analiza aiqbench
This research proposes a novel quantization strategy specifically for tabular foundation models, prioritizing attention mechanism optimization over standard weight or KV cache reduction common in LLMs. By aligning quantization errors between training and test rows and leveraging FP8 matrix multiplication instructions, the approach achieves up to 1.7x speedup without sacrificing accuracy on key benchmarks. This work addresses a critical efficiency gap for deploying large-scale tabular models in production environments.
Attention Quantization for Tabular Foundation Models
arXiv:2609.13031v1 Announce Type: new Abstract: With the recent rise and adoption of tabular foundation models, optimizing their inference performance becomes an emerging field for efficiency research. While the models are architecturally similar to transformer-based large language models (LLMs), the size and serving patterns differ significantly. We show that the focus should be on the attention calculation and less on weight or KV cache quantization, which are more popular in LLMs. We develop a quantization strategy for queries, keys, and values to FP8 and use explicit FP8 matrix multiplication instructions to speed up the attention calculation. We find that it is crucial to align the quantization error in the test rows with the quantization error in the training rows, as otherwise the accuracy drops drastically. Our Triton kernel achieves a speedup up to 1.7x over regular 16-bit kernels, and we show that on TabPFN-v3 and TabICLv2 there is no relevant accuracy loss across TabArena and BeyondArena.
Abstract:With the recent rise and adoption of tabular foundation models, optimizing their inference performance becomes an emerging field for efficiency research. While the models are architecturally similar to transformer-based large language models (LLMs), the size and serving patterns differ significantly. We show that the focus should be on the attention calculation and less on weight or KV cache quantization, which are more popular in LLMs. We develop a quantization strategy for queries, keys, and values to FP8 and use explicit FP8 matrix multiplication instructions to speed up the attention calculation. We find that it is crucial to align the quantization error in the test rows with the quantization error in the training rows, as otherwise the accuracy drops drastically. Our Triton kernel achieves a speedup up to 1.7x over regular 16-bit kernels, and we show that on TabPFN-v3 and TabICLv2 there is no relevant accuracy loss across TabArena and BeyondArena.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2609.13031 [cs.LG]
(or arXiv:2609.13031v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.13031 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 247 cuvinte
Subiecte conexe
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model
Meta Engineering · 3 aug. 2026
Build Your Own Transaction Foundation Model for Financial Intelligence
NVIDIA · 16 iun. 2026
[AINews] The Future of Latent Space
Latent Space · 25 sept. 2026