# Pyre: beating nvforest in forest inference **URL:** \ **Category:** Community Showcase\ **Tags:** gpu, mojo-1\_0\ **Created:** [September 27, 2026, 8:52am UTC](https://forum.modular.com/t/pyre-beating-nvforest-in-forest-inference/3530 "2026-09-27T08:52:22Z")\ **Posts on this page:** 1\ **Page:** 1
**Author:** ![StrGenIx](https://sea1.discourse-cdn.com/flex001/user_avatar/forum.modular.com/strgenix/32/1289_2.png) [@StrGenIx](https://forum.modular.com/u/StrGenIx)\ **Post date:** [September 27, 2026, 8:52am UTC](https://forum.modular.com/t/pyre-beating-nvforest-in-forest-inference/3530/1 "2026-09-27T08:52:22Z")
Hi all, Last week I open-sourced Balsa: [Balsa, mojo-native Treelite reader / writer](https://forum.modular.com/t/balsa-mojo-native-treelite-reader-writer/3518) but of course a Treelite reader/writer is a building block for something bigger. That bigger project has been published in open source yesterday. It’s called Pyre [GitHub - idlab-discover/pyre: Native Mojo CPU and GPU tree-ensemble inference with Balsa models and NuMojo arrays. · GitHub](https://github.com/idlab-discover/pyre) and it’s an inference library for tree ensembles. The main objective was parity or outperformance against NVForest [GitHub - rapidsai/nvforest: Highly optimized and lightweight library for decision tree inference on NVIDIA GPUs and CPUs · GitHub](https://github.com/rapidsai/nvforest) which is a member of the Nvidia Rapids ecosystem. That objective has been reached! Now that Mojo has entered 1.x.x territory, I feel more comfortable building on other Mojo projects so this project has 2 dependencies, Balsa & NuMojo. Support has been validated for models exported by scikit-learn, xgboost, catboost and lightgbm. Those libraries have their own inference back-ends and when comparing those against Pyre, it comes out on-top very often. There is still some work to do towards saturating the CPU and GPU pipelines but that will be for 0.2.0.