HYBRID JIT–CUDA GRAPH OPTIMIZATION FOR LOW-LATENCY LARGE LANGUAGE MODEL INFERENCE

Abstract

Large Language Models (LLMs) deliver strong performance but suffer from high inference latency and unstable kernel-launch overhead. This the- sis introduces a Hybrid JIT–CUDA Graph Runtime that combines static CUDA Graph replay with JIT-compiled dynamic kernels to achieve deter- ministic, low-latency autoregressive decoding. A rolling graph-generation mechanism enables variable sequence lengths while preserving static-shape constraints. Evaluated on an NVIDIA H100 GPU, the system reduces Time- to-First-Token (TTFT) by 45–80%, achieves the lowest and most stable P99 latency, and provides the fastest end-to-end 500-token generation compared with HuggingFace (PyTorch Eager) and TensorRT–LLM. The results demon- strate that integrating JIT flexibility with CUDA Graph determinism offers an effective approach for high-performance LLM inference.