[EM] Return a full DMatrix instead of a Ellpack from the GPU sampler. (#10753)

.gitignore

@@ -63,6 +63,7 @@ java/xgboost4j-demo/data/
 java/xgboost4j-demo/tmp/
 java/xgboost4j-demo/model/
 nb-configuration*
+
 # Eclipse
 .project
 .cproject
@@ -154,3 +155,6 @@ model*.json
 *.rds
 Rplots.pdf
 *.zip
+
+# nsys
+*.nsys-rep

include/xgboost/data.h

@@ -110,8 +110,15 @@ class MetaInfo {
    * @brief Validate all metainfo.
    */
   void Validate(DeviceOrd device) const;
-
+  /**
-  MetaInfo Slice(common::Span<int32_t const> ridxs) const;
+   * @brief Slice the meta info.
+   *
+   * The device of ridxs is specified by the ctx object.
+   *
+   * @param ridxs Index of selected rows.
+   * @param nnz   The number of non-missing values.
+   */
+  MetaInfo Slice(Context const* ctx, common::Span<bst_idx_t const> ridxs, bst_idx_t nnz) const;
 
   MetaInfo Copy() const;
   /**

src/common/device_helpers.cuh

@@ -508,6 +508,11 @@ xgboost::common::Span<T> ToSpan(DeviceUVector<T> &vec) {
   return {vec.data(), vec.size()};
 }
 
+template <typename T>
+xgboost::common::Span<std::add_const_t<T>> ToSpan(DeviceUVector<T> const &vec) {
+  return {vec.data(), vec.size()};
+}
+
 // thrust begin, similiar to std::begin
 template <typename T>
 thrust::device_ptr<T> tbegin(xgboost::HostDeviceVector<T>& vector) {  // NOLINT

src/common/linalg_op.cuh

@@ -76,7 +76,7 @@ struct IterOp {
 // returns a thrust iterator for a tensor view.
 template <typename T, std::int32_t kDim>
 auto tcbegin(TensorView<T, kDim> v) {  // NOLINT
-  return dh::MakeTransformIterator<T>(
+  return thrust::make_transform_iterator(
       thrust::make_counting_iterator(0ul),
       detail::IterOp<std::add_const_t<std::remove_const_t<T>>, kDim>{v});
 }
@@ -85,5 +85,16 @@ template <typename T, std::int32_t kDim>
 auto tcend(TensorView<T, kDim> v) {  // NOLINT
   return tcbegin(v) + v.Size();
 }
+
+template <typename T, std::int32_t kDim>
+auto tbegin(TensorView<T, kDim> v) {  // NOLINT
+  return thrust::make_transform_iterator(thrust::make_counting_iterator(0ul),
+                                         detail::IterOp<std::remove_const_t<T>, kDim>{v});
+}
+
+template <typename T, std::int32_t kDim>
+auto tend(TensorView<T, kDim> v) {  // NOLINT
+  return tbegin(v) + v.Size();
+}
 }  // namespace xgboost::linalg
 #endif  // XGBOOST_COMMON_LINALG_OP_CUH_

src/data/data.cc

@@ -351,8 +351,10 @@ void MetaInfo::LoadBinary(dmlc::Stream *fi) {
   this->has_categorical_ = LoadFeatureType(feature_type_names, &feature_types.HostVector());
 }
 
+namespace {
 template <typename T>
-std::vector<T> Gather(const std::vector<T> &in, common::Span<int const> ridxs, size_t stride = 1) {
+std::vector<T> Gather(const std::vector<T>& in, common::Span<bst_idx_t const> ridxs,
+                      size_t stride = 1) {
   if (in.empty()) {
     return {};
   }
@@ -366,11 +368,51 @@ std::vector<T> Gather(const std::vector<T> &in, common::Span<int const> ridxs, s
   }
   return out;
 }
+}  // namespace
 
-MetaInfo MetaInfo::Slice(common::Span<int32_t const> ridxs) const {
+namespace cuda_impl {
+void SliceMetaInfo(Context const* ctx, MetaInfo const& info, common::Span<bst_idx_t const> ridx,
+                   MetaInfo* p_out);
+#if !defined(XGBOOST_USE_CUDA)
+void SliceMetaInfo(Context const*, MetaInfo const&, common::Span<bst_idx_t const>, MetaInfo*) {
+  common::AssertGPUSupport();
+}
+#endif
+}  // namespace cuda_impl
+
+MetaInfo MetaInfo::Slice(Context const* ctx, common::Span<bst_idx_t const> ridxs,
+                         bst_idx_t nnz) const {
+  /**
+   * Shape
+   */
   MetaInfo out;
   out.num_row_ = ridxs.size();
   out.num_col_ = this->num_col_;
+  out.num_nonzero_ = nnz;
+
+  /**
+   * Feature Info
+   */
+  out.feature_weights.SetDevice(ctx->Device());
+  out.feature_weights.Resize(this->feature_weights.Size());
+  out.feature_weights.Copy(this->feature_weights);
+
+  out.feature_names = this->feature_names;
+
+  out.feature_types.SetDevice(ctx->Device());
+  out.feature_types.Resize(this->feature_types.Size());
+  out.feature_types.Copy(this->feature_types);
+
+  out.feature_type_names = this->feature_type_names;
+
+  /**
+   * Sample Info
+   */
+  if (ctx->IsCUDA()) {
+    cuda_impl::SliceMetaInfo(ctx, *this, ridxs, &out);
+    return out;
+  }
+
   // Groups is maintained by a higher level Python function.  We should aim at deprecating
   // the slice function.
   if (this->labels.Size() != this->num_row_) {
@@ -386,10 +428,8 @@ MetaInfo MetaInfo::Slice(common::Span<int32_t const> ridxs) const {
     });
   }
 
-  out.labels_upper_bound_.HostVector() =
+  out.labels_upper_bound_.HostVector() = Gather(this->labels_upper_bound_.HostVector(), ridxs);
-      Gather(this->labels_upper_bound_.HostVector(), ridxs);
+  out.labels_lower_bound_.HostVector() = Gather(this->labels_lower_bound_.HostVector(), ridxs);
-  out.labels_lower_bound_.HostVector() =
-      Gather(this->labels_lower_bound_.HostVector(), ridxs);
   // weights
   if (this->weights_.Size() + 1 == this->group_ptr_.size()) {
     auto& h_weights = out.weights_.HostVector();
@@ -414,14 +454,6 @@ MetaInfo MetaInfo::Slice(common::Span<int32_t const> ridxs) const {
     });
   }
 
-  out.feature_weights.Resize(this->feature_weights.Size());
-  out.feature_weights.Copy(this->feature_weights);
-
-  out.feature_names = this->feature_names;
-  out.feature_types.Resize(this->feature_types.Size());
-  out.feature_types.Copy(this->feature_types);
-  out.feature_type_names = this->feature_type_names;
-
   return out;
 }
 

src/data/data.cu

@@ -1,9 +1,11 @@
 /**
- * Copyright 2019-2022 by XGBoost Contributors
+ * Copyright 2019-2024, XGBoost Contributors
  *
  * \file data.cu
  * \brief Handles setting metainfo from array interface.
  */
+#include <thrust/gather.h>  // for gather
+
 #include "../common/cuda_context.cuh"
 #include "../common/device_helpers.cuh"
 #include "../common/linalg_op.cuh"
@@ -169,6 +171,62 @@ void MetaInfo::SetInfoFromCUDA(Context const& ctx, StringView key, Json array) {
   }
 }
 
+namespace {
+void Gather(Context const* ctx, linalg::MatrixView<float const> in,
+            common::Span<bst_idx_t const> ridx, linalg::Matrix<float>* p_out) {
+  if (in.Empty()) {
+    return;
+  }
+  auto& out = *p_out;
+  out.Reshape(ridx.size(), in.Shape(1));
+  auto d_out = out.View(ctx->Device());
+
+  auto cuctx = ctx->CUDACtx();
+  auto map_it = thrust::make_transform_iterator(thrust::make_counting_iterator(0ull),
+                                                [=] XGBOOST_DEVICE(bst_idx_t i) {
+                                                  auto [r, c] = linalg::UnravelIndex(i, in.Shape());
+                                                  return (ridx[r] * in.Shape(1)) + c;
+                                                });
+  CHECK_NE(in.Shape(1), 0);
+  thrust::gather(cuctx->TP(), map_it, map_it + out.Size(), linalg::tcbegin(in),
+                 linalg::tbegin(d_out));
+}
+
+template <typename T>
+void Gather(Context const* ctx, HostDeviceVector<T> const& in, common::Span<bst_idx_t const> ridx,
+            HostDeviceVector<T>* p_out) {
+  if (in.Empty()) {
+    return;
+  }
+  in.SetDevice(ctx->Device());
+
+  auto& out = *p_out;
+  out.SetDevice(ctx->Device());
+  out.Resize(ridx.size());
+  auto d_out = out.DeviceSpan();
+
+  auto cuctx = ctx->CUDACtx();
+  auto d_in = in.ConstDeviceSpan();
+  thrust::gather(cuctx->TP(), dh::tcbegin(ridx), dh::tcend(ridx), dh::tcbegin(d_in),
+                 dh::tbegin(d_out));
+}
+}  // anonymous namespace
+
+namespace cuda_impl {
+void SliceMetaInfo(Context const* ctx, MetaInfo const& info, common::Span<bst_idx_t const> ridx,
+                   MetaInfo* p_out) {
+  auto& out = *p_out;
+
+  Gather(ctx, info.labels.View(ctx->Device()), ridx, &p_out->labels);
+  Gather(ctx, info.base_margin_.View(ctx->Device()), ridx, &p_out->base_margin_);
+
+  Gather(ctx, info.labels_lower_bound_, ridx, &out.labels_lower_bound_);
+  Gather(ctx, info.labels_upper_bound_, ridx, &out.labels_upper_bound_);
+
+  Gather(ctx, info.weights_, ridx, &out.weights_);
+}
+}  // namespace cuda_impl
+
 template <typename AdapterT>
 DMatrix* DMatrix::Create(AdapterT* adapter, float missing, int nthread,
                          const std::string& cache_prefix, DataSplitMode data_split_mode) {

src/data/ellpack_page.cu

@@ -1,6 +1,7 @@
 /**
  * Copyright 2019-2024, XGBoost contributors
  */
+#include <cuda/functional>  // for proclaim_return_type
 #include <thrust/iterator/discard_iterator.h>
 #include <thrust/iterator/transform_output_iterator.h>
 
@@ -576,4 +577,17 @@ EllpackDeviceAccessor EllpackPageImpl::GetHostAccessor(
           common::CompressedIterator<uint32_t>(h_gidx_buffer->data(), NumSymbols()),
           feature_types};
 }
+
+[[nodiscard]] bst_idx_t EllpackPageImpl::NumNonMissing(
+    Context const* ctx, common::Span<FeatureType const> feature_types) const {
+  auto d_acc = this->GetDeviceAccessor(ctx->Device(), feature_types);
+  using T = typename decltype(d_acc.gidx_iter)::value_type;
+  auto it = thrust::make_transform_iterator(
+      thrust::make_counting_iterator(0ull),
+      cuda::proclaim_return_type<T>([=] __device__(std::size_t i) { return d_acc.gidx_iter[i]; }));
+  auto nnz = thrust::count_if(ctx->CUDACtx()->CTP(), it, it + d_acc.row_stride * d_acc.n_rows,
+                              cuda::proclaim_return_type<bool>(
+                                  [=] __device__(T gidx) { return gidx != d_acc.NullValue(); }));
+  return nnz;
+}
 }  // namespace xgboost

src/data/ellpack_page.cuh

@@ -236,6 +236,11 @@ class EllpackPageImpl {
   [[nodiscard]] EllpackDeviceAccessor GetHostAccessor(
       Context const* ctx, std::vector<common::CompressedByteT>* h_gidx_buffer,
       common::Span<FeatureType const> feature_types = {}) const;
+  /**
+   * @brief Calculate the number of non-missing values.
+   */
+  [[nodiscard]] bst_idx_t NumNonMissing(Context const* ctx,
+                                        common::Span<FeatureType const> feature_types) const;
 
  private:
   /**

src/data/iterative_dmatrix.cu

@@ -101,6 +101,17 @@ void IterativeDMatrix::InitFromCUDA(Context const* ctx, BatchParam const& p,
   // Synchronise worker columns
 }
 
+IterativeDMatrix::IterativeDMatrix(std::shared_ptr<EllpackPage> ellpack, MetaInfo const& info,
+                                   BatchParam batch) {
+  this->ellpack_ = ellpack;
+  CHECK_EQ(this->Info().num_row_, 0);
+  CHECK_EQ(this->Info().num_col_, 0);
+  this->Info().Extend(info, true, true);
+  this->Info().num_nonzero_ = info.num_nonzero_;
+  CHECK_EQ(this->Info().num_row_, info.num_row_);
+  this->batch_ = batch;
+}
+
 BatchSet<EllpackPage> IterativeDMatrix::GetEllpackBatches(Context const* ctx,
                                                           BatchParam const& param) {
   if (param.Initialized()) {

src/data/iterative_dmatrix.h

@@ -48,6 +48,11 @@ class IterativeDMatrix : public QuantileDMatrix {
                             std::shared_ptr<DMatrix> ref, DataIterResetCallback *reset,
                             XGDMatrixCallbackNext *next, float missing, int nthread,
                             bst_bin_t max_bin);
+  /**
+   * @param Directly construct a QDM from an existing one.
+   */
+  IterativeDMatrix(std::shared_ptr<EllpackPage> ellpack, MetaInfo const &info, BatchParam batch);
+
   ~IterativeDMatrix() override = default;
 
   bool EllpackExists() const override { return static_cast<bool>(ellpack_); }

src/data/simple_dmatrix.cc

@@ -31,6 +31,9 @@ const MetaInfo& SimpleDMatrix::Info() const { return info_; }
 DMatrix* SimpleDMatrix::Slice(common::Span<int32_t const> ridxs) {
   auto out = new SimpleDMatrix;
   SparsePage& out_page = *out->sparse_page_;
+  // Convert to uint64 to avoid a breaking change in the C API. The performance impact is
+  // small since we have to iteratve through the sparse page.
+  std::vector<bst_idx_t> h_ridx(ridxs.data(), ridxs.data() + ridxs.size());
   for (auto const& page : this->GetBatches<SparsePage>()) {
     auto batch = page.GetView();
     auto& h_data = out_page.data.HostVector();
@@ -42,8 +45,8 @@ DMatrix* SimpleDMatrix::Slice(common::Span<int32_t const> ridxs) {
       std::copy(inst.begin(), inst.end(), std::back_inserter(h_data));
       h_offset.emplace_back(rptr);
     }
-    out->Info() = this->Info().Slice(ridxs);
+    auto ctx = this->fmat_ctx_.MakeCPU();
-    out->Info().num_nonzero_ = h_offset.back();
+    out->Info() = this->Info().Slice(&ctx, h_ridx, h_offset.back());
   }
   out->fmat_ctx_ = this->fmat_ctx_;
   return out;

src/tree/gpu_hist/gradient_based_sampler.cu

@@ -14,12 +14,12 @@
 
 #include "../../common/cuda_context.cuh"  // for CUDAContext
 #include "../../common/random.h"
+#include "../../data/ellpack_page.cuh"     // for EllpackPageImpl
+#include "../../data/iterative_dmatrix.h"  // for IterativeDMatrix
 #include "../param.h"
 #include "gradient_based_sampler.cuh"
 
-namespace xgboost {
+namespace xgboost::tree {
-namespace tree {
-
 /*! \brief A functor that returns random weights. */
 class RandomWeight : public thrust::unary_function<size_t, float> {
  public:
@@ -58,12 +58,14 @@ struct IsNonZero : public thrust::unary_function<GradientPair, bool> {
 };
 
 /*! \brief A functor that clears the row indexes with empty gradient. */
-struct ClearEmptyRows : public thrust::binary_function<GradientPair, size_t, size_t> {
+struct ClearEmptyRows : public thrust::binary_function<GradientPair, bst_idx_t, bst_idx_t> {
+  static constexpr bst_idx_t InvalidRow() { return std::numeric_limits<std::size_t>::max(); }
+
   XGBOOST_DEVICE size_t operator()(const GradientPair& gpair, size_t row_index) const {
     if (gpair.GetGrad() != 0 || gpair.GetHess() != 0) {
       return row_index;
     } else {
-      return std::numeric_limits<std::size_t>::max();
+      return InvalidRow();
     }
   }
 };
@@ -148,10 +150,9 @@ class PoissonSampling : public thrust::binary_function<GradientPair, size_t, Gra
 
 NoSampling::NoSampling(BatchParam batch_param) : batch_param_(std::move(batch_param)) {}
 
-GradientBasedSample NoSampling::Sample(Context const* ctx, common::Span<GradientPair> gpair,
+GradientBasedSample NoSampling::Sample(Context const*, common::Span<GradientPair> gpair,
                                        DMatrix* dmat) {
-  auto page = (*dmat->GetBatches<EllpackPage>(ctx, batch_param_).begin()).Impl();
+  return {dmat->Info().num_row_, dmat, gpair};
-  return {dmat->Info().num_row_, page, gpair};
 }
 
 ExternalMemoryNoSampling::ExternalMemoryNoSampling(BatchParam batch_param)
@@ -159,37 +160,39 @@ ExternalMemoryNoSampling::ExternalMemoryNoSampling(BatchParam batch_param)
 
 GradientBasedSample ExternalMemoryNoSampling::Sample(Context const* ctx,
                                                      common::Span<GradientPair> gpair,
-                                                     DMatrix* dmat) {
+                                                     DMatrix* p_fmat) {
+  std::shared_ptr<EllpackPage> new_page;
   if (!page_concatenated_) {
     // Concatenate all the external memory ELLPACK pages into a single in-memory page.
-    page_.reset(nullptr);
     bst_idx_t offset = 0;
-    for (auto& batch : dmat->GetBatches<EllpackPage>(ctx, batch_param_)) {
+    for (auto& batch : p_fmat->GetBatches<EllpackPage>(ctx, batch_param_)) {
       auto page = batch.Impl();
-      if (!page_) {
+      if (!new_page) {
-        page_ = std::make_unique<EllpackPageImpl>(ctx, page->CutsShared(), page->is_dense,
+        new_page = std::make_shared<EllpackPage>();
-                                                  page->row_stride, dmat->Info().num_row_);
+        *new_page->Impl() = EllpackPageImpl(ctx, page->CutsShared(), page->is_dense,
+                                            page->row_stride, p_fmat->Info().num_row_);
       }
-      bst_idx_t num_elements = page_->Copy(ctx, page, offset);
+      bst_idx_t num_elements = new_page->Impl()->Copy(ctx, page, offset);
       offset += num_elements;
     }
     page_concatenated_ = true;
+    this->p_fmat_new_ =
+        std::make_unique<data::IterativeDMatrix>(new_page, p_fmat->Info(), batch_param_);
   }
-  return {dmat->Info().num_row_, page_.get(), gpair};
+  return {p_fmat->Info().num_row_, this->p_fmat_new_.get(), gpair};
 }
 
 UniformSampling::UniformSampling(BatchParam batch_param, float subsample)
-    : batch_param_{std::move(batch_param)}, subsample_(subsample) {}
+    : batch_param_{std::move(batch_param)}, subsample_{subsample} {}
 
 GradientBasedSample UniformSampling::Sample(Context const* ctx, common::Span<GradientPair> gpair,
-                                            DMatrix* dmat) {
+                                            DMatrix* p_fmat) {
   // Set gradient pair to 0 with p = 1 - subsample
   auto cuctx = ctx->CUDACtx();
   thrust::replace_if(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair),
                      thrust::counting_iterator<std::size_t>(0),
                      BernoulliTrial(common::GlobalRandom()(), subsample_), GradientPair());
-  auto page = (*dmat->GetBatches<EllpackPage>(ctx, batch_param_).begin()).Impl();
+  return {p_fmat->Info().num_row_, p_fmat, gpair};
-  return {dmat->Info().num_row_, page, gpair};
 }
 
 ExternalMemoryUniformSampling::ExternalMemoryUniformSampling(size_t n_rows,
@@ -203,13 +206,17 @@ GradientBasedSample ExternalMemoryUniformSampling::Sample(Context const* ctx,
                                                           common::Span<GradientPair> gpair,
                                                           DMatrix* dmat) {
   auto cuctx = ctx->CUDACtx();
+
+  std::shared_ptr<EllpackPage> new_page = std::make_shared<EllpackPage>();
+  auto page = new_page->Impl();
+
   // Set gradient pair to 0 with p = 1 - subsample
   thrust::replace_if(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair),
                      thrust::counting_iterator<std::size_t>(0),
                      BernoulliTrial(common::GlobalRandom()(), subsample_), GradientPair{});
 
   // Count the sampled rows.
-  size_t sample_rows =
+  bst_idx_t sample_rows =
       thrust::count_if(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair), IsNonZero{});
 
   // Compact gradient pairs.
@@ -227,17 +234,25 @@ GradientBasedSample ExternalMemoryUniformSampling::Sample(Context const* ctx,
   auto batch_iterator = dmat->GetBatches<EllpackPage>(ctx, batch_param_);
   auto first_page = (*batch_iterator.begin()).Impl();
   // Create a new ELLPACK page with empty rows.
-  page_.reset();  // Release the device memory first before reallocating
+  *page = EllpackPageImpl{ctx, first_page->CutsShared(), first_page->is_dense,
-  page_.reset(new EllpackPageImpl(ctx, first_page->CutsShared(), first_page->is_dense,
+                          first_page->row_stride, sample_rows};
-                                  first_page->row_stride, sample_rows));
 
   // Compact the ELLPACK pages into the single sample page.
-  thrust::fill(cuctx->CTP(), page_->gidx_buffer.begin(), page_->gidx_buffer.end(), 0);
+  thrust::fill(cuctx->CTP(), page->gidx_buffer.begin(), page->gidx_buffer.end(), 0);
   for (auto& batch : batch_iterator) {
-    page_->Compact(ctx, batch.Impl(), dh::ToSpan(sample_row_index_));
+    page->Compact(ctx, batch.Impl(), dh::ToSpan(sample_row_index_));
   }
-
+  // Select the metainfo
-  return {sample_rows, page_.get(), dh::ToSpan(gpair_)};
+  dmat->Info().feature_types.SetDevice(ctx->Device());
+  auto nnz = page->NumNonMissing(ctx, dmat->Info().feature_types.ConstDeviceSpan());
+  compact_row_index_.resize(sample_rows);
+  thrust::copy_if(
+      cuctx->TP(), sample_row_index_.cbegin(), sample_row_index_.cend(), compact_row_index_.begin(),
+      [] XGBOOST_DEVICE(std::size_t idx) { return idx != ClearEmptyRows::InvalidRow(); });
+  // Create the new DMatrix
+  this->p_fmat_new_ = std::make_unique<data::IterativeDMatrix>(
+      new_page, dmat->Info().Slice(ctx, dh::ToSpan(compact_row_index_), nnz), batch_param_);
+  return {sample_rows, this->p_fmat_new_.get(), dh::ToSpan(gpair_)};
 }
 
 GradientBasedSampling::GradientBasedSampling(std::size_t n_rows, BatchParam batch_param,
@@ -254,14 +269,12 @@ GradientBasedSample GradientBasedSampling::Sample(Context const* ctx,
   size_t threshold_index = GradientBasedSampler::CalculateThresholdIndex(
       ctx, gpair, dh::ToSpan(threshold_), dh::ToSpan(grad_sum_), n_rows * subsample_);
 
-  auto page = (*dmat->GetBatches<EllpackPage>(ctx, batch_param_).begin()).Impl();
-
   // Perform Poisson sampling in place.
   thrust::transform(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair),
                     thrust::counting_iterator<size_t>(0), dh::tbegin(gpair),
                     PoissonSampling(dh::ToSpan(threshold_), threshold_index,
                                     RandomWeight(common::GlobalRandom()())));
-  return {n_rows, page, gpair};
+  return {n_rows, dmat, gpair};
 }
 
 ExternalMemoryGradientBasedSampling::ExternalMemoryGradientBasedSampling(size_t n_rows,
@@ -277,6 +290,8 @@ GradientBasedSample ExternalMemoryGradientBasedSampling::Sample(Context const* c
                                                                 common::Span<GradientPair> gpair,
                                                                 DMatrix* dmat) {
   auto cuctx = ctx->CUDACtx();
+  std::shared_ptr<EllpackPage> new_page = std::make_shared<EllpackPage>();
+  auto page = new_page->Impl();
   bst_idx_t n_rows = dmat->Info().num_row_;
   size_t threshold_index = GradientBasedSampler::CalculateThresholdIndex(
       ctx, gpair, dh::ToSpan(threshold_), dh::ToSpan(grad_sum_), n_rows * subsample_);
@@ -293,24 +308,33 @@ GradientBasedSample ExternalMemoryGradientBasedSampling::Sample(Context const* c
   thrust::copy_if(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair), gpair_.begin(), IsNonZero());
   // Index the sample rows.
   thrust::transform(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair), sample_row_index_.begin(),
-                    IsNonZero());
+                    IsNonZero{});
   thrust::exclusive_scan(cuctx->CTP(), sample_row_index_.begin(), sample_row_index_.end(),
                          sample_row_index_.begin());
   thrust::transform(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair), sample_row_index_.begin(),
-                    sample_row_index_.begin(), ClearEmptyRows());
+                    sample_row_index_.begin(), ClearEmptyRows{});
   auto batch_iterator = dmat->GetBatches<EllpackPage>(ctx, batch_param_);
   auto first_page = (*batch_iterator.begin()).Impl();
   // Create a new ELLPACK page with empty rows.
-  page_.reset();  // Release the device memory first before reallocating
-  page_.reset(new EllpackPageImpl{ctx, first_page->CutsShared(), dmat->IsDense(),
-                                  first_page->row_stride, sample_rows});
-  // Compact the ELLPACK pages into the single sample page.
-  thrust::fill(cuctx->CTP(), page_->gidx_buffer.begin(), page_->gidx_buffer.end(), 0);
-  for (auto& batch : batch_iterator) {
-    page_->Compact(ctx, batch.Impl(), dh::ToSpan(sample_row_index_));
-  }
 
-  return {sample_rows, page_.get(), dh::ToSpan(gpair_)};
+  *page = EllpackPageImpl{ctx, first_page->CutsShared(), dmat->IsDense(), first_page->row_stride,
+                          sample_rows};
+  // Compact the ELLPACK pages into the single sample page.
+  thrust::fill(cuctx->CTP(), page->gidx_buffer.begin(), page->gidx_buffer.end(), 0);
+  for (auto& batch : batch_iterator) {
+    page->Compact(ctx, batch.Impl(), dh::ToSpan(sample_row_index_));
+  }
+  // Select the metainfo
+  dmat->Info().feature_types.SetDevice(ctx->Device());
+  auto nnz = page->NumNonMissing(ctx, dmat->Info().feature_types.ConstDeviceSpan());
+  compact_row_index_.resize(sample_rows);
+  thrust::copy_if(
+      cuctx->TP(), sample_row_index_.cbegin(), sample_row_index_.cend(), compact_row_index_.begin(),
+      [] XGBOOST_DEVICE(std::size_t idx) { return idx != ClearEmptyRows::InvalidRow(); });
+  // Create the new DMatrix
+  this->p_fmat_new_ = std::make_unique<data::IterativeDMatrix>(
+      new_page, dmat->Info().Slice(ctx, dh::ToSpan(compact_row_index_), nnz), batch_param_);
+  return {sample_rows, this->p_fmat_new_.get(), dh::ToSpan(gpair_)};
 }
 
 GradientBasedSampler::GradientBasedSampler(Context const* /*ctx*/, size_t n_rows,
@@ -378,5 +402,4 @@ size_t GradientBasedSampler::CalculateThresholdIndex(Context const* ctx,
       thrust::min_element(cuctx->CTP(), dh::tbegin(grad_sum), dh::tend(grad_sum));
   return thrust::distance(dh::tbegin(grad_sum), min) + 1;
 }
-};  // namespace tree
+};  // namespace xgboost::tree
-};  // namespace xgboost

src/tree/gpu_hist/gradient_based_sampler.cuh

@@ -5,7 +5,7 @@
 #include <cstddef>  // for size_t
 
 #include "../../common/device_vector.cuh"  // for device_vector, caching_device_vector
-#include "../../data/ellpack_page.cuh"     // for EllpackPageImpl
+#include "../../common/timer.h"            // for Monitor
 #include "xgboost/base.h"                  // for GradientPair
 #include "xgboost/data.h"                  // for BatchParam
 #include "xgboost/span.h"                  // for Span
@@ -13,11 +13,11 @@
 namespace xgboost::tree {
 struct GradientBasedSample {
   /*!\brief Number of sampled rows. */
-  std::size_t sample_rows;
+  bst_idx_t sample_rows;
   /*!\brief Sampled rows in ELLPACK format. */
-  EllpackPageImpl const* page;
+  DMatrix* p_fmat;
   /*!\brief Gradient pairs for the sampled rows. */
-  common::Span<GradientPair> gpair;
+  common::Span<GradientPair const> gpair;
 };
 
 class SamplingStrategy {
@@ -48,7 +48,7 @@ class ExternalMemoryNoSampling : public SamplingStrategy {
 
  private:
   BatchParam batch_param_;
-  std::unique_ptr<EllpackPageImpl> page_{nullptr};
+  std::unique_ptr<DMatrix> p_fmat_new_{nullptr};
   bool page_concatenated_{false};
 };
 
@@ -74,9 +74,10 @@ class ExternalMemoryUniformSampling : public SamplingStrategy {
  private:
   BatchParam batch_param_;
   float subsample_;
-  std::unique_ptr<EllpackPageImpl> page_;
+  std::unique_ptr<DMatrix> p_fmat_new_{nullptr};
   dh::device_vector<GradientPair> gpair_{};
-  dh::caching_device_vector<size_t> sample_row_index_;
+  dh::caching_device_vector<bst_idx_t> sample_row_index_;
+  dh::device_vector<bst_idx_t> compact_row_index_;
 };
 
 /*! \brief Gradient-based sampling in in-memory mode.. */
@@ -105,9 +106,10 @@ class ExternalMemoryGradientBasedSampling : public SamplingStrategy {
   float subsample_;
   dh::device_vector<float> threshold_;
   dh::device_vector<float> grad_sum_;
-  std::unique_ptr<EllpackPageImpl> page_;
+  std::unique_ptr<DMatrix> p_fmat_new_{nullptr};
   dh::device_vector<GradientPair> gpair_;
-  dh::device_vector<size_t> sample_row_index_;
+  dh::device_vector<bst_idx_t> sample_row_index_;
+  dh::device_vector<bst_idx_t> compact_row_index_;
 };
 
 /*! \brief Draw a sample of rows from a DMatrix.

src/tree/updater_gpu_common.cuh

@@ -119,9 +119,9 @@ struct DeviceSplitCandidate {
 };
 
 namespace cuda_impl {
-inline BatchParam HistBatch(TrainParam const& param, bool prefetch_copy = true) {
+inline BatchParam HistBatch(TrainParam const& param) {
   auto p = BatchParam{param.max_bin, TrainParam::DftSparseThreshold()};
-  p.prefetch_copy = prefetch_copy;
+  p.prefetch_copy = true;
   p.n_prefetch_batches = 1;
   return p;
 }
@@ -134,6 +134,14 @@ inline BatchParam ApproxBatch(TrainParam const& p, common::Span<float const> hes
                               ObjInfo const& task) {
   return BatchParam{p.max_bin, hess, !task.const_hess};
 }
+
+// Empty parameter to prevent regen, only used to control external memory prefetching.
+inline BatchParam StaticBatch(bool prefetch_copy) {
+  BatchParam p;
+  p.prefetch_copy = prefetch_copy;
+  p.n_prefetch_batches = 1;
+  return p;
+}
 }  // namespace cuda_impl
 
 template <typename T>

src/tree/updater_gpu_hist.cu

@@ -52,6 +52,11 @@ DMLC_REGISTRY_FILE_TAG(updater_gpu_hist);
 using cuda_impl::ApproxBatch;
 using cuda_impl::HistBatch;
 
+// Both the approx and hist initializes the DMatrix before creating the actual
+// implementation (InitDataOnce). Therefore, the `GPUHistMakerDevice` can use an empty
+// parameter to avoid any regen.
+using cuda_impl::StaticBatch;
+
 // GPU tree updater implementation.
 struct GPUHistMakerDevice {
  private:
@@ -64,6 +69,7 @@ struct GPUHistMakerDevice {
   // node idx for each sample
   dh::device_vector<bst_node_t> positions_;
   std::unique_ptr<RowPartitioner> row_partitioner_;
+  std::shared_ptr<common::HistogramCuts const> cuts_{nullptr};
 
  public:
   // Extra data for each node that is passed to the update position function
@@ -75,13 +81,12 @@ struct GPUHistMakerDevice {
   static_assert(std::is_trivially_copyable_v<NodeSplitData>);
 
  public:
-  EllpackPageImpl const* page{nullptr};
   common::Span<FeatureType const> feature_types;
 
   DeviceHistogramStorage<> hist{};
 
   dh::device_vector<GradientPair> d_gpair;  // storage for gpair;
-  common::Span<GradientPair> gpair;
+  common::Span<GradientPair const> gpair;
 
   dh::device_vector<int> monotone_constraints;
 
@@ -99,18 +104,20 @@ struct GPUHistMakerDevice {
 
   std::unique_ptr<FeatureGroups> feature_groups;
 
-  GPUHistMakerDevice(Context const* ctx, bool is_external_memory,
+  GPUHistMakerDevice(Context const* ctx, std::shared_ptr<common::HistogramCuts const> cuts,
-                     common::Span<FeatureType const> _feature_types, bst_idx_t _n_rows,
+                     bool is_external_memory, common::Span<FeatureType const> _feature_types,
                      TrainParam _param, std::shared_ptr<common::ColumnSampler> column_sampler,
-                     uint32_t n_features, BatchParam batch_param, MetaInfo const& info)
+                     BatchParam batch_param, MetaInfo const& info)
-      : evaluator_{_param, n_features, ctx->Device()},
+      : evaluator_{_param, static_cast<bst_feature_t>(info.num_col_), ctx->Device()},
         ctx_(ctx),
         feature_types{_feature_types},
         param(std::move(_param)),
         column_sampler_(std::move(column_sampler)),
-        interaction_constraints(param, n_features),
+        interaction_constraints(param, info.num_col_),
-        info_{info} {
+        info_{info},
-    sampler = std::make_unique<GradientBasedSampler>(ctx, _n_rows, batch_param, param.subsample,
+        cuts_{std::move(cuts)} {
+    sampler =
+        std::make_unique<GradientBasedSampler>(ctx, info.num_row_, batch_param, param.subsample,
                                                param.sampling_method, is_external_memory);
     if (!param.monotone_constraints.empty()) {
       // Copy assigning an empty vector causes an exception in MSVC debug builds
@@ -123,19 +130,19 @@ struct GPUHistMakerDevice {
 
   ~GPUHistMakerDevice() = default;
 
-  void InitFeatureGroupsOnce() {
+  void InitFeatureGroupsOnce(MetaInfo const& info) {
     if (!feature_groups) {
-      CHECK(page);
+      CHECK(cuts_);
-      feature_groups = std::make_unique<FeatureGroups>(page->Cuts(), page->is_dense,
+      feature_groups = std::make_unique<FeatureGroups>(*cuts_, info.IsDense(),
                                                        dh::MaxSharedMemoryOptin(ctx_->Ordinal()),
                                                        sizeof(GradientPairPrecise));
     }
   }
 
   // Reset values for each update iteration
-  void Reset(HostDeviceVector<GradientPair>* dh_gpair, DMatrix* dmat, int64_t num_columns) {
+  [[nodiscard]] DMatrix* Reset(HostDeviceVector<GradientPair>* dh_gpair, DMatrix* p_fmat) {
-    auto const& info = dmat->Info();
+    auto const& info = p_fmat->Info();
-    this->column_sampler_->Init(ctx_, num_columns, info.feature_weights.HostVector(),
+    this->column_sampler_->Init(ctx_, p_fmat->Info().num_col_, info.feature_weights.HostVector(),
                                 param.colsample_bynode, param.colsample_bylevel,
                                 param.colsample_bytree);
     dh::safe_cuda(cudaSetDevice(ctx_->Ordinal()));
@@ -148,54 +155,54 @@ struct GPUHistMakerDevice {
     dh::safe_cuda(cudaMemcpyAsync(d_gpair.data().get(), dh_gpair->ConstDevicePointer(),
                                   dh_gpair->Size() * sizeof(GradientPair),
                                   cudaMemcpyDeviceToDevice));
-    auto sample = sampler->Sample(ctx_, dh::ToSpan(d_gpair), dmat);
+    auto sample = sampler->Sample(ctx_, dh::ToSpan(d_gpair), p_fmat);
-    page = sample.page;
+    this->gpair = sample.gpair;
-    gpair = sample.gpair;
+    p_fmat = sample.p_fmat;
+    CHECK(p_fmat->SingleColBlock());
 
-    this->evaluator_.Reset(page->Cuts(), feature_types, dmat->Info().num_col_, param,
+    this->evaluator_.Reset(*cuts_, feature_types, p_fmat->Info().num_col_, param,
-                           dmat->Info().IsColumnSplit(), ctx_->Device());
+                           p_fmat->Info().IsColumnSplit(), ctx_->Device());
 
-    quantiser = std::make_unique<GradientQuantiser>(ctx_, this->gpair, dmat->Info());
+    quantiser = std::make_unique<GradientQuantiser>(ctx_, this->gpair, p_fmat->Info());
 
     if (!row_partitioner_) {
       row_partitioner_ = std::make_unique<RowPartitioner>();
     }
-    row_partitioner_->Reset(ctx_, sample.sample_rows, page->base_rowid);
+    row_partitioner_->Reset(ctx_, sample.sample_rows, 0);
-    CHECK_EQ(page->base_rowid, 0);
 
     // Init histogram
-    hist.Init(ctx_->Device(), page->Cuts().TotalBins());
+    hist.Init(ctx_->Device(), this->cuts_->TotalBins());
     hist.Reset(ctx_);
 
-    this->InitFeatureGroupsOnce();
+    this->InitFeatureGroupsOnce(info);
 
     this->histogram_.Reset(ctx_, feature_groups->DeviceAccessor(ctx_->Device()), false);
+    return p_fmat;
   }
 
-  GPUExpandEntry EvaluateRootSplit(GradientPairInt64 root_sum) {
+  GPUExpandEntry EvaluateRootSplit(DMatrix const * p_fmat, GradientPairInt64 root_sum) {
     int nidx = RegTree::kRoot;
     GPUTrainingParam gpu_param(param);
     auto sampled_features = column_sampler_->GetFeatureSet(0);
     sampled_features->SetDevice(ctx_->Device());
     common::Span<bst_feature_t> feature_set =
         interaction_constraints.Query(sampled_features->DeviceSpan(), nidx);
-    auto matrix = page->GetDeviceAccessor(ctx_->Device());
     EvaluateSplitInputs inputs{nidx, 0, root_sum, feature_set, hist.GetNodeHistogram(nidx)};
     EvaluateSplitSharedInputs shared_inputs{
         gpu_param,
         *quantiser,
         feature_types,
-        matrix.feature_segments,
+        cuts_->cut_ptrs_.ConstDeviceSpan(),
-        matrix.gidx_fvalue_map,
+        cuts_->cut_values_.ConstDeviceSpan(),
-        matrix.min_fvalue,
+        cuts_->min_vals_.ConstDeviceSpan(),
-        matrix.is_dense && !collective::IsDistributed()
+        p_fmat->IsDense() && !collective::IsDistributed()
     };
     auto split = this->evaluator_.EvaluateSingleSplit(ctx_, inputs, shared_inputs);
     return split;
   }
 
-  void EvaluateSplits(const std::vector<GPUExpandEntry>& candidates, const RegTree& tree,
+  void EvaluateSplits(DMatrix const* p_fmat, const std::vector<GPUExpandEntry>& candidates,
-                               common::Span<GPUExpandEntry> pinned_candidates_out) {
+                      const RegTree& tree, common::Span<GPUExpandEntry> pinned_candidates_out) {
     if (candidates.empty()) {
       return;
     }
@@ -204,12 +211,11 @@ struct GPUHistMakerDevice {
     dh::TemporaryArray<DeviceSplitCandidate> splits_out(2 * candidates.size());
     std::vector<bst_node_t> nidx(2 * candidates.size());
     auto h_node_inputs = pinned2.GetSpan<EvaluateSplitInputs>(2 * candidates.size());
-    auto matrix = page->GetDeviceAccessor(ctx_->Device());
+    EvaluateSplitSharedInputs shared_inputs{
-    EvaluateSplitSharedInputs shared_inputs{GPUTrainingParam{param}, *quantiser, feature_types,
+        GPUTrainingParam{param}, *quantiser, feature_types, cuts_->cut_ptrs_.ConstDeviceSpan(),
-                                            matrix.feature_segments, matrix.gidx_fvalue_map,
+        cuts_->cut_values_.ConstDeviceSpan(), cuts_->min_vals_.ConstDeviceSpan(),
-                                            matrix.min_fvalue,
         // is_dense represents the local data
-                                            matrix.is_dense && !collective::IsDistributed()};
+        p_fmat->IsDense() && !collective::IsDistributed()};
     dh::TemporaryArray<GPUExpandEntry> entries(2 * candidates.size());
     // Store the feature set ptrs so they dont go out of scope before the kernel is called
     std::vector<std::shared_ptr<HostDeviceVector<bst_feature_t>>> feature_sets;
@@ -254,7 +260,7 @@ struct GPUHistMakerDevice {
     this->monitor.Stop(__func__);
   }
 
-  void BuildHist(int nidx) {
+  void BuildHist(EllpackPageImpl const* page, int nidx) {
     auto d_node_hist = hist.GetNodeHistogram(nidx);
     auto d_ridx = row_partitioner_->GetRows(nidx);
     this->histogram_.BuildHistogram(ctx_->CUDACtx(), page->GetDeviceAccessor(ctx_->Device()),
@@ -272,9 +278,8 @@ struct GPUHistMakerDevice {
     auto d_node_hist_histogram = hist.GetNodeHistogram(nidx_histogram);
     auto d_node_hist_subtraction = hist.GetNodeHistogram(nidx_subtraction);
 
-    dh::LaunchN(page->Cuts().TotalBins(), [=] __device__(size_t idx) {
+    dh::LaunchN(cuts_->TotalBins(), [=] __device__(size_t idx) {
-      d_node_hist_subtraction[idx] =
+      d_node_hist_subtraction[idx] = d_node_hist_parent[idx] - d_node_hist_histogram[idx];
-          d_node_hist_parent[idx] - d_node_hist_histogram[idx];
     });
     return true;
   }
@@ -366,7 +371,8 @@ struct GPUHistMakerDevice {
     }
   };
 
-  void UpdatePosition(std::vector<GPUExpandEntry> const& candidates, RegTree* p_tree) {
+  void UpdatePosition(DMatrix* p_fmat, std::vector<GPUExpandEntry> const& candidates,
+                      RegTree* p_tree) {
     if (candidates.empty()) {
       return;
     }
@@ -390,7 +396,8 @@ struct GPUHistMakerDevice {
       CHECK_EQ(split_type == FeatureType::kCategorical, e.split.is_cat);
     }
 
-    auto d_matrix = page->GetDeviceAccessor(ctx_->Device());
+    for (auto const& page : p_fmat->GetBatches<EllpackPage>(ctx_, StaticBatch(true))) {
+      auto d_matrix = page.Impl()->GetDeviceAccessor(ctx_->Device());
 
       if (info_.IsColumnSplit()) {
         UpdatePositionColumnSplit(d_matrix, split_data, nidx, left_nidx, right_nidx);
@@ -402,18 +409,20 @@ struct GPUHistMakerDevice {
           nidx, left_nidx, right_nidx, split_data,
           [=] __device__(cuda_impl::RowIndexT ridx, int /*nidx_in_batch*/,
                          const NodeSplitData& data) { return go_left(ridx, data); });
+    }
+
     monitor.Stop(__func__);
   }
 
   // After tree update is finished, update the position of all training
   // instances to their final leaf. This information is used later to update the
   // prediction cache
-  void FinalisePosition(RegTree const* p_tree, DMatrix* p_fmat, ObjInfo task,
+  void FinalisePosition(DMatrix* p_fmat, RegTree const* p_tree, ObjInfo task, bst_idx_t n_samples,
                         HostDeviceVector<bst_node_t>* p_out_position) {
     if (!p_fmat->SingleColBlock() && task.UpdateTreeLeaf()) {
       LOG(FATAL) << "Current objective function can not be used with external memory.";
     }
-    if (p_fmat->Info().num_row_ != row_partitioner_->GetRows().size()) {
+    if (p_fmat->Info().num_row_ != n_samples) {
       // Subsampling with external memory. Not supported.
       p_out_position->Resize(0);
       positions_.clear();
@@ -438,11 +447,12 @@ struct GPUHistMakerDevice {
     }
 
     dh::caching_device_vector<uint32_t> categories;
-    dh::CopyToD(p_tree->GetSplitCategories(), &categories);
+    dh::CopyTo(p_tree->GetSplitCategories(), &categories);
     auto const& cat_segments = p_tree->GetSplitCategoriesPtr();
     auto d_categories = dh::ToSpan(categories);
 
-    auto d_matrix = page->GetDeviceAccessor(ctx_->Device());
+    for (auto const& page : p_fmat->GetBatches<EllpackPage>(ctx_, StaticBatch(true))) {
+      auto d_matrix = page.Impl()->GetDeviceAccessor(ctx_->Device());
 
       std::vector<NodeSplitData> split_data(p_tree->NumNodes());
       auto const& tree = *p_tree;
@@ -469,6 +479,8 @@ struct GPUHistMakerDevice {
                                            }
                                            return encode_op(row_id, nidx);
                                          });
+    }
+
     dh::CopyTo(d_out_position, &positions_);
   }
 
@@ -508,7 +520,7 @@ struct GPUHistMakerDevice {
     auto rc = collective::GlobalSum(
         ctx_, info_,
         linalg::MakeVec(reinterpret_cast<ReduceT*>(d_node_hist),
-                        page->Cuts().TotalBins() * 2 * num_histograms, ctx_->Device()));
+                        cuts_->TotalBins() * 2 * num_histograms, ctx_->Device()));
     SafeColl(rc);
 
     monitor.Stop("AllReduce");
@@ -517,7 +529,8 @@ struct GPUHistMakerDevice {
   /**
    * \brief Build GPU local histograms for the left and right child of some parent node
    */
-  void BuildHistLeftRight(std::vector<GPUExpandEntry> const& candidates, const RegTree& tree) {
+  void BuildHistLeftRight(DMatrix* p_fmat, std::vector<GPUExpandEntry> const& candidates,
+                          const RegTree& tree) {
     if (candidates.empty()) {
       return;
     }
@@ -544,8 +557,10 @@ struct GPUHistMakerDevice {
     // Guaranteed contiguous memory
     hist.AllocateHistograms(ctx_, all_new);
 
+    for (auto const& page : p_fmat->GetBatches<EllpackPage>(ctx_, StaticBatch(true))) {
       for (auto nidx : hist_nidx) {
-      this->BuildHist(nidx);
+        this->BuildHist(page.Impl(), nidx);
+      }
     }
 
     // Reduce all in one go
@@ -560,7 +575,9 @@ struct GPUHistMakerDevice {
 
       if (!this->SubtractionTrick(parent_nidx, build_hist_nidx, subtraction_trick_nidx)) {
         // Calculate other histogram manually
-        this->BuildHist(subtraction_trick_nidx);
+        for (auto const& page : p_fmat->GetBatches<EllpackPage>(ctx_, StaticBatch(true))) {
+          this->BuildHist(page.Impl(), subtraction_trick_nidx);
+        }
         this->AllReduceHist(subtraction_trick_nidx, 1);
       }
     }
@@ -595,7 +612,7 @@ struct GPUHistMakerDevice {
       std::vector<common::CatBitField::value_type> split_cats;
 
       auto h_cats = this->evaluator_.GetHostNodeCats(candidate.nid);
-      auto n_bins_feature = page->Cuts().FeatureBins(candidate.split.findex);
+      auto n_bins_feature = cuts_->FeatureBins(candidate.split.findex);
       split_cats.resize(common::CatBitField::ComputeStorageSize(n_bins_feature), 0);
       CHECK_LE(split_cats.size(), h_cats.size());
       std::copy(h_cats.data(), h_cats.data() + split_cats.size(), split_cats.data());
@@ -618,7 +635,7 @@ struct GPUHistMakerDevice {
                                   parent.RightChild());
   }
 
-  GPUExpandEntry InitRoot(RegTree* p_tree) {
+  GPUExpandEntry InitRoot(DMatrix* p_fmat, RegTree* p_tree) {
     constexpr bst_node_t kRootNIdx = 0;
     dh::XGBCachingDeviceAllocator<char> alloc;
     auto quantiser = *this->quantiser;
@@ -635,7 +652,9 @@ struct GPUHistMakerDevice {
     collective::SafeColl(rc);
 
     hist.AllocateHistograms(ctx_, {kRootNIdx});
-    this->BuildHist(kRootNIdx);
+    for (auto const& page : p_fmat->GetBatches<EllpackPage>(ctx_, StaticBatch(true))) {
+      this->BuildHist(page.Impl(), kRootNIdx);
+    }
     this->AllReduceHist(kRootNIdx, 1);
 
     // Remember root stats
@@ -646,24 +665,25 @@ struct GPUHistMakerDevice {
     (*p_tree)[kRootNIdx].SetLeaf(param.learning_rate * weight);
 
     // Generate first split
-    auto root_entry = this->EvaluateRootSplit(root_sum_quantised);
+    auto root_entry = this->EvaluateRootSplit(p_fmat, root_sum_quantised);
     return root_entry;
   }
 
   void UpdateTree(HostDeviceVector<GradientPair>* gpair_all, DMatrix* p_fmat, ObjInfo const* task,
                   RegTree* p_tree, HostDeviceVector<bst_node_t>* p_out_position) {
     bool const is_single_block = p_fmat->SingleColBlock();
+    bst_idx_t const n_samples = p_fmat->Info().num_row_;
 
     auto& tree = *p_tree;
     // Process maximum 32 nodes at a time
     Driver<GPUExpandEntry> driver(param, 32);
 
     monitor.Start("Reset");
-    this->Reset(gpair_all, p_fmat, p_fmat->Info().num_col_);
+    p_fmat = this->Reset(gpair_all, p_fmat);
     monitor.Stop("Reset");
 
     monitor.Start("InitRoot");
-    driver.Push({this->InitRoot(p_tree)});
+    driver.Push({this->InitRoot(p_fmat, p_tree)});
     monitor.Stop("InitRoot");
 
     // The set of leaves that can be expanded asynchronously
@@ -683,11 +703,11 @@ struct GPUHistMakerDevice {
       // Update all the nodes if working with external memory, this saves us from working
       // with the finalize position call, which adds an additional iteration and requires
       // special handling for row index.
-      this->UpdatePosition(is_single_block ? filtered_expand_set : expand_set, p_tree);
+      this->UpdatePosition(p_fmat, is_single_block ? filtered_expand_set : expand_set, p_tree);
 
-      this->BuildHistLeftRight(filtered_expand_set, tree);
+      this->BuildHistLeftRight(p_fmat, filtered_expand_set, tree);
 
-      this->EvaluateSplits(filtered_expand_set, *p_tree, new_candidates);
+      this->EvaluateSplits(p_fmat, filtered_expand_set, *p_tree, new_candidates);
       dh::DefaultStream().Sync();
 
       driver.Push(new_candidates.begin(), new_candidates.end());
@@ -701,7 +721,7 @@ struct GPUHistMakerDevice {
     if (is_single_block) {
       CHECK_GE(p_tree->NumNodes(), this->row_partitioner_->GetNumNodes());
     }
-    this->FinalisePosition(p_tree, p_fmat, *task, p_out_position);
+    this->FinalisePosition(p_fmat, p_tree, *task, n_samples, p_out_position);
   }
 };
 
@@ -750,9 +770,8 @@ class GPUHistMaker : public TreeUpdater {
     monitor_.Stop(__func__);
   }
 
-  void InitDataOnce(TrainParam const* param, DMatrix* dmat) {
+  void InitDataOnce(TrainParam const* param, DMatrix* p_fmat) {
     CHECK_GE(ctx_->Ordinal(), 0) << "Must have at least one device";
-    info_ = &dmat->Info();
 
     // Synchronise the column sampling seed
     uint32_t column_sampling_seed = common::GlobalRandom()();
@@ -761,13 +780,19 @@ class GPUHistMaker : public TreeUpdater {
     SafeColl(rc);
     this->column_sampler_ = std::make_shared<common::ColumnSampler>(column_sampling_seed);
 
-    dh::safe_cuda(cudaSetDevice(ctx_->Ordinal()));
+    std::shared_ptr<common::HistogramCuts const> cuts;
-    info_->feature_types.SetDevice(ctx_->Device());
+    auto batch = HistBatch(*param);
-    maker = std::make_unique<GPUHistMakerDevice>(
+    for (auto const& page : p_fmat->GetBatches<EllpackPage>(ctx_, HistBatch(*param))) {
-        ctx_, !dmat->SingleColBlock(), info_->feature_types.ConstDeviceSpan(), info_->num_row_,
+      cuts = page.Impl()->CutsShared();
-        *param, column_sampler_, info_->num_col_, HistBatch(*param), dmat->Info());
+    }
 
-    p_last_fmat_ = dmat;
+    dh::safe_cuda(cudaSetDevice(ctx_->Ordinal()));
+    p_fmat->Info().feature_types.SetDevice(ctx_->Device());
+    maker = std::make_unique<GPUHistMakerDevice>(ctx_, cuts, !p_fmat->SingleColBlock(),
+                                                 p_fmat->Info().feature_types.ConstDeviceSpan(),
+                                                 *param, column_sampler_, batch, p_fmat->Info());
+
+    p_last_fmat_ = p_fmat;
     initialised_ = true;
   }
 
@@ -801,8 +826,6 @@ class GPUHistMaker : public TreeUpdater {
     return result;
   }
 
-  MetaInfo* info_{};  // NOLINT
-
   std::unique_ptr<GPUHistMakerDevice> maker;  // NOLINT
 
   [[nodiscard]] char const* Name() const override { return "grow_gpu_hist"; }
@@ -873,9 +896,15 @@ class GPUGlobalApproxMaker : public TreeUpdater {
 
     auto const& info = p_fmat->Info();
     info.feature_types.SetDevice(ctx_->Device());
-    maker_ = std::make_unique<GPUHistMakerDevice>(
+    std::shared_ptr<common::HistogramCuts const> cuts;
-        ctx_, !p_fmat->SingleColBlock(), info.feature_types.ConstDeviceSpan(), info.num_row_,
+    auto batch = ApproxBatch(*param, hess, *task_);
-        *param, column_sampler_, info.num_col_, ApproxBatch(*param, hess, *task_), p_fmat->Info());
+    for (auto const& page : p_fmat->GetBatches<EllpackPage>(ctx_, batch)) {
+      cuts = page.Impl()->CutsShared();
+    }
+    batch.regen = false;  // Regen only at the beginning of the iteration.
+    maker_ = std::make_unique<GPUHistMakerDevice>(ctx_, cuts, !p_fmat->SingleColBlock(),
+                                                  info.feature_types.ConstDeviceSpan(), *param,
+                                                  column_sampler_, batch, p_fmat->Info());
 
     std::size_t t_idx{0};
     for (xgboost::RegTree* tree : trees) {

tests/cpp/common/test_linalg.cu

@@ -3,6 +3,7 @@
  */
 #include <gtest/gtest.h>
 #include <thrust/equal.h>                       // for equal
+#include <thrust/iterator/constant_iterator.h>  // for make_constant_iterator
 #include <thrust/sequence.h>                    // for sequence
 
 #include "../../../src/common/cuda_context.cuh"
@@ -83,6 +84,14 @@ void TestSlice() {
     }
   });
 }
+
+void TestWriteAccess(CUDAContext const* cuctx, linalg::TensorView<double, 3> t) {
+  thrust::for_each(cuctx->CTP(), linalg::tbegin(t), linalg::tend(t),
+                   [=] XGBOOST_DEVICE(double& v) { v = 0; });
+  auto eq = thrust::equal(cuctx->CTP(), linalg::tcbegin(t), linalg::tcend(t),
+                          thrust::make_constant_iterator<double>(0.0), thrust::equal_to<>{});
+  ASSERT_TRUE(eq);
+}
 }  // anonymous namespace
 
 TEST(Linalg, GPUElementWise) { TestElementWiseKernel(); }
@@ -106,5 +115,7 @@ TEST(Linalg, GPUIter) {
 
   bool eq = thrust::equal(cuctx->CTP(), data.cbegin(), data.cend(), linalg::tcbegin(t));
   ASSERT_TRUE(eq);
+
+  TestWriteAccess(cuctx, t);
 }
 }  // namespace xgboost::linalg

tests/cpp/data/test_ellpack_page.cu

@@ -1,5 +1,5 @@
 /**
- * Copyright 2019-2023, XGBoost contributors
+ * Copyright 2019-2024, XGBoost contributors
  */
 #include <xgboost/base.h>
 
@@ -15,7 +15,6 @@
 #include "gtest/gtest.h"
 
 namespace xgboost {
-
 TEST(EllpackPage, EmptyDMatrix) {
   constexpr int kNRows = 0, kNCols = 0, kMaxBin = 256;
   constexpr float kSparsity = 0;
@@ -242,7 +241,7 @@ TEST(EllpackPage, Compact) {
 namespace {
 class EllpackPageTest : public testing::TestWithParam<float> {
  protected:
-  void Run(float sparsity) {
+  void TestFromGHistIndex(float sparsity) const {
     // Only testing with small sample size as the cuts might be different between host and
     // device.
     size_t n_samples{128}, n_features{13};
@@ -273,9 +272,25 @@ class EllpackPageTest : public testing::TestWithParam<float> {
       }
     }
   }
+
+  void TestNumNonMissing(float sparsity) const {
+    size_t n_samples{1024}, n_features{13};
+    auto ctx = MakeCUDACtx(0);
+    auto p_fmat = RandomDataGenerator{n_samples, n_features, sparsity}.GenerateDMatrix(true);
+    auto nnz = p_fmat->Info().num_nonzero_;
+    for (auto const& page : p_fmat->GetBatches<EllpackPage>(
+             &ctx, BatchParam{17, tree::TrainParam::DftSparseThreshold()})) {
+      auto ellpack_nnz =
+          page.Impl()->NumNonMissing(&ctx, p_fmat->Info().feature_types.ConstDeviceSpan());
+      ASSERT_EQ(nnz, ellpack_nnz);
+    }
+  }
 };
 }  // namespace
 
-TEST_P(EllpackPageTest, FromGHistIndex) { this->Run(GetParam()); }
+TEST_P(EllpackPageTest, FromGHistIndex) { this->TestFromGHistIndex(GetParam()); }
+
+TEST_P(EllpackPageTest, NumNonMissing) { this->TestNumNonMissing(this->GetParam()); }
+
 INSTANTIATE_TEST_SUITE_P(EllpackPage, EllpackPageTest, testing::Values(.0f, .2f, .4f, .8f));
 }  // namespace xgboost

tests/cpp/data/test_metainfo.cc

@@ -355,4 +355,70 @@ TEST(MetaInfo, HostExtend) {
 }
 
 TEST(MetaInfo, CPUStridedData) { TestMetaInfoStridedData(DeviceOrd::CPU()); }
+
+namespace {
+class TestMetaInfo : public ::testing::TestWithParam<std::tuple<bst_target_t, bool>> {
+ public:
+  void Run(Context const *ctx, bst_target_t n_targets) {
+    MetaInfo info;
+    info.num_row_ = 128;
+    info.num_col_ = 3;
+    info.feature_names.resize(info.num_col_, "a");
+    info.labels.Reshape(info.num_row_, n_targets);
+
+    HostDeviceVector<bst_idx_t> ridx(info.num_row_ / 2, 0);
+    ridx.SetDevice(ctx->Device());
+    auto h_ridx = ridx.HostSpan();
+    for (std::size_t i = 0, j = 0; i < ridx.Size(); i++, j += 2) {
+      h_ridx[i] = j;
+    }
+
+    {
+      info.weights_.Resize(info.num_row_);
+      auto h_w = info.weights_.HostSpan();
+      std::iota(h_w.begin(), h_w.end(), 0);
+    }
+
+    auto out = info.Slice(ctx, ctx->IsCPU() ? h_ridx : ridx.ConstDeviceSpan(), /*nnz=*/256);
+
+    ASSERT_EQ(info.labels.Device(), ctx->Device());
+    auto h_y = info.labels.HostView();
+    auto h_y_out = out.labels.HostView();
+    ASSERT_EQ(h_y_out.Shape(0), ridx.Size());
+    ASSERT_EQ(h_y_out.Shape(1), n_targets);
+
+    auto h_w = info.weights_.ConstHostSpan();
+    auto h_w_out = out.weights_.ConstHostSpan();
+    ASSERT_EQ(h_w_out.size(), ridx.Size());
+
+    for (std::size_t i = 0; i < ridx.Size(); ++i) {
+      for (bst_target_t t = 0; t < n_targets; ++t) {
+        ASSERT_EQ(h_y_out(i, t), h_y(h_ridx[i], t));
+      }
+      ASSERT_EQ(h_w_out[i], h_w[h_ridx[i]]);
+    }
+
+    for (auto v : info.feature_names) {
+      ASSERT_EQ(v, "a");
+    }
+  }
+};
+}  // anonymous namespace
+
+TEST_P(TestMetaInfo, Slice) {
+  Context ctx;
+  auto [n_targets, is_cuda] = this->GetParam();
+  if (is_cuda) {
+    ctx = MakeCUDACtx(0);
+  }
+  this->Run(&ctx, n_targets);
+}
+
+INSTANTIATE_TEST_SUITE_P(Cpu, TestMetaInfo,
+                         ::testing::Values(std::tuple{1u, false}, std::tuple{3u, false}));
+
+#if defined(XGBOOST_USE_CUDA)
+INSTANTIATE_TEST_SUITE_P(Gpu, TestMetaInfo,
+                         ::testing::Values(std::tuple{1u, true}, std::tuple{3u, true}));
+#endif  // defined(XGBOOST_USE_CUDA)
 }  // namespace xgboost

tests/cpp/data/test_metainfo.h

@@ -1,5 +1,5 @@
-/*!
+/**
- * Copyright 2021 by XGBoost Contributors
+ * Copyright 2021-2024, XGBoost Contributors
  */
 #ifndef XGBOOST_TESTS_CPP_DATA_TEST_METAINFO_H_
 #define XGBOOST_TESTS_CPP_DATA_TEST_METAINFO_H_
@@ -11,7 +11,6 @@
 #include <numeric>
 
 #include "../../../src/common/linalg_op.h"
-#include "../../../src/data/array_interface.h"
 
 namespace xgboost {
 inline void TestMetaInfoStridedData(DeviceOrd device) {

tests/cpp/tree/gpu_hist/test_gradient_based_sampler.cu

@@ -39,11 +39,11 @@ void VerifySampling(size_t page_size, float subsample, int sampling_method,
 
   if (fixed_size_sampling) {
     EXPECT_EQ(sample.sample_rows, kRows);
-    EXPECT_EQ(sample.page->n_rows, kRows);
+    EXPECT_EQ(sample.p_fmat->Info().num_row_, kRows);
     EXPECT_EQ(sample.gpair.size(), kRows);
   } else {
     EXPECT_NEAR(sample.sample_rows, sample_rows, kRows * 0.03);
-    EXPECT_NEAR(sample.page->n_rows, sample_rows, kRows * 0.03f);
+    EXPECT_NEAR(sample.p_fmat->Info().num_row_, sample_rows, kRows * 0.03f);
     EXPECT_NEAR(sample.gpair.size(), sample_rows, kRows * 0.03f);
   }
 
@@ -88,14 +88,16 @@ TEST(GradientBasedSampler, NoSamplingExternalMemory) {
 
   GradientBasedSampler sampler(&ctx, kRows, param, kSubsample, TrainParam::kUniform, true);
   auto sample = sampler.Sample(&ctx, gpair.DeviceSpan(), dmat.get());
-  auto sampled_page = sample.page;
+  auto p_fmat = sample.p_fmat;
   EXPECT_EQ(sample.sample_rows, kRows);
   EXPECT_EQ(sample.gpair.size(), gpair.Size());
   EXPECT_EQ(sample.gpair.data(), gpair.DevicePointer());
-  EXPECT_EQ(sampled_page->n_rows, kRows);
+  EXPECT_EQ(p_fmat->Info().num_row_, kRows);
 
+  ASSERT_EQ(p_fmat->NumBatches(), 1);
+  for (auto const& sampled_page : p_fmat->GetBatches<EllpackPage>(&ctx, param)) {
     std::vector<common::CompressedByteT> h_gidx_buffer;
-  auto h_accessor = sampled_page->GetHostAccessor(&ctx, &h_gidx_buffer);
+    auto h_accessor = sampled_page.Impl()->GetHostAccessor(&ctx, &h_gidx_buffer);
 
     std::size_t offset = 0;
     for (auto& batch : dmat->GetBatches<EllpackPage>(&ctx, param)) {
@@ -109,6 +111,7 @@ TEST(GradientBasedSampler, NoSamplingExternalMemory) {
       offset += num_elements;
     }
   }
+}
 
 TEST(GradientBasedSampler, UniformSampling) {
   constexpr size_t kPageSize = 0;