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

Disable dense optimization in hist for distributed training. (#9272)

Browse Source
This commit is contained in:
Jiaming Yuan
2023-06-10 02:31:34 +08:00
committed by GitHub
parent 8c1065f645
commit ea0deeca68
5 changed files with 44 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/tree/updater_gpu_hist.cu
View File

@@ -285,7 +285,7 @@ struct GPUHistMakerDevice {
matrix.feature_segments,
matrix.gidx_fvalue_map,
matrix.min_fvalue,
matrix.is_dense
matrix.is_dense && !collective::IsDistributed()
};
auto split = this->evaluator_.EvaluateSingleSplit(inputs, shared_inputs);
return split;
@@ -299,11 +299,11 @@ struct GPUHistMakerDevice {
std::vector<bst_node_t> nidx(2 * candidates.size());
auto h_node_inputs = pinned2.GetSpan<EvaluateSplitInputs>(2 * candidates.size());
auto matrix = page->GetDeviceAccessor(ctx_->gpu_id);
EvaluateSplitSharedInputs shared_inputs{
GPUTrainingParam{param}, *quantiser, feature_types, matrix.feature_segments,
matrix.gidx_fvalue_map, matrix.min_fvalue,
matrix.is_dense
};
EvaluateSplitSharedInputs shared_inputs{GPUTrainingParam{param}, *quantiser, feature_types,
matrix.feature_segments, matrix.gidx_fvalue_map,
matrix.min_fvalue,
// is_dense represents the local data
matrix.is_dense && !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;

2
src/tree/updater_quantile_hist.cc
View File

@@ -435,7 +435,7 @@ class HistBuilder {
{
GradientPairPrecise grad_stat;
if (p_fmat->IsDense()) {
if (p_fmat->IsDense() && !collective::IsDistributed()) {
/**
* Specialized code for dense data: For dense data (with no missing value), the sum
* of gradient histogram is equal to snode[nid]