshold_type{});
+            add_node(typename node_type::threshold_type{}, std::nullopt);
           }
         }
       }
@@ -81,6 +81,7 @@ struct decision_forest_builder {
   void add_categorical_node(
     iter_t vec_begin,
     iter_t vec_end,
+    std::optional<int> tl_node_id                     = std::nullopt,
     bool default_to_distant_child                     = false,
     typename node_type::metadata_storage_type feature = typename node_type::metadata_storage_type{},
     typename node_type::offset_type offset            = typename node_type::offset_type{})
@@ -100,12 +101,14 @@ struct decision_forest_builder {
     auto set = bitset{set_storage, max_node_categories};
     std::for_each(vec_begin, vec_end, [&set](auto&& cat_index) { set.set(cat_index); });
 
-    add_node(node_value, false, default_to_distant_child, true, feature, offset, false);
+    add_node(node_value, tl_node_id, false, default_to_distant_child, true, feature, offset, false);
   }
 
   /* Add a leaf node with vector output */
   template <typename iter_t>
-  void add_leaf_vector_node(iter_t vec_begin, iter_t vec_end)
+  void add_leaf_vector_node(iter_t vec_begin,
+                            iter_t vec_end,
+                            std::optional<int> tl_node_id = std::nullopt)
   {
     auto leaf_index = typename node_type::index_type(vector_output_.size() / output_size_);
     std::copy(vec_begin, vec_end, std::back_inserter(vector_output_));
@@ -115,13 +118,16 @@ struct decision_forest_builder {
                         false,
                         typename node_type::metadata_storage_type{},
                         typename node_type::offset_type{});
+    // 0 indicates the lack of ID mapping for a particular node
+    node_id_mapping_.push_back(static_cast<index_type>(tl_node_id.value_or(0)));
     ++cur_tree_size_;
   }
 
   /* Add a node to the model */
   template <typename value_t>
   void add_node(
     value_t val,
+    std::optional<int> tl_node_id                     = std::nullopt,
     bool is_leaf_node                                 = true,
     bool default_to_distant_child                     = false,
     bool is_categorical_node                          = false,
@@ -132,6 +138,8 @@ struct decision_forest_builder {
     if (is_inclusive) { val = std::nextafter(val, std::numeric_limits<value_t>::infinity()); }
     nodes_.emplace_back(
       val, is_leaf_node, default_to_distant_child, is_categorical_node, feature, offset);
+    // 0 indicates the lack of ID mapping for a particular node
+    node_id_mapping_.push_back(static_cast<index_type>(tl_node_id.value_or(0)));
     ++cur_tree_size_;
   }
 
@@ -192,6 +200,10 @@ struct decision_forest_builder {
                          mem_type,
                          device,
                          stream},
+      raft_proto::buffer{raft_proto::buffer{node_id_mapping_.data(), node_id_mapping_.size()},
+                         mem_type,
+                         device,
+                         stream},
       num_feature,
       num_class,
       max_num_categories_ != 0,
@@ -234,6 +246,7 @@ struct decision_forest_builder {
   std::vector<index_type> root_node_indexes_;
   std::vector<typename node_type::threshold_type> vector_output_;
   std::vector<typename node_type::index_type> categorical_storage_;
+  std::vector<index_type> node_id_mapping_;
 };
 
 }  // namespace detail

cpp/include/cuml/experimental/fil/detail/evaluate_tree.hpp

@@ -15,6 +15,7 @@
  */
 #pragma once
 #include <stdint.h>
+#include <type_traits>
 #ifndef __CUDACC__
 #include <math.h>
 #endif
@@ -26,19 +27,33 @@ namespace fil {
 namespace detail {
 
 /*
- * Evaluate a single tree on a single row
+ * Evaluate a single tree on a single row.
+ * If node_id_mapping is not-nullptr, this kernel outputs leaf node's ID
+ * instead of the leaf value.
  *
  * @tparam has_vector_leaves Whether or not this tree has vector leaves
- * @tparam has_categorical nodes Whether or not this tree has any nodes with
+ * @tparam has_categorical_nodes Whether or not this tree has any nodes with
  * categorical splits
  * @tparam node_t The type of nodes in this tree
  * @tparam io_t The type used for input to and output from this tree (typically
  * either floats or doubles)
+ * @tparam node_id_mapping_t If non-nullptr_t, this indicates the type we expect for
+ * node_id_mapping.
  * @param node Pointer to the root node of this tree
  * @param row Pointer to the input data for this row
+ * @param first_root_node Pointer to the root node of the first tree.
+ * @param node_id_mapping Array representing the mapping from internal node IDs to
+ * final leaf ID outputs
  */
-template <bool has_vector_leaves, bool has_categorical_nodes, typename node_t, typename io_t>
-HOST DEVICE auto evaluate_tree(node_t const* __restrict__ node, io_t const* __restrict__ row)
+template <bool has_vector_leaves,
+          bool has_categorical_nodes,
+          typename node_t,
+          typename io_t,
+          typename node_id_mapping_t = std::nullptr_t>
+HOST DEVICE auto evaluate_tree_impl(node_t const* __restrict__ node,
+                                    io_t const* __restrict__ row,
+                                    node_t const* __restrict__ first_root_node = nullptr,
+                                    node_id_mapping_t node_id_mapping          = nullptr)
 {
   using categorical_set_type = bitset<uint32_t, typename node_t::index_type const>;
   auto cur_node              = *node;
@@ -60,12 +75,18 @@ HOST DEVICE auto evaluate_tree(node_t const* __restrict__ node, io_t const* __re
     node += cur_node.child_offset(condition);
     cur_node = *node;
   } while (!cur_node.is_leaf());
-  return cur_node.template output<has_vector_leaves>();
+  if constexpr (std::is_same_v<node_id_mapping_t, std::nullptr_t>) {
+    return cur_node.template output<has_vector_leaves>();
+  } else {
+    return node_id_mapping[node - first_root_node];
+  }
 }
 
 /*
  * Evaluate a single tree which requires external categorical storage on a
- * single node
+ * single node.
+ * If node_id_mapping is not-nullptr, this kernel outputs leaf node's ID
+ * instead of the leaf value.
  *
  * For non-categorical models and models with a relatively small number of
  * categories for any feature, all information necessary for model evaluation
@@ -81,15 +102,23 @@ HOST DEVICE auto evaluate_tree(node_t const* __restrict__ node, io_t const* __re
  * either floats or doubles)
  * @tparam categorical_storage_t The underlying type used for storing
  * categorical data (typically char)
+ * @tparam node_id_mapping_t If non-nullptr_t, this indicates the type we expect for
+ * node_id_mapping.
  * @param node Pointer to the root node of this tree
  * @param row Pointer to the input data for this row
  * @param categorical_storage Pointer to where categorical split data is
  * stored.
  */
-template <bool has_vector_leaves, typename node_t, typename io_t, typename categorical_storage_t>
-HOST DEVICE auto evaluate_tree(node_t const* __restrict__ node,
-                               io_t const* __restrict__ row,
-                               categorical_storage_t const* __restrict__ categorical_storage)
+template <bool has_vector_leaves,
+          typename node_t,
+          typename io_t,
+          typename categorical_storage_t,
+          typename node_id_mapping_t = std::nullptr_t>
+HOST DEVICE auto evaluate_tree_impl(node_t const* __restrict__ node,
+                                    io_t const* __restrict__ row,
+                                    categorical_storage_t const* __restrict__ categorical_storage,
+                                    node_t const* __restrict__ first_root_node = nullptr,
+                                    node_id_mapping_t node_id_mapping          = nullptr)
 {
   using categorical_set_type = bitset<uint32_t, categorical_storage_t const>;
   auto cur_node              = *node;
@@ -109,7 +138,73 @@ HOST DEVICE auto evaluate_tree(node_t const* __restrict__ node,
     node += cur_node.child_offset(condition);
     cur_node = *node;
   } while (!cur_node.is_leaf());
-  return cur_node.template output<has_vector_leaves>();
+  if constexpr (std::is_same_v<node_id_mapping_t, std::nullptr_t>) {
+    return cur_node.template output<has_vector_leaves>();
+  } else {
+    return node_id_mapping[node - first_root_node];
+  }
+}
+
+/**
+ * Dispatch to an appropriate version of evaluate_tree kernel.
+ *
+ * @tparam has_vector_leaves Whether or not this tree has vector leaves
+ * @tparam has_categorical_nodes Whether or not this tree has any nodes with
+ * categorical splits
+ * @tparam has_nonlocal_categories Whether or not this tree has any nodes that store
+ * categorical split data externally
+ * @tparam predict_leaf Whether to predict leaf IDs
+ * @tparam forest_t The type of forest
+ * @tparam io_t The type used for input to and output from this tree (typically
+ * either floats or doubles)
+ * @tparam categorical_data_t The type for non-local categorical data storage.
+ * @param forest The forest used to perform inference
+ * @param tree_index The index of the tree we are evaluating
+ * @param row The data row we are evaluating
+ * @param categorical_data The pointer to where non-local data on categorical splits are stored.
+ */
+template <bool has_vector_leaves,
+          bool has_categorical_nodes,
+          bool has_nonlocal_categories,
+          bool predict_leaf,
+          typename forest_t,
+          typename io_t,
+          typename categorical_data_t>
+HOST DEVICE auto evaluate_tree(forest_t const& forest,
+                               index_type tree_index,
+                               io_t const* __restrict__ row,
+                               categorical_data_t categorical_data)
+{
+  using node_t = typename forest_t::node_type;
+  if constexpr (predict_leaf) {
+    auto leaf_node_id = index_type{};
+    if constexpr (has_nonlocal_categories) {
+      leaf_node_id = evaluate_tree_impl<has_vector_leaves>(forest.get_tree_root(tree_index),
+                                                           row,
+                                                           categorical_data,
+                                                           forest.get_tree_root(0),
+                                                           forest.get_node_id_mapping());
+    } else {
+      leaf_node_id = evaluate_tree_impl<has_vector_leaves, has_categorical_nodes>(
+        forest.get_tree_root(tree_index),
+        row,
+        forest.get_tree_root(0),
+        forest.get_node_id_mapping());
+    }
+    return leaf_node_id;
+  } else {
+    auto tree_output = std::conditional_t<has_vector_leaves,
+                                          typename node_t::index_type,
+                                          typename node_t::threshold_type>{};
+    if constexpr (has_nonlocal_categories) {
+      tree_output = evaluate_tree_impl<has_vector_leaves>(
+        forest.get_tree_root(tree_index), row, categorical_data);
+    } else {
+      tree_output = evaluate_tree_impl<has_vector_leaves, has_categorical_nodes>(
+        forest.get_tree_root(tree_index), row);
+    }
+    return tree_output;
+  }
 }
 
 }  // namespace detail

cpp/include/cuml/experimental/fil/detail/forest.hpp

@@ -41,10 +41,12 @@ struct forest {
 
   HOST DEVICE forest(node_type* forest_nodes,
                      index_type* forest_root_indexes,
+                     index_type* node_id_mapping,
                      index_type num_trees,
                      index_type num_outputs)
     : nodes_{forest_nodes},
       root_node_indexes_{forest_root_indexes},
+      node_id_mapping_{node_id_mapping},
       num_trees_{num_trees},
       num_outputs_{num_outputs}
   {
@@ -56,6 +58,10 @@ struct forest {
     return nodes_ + root_node_indexes_[tree_index];
   }
 
+  /* Return pointer to the mapping from internal node IDs to final node ID outputs.
+   * Only used when infer_type == infer_kind::leaf_id */
+  HOST DEVICE const auto* get_node_id_mapping() const { return node_id_mapping_; }
+
   /* Return the number of trees in this forest */
   HOST DEVICE auto tree_count() const { return num_trees_; }
 
@@ -66,6 +72,7 @@ struct forest {
  private:
   node_type* nodes_;
   index_type* root_node_indexes_;
+  index_type* node_id_mapping_;
   index_type num_trees_;
   index_type num_outputs_;
 };

cpp/include/cuml/experimental/fil/detail/infer.hpp

@@ -55,7 +55,8 @@ namespace detail {
  * required
  * @param infer_type Type of inference to perform. Defaults to summing the outputs of all trees
  * and produce an output per row. If set to "per_tree", we will instead output all outputs of
- * individual trees.
+ * individual trees. If set to "leaf_id", we will output the integer ID of the leaf node
+ * for each tree.
  * @param specified_chunk_size If non-nullopt, the size of "mini-batches"
  * used for distributing work across threads
  * @param device The device on which to execute evaluation