use inplace_predict for xgboost to bypass dmatrix #48

nikromen · 2026-01-28T18:20:57Z

few adjustments like reducing size of the models by penalizing large models

This should with a bit of change in parameters prevent mainly the xgboost model to get enormously big with much improvement.

gemini-code-assist · 2026-01-28T18:21:27Z

Summary of Changes

Hello @nikromen, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request enhances the RPMeta project by optimizing XGBoost prediction performance through the use of inplace_predict, which resolves a known efficiency issue. It also introduces a new feature to penalize model size during training for both XGBoost and LightGBM, promoting the creation of more compact models. Furthermore, the PR improves the developer experience by providing an interactive shell for model inspection and refactors the core model handling logic for better maintainability.

Highlights

XGBoost Prediction Optimization: Implemented inplace_predict for XGBoost to significantly reduce prediction overhead by avoiding repeated DMatrix creation, directly addressing issue predicting on xgboost is too inefficient #45.
Model Size Management: Introduced configurable size penalization for both XGBoost and LightGBM models within the Optuna objective, allowing for the training of smaller, more efficient models.
Interactive Debugging Shell: Added a new CLI command (rpmeta shell) to launch an interactive Python shell, pre-loaded with the model and configuration, for easier debugging and exploration.
Refactored Model Handling: Centralized model loading and prediction logic within the Model base class, improving encapsulation and simplifying the Predictor interface.
XGBoost Hyperparameter Tuning Enhancements: Updated XGBoost hyperparameter search space in Optuna, transitioning from max_depth to max_leaves and max_bin for finer control over tree complexity, and adjusting ranges for other parameters.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces two main improvements: it optimizes XGBoost prediction performance by using inplace_predict to avoid DMatrix creation overhead, and it adds a mechanism to penalize large models during hyperparameter tuning to encourage smaller, more efficient models. The refactoring of the Model class to encapsulate the loaded regressor is a good architectural improvement that makes the code cleaner and more object-oriented. The changes are well-tested, including new tests to verify the correctness of the optimized prediction path. I have a few suggestions to improve robustness and fix a potential bug.

gemini-code-assist · 2026-01-28T18:22:45Z

rpmeta/model.py

+    def compute_size_penalty(self, regressor: Any, trial: Any = None) -> float:
+        if not self.config.model.lightgbm.size_penalty_enabled:
+            return 0.0
+


The method call booster.num_leaves() will likely raise an AttributeError at runtime because the lightgbm.Booster object does not have this method. To get the total number of leaves, you would need to parse the model structure from booster.dump_model().

Additionally, for consistency with the XGBoost implementation and the configuration comments which refer to nodes, you might consider calculating the penalty based on the total number of nodes instead of leaves.

Suggested change

model_dump = booster.dump_model()

n_leaves = sum(tree_info["num_leaves"] for tree_info in model_dump["tree_info"])

gemini-code-assist · 2026-01-28T18:22:45Z

rpmeta/model.py

+        raw_pred = booster.inplace_predict(data)
+        # lost the inverse transformation here because of the inplace_predict :(
+        # https://xgboost.readthedocs.io/en/stable/prediction.html#in-place-prediction
+        return np.expm1(raw_pred)


Hardcoding np.expm1 here makes the prediction logic brittle. If the transformation function (func=np.log1p in create_regressor) is ever changed, this will silently produce incorrect predictions. It's more robust to use the inverse_func stored on the TransformedTargetRegressor instance, which is available via self.regressor.

Suggested change

return np.expm1(raw_pred)

return self.regressor.inverse_func(raw_pred)

gemini-code-assist · 2026-01-28T18:22:45Z

rpmeta/trainer/base.py

+        model_size = sys.getsizeof(best_regressor) / (1024 * 1024)
+        logger.info("Model size is %.2f MB", model_size)


Using sys.getsizeof() to determine the model size is likely inaccurate. This function doesn't account for the memory used by the objects the model contains, so the logged size will be misleadingly small. A more reliable method is to measure the size of the model once it's serialized to disk, for example by summing the sizes of the files created by save_regressor.

nikromen · 2026-01-28T18:32:28Z

/gemini review

gemini-code-assist

Code Review

This pull request introduces a significant performance improvement for XGBoost predictions by using inplace_predict, bypassing the overhead of DMatrix creation. It also refactors the model loading logic for better encapsulation, making the Model class stateful. Additionally, it adds a mechanism to penalize large models during hyperparameter tuning to encourage smaller, more efficient models. The changes are well-structured and accompanied by thorough tests, including new tests to validate the optimized prediction path. A new shell command is also added for easier debugging. My review found a minor issue with a duplicated test case, which I've commented on.

gemini-code-assist · 2026-01-28T18:33:38Z

test/unit/test_predictor.py

+def test_predict_uses_model_predict(example_config):
+    category_maps = {"package_name": ["pkg1"], "feature": ["a"]}
+    mock_input = MagicMock()
+    mock_input.package_name = "pkg1"
+    df_dict = {
+        "feature": "a",
+        "package_name": "pkg1",
+        "ram": 1000,
+        "swap": 2000,
+    }
+    mock_input.to_data_frame.return_value = pd.DataFrame([df_dict])
+    mock_model = MagicMock()
+    mock_model.predict.return_value = np.array([42])
+
+    predictor = Predictor(mock_model, category_maps, example_config)
+    result = predictor.predict(mock_input, example_config.model.behavior)
+
+    assert result == 42
+    mock_model.predict.assert_called_once()


This test test_predict_uses_model_predict is a duplicate of test_predict_returns_prediction (lines 34-52). Having duplicate tests increases maintenance overhead. Please remove this redundant test.

Bypass sklearn's predict() which creates a new DMatrix on every call, causing expensive memmove operations. Use Booster.inplace_predict() directly with manual expm1() inverse transformation.

nikromen added 2 commits January 28, 2026 17:35

feat: Implement compute size penalty and interactive shell cmd for debug

e1d665f

This should with a bit of change in parameters prevent mainly the xgboost model to get enormously big with much improvement.

update example config

b464e8f

nikromen force-pushed the wip branch from c3b2a07 to 23d10e8 Compare January 28, 2026 18:21

gemini-code-assist bot reviewed Jan 28, 2026

View reviewed changes

nikromen added 2 commits January 28, 2026 19:36

fix: use XGBoost inplace_predict for faster inference

802e365

Bypass sklearn's predict() which creates a new DMatrix on every call, causing expensive memmove operations. Use Booster.inplace_predict() directly with manual expm1() inverse transformation.

apply gemini suggestions

0554176

nikromen force-pushed the wip branch from 4a3f779 to 0554176 Compare January 28, 2026 18:37

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

use inplace_predict for xgboost to bypass dmatrix #48

use inplace_predict for xgboost to bypass dmatrix #48

Uh oh!

nikromen commented Jan 28, 2026

Uh oh!

gemini-code-assist bot commented Jan 28, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Jan 28, 2026

Uh oh!

gemini-code-assist bot Jan 28, 2026

Uh oh!

gemini-code-assist bot Jan 28, 2026

Uh oh!

nikromen commented Jan 28, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Jan 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant


	model_dump = booster.dump_model()
	n_leaves = sum(tree_info["num_leaves"] for tree_info in model_dump["tree_info"])

	return np.expm1(raw_pred)
	return self.regressor.inverse_func(raw_pred)

		model_size = sys.getsizeof(best_regressor) / (1024 * 1024)
		logger.info("Model size is %.2f MB", model_size)

use inplace_predict for xgboost to bypass dmatrix #48

Are you sure you want to change the base?

use inplace_predict for xgboost to bypass dmatrix #48

Uh oh!

Conversation

nikromen commented Jan 28, 2026

Uh oh!

gemini-code-assist bot commented Jan 28, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jan 28, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 28, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 28, 2026

Choose a reason for hiding this comment

Uh oh!

nikromen commented Jan 28, 2026

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jan 28, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant