Saved in:
Bibliographic Details
Main Authors: Zhang, Zixi, Mo, Zhiwen, Zhao, Yiren, Mullins, Robert
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.11808
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917270415998976
author Zhang, Zixi
Mo, Zhiwen
Zhao, Yiren
Mullins, Robert
author_facet Zhang, Zixi
Mo, Zhiwen
Zhao, Yiren
Mullins, Robert
contents Agentic LLM inference with long contexts is increasingly limited by memory bandwidth rather than compute. In this setting, SwiGLU MLP blocks, whose large weights exceed cache capacity, become a major yet under-optimized bottleneck. We propose DeepFusionKernel, a deeply fused kernel that cuts HBM traffic and boosts cache reuse, delivering up to 13.2% speedup on H100 and 9.7% on A100 over SGLang. Integrated with SGLang and paired with a kernel scheduler, DeepFusionKernel ensures consistent accelerations over generation lengths, while remaining adaptable to diverse models, inference configurations, and hardware platforms.
format Preprint
id arxiv_https___arxiv_org_abs_2602_11808
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Deep Kernel Fusion for Transformers
Zhang, Zixi
Mo, Zhiwen
Zhao, Yiren
Mullins, Robert
Machine Learning
Agentic LLM inference with long contexts is increasingly limited by memory bandwidth rather than compute. In this setting, SwiGLU MLP blocks, whose large weights exceed cache capacity, become a major yet under-optimized bottleneck. We propose DeepFusionKernel, a deeply fused kernel that cuts HBM traffic and boosts cache reuse, delivering up to 13.2% speedup on H100 and 9.7% on A100 over SGLang. Integrated with SGLang and paired with a kernel scheduler, DeepFusionKernel ensures consistent accelerations over generation lengths, while remaining adaptable to diverse models, inference configurations, and hardware platforms.
title Deep Kernel Fusion for Transformers
topic Machine Learning
url https://arxiv.org/abs/2602.11808