Flide speeds up decision model reasoning with exact GPU execution
Flide: Accelerating Decision Inference with Exact State Forking
Computation and Language
Summary
Making decisions quickly and reliably in AI systems is tricky because existing methods waste time redoing calculations or give unclear answers. The authors present Flide, a new tool that organizes decision-making steps efficiently inside a graphics processor, avoiding repeated work and ensuring exact, consistent results. Flide can handle complex question patterns and remembers past work to speed up future answers. Tests show it works up to 9.5 times faster than some current tools while staying very accurate.
What this means in practice
- •For ai system builders: Accelerate decision-making components in systems using large language models with improved speed and accuracy on GPUs.
- •For real-time data analysts: Implement fast, precise reasoning engines to respond rapidly to complex categorical queries within latency-sensitive environments.
Authors
Bobo Li, Jiajia Song, Hao Fei, Wynne Hsu, Mong Li Lee
Abstract
Decision models such as Jev have emerged as a highly promising paradigm for LLM agent systems, providing lightweight and reliable categorical reasoning for rapid System One execution. However, existing generation-oriented inference engines fail to support these workloads effectively, suffering from severe prefix recomputation overheads or distorted output probabilities. To address this challenge, we propose Flide, an efficient and exact inference engine that decouples the logical tree structure of a decision request from its physical execution within a single flattened token stream. Through ForkAttention, Flide navigates request dependencies using GPU-resident indices across both standard attention and recurrent Gated DeltaNet layers, guaranteeing in exact arithmetic that every query branch evaluates identically to an isolated forward pass. By expressing branch topologies directly in GPU-resident metadata, Flide accommodates diverse request mixtures with a compact set of pre-recorded execution graphs while enabling expanding agent histories to resume from checkpoints saved in earlier passes. Extensive evaluations on public hybrid and decision models demonstrate that Flide answers queries up to 9.5 times faster than vLLM and SGLang while staying closest to a float32 reference in total variation distance. By unifying physical execution throughput with mathematical exactness, Flide establishes a principled foundation for scalable, latency-critical decision intelligence in agent systems.