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

[backport] Disable dense opt for distributed training. (#9272) (#9288)

Browse Source
This commit is contained in:
Jiaming Yuan 2023-06-11 11:08:45 +08:00 committed by GitHub
parent a962611de7
commit 3218f6cd3c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
src/tree/updater_gpu_hist.cu
View File

@ -286,7 +286,7 @@ struct GPUHistMakerDevice {
matrix.feature_segments, matrix.feature_segments,
matrix.gidx_fvalue_map, matrix.gidx_fvalue_map,
matrix.min_fvalue, matrix.min_fvalue,
matrix.is_dense matrix.is_dense && !collective::IsDistributed()
}; };
auto split = this->evaluator_.EvaluateSingleSplit(inputs, shared_inputs); auto split = this->evaluator_.EvaluateSingleSplit(inputs, shared_inputs);
return split; return split;
@ -300,11 +300,11 @@ struct GPUHistMakerDevice {
std::vector<bst_node_t> nidx(2 * candidates.size()); std::vector<bst_node_t> nidx(2 * candidates.size());
auto h_node_inputs = pinned2.GetSpan<EvaluateSplitInputs>(2 * candidates.size()); auto h_node_inputs = pinned2.GetSpan<EvaluateSplitInputs>(2 * candidates.size());
auto matrix = page->GetDeviceAccessor(ctx_->gpu_id); auto matrix = page->GetDeviceAccessor(ctx_->gpu_id);
EvaluateSplitSharedInputs shared_inputs{ EvaluateSplitSharedInputs shared_inputs{GPUTrainingParam{param}, *quantiser, feature_types,
GPUTrainingParam{param}, *quantiser, feature_types, matrix.feature_segments, matrix.feature_segments, matrix.gidx_fvalue_map,
matrix.gidx_fvalue_map, matrix.min_fvalue, matrix.min_fvalue,
matrix.is_dense // is_dense represents the local data
}; matrix.is_dense && !collective::IsDistributed()};
dh::TemporaryArray<GPUExpandEntry> entries(2 * candidates.size()); dh::TemporaryArray<GPUExpandEntry> entries(2 * candidates.size());
// Store the feature set ptrs so they dont go out of scope before the kernel is called // 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; std::vector<std::shared_ptr<HostDeviceVector<bst_feature_t>>> feature_sets;

2
src/tree/updater_quantile_hist.cc
View File

@ -78,7 +78,7 @@ CPUExpandEntry QuantileHistMaker::Builder::InitRoot(
{ {
GradientPairPrecise grad_stat; 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 * Specialized code for dense data: For dense data (with no missing value), the sum
* of gradient histogram is equal to snode[nid] * of gradient histogram is equal to snode[nid]