Possible bug in feature_parallel_tree_learner.cpp

[celestinoxp]

Description

I identified a possible bug in LightGBM's feature-parallel mode that directly affects the quality of cuts in distributed scenarios. Specifically, the SyncUpGlobalBestSplit function is receiving the same input buffer twice, instead of also using the output buffer, which prevents proper synchronization of “best splits” across machines and can result in suboptimal models

File: src/treelearner/feature_parallel_tree_learner.cpp
Line 71

SyncUpGlobalBestSplit(
    input_buffer_.data(),
    input_buffer_.data(),  // ← should be output_buffer_.data()
    &smaller_best_split,
    &larger_best_split,
    this->config_->max_cat_threshold
);

[jameslamb]

Thanks for using LightGBM and taking the time to report this.

In the future, we'd appreciate commit-anchored links instead of just a file name and line number... that file name and line number will not point to the same code in future versions of this project.

Adding one here:

LightGBM/src/treelearner/feature_parallel_tree_learner.cpp

Line 71 in ec5492f

SyncUpGlobalBestSplit(input_buffer_.data(), input_buffer_.data(), &smaller_best_split, &larger_best_split, this->config_->max_cat_threshold);

---

@shiyu1994 could you take a look at this?

If you can't tell by quick reading of the code whether this is really a bug, we can ask @celestinoxp for more details.