Unify the hist tree method for different devices. (#9363)

R-package/src/Makevars.in

@@ -82,6 +82,7 @@ OBJECTS= \
     $(PKGROOT)/src/common/charconv.o \
     $(PKGROOT)/src/common/column_matrix.o \
     $(PKGROOT)/src/common/common.o \
+    $(PKGROOT)/src/common/error_msg.o \
     $(PKGROOT)/src/common/hist_util.o \
     $(PKGROOT)/src/common/host_device_vector.o \
     $(PKGROOT)/src/common/io.o \

R-package/src/Makevars.win

@@ -82,6 +82,7 @@ OBJECTS= \
     $(PKGROOT)/src/common/charconv.o \
     $(PKGROOT)/src/common/column_matrix.o \
     $(PKGROOT)/src/common/common.o \
+    $(PKGROOT)/src/common/error_msg.o \
     $(PKGROOT)/src/common/hist_util.o \
     $(PKGROOT)/src/common/host_device_vector.o \
     $(PKGROOT)/src/common/io.o \

src/common/error_msg.cc

@@ -0,0 +1,36 @@
+/**
+ * Copyright 2023 by XGBoost contributors
+ */
+#include "error_msg.h"
+
+#include "xgboost/logging.h"
+
+namespace xgboost::error {
+void WarnDeprecatedGPUHist() {
+  bool static thread_local logged{false};
+  if (logged) {
+    return;
+  }
+  auto msg =
+      "The tree method `gpu_hist` is deprecated since 2.0.0. To use GPU training, set the `device` "
+      R"(parameter to CUDA instead.
+
+    E.g. tree_method = "hist", device = "CUDA"
+
+)";
+  LOG(WARNING) << msg;
+  logged = true;
+}
+
+void WarnManualUpdater() {
+  bool static thread_local logged{false};
+  if (logged) {
+    return;
+  }
+  LOG(WARNING)
+      << "You have manually specified the `updater` parameter. The `tree_method` parameter "
+         "will be ignored. Incorrect sequence of updaters will produce undefined "
+         "behavior. For common uses, we recommend using `tree_method` parameter instead.";
+  logged = true;
+}
+}  // namespace xgboost::error

src/common/error_msg.h

@@ -75,9 +75,12 @@ inline void WarnOldSerialization() {
   if (logged) {
     return;
   }
-
   LOG(WARNING) << OldSerialization();
   logged = true;
 }
+
+void WarnDeprecatedGPUHist();
+
+void WarnManualUpdater();
 }  // namespace xgboost::error
 #endif  // XGBOOST_COMMON_ERROR_MSG_H_

src/gbm/gbtree.cc

@@ -9,7 +9,7 @@
 #include <dmlc/omp.h>
 #include <dmlc/parameter.h>
 
-#include <algorithm>
+#include <algorithm>  // for equal
 #include <cinttypes>  // for uint32_t
 #include <limits>
 #include <memory>
@@ -40,8 +40,53 @@
 namespace xgboost::gbm {
 DMLC_REGISTRY_FILE_TAG(gbtree);
 
+namespace {
+/** @brief Map the `tree_method` parameter to the `updater` parameter. */
+std::string MapTreeMethodToUpdaters(Context const* ctx_, TreeMethod tree_method) {
+  // Choose updaters according to tree_method parameters
+  switch (tree_method) {
+    case TreeMethod::kAuto:  // Use hist as default in 2.0
+    case TreeMethod::kHist: {
+      return ctx_->DispatchDevice([] { return "grow_quantile_histmaker"; },
+                                  [] {
+                                    common::AssertGPUSupport();
+                                    return "grow_gpu_hist";
+                                  });
+    }
+    case TreeMethod::kApprox:
+      CHECK(ctx_->IsCPU()) << "The `approx` tree method is not supported on GPU.";
+      return "grow_histmaker";
+    case TreeMethod::kExact:
+      CHECK(ctx_->IsCPU()) << "The `exact` tree method is not supported on GPU.";
+      return "grow_colmaker,prune";
+    case TreeMethod::kGPUHist: {
+      common::AssertGPUSupport();
+      error::WarnDeprecatedGPUHist();
+      return "grow_gpu_hist";
+    }
+    default:
+      auto tm = static_cast<std::underlying_type_t<TreeMethod>>(tree_method);
+      LOG(FATAL) << "Unknown tree_method: `" << tm << "`.";
+  }
+
+  LOG(FATAL) << "unreachable";
+  return "";
+}
+
+bool UpdatersMatched(std::vector<std::string> updater_seq,
+                     std::vector<std::unique_ptr<TreeUpdater>> const& updaters) {
+  if (updater_seq.size() != updaters.size()) {
+    return false;
+  }
+
+  return std::equal(updater_seq.cbegin(), updater_seq.cend(), updaters.cbegin(),
+                    [](std::string const& name, std::unique_ptr<TreeUpdater> const& up) {
+                      return name == up->Name();
+                    });
+}
+}  // namespace
+
 void GBTree::Configure(Args const& cfg) {
-  std::string updater_seq = tparam_.updater_seq;
   tparam_.UpdateAllowUnknown(cfg);
   tree_param_.UpdateAllowUnknown(cfg);
 
@@ -54,8 +99,7 @@ void GBTree::Configure(Args const& cfg) {
 
   // configure predictors
   if (!cpu_predictor_) {
-    cpu_predictor_ = std::unique_ptr<Predictor>(
+    cpu_predictor_ = std::unique_ptr<Predictor>(Predictor::Create("cpu_predictor", this->ctx_));
-        Predictor::Create("cpu_predictor", this->ctx_));
   }
   cpu_predictor_->Configure(cfg);
 #if defined(XGBOOST_USE_CUDA)
@@ -70,26 +114,17 @@ void GBTree::Configure(Args const& cfg) {
 
 #if defined(XGBOOST_USE_ONEAPI)
   if (!oneapi_predictor_) {
-    oneapi_predictor_ = std::unique_ptr<Predictor>(
+    oneapi_predictor_ =
-        Predictor::Create("oneapi_predictor", this->ctx_));
+        std::unique_ptr<Predictor>(Predictor::Create("oneapi_predictor", this->ctx_));
   }
   oneapi_predictor_->Configure(cfg);
 #endif  // defined(XGBOOST_USE_ONEAPI)
 
-  monitor_.Init("GBTree");
+  // `updater` parameter was manually specified
-
+  specified_updater_ =
-  specified_updater_ = std::any_of(
+      std::any_of(cfg.cbegin(), cfg.cend(), [](auto const& arg) { return arg.first == "updater"; });
-      cfg.cbegin(), cfg.cend(),
+  if (specified_updater_) {
-      [](std::pair<std::string, std::string> const& arg) { return arg.first == "updater"; });
+    error::WarnManualUpdater();
-
-  if (specified_updater_ && !showed_updater_warning_) {
-    LOG(WARNING) << "DANGER AHEAD: You have manually specified `updater` "
-        "parameter. The `tree_method` parameter will be ignored. "
-        "Incorrect sequence of updaters will produce undefined "
-        "behavior. For common uses, we recommend using "
-        "`tree_method` parameter instead.";
-    // Don't drive users to silent XGBOost.
-    showed_updater_warning_ = true;
   }
 
   if (model_.learner_model_param->IsVectorLeaf()) {
@@ -97,50 +132,28 @@ void GBTree::Configure(Args const& cfg) {
         << "Only the hist tree method is supported for building multi-target trees with vector "
            "leaf.";
   }
-  LOG(DEBUG) << "Using tree method: " << static_cast<int>(tparam_.tree_method);
-  this->ConfigureUpdaters();
 
-  if (updater_seq != tparam_.updater_seq) {
+  LOG(DEBUG) << "Using tree method: " << static_cast<int>(tparam_.tree_method);
+
+  if (!specified_updater_) {
+    this->tparam_.updater_seq = MapTreeMethodToUpdaters(ctx_, tparam_.tree_method);
+  }
+
+  auto up_names = common::Split(tparam_.updater_seq, ',');
+  if (!UpdatersMatched(up_names, updaters_)) {
     updaters_.clear();
-    this->InitUpdater(cfg);
+    for (auto const& name : up_names) {
-  } else {
+      std::unique_ptr<TreeUpdater> up(
+          TreeUpdater::Create(name.c_str(), ctx_, &model_.learner_model_param->task));
+      updaters_.push_back(std::move(up));
+    }
+  }
+
   for (auto& up : updaters_) {
     up->Configure(cfg);
   }
 }
 
-  configured_ = true;
-}
-
-void GBTree::ConfigureUpdaters() {
-  if (specified_updater_) {
-    return;
-  }
-  // `updater` parameter was manually specified
-  /* Choose updaters according to tree_method parameters */
-  switch (tparam_.tree_method) {
-    case TreeMethod::kAuto:  // Use hist as default in 2.0
-    case TreeMethod::kHist: {
-      tparam_.updater_seq = "grow_quantile_histmaker";
-      break;
-    }
-    case TreeMethod::kApprox:
-      tparam_.updater_seq = "grow_histmaker";
-      break;
-    case TreeMethod::kExact:
-      tparam_.updater_seq = "grow_colmaker,prune";
-      break;
-    case TreeMethod::kGPUHist: {
-      common::AssertGPUSupport();
-      tparam_.updater_seq = "grow_gpu_hist";
-      break;
-    }
-    default:
-      LOG(FATAL) << "Unknown tree_method (" << static_cast<int>(tparam_.tree_method)
-                 << ") detected";
-  }
-}
-
 void GPUCopyGradient(HostDeviceVector<GradientPair> const*, bst_group_t, bst_group_t,
                      HostDeviceVector<GradientPair>*)
 #if defined(XGBOOST_USE_CUDA)
@@ -195,14 +208,8 @@ void GBTree::DoBoost(DMatrix* p_fmat, HostDeviceVector<GradientPair>* in_gpair,
   bst_target_t const n_groups = model_.learner_model_param->OutputLength();
   monitor_.Start("BoostNewTrees");
 
-  // Weird case that tree method is cpu-based but gpu_id is set.  Ideally we should let
+  auto out = linalg::MakeTensorView(ctx_, &predt->predictions, p_fmat->Info().num_row_,
-  // `gpu_id` be the single source of determining what algorithms to run, but that will
+                                    model_.learner_model_param->OutputLength());
-  // break a lots of existing code.
-  auto device = tparam_.tree_method != TreeMethod::kGPUHist ? Context::kCpuId : ctx_->gpu_id;
-  auto out = linalg::MakeTensorView(
-      device,
-      device == Context::kCpuId ? predt->predictions.HostSpan() : predt->predictions.DeviceSpan(),
-      p_fmat->Info().num_row_, model_.learner_model_param->OutputLength());
   CHECK_NE(n_groups, 0);
 
   if (!p_fmat->SingleColBlock() && obj->Task().UpdateTreeLeaf()) {
@@ -261,47 +268,6 @@ void GBTree::DoBoost(DMatrix* p_fmat, HostDeviceVector<GradientPair>* in_gpair,
   this->CommitModel(std::move(new_trees));
 }
 
-void GBTree::InitUpdater(Args const& cfg) {
-  std::string tval = tparam_.updater_seq;
-  std::vector<std::string> ups = common::Split(tval, ',');
-
-  if (updaters_.size() != 0) {
-    // Assert we have a valid set of updaters.
-    CHECK_EQ(ups.size(), updaters_.size());
-    for (auto const& up : updaters_) {
-      bool contains = std::any_of(ups.cbegin(), ups.cend(),
-                        [&up](std::string const& name) {
-                          return name == up->Name();
-                        });
-      if (!contains) {
-        std::stringstream ss;
-        ss << "Internal Error: " << " mismatched updater sequence.\n";
-        ss << "Specified updaters: ";
-        std::for_each(ups.cbegin(), ups.cend(),
-                      [&ss](std::string const& name){
-                        ss << name << " ";
-                      });
-        ss << "\n" << "Actual updaters: ";
-        std::for_each(updaters_.cbegin(), updaters_.cend(),
-                      [&ss](std::unique_ptr<TreeUpdater> const& updater){
-                        ss << updater->Name() << " ";
-                      });
-        LOG(FATAL) << ss.str();
-      }
-    }
-    // Do not push new updater in.
-    return;
-  }
-
-  // create new updaters
-  for (const std::string& pstr : ups) {
-    std::unique_ptr<TreeUpdater> up(
-        TreeUpdater::Create(pstr.c_str(), ctx_, &model_.learner_model_param->task));
-    up->Configure(cfg);
-    updaters_.push_back(std::move(up));
-  }
-}
-
 void GBTree::BoostNewTrees(HostDeviceVector<GradientPair>* gpair, DMatrix* p_fmat, int bst_group,
                            std::vector<HostDeviceVector<bst_node_t>>* out_position,
                            TreesOneGroup* ret) {
@@ -310,6 +276,7 @@ void GBTree::BoostNewTrees(HostDeviceVector<GradientPair>* gpair, DMatrix* p_fma
   // create the trees
   for (int i = 0; i < model_.param.num_parallel_tree; ++i) {
     if (tparam_.process_type == TreeProcessType::kDefault) {
+      CHECK(!updaters_.empty());
       CHECK(!updaters_.front()->CanModifyTree())
           << "Updater: `" << updaters_.front()->Name() << "` "
           << "can not be used to create new trees. "
@@ -465,7 +432,6 @@ void GBTree::SaveModel(Json* p_out) const {
 
 void GBTree::Slice(bst_layer_t begin, bst_layer_t end, bst_layer_t step, GradientBooster* out,
                    bool* out_of_bound) const {
-  CHECK(configured_);
   CHECK(out);
 
   auto p_gbtree = dynamic_cast<GBTree*>(out);
@@ -517,7 +483,6 @@ void GBTree::Slice(bst_layer_t begin, bst_layer_t end, bst_layer_t step, Gradien
 
 void GBTree::PredictBatchImpl(DMatrix* p_fmat, PredictionCacheEntry* out_preds, bool is_training,
                               bst_layer_t layer_begin, bst_layer_t layer_end) const {
-  CHECK(configured_);
   if (layer_end == 0) {
     layer_end = this->BoostedRounds();
   }
@@ -577,7 +542,6 @@ void GBTree::PredictBatch(DMatrix* p_fmat, PredictionCacheEntry* out_preds, bool
 void GBTree::InplacePredict(std::shared_ptr<DMatrix> p_m, float missing,
                             PredictionCacheEntry* out_preds, bst_layer_t layer_begin,
                             bst_layer_t layer_end) const {
-  CHECK(configured_);
   auto [tree_begin, tree_end] = detail::LayerToTree(model_, layer_begin, layer_end);
   CHECK_LE(tree_end, model_.trees.size()) << "Invalid number of trees.";
   if (p_m->Ctx()->Device() != this->ctx_->Device()) {
@@ -606,8 +570,6 @@ void GBTree::InplacePredict(std::shared_ptr<DMatrix> p_m, float missing,
 
 [[nodiscard]] std::unique_ptr<Predictor> const& GBTree::GetPredictor(
     bool is_training, HostDeviceVector<float> const* out_pred, DMatrix* f_dmat) const {
-  CHECK(configured_);
-
   // Data comes from SparsePageDMatrix. Since we are loading data in pages, no need to
   // prevent data copy.
   if (f_dmat && !f_dmat->SingleColBlock()) {
@@ -914,7 +876,6 @@ class Dart : public GBTree {
   void PredictContribution(DMatrix* p_fmat, HostDeviceVector<bst_float>* out_contribs,
                            bst_layer_t layer_begin, bst_layer_t layer_end,
                            bool approximate) override {
-    CHECK(configured_);
     auto [tree_begin, tree_end] = detail::LayerToTree(model_, layer_begin, layer_end);
     cpu_predictor_->PredictContribution(p_fmat, out_contribs, model_, tree_end, &weight_drop_,
                                         approximate);
@@ -923,7 +884,6 @@ class Dart : public GBTree {
   void PredictInteractionContributions(DMatrix* p_fmat, HostDeviceVector<float>* out_contribs,
                                        bst_layer_t layer_begin, bst_layer_t layer_end,
                                        bool approximate) override {
-    CHECK(configured_);
     auto [tree_begin, tree_end] = detail::LayerToTree(model_, layer_begin, layer_end);
     cpu_predictor_->PredictInteractionContributions(p_fmat, out_contribs, model_, tree_end,
                                                     &weight_drop_, approximate);

src/gbm/gbtree.h

@@ -59,9 +59,7 @@ struct GBTreeTrainParam : public XGBoostParameter<GBTreeTrainParam> {
   TreeMethod tree_method;
   // declare parameters
   DMLC_DECLARE_PARAMETER(GBTreeTrainParam) {
-    DMLC_DECLARE_FIELD(updater_seq)
+    DMLC_DECLARE_FIELD(updater_seq).describe("Tree updater sequence.").set_default("");
-        .set_default("grow_colmaker,prune")
-        .describe("Tree updater sequence.");
     DMLC_DECLARE_FIELD(process_type)
         .set_default(TreeProcessType::kDefault)
         .add_enum("default", TreeProcessType::kDefault)
@@ -170,22 +168,21 @@ bool SliceTrees(bst_layer_t begin, bst_layer_t end, bst_layer_t step, GBTreeMode
 class GBTree : public GradientBooster {
  public:
   explicit GBTree(LearnerModelParam const* booster_config, Context const* ctx)
-      : GradientBooster{ctx}, model_(booster_config, ctx_) {}
+      : GradientBooster{ctx}, model_(booster_config, ctx_) {
-
+    monitor_.Init(__func__);
-  void Configure(const Args& cfg) override;
+  }
-  /*! \brief Map `tree_method` parameter to `updater` parameter */
-  void ConfigureUpdaters();
 
+  void Configure(Args const& cfg) override;
   /**
-   * \brief Optionally update the leaf value.
+   * @brief Optionally update the leaf value.
    */
   void UpdateTreeLeaf(DMatrix const* p_fmat, HostDeviceVector<float> const& predictions,
-                      ObjFunction const* obj,
+                      ObjFunction const* obj, std::int32_t group_idx,
-                      std::int32_t group_idx,
                       std::vector<HostDeviceVector<bst_node_t>> const& node_position,
                       std::vector<std::unique_ptr<RegTree>>* p_trees);
-
+  /**
-  /*! \brief Carry out one iteration of boosting */
+   * @brief Carry out one iteration of boosting.
+   */
   void DoBoost(DMatrix* p_fmat, HostDeviceVector<GradientPair>* in_gpair,
                PredictionCacheEntry* predt, ObjFunction const* obj) override;
 
@@ -289,7 +286,6 @@ class GBTree : public GradientBooster {
 
   void PredictInstance(const SparsePage::Inst& inst, std::vector<bst_float>* out_preds,
                        uint32_t layer_begin, uint32_t layer_end) override {
-    CHECK(configured_);
     std::uint32_t _, tree_end;
     std::tie(_, tree_end) = detail::LayerToTree(model_, layer_begin, layer_end);
     cpu_predictor_->PredictInstance(inst, out_preds, model_, tree_end);
@@ -307,7 +303,6 @@ class GBTree : public GradientBooster {
   void PredictContribution(DMatrix* p_fmat, HostDeviceVector<float>* out_contribs,
                            bst_layer_t layer_begin, bst_layer_t layer_end,
                            bool approximate) override {
-    CHECK(configured_);
     auto [tree_begin, tree_end] = detail::LayerToTree(model_, layer_begin, layer_end);
     CHECK_EQ(tree_begin, 0) << "Predict contribution supports only iteration end: (0, "
                                "n_iteration), using model slicing instead.";
@@ -318,7 +313,6 @@ class GBTree : public GradientBooster {
   void PredictInteractionContributions(DMatrix* p_fmat, HostDeviceVector<float>* out_contribs,
                                        bst_layer_t layer_begin, bst_layer_t layer_end,
                                        bool approximate) override {
-    CHECK(configured_);
     auto [tree_begin, tree_end] = detail::LayerToTree(model_, layer_begin, layer_end);
     CHECK_EQ(tree_begin, 0) << "Predict interaction contribution supports only iteration end: (0, "
                                "n_iteration), using model slicing instead.";
@@ -332,9 +326,6 @@ class GBTree : public GradientBooster {
   }
 
  protected:
-  // initialize updater before using them
-  void InitUpdater(Args const& cfg);
-
   void BoostNewTrees(HostDeviceVector<GradientPair>* gpair, DMatrix* p_fmat, int bst_group,
                      std::vector<HostDeviceVector<bst_node_t>>* out_position,
                      std::vector<std::unique_ptr<RegTree>>* ret);
@@ -352,10 +343,7 @@ class GBTree : public GradientBooster {
   GBTreeTrainParam tparam_;
   // Tree training parameter
   tree::TrainParam tree_param_;
-  // ----training fields----
-  bool showed_updater_warning_ {false};
   bool specified_updater_   {false};
-  bool configured_ {false};
   // the updaters that can be applied to each of tree
   std::vector<std::unique_ptr<TreeUpdater>> updaters_;
   // Predictors

tests/cpp/gbm/test_gbtree.cc

@@ -4,10 +4,12 @@
 #include <gtest/gtest.h>
 #include <xgboost/context.h>
 #include <xgboost/host_device_vector.h>  // for HostDeviceVector
+#include <xgboost/json.h>                // for Json, Object
 #include <xgboost/learner.h>             // for Learner
 
 #include <limits>    // for numeric_limits
 #include <memory>    // for shared_ptr
+#include <optional>  // for optional
 #include <string>    // for string
 
 #include "../../../src/data/proxy_dmatrix.h"  // for DMatrixProxy
@@ -165,6 +167,115 @@ TEST(GBTree, ChoosePredictor) {
   // data is not pulled back into host
   ASSERT_FALSE(data.HostCanWrite());
 }
+
+TEST(GBTree, ChooseTreeMethod) {
+  bst_row_t n_samples{128};
+  bst_feature_t n_features{64};
+  auto Xy = RandomDataGenerator{n_samples, n_features, 0.5f}.GenerateDMatrix(true);
+
+  auto with_update = [&](std::optional<std::string> device,
+                         std::optional<std::string> tree_method) {
+    auto learner = std::unique_ptr<Learner>(Learner::Create({Xy}));
+    if (tree_method.has_value()) {
+      learner->SetParam("tree_method", tree_method.value());
+    }
+    if (device.has_value()) {
+      learner->SetParam("gpu_id", device.value());
+    }
+    learner->Configure();
+    for (std::int32_t i = 0; i < 3; ++i) {
+      learner->UpdateOneIter(0, Xy);
+    }
+    Json config{Object{}};
+    learner->SaveConfig(&config);
+    auto updater = config["learner"]["gradient_booster"]["updater"];
+    CHECK(!IsA<Null>(updater));
+    return updater;
+  };
+
+  auto with_boost = [&](std::optional<std::string> device, std::optional<std::string> tree_method) {
+    auto learner = std::unique_ptr<Learner>(Learner::Create({Xy}));
+    if (tree_method.has_value()) {
+      learner->SetParam("tree_method", tree_method.value());
+    }
+    if (device.has_value()) {
+      learner->SetParam("gpu_id", device.value());
+    }
+    learner->Configure();
+    for (std::int32_t i = 0; i < 3; ++i) {
+      HostDeviceVector<GradientPair> gpair{GenerateRandomGradients(Xy->Info().num_row_)};
+      learner->BoostOneIter(0, Xy, &gpair);
+    }
+
+    Json config{Object{}};
+    learner->SaveConfig(&config);
+    auto updater = config["learner"]["gradient_booster"]["updater"];
+    return updater;
+  };
+
+  // |        | hist    | gpu_hist | exact | NA  |
+  // |--------+---------+----------+-------+-----|
+  // | CUDA:0 | GPU     | GPU (w)  | Err   | GPU | # not yet tested
+  // | CPU    | CPU     | Err      | CPU   | CPU | # not yet tested
+  // |--------+---------+----------+-------+-----|
+  // | -1     | CPU     | GPU (w)  | CPU   | CPU |
+  // | 0      | GPU     | GPU (w)  | Err   | GPU |
+  // | NA     | CPU     | GPU (w)  | CPU   | CPU |
+  //
+  // - (w): warning
+  // - CPU: Run on CPU.
+  // - GPU: Run on CUDA.
+  // - Err: Not feasible.
+  // - NA:  Parameter is not specified.
+
+  // When GPU hist is specified with a CPU context, we should emit an error. However, it's
+  // quite difficult to detect whether the CPU context is being used because it's the
+  // default or because it's specified by the user.
+
+  std::map<std::pair<std::optional<std::string>, std::optional<std::string>>, std::string>
+      expectation{
+          // hist
+          {{"hist", "-1"}, "grow_quantile_histmaker"},
+          {{"hist", "0"}, "grow_gpu_hist"},
+          {{"hist", std::nullopt}, "grow_quantile_histmaker"},
+          // gpu_hist
+          {{"gpu_hist", "-1"}, "grow_gpu_hist"},
+          {{"gpu_hist", "0"}, "grow_gpu_hist"},
+          {{"gpu_hist", std::nullopt}, "grow_gpu_hist"},
+          // exact
+          {{"exact", "-1"}, "grow_colmaker,prune"},
+          {{"exact", "0"}, "err"},
+          {{"exact", std::nullopt}, "grow_colmaker,prune"},
+          // NA
+          {{std::nullopt, "-1"}, "grow_quantile_histmaker"},
+          {{std::nullopt, "0"}, "grow_gpu_hist"},  // default to hist
+          {{std::nullopt, std::nullopt}, "grow_quantile_histmaker"},
+      };
+
+  auto run_test = [&](auto fn) {
+    for (auto const& kv : expectation) {
+      auto device = kv.first.second;
+      auto tm = kv.first.first;
+
+      if (kv.second == "err") {
+        ASSERT_THROW({ fn(device, tm); }, dmlc::Error)
+            << " device:" << device.value_or("NA") << " tm:" << tm.value_or("NA");
+        continue;
+      }
+      auto up = fn(device, tm);
+      auto ups = get<Array const>(up);
+      auto exp_names = common::Split(kv.second, ',');
+      ASSERT_EQ(exp_names.size(), ups.size());
+      for (std::size_t i = 0; i < exp_names.size(); ++i) {
+        ASSERT_EQ(get<String const>(ups[i]["name"]), exp_names[i])
+            << " device:" << device.value_or("NA") << " tm:" << tm.value_or("NA");
+      }
+    }
+  };
+
+  run_test(with_update);
+  run_test(with_boost);
+}
 #endif  // XGBOOST_USE_CUDA
 
 // Some other parts of test are in `Tree.JsonIO'.

tests/python/test_training_continuation.py

@@ -57,12 +57,12 @@ class TestTrainingContinuation:
 
         gbdt_02 = xgb.train(xgb_params_01, dtrain_2class,
                             num_boost_round=0)
-        gbdt_02.save_model('xgb_tc.model')
+        gbdt_02.save_model('xgb_tc.json')
 
         gbdt_02a = xgb.train(xgb_params_01, dtrain_2class,
                              num_boost_round=10, xgb_model=gbdt_02)
         gbdt_02b = xgb.train(xgb_params_01, dtrain_2class,
-                             num_boost_round=10, xgb_model="xgb_tc.model")
+                             num_boost_round=10, xgb_model="xgb_tc.json")
         ntrees_02a = len(gbdt_02a.get_dump())
         ntrees_02b = len(gbdt_02b.get_dump())
         assert ntrees_02a == 10
@@ -78,18 +78,18 @@ class TestTrainingContinuation:
 
         gbdt_03 = xgb.train(xgb_params_01, dtrain_2class,
                             num_boost_round=3)
-        gbdt_03.save_model('xgb_tc.model')
+        gbdt_03.save_model('xgb_tc.json')
 
         gbdt_03a = xgb.train(xgb_params_01, dtrain_2class,
                              num_boost_round=7, xgb_model=gbdt_03)
         gbdt_03b = xgb.train(xgb_params_01, dtrain_2class,
-                             num_boost_round=7, xgb_model="xgb_tc.model")
+                             num_boost_round=7, xgb_model="xgb_tc.json")
         ntrees_03a = len(gbdt_03a.get_dump())
         ntrees_03b = len(gbdt_03b.get_dump())
         assert ntrees_03a == 10
         assert ntrees_03b == 10
 
-        os.remove('xgb_tc.model')
+        os.remove('xgb_tc.json')
 
         res1 = mean_squared_error(y_2class, gbdt_03a.predict(dtrain_2class))
         res2 = mean_squared_error(y_2class, gbdt_03b.predict(dtrain_2class))