Set device in device dmatrix. (#5596)

include/xgboost/data.h

@@ -102,7 +102,7 @@ class MetaInfo {
   /*!
    * \brief Validate all metainfo.
    */
-  void Validate() const;
+  void Validate(int32_t device) const;
 
   MetaInfo Slice(common::Span<int32_t const> ridxs) const;
   /*!

src/data/data.cc

@@ -338,7 +338,7 @@ void MetaInfo::SetInfo(const char* key, const void* dptr, DataType dtype, size_t
   }
 }
 
-void MetaInfo::Validate() const {
+void MetaInfo::Validate(int32_t device) const {
   if (group_ptr_.size() != 0 && weights_.Size() != 0) {
     CHECK_EQ(group_ptr_.size(), weights_.Size() + 1)
         << "Size of weights must equal to number of groups when ranking "
@@ -350,30 +350,44 @@ void MetaInfo::Validate() const {
         << "Invalid group structure.  Number of rows obtained from groups "
            "doesn't equal to actual number of rows given by data.";
   }
+  auto check_device = [device](HostDeviceVector<float> const &v) {
+    CHECK(v.DeviceIdx() == GenericParameter::kCpuId ||
+          device  == GenericParameter::kCpuId ||
+          v.DeviceIdx() == device)
+        << "Data is resided on a different device than `gpu_id`. "
+        << "Device that data is on: " << v.DeviceIdx() << ", "
+        << "`gpu_id` for XGBoost: " << device;
+  };
+
   if (weights_.Size() != 0) {
     CHECK_EQ(weights_.Size(), num_row_)
         << "Size of weights must equal to number of rows.";
+    check_device(weights_);
     return;
   }
   if (labels_.Size() != 0) {
     CHECK_EQ(labels_.Size(), num_row_)
         << "Size of labels must equal to number of rows.";
+    check_device(labels_);
     return;
   }
   if (labels_lower_bound_.Size() != 0) {
     CHECK_EQ(labels_lower_bound_.Size(), num_row_)
         << "Size of label_lower_bound must equal to number of rows.";
+    check_device(labels_lower_bound_);
     return;
   }
   if (labels_upper_bound_.Size() != 0) {
     CHECK_EQ(labels_upper_bound_.Size(), num_row_)
         << "Size of label_upper_bound must equal to number of rows.";
+    check_device(labels_upper_bound_);
     return;
   }
   CHECK_LE(num_nonzero_, num_col_ * num_row_);
   if (base_margin_.Size() != 0) {
     CHECK_EQ(base_margin_.Size() % num_row_, 0)
         << "Size of base margin must be a multiple of number of rows.";
+    check_device(base_margin_);
   }
 }
 

src/data/device_dmatrix.cu

@@ -201,6 +201,7 @@ template <typename AdapterT>
 DeviceDMatrix::DeviceDMatrix(AdapterT* adapter, float missing, int nthread, int max_bin) {
   common::HistogramCuts cuts =
       common::AdapterDeviceSketch(adapter, max_bin, missing);
+  dh::safe_cuda(cudaSetDevice(adapter->DeviceIdx()));
   auto& batch = adapter->Value();
   // Work out how many valid entries we have in each row
   dh::caching_device_vector<size_t> row_counts(adapter->NumRows() + 1, 0);

src/data/simple_dmatrix.cu

@@ -99,6 +99,7 @@ void CopyDataRowMajor(AdapterT* adapter, common::Span<Entry> data,
 // be supported in future. Does not currently support inferring row/column size
 template <typename AdapterT>
 SimpleDMatrix::SimpleDMatrix(AdapterT* adapter, float missing, int nthread) {
+  dh::safe_cuda(cudaSetDevice(adapter->DeviceIdx()));
   CHECK(adapter->NumRows() != kAdapterUnknownSize);
   CHECK(adapter->NumColumns() != kAdapterUnknownSize);
 

src/learner.cc

@@ -1052,7 +1052,7 @@ class LearnerImpl : public LearnerIO {
 
   void ValidateDMatrix(DMatrix* p_fmat) const {
     MetaInfo const& info = p_fmat->Info();
-    info.Validate();
+    info.Validate(generic_parameters_.gpu_id);
 
     auto const row_based_split = [this]() {
       return tparam_.dsplit == DataSplitMode::kRow ||

tests/cpp/data/test_metainfo.cc

@@ -149,9 +149,17 @@ TEST(MetaInfo, Validate) {
   info.num_col_ = 3;
   std::vector<xgboost::bst_group_t> groups (11);
   info.SetInfo("group", groups.data(), xgboost::DataType::kUInt32, 11);
-  EXPECT_THROW(info.Validate(), dmlc::Error);
+  EXPECT_THROW(info.Validate(0), dmlc::Error);
 
   std::vector<float> labels(info.num_row_ + 1);
   info.SetInfo("label", labels.data(), xgboost::DataType::kFloat32, info.num_row_ + 1);
-  EXPECT_THROW(info.Validate(), dmlc::Error);
+  EXPECT_THROW(info.Validate(0), dmlc::Error);
+
+#if defined(XGBOOST_USE_CUDA)
+  info.group_ptr_.clear();
+  labels.resize(info.num_row_);
+  info.SetInfo("label", labels.data(), xgboost::DataType::kFloat32, info.num_row_);
+  info.labels_.SetDevice(0);
+  EXPECT_THROW(info.Validate(1), dmlc::Error);
+#endif  // defined(XGBOOST_USE_CUDA)
 }

tests/python-gpu/test_from_cupy.py

@@ -136,3 +136,14 @@ def test_dlpack_device_dmat(self):
         n = 100
         X = cp.random.random((n, 2))
         xgb.DeviceQuantileDMatrix(X.toDlpack())
+
+    @pytest.mark.skipif(**tm.no_cupy())
+    @pytest.mark.mgpu
+    def test_specified_device(self):
+        import cupy as cp
+        cp.cuda.runtime.setDevice(0)
+        dtrain = dmatrix_from_cupy(
+            np.float32, xgb.DeviceQuantileDMatrix, np.nan)
+        with pytest.raises(xgb.core.XGBoostError):
+            xgb.train({'tree_method': 'gpu_hist', 'gpu_id': 1},
+                      dtrain, num_boost_round=10)

tests/python-gpu/test_gpu_prediction.py

@@ -121,6 +121,7 @@ def test_sklearn(self):
     @pytest.mark.skipif(**tm.no_cupy())
     def test_inplace_predict_cupy(self):
         import cupy as cp
+        cp.cuda.runtime.setDevice(0)
         rows = 1000
         cols = 10
         cp_rng = cp.random.RandomState(1994)