Deep Kernel Fusion for Transformers

2026-02-12

AI Summary

This work addresses the performance bottleneck in long-context large language model inference caused by the SwiGLU MLP module, whose large weights induce frequent high-bandwidth memory (HBM) accesses under memory-bandwidth-constrained conditions. To mitigate this issue, the authors propose DeepFusionKernel, a novel approach that integrates operator fusion, optimized cache reuse, and reduced HBM traffic to substantially lower memory access overhead. The method features adaptable fused kernels compatible with diverse models, inference configurations, and hardware platforms, coupled with a scheduling mechanism that delivers consistent acceleration independent of generation length. Experimental results demonstrate up to 13.2% and 9.7% inference speedup on H100 and A100 GPUs, respectively, significantly outperforming the SGLang baseline.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Natural Language Processing: (Large) Language Models
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

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.

Problem

• memory bandwidth bottleneck
• SwiGLU MLP
• long-context LLM inference
• cache capacity
• HBM traffic

Innovation

• Deep Kernel Fusion
• SwiGLU Optimization
• Memory Bandwidth Efficiency
• Cache Reuse
• Transformer Inference Acceleration