Hybrid JIT-CUDA Graph Optimization for Low-Latency Large Language Model Inference

• URL: http://arxiv.org/abs/2604.23467v1
• Date: Sat, 25 Apr 2026 23:19:53 GMT
• Title: Hybrid JIT-CUDA Graph Optimization for Low-Latency Large Language Model Inference
• Abstract summary: This paper presents a hybrid framework that combines Just-In-Time (JIT) compilation with runtime Graph execution to reduce launch overhead.<n>We evaluate the proposed approach on LLaMA-2 7B using single-GPU, batch-size-one inference across prompt lengths from 10 to 500 tokens.
• Score: 0.3580891736370874
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large Language Models (LLMs) have achieved strong performance across natural language and multimodal tasks, yet their practical deployment remains constrained by inference latency and kernel launch overhead, particularly in interactive, short-sequence settings. This paper presents a hybrid runtime framework that combines Just-In-Time (JIT) compilation with CUDA Graph execution to reduce launch overhead while preserving runtime flexibility during autoregressive decoding. The framework partitions transformer inference into static components executed via CUDA Graph replay and dynamic components handled through JIT-compiled kernels, enabling asynchronous graph capture and reuse across decoding steps. We evaluate the proposed approach on LLaMA-2 7B using single-GPU, batch-size-one inference across prompt lengths from 10 to 500 tokens. Experimental results show that the hybrid runtime reduces Time-to-First-Token (TTFT) by up to 66.0% and achieves lower P99 latency compared with TensorRT-LLM in this regime. These results indicate that hybrid JIT-CUDA Graph execution can effectively reduce inference latency and variance for short-sequence LLM workloads, making it a practical optimization strategy for latency-sensitive AI applications.

Related papers

• CTA-Pipelining: A Latency-Oriented Spatial Scaling Method for Multi-GPU Systems [2.870762512009438]
  We introduce CTA-pipelining, an execution paradigm designed to exploit shared-memory multi-GPU systems.<n>As a latency-oriented spatial scaling technique, CTA-pipelining leverages the Cooperative Thread Array level.<n>Results show CTA-pipelining reduces latency by up to 31.8% compared to micro-batching.
  arXiv  Detail & Related papers  (2026-07-08T18:54:08Z)
• DiscreteRTC: Discrete Diffusion Policies are Natural Asynchronous Executors [57.944744187489185]
  We propose DiscreteRTC, which replaces external corrections with native unmasking.<n>DiscreteRTC is simpler to implement with 0 lines of code for async inpainting, faster at inference with only 0.7x computation compared with generating actions from scratch, and better at execution with 50% higher success rate in real-world dynamic pick task compared with flow-based RTC.
  arXiv  Detail & Related papers  (2026-04-27T23:04:03Z)
• Compiling Code LLMs into Lightweight Executables [18.677096682203697]
  Ditto is a novel method for optimizing both the model size of Code LLMs and their inference programs.<n>We evaluate Ditto on three popular Code LLMs, achieving up to 10.5$times$ lower memory usage compared with their original inference pipeline.
  arXiv  Detail & Related papers  (2026-03-31T14:40:19Z)
• Canzona: A Unified, Asynchronous, and Load-Balanced Framework for Distributed Matrix-based Optimizers [36.650880799066215]
  Asynchronous approaches suffer from computational redundancy, while layer-wise partitioning fails to reconcile this conflict.<n>For Data Parallelism, we introduce an alpha-Balanced Static Partitioning strategy that respects atomicity while neutralizing the load imbalance.<n>Our approach achieves a 1.57x speedup in end-to-end time and reducing step latency by 5.8x compared to the baseline.
  arXiv  Detail & Related papers  (2026-02-04T07:38:24Z)
• Yggdrasil: Bridging Dynamic Speculation and Static Runtime for Latency-Optimal Tree-Based LLM Decoding [49.16776388429616]
  Yggdrasil is a system that enables latency-optimal speculative decoding through context-aware tree drafting and compiler-friendly execution.<n>Yggdrasil supports unmodified LLMs and achieves up to $3.98times$ speedup over state-of-the-art baselines across multiple hardware setups.
  arXiv  Detail & Related papers  (2025-12-29T20:51:38Z)
• Eliminating Multi-GPU Performance Taxes: A Systems Approach to Efficient Distributed LLMs [61.953548065938385]
  We introduce the ''Three Taxes'' (Bulk Synchronous, Inter- Kernel Data Locality, and Kernel Launch Overhead) as an analytical framework.<n>We propose moving beyond the rigid BSP model to address key inefficiencies in distributed GPU execution.<n>We observe a 10-20% speedup in end-to-end latency over BSP-based approaches.
  arXiv  Detail & Related papers  (2025-11-04T01:15:44Z)
• JEDI-linear: Fast and Efficient Graph Neural Networks for Jet Tagging on FPGAs [36.158374493924455]
  Graph Neural Networks (GNNs) have shown exceptional performance for jet tagging at the CERN High Luminosity Large Hadron Collider (HLLHC)<n>We propose JEDI-linear, a novel GNN architecture with linear computational complexity that eliminates explicit pairwise interactions.<n>This is the first interaction-based GNN to achieve less than 60ns latency and currently meets the requirements for use in the HL-LHC CMS Level-1 trigger system.
  arXiv  Detail & Related papers  (2025-08-21T11:40:49Z)
• CSGO: Generalized Optimization for Cold Start in Wireless Collaborative Edge LLM Systems [62.24576366776727]
  We propose a latency-aware scheduling framework to minimize total inference latency.<n>We show that the proposed method significantly reduces cold-start latency compared to baseline strategies.
  arXiv  Detail & Related papers  (2025-08-15T07:49:22Z)
• Latency-aware Unified Dynamic Networks for Efficient Image Recognition [72.8951331472913]
  LAUDNet is a framework to bridge the theoretical and practical efficiency gap in dynamic networks. It integrates three primary dynamic paradigms-spatially adaptive computation, dynamic layer skipping, and dynamic channel skipping. It can notably reduce the latency of models like ResNet by over 50% on platforms such as V100,3090, and TX2 GPUs.
  arXiv  Detail & Related papers  (2023-08-30T10:57:41Z)
• EdgeBERT: Sentence-Level Energy Optimizations for Latency-Aware Multi-Task NLP Inference [82.1584439276834]
  Transformer-based language models such as BERT provide significant accuracy improvement for a multitude of natural language processing (NLP) tasks. We present EdgeBERT, an in-depth algorithm- hardware co-design for latency-aware energy optimization for multi-task NLP.
  arXiv  Detail & Related papers  (2020-11-28T19:21:47Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.