# LaCache: Exact Caching and Precision-Adaptive Inference for Diffusion Large Language Models

arXiv: https://arxiv.org/abs/2607.16339 · HTML: https://mlsanity.com/paper/2607.16339/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 26 of its 40 references that are in the map.
- First posted on arXiv 16 Jul 2026.

## Builds on (most cited first, 26 of 26)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [Evaluating Large Language Models Trained on Code](https://mlsanity.com/paper/2107.03374/) (2021)
- [Measuring Mathematical Problem Solving With the MATH Dataset](https://mlsanity.com/paper/2103.03874/) (2021)
- [Program Synthesis with Large Language Models](https://mlsanity.com/paper/2108.07732/) (2021)
- [Mixed Precision Training](https://mlsanity.com/paper/1710.03740/) (2017)
- [A Survey on Large Language Models for Code Generation](https://mlsanity.com/paper/2406.00515/) (2024)
- [Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding](https://mlsanity.com/paper/2505.22618/) (2025)
- [Mercury: Ultra-Fast Language Models Based on Diffusion](https://mlsanity.com/paper/2506.17298/) (2025)
- [A Study of BFLOAT16 for Deep Learning Training](https://mlsanity.com/paper/1905.12322/) (2019)
- [Scaling up Masked Diffusion Models on Text](https://mlsanity.com/paper/2410.18514/) (2024)
- [Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI](https://mlsanity.com/paper/2407.06886/) (2024)
- [dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Caching](https://mlsanity.com/paper/2506.06295/) (2025)
- [DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs](https://mlsanity.com/paper/2406.01721/) (2024)
- [DPad: Efficient Diffusion Language Models with Suffix Dropout](https://mlsanity.com/paper/2508.14148/) (2025)
- [Plan for Speed: Dilated Scheduling for Masked Diffusion Language Models](https://mlsanity.com/paper/2506.19037/) (2025)
- [COAT: Compressing Optimizer states and Activation for Memory-Efficient FP8 Training](https://mlsanity.com/paper/2410.19313/) (2024)
- [SonicMoE: Accelerating MoE with IO and Tile-aware Optimizations](https://mlsanity.com/paper/2512.14080/) (2025)
- [Multi-User Chat Assistant (MUCA): a Framework Using LLMs to Facilitate Group Conversations](https://mlsanity.com/paper/2401.04883/) (2024)
- [TrAME: Trajectory-Anchored Multi-View Editing for Text-Guided 3D Gaussian Splatting Manipulation](https://mlsanity.com/paper/2407.02034/) (2024)
- [Adams Bashforth Moulton Solver for Inversion and Editing in Rectified Flow](https://mlsanity.com/paper/2503.16522/) (2025)
- [QR-LoRA: Efficient and Disentangled Fine-tuning via QR Decomposition for Customized Generation](https://mlsanity.com/paper/2507.04599/) (2025)
- [TV-3DG: Mastering Text-to-3D Customized Generation with Visual Prompt](https://mlsanity.com/paper/2410.21299/) (2024)
- [DiTalker: A Unified DiT-based Framework for High-Quality and Speaking Styles Controllable Portrait Animation](https://mlsanity.com/paper/2508.06511/) (2025)
- [MOSS: Efficient and Accurate FP8 LLM Training with Microscaling and Automatic Scaling](https://mlsanity.com/paper/2511.05811/) (2025)
- [CogPortrait: Fine-Grained Eye-Region Control in Portrait Animation via Hierarchical Agent Planning](https://mlsanity.com/paper/2605.28056/) (2026)

## Abstract (arXiv)

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: (i) EmbedCache for embedding outputs, (ii) RoPECache for token-wise pre-attention states, and (iii) 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 inegrates 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.