NeFut Logo NeFut
Admin Login

[CS.AI] LaCache: Precision-Adaptive Inference for Diffusion Models

Published at: 2026-07-21 22:00 Last updated: 2026-07-22 01:01
#Machine Learning #optimization #C++

Abstract

Diffusion-based Large Language Models (DLLMs) enable parallel generation via Semi-Autoregressive (SAR) decoding in text generation. However, current methods suffer from severe operator-level redundancy: they recompute the entire sequence during denoising steps, ignoring that the prefix and masked suffix remain invariant within a block. We propose LaCache, a training-free acceleration framework that alleviates this redundancy through lossless caching and mixed precision.

Specifically, LaCache employs Lossless State Memoization (LSM) by caching three types of intermediate results:

  1. EmbedCache for embedding outputs,
  2. RoPECache for token-wise pre-attention states, and
  3. FACache for the online softmax statistics within FlashAttention.

These caches allow the model to skip redundant computation on unchanged tokens without altering the output. To further alleviate memory-bandwidth bottlenecks, LaCache integrates a per-group FP8 quantization strategy for FFN layers, tailored to step-dependent activation distributions across the diffusion process.

Experiments demonstrate that LaCache alone achieves approximately 1.3X end-to-end speedup over vanilla DLLM. When combined with existing acceleration methods, LaCache reaches up to 40.2X end-to-end speedup while maintaining comparable task accuracy.

Blogger's Review: The introduction of LaCache not only effectively addresses the redundancy in computation for current DLLMs but also achieves significant performance improvements through lossless caching and mixed precision strategies. This presents a new perspective for efficient inference in large language models, making it worthy of attention and further exploration.

Original Source: https://arxiv.org/abs/2607.16339

[h] Back to Home