Darius Baruar
June 5, 2025 07:57
NVIDIA’s Cuml 25.04 introduces forest reasoning library improvements to improve tree -based model reasoning performance with new features and optimization.
NVIDIA has announced a significant update for the forest onference library (FIL) as part of the Cuml 25.04 release, aiming to overlook the performance of the Tree -based model reasoning. According to NVIDIA, this improvement focuses on achieving faster and more efficient reasoning of gradient boost trees and random forests.
New functions and optimization
One of the main updates includes a re -designed C ++ implementation that supports reasoning in both GPUs and CPUs. The updated FIL boasts an optimization () function for the tuning reasoning model and introduces advanced reasoning APIs such as Predict_per_tree and application. In particular, the new version promises up to four times the amount of GPU throughput compared to the previous FIL version.
The automatic optimization function is noticeable and simplifies the micro -adjustment performance process with a built -in method that adjusts the hyper parameter for optimal performance based on the batch size. This is especially advantageous for users who want to take advantage of the function of FIL without a wide range of manual configuration.
Performance benchmark
In the performance test, Cuml 25.04 has been significantly improved than the predecessor. Over the size of various model parameters and batch sizes, the new FIL version is more than 1.16 times the intermediate speed of 75%of the scenario. Improvements were clear in scenarios that require one -sized performance and maximum throughput.
Compared to the basic execution of Scikit-Learn, the performance of the FIL was significantly excellent, and the speed of 13.9x to 882x was improved depending on the model and placement size. This improvement emphasizes the potential of FIL, which can replace more resource -intensive CPU settings with a single GPU, providing both speed and cost efficiency.
A wide range of applicability and future development
In Cuml 25.04, the diversity of the FIL enables local testing and distribution flexibility because of the ability to operate on the system without the NVIDIA GPU. This library supports both GPU and CPU environments, so it is suitable for a wide range of applications from mass deployment to hybrid deployment scenarios.
In the future, NVIDIA plans to integrate these features into future releases of the Triton reasoning server to further expand the reach and usefulness of the FIL. The user can explore this improvement by downloading the Cuml 25.04 release, and the upcoming blog post is expected to provide additional benchmarks by digging deeper technical details.
For more information on the Forest reasoning library and its functions, the stakeholder can refer to the Cuml Fil document.
Image Source: Shutter Stock