Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[EM] Support ExtMemQdm in the GPU predictor. (#10694)

Browse Source
This commit is contained in:
Jiaming Yuan
2024-08-13 12:21:11 +08:00
committed by GitHub
parent 43704549a2
commit 2ecc85ffad
6 changed files with 124 additions and 129 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
tests/cpp/predictor/test_gpu_predictor.cu
View File

@@ -147,39 +147,54 @@ TEST(GPUPredictor, EllpackTraining) {
TestTrainingPrediction(&ctx, kRows, kBins, p_full, p_ellpack);
}
TEST(GPUPredictor, ExternalMemoryTest) {
auto lparam = MakeCUDACtx(0);
namespace {
template <typename Create>
void TestDecisionStumpExternalMemory(Context const* ctx, bst_feature_t n_features,
Create create_fn) {
std::int32_t n_classes = 3;
LearnerModelParam mparam{MakeMP(n_features, .5, n_classes, ctx->Device())};
auto model = CreateTestModel(&mparam, ctx, n_classes);
std::unique_ptr<Predictor> gpu_predictor =
std::unique_ptr<Predictor>(Predictor::Create("gpu_predictor", &lparam));
std::unique_ptr<Predictor>(Predictor::Create("gpu_predictor", ctx));
gpu_predictor->Configure({});
const int n_classes = 3;
Context ctx = MakeCUDACtx(0);
LearnerModelParam mparam{MakeMP(5, .5, n_classes, ctx.Device())};
gbm::GBTreeModel model = CreateTestModel(&mparam, &ctx, n_classes);
std::vector<std::unique_ptr<DMatrix>> dmats;
dmats.push_back(CreateSparsePageDMatrix(400));
dmats.push_back(CreateSparsePageDMatrix(800));
dmats.push_back(CreateSparsePageDMatrix(8000));
for (const auto& dmat: dmats) {
dmat->Info().base_margin_ = decltype(dmat->Info().base_margin_){
{dmat->Info().num_row_, static_cast<size_t>(n_classes)}, DeviceOrd::CUDA(0)};
dmat->Info().base_margin_.Data()->Fill(0.5);
for (auto p_fmat : {create_fn(400), create_fn(800), create_fn(2048)}) {
p_fmat->Info().base_margin_ = linalg::Constant(ctx, 0.5f, p_fmat->Info().num_row_, n_classes);
PredictionCacheEntry out_predictions;
gpu_predictor->InitOutPredictions(dmat->Info(), &out_predictions.predictions, model);
gpu_predictor->PredictBatch(dmat.get(), &out_predictions, model, 0);
EXPECT_EQ(out_predictions.predictions.Size(), dmat->Info().num_row_ * n_classes);
const std::vector<float> &host_vector = out_predictions.predictions.ConstHostVector();
for (size_t i = 0; i < host_vector.size() / n_classes; i++) {
ASSERT_EQ(host_vector[i * n_classes], 2.0);
ASSERT_EQ(host_vector[i * n_classes + 1], 0.5);
ASSERT_EQ(host_vector[i * n_classes + 2], 0.5);
gpu_predictor->InitOutPredictions(p_fmat->Info(), &out_predictions.predictions, model);
gpu_predictor->PredictBatch(p_fmat.get(), &out_predictions, model, 0);
ASSERT_EQ(out_predictions.predictions.Size(), p_fmat->Info().num_row_ * n_classes);
auto const& h_predt = out_predictions.predictions.ConstHostVector();
for (size_t i = 0; i < h_predt.size() / n_classes; i++) {
ASSERT_EQ(h_predt[i * n_classes], 2.0);
ASSERT_EQ(h_predt[i * n_classes + 1], 0.5);
ASSERT_EQ(h_predt[i * n_classes + 2], 0.5);
}
}
}
} // namespace
TEST(GPUPredictor, ExternalMemory) {
auto ctx = MakeCUDACtx(0);
bst_bin_t max_bin = 128;
bst_feature_t n_features = 32;
TestDecisionStumpExternalMemory(&ctx, n_features, [&](bst_idx_t n_samples) {
return RandomDataGenerator{n_samples, n_features, 0.0f}
.Batches(4)
.Device(ctx.Device())
.Bins(max_bin)
.GenerateSparsePageDMatrix("temp", false);
});
TestDecisionStumpExternalMemory(&ctx, n_features, [&](bst_idx_t n_samples) {
return RandomDataGenerator{n_samples, n_features, 0.0f}
.Batches(4)
.Device(ctx.Device())
.Bins(max_bin)
.GenerateExtMemQuantileDMatrix("temp", false);
});
}
TEST(GPUPredictor, InplacePredictCupy) {
auto ctx = MakeCUDACtx(0);