# Taming the Memory Footprint Crisis: System Design for Production Diffusion LLM Serving

arXiv: https://arxiv.org/abs/2512.17077 · HTML: https://mlsanity.com/paper/2512.17077/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- Placed in Serving systems and request scheduling, one of 12 directions in Efficient LLM inference and serving.
- Builds on 8 of its 20 references that are in the map.
- First posted on arXiv 18 Dec 2025.

## Placed in

- Efficient LLM inference and serving / [Serving systems and request scheduling](https://mlsanity.com/inference/serving-systems-and-request-scheduling/)

## Builds on (most cited first, 8 of 8)

- [FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning](https://mlsanity.com/paper/2307.08691/) (2023)
- [FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU](https://mlsanity.com/paper/2303.06865/) (2023)
- [Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding](https://mlsanity.com/paper/2505.22618/) (2025)
- [Accelerating Diffusion LLMs via Adaptive Parallel Decoding](https://mlsanity.com/paper/2506.00413/) (2025)
- [Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Eviction](https://mlsanity.com/paper/2508.02558/) (2025)
- [AdaBlock-dLLM: Semantic-Aware Diffusion LLM Inference via Adaptive Block Size](https://mlsanity.com/paper/2509.26432/) (2025)
- [SparseD: Sparse Attention for Diffusion Language Models](https://mlsanity.com/paper/2509.24014/) (2025)
- [Mask Tokens as Prophet: Fine-Grained Cache Eviction for Efficient dLLM Inference](https://mlsanity.com/paper/2510.09309/) (2025)

## Built on by (most cited first, 5 of 5)

- [TIDE: Efficient and Lossless MoE Diffusion LLM Inference with I/O-aware Expert Offload](https://mlsanity.com/paper/2605.20179/) (2026)
- [Sangam: Efficiently Serving Diffusion LLMs with the AR Stack](https://mlsanity.com/paper/2607.04206/) (2026)
- [BlockServe: Block-Grained Continuous Batching for High-Throughput Diffusion LLM Serving](https://mlsanity.com/paper/2607.08930/) (2026)
- [Serving Masked Diffusion LLMs: Characterization and Design Principles from Real Hardware](https://mlsanity.com/paper/2608.23807/) (2026)
- [Affix Cache for Diffusion Large Language Models](https://mlsanity.com/paper/2608.26140/) (2026)

## Abstract (arXiv)

Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to Autoregressive Models (ARMs), utilizing parallel decoding to overcome sequential bottlenecks. However, existing research focuses primarily on kernel-level optimizations, lacking a holistic serving framework that addresses the unique memory dynamics of diffusion processes in production. We identify a critical "memory footprint crisis" specific to dLLMs, driven by monolithic logit tensors and the severe resource oscillation between compute-bound "Refresh" phases and bandwidth-bound "Reuse" phases. To bridge this gap, we present dLLM-Serve, an efficient dLLM serving system that co-optimizes memory footprint, computational scheduling, and generation quality. dLLM-Serve introduces Logit-Aware Activation Budgeting to decompose transient tensor peaks, a Phase-Multiplexed Scheduler to interleave heterogeneous request phases, and Head-Centric Sparse Attention to decouple logical sparsity from physical storage. We evaluate dLLM-Serve on diverse workloads (LiveBench, Burst, OSC) and GPUs (RTX 4090, L40S). Relative to the state-of-the-art baseline, dLLM-Serve improves throughput by 1.61×-1.81× on the consumer-grade RTX 4090 and 1.60×-1.74× on the server-grade NVIDIA L40S, while reducing tail latency by nearly 4× under heavy contention. dLLM-Serve establishes the first blueprint for scalable dLLM inference, converting theoretical algorithmic sparsity into tangible wall-clock acceleration across heterogeneous hardware. The code is available at https://github.com/chosen-ox/dLLM-Serve.