Block diffusion language models keep a large key-value (KV) cache throughout generation and attend to it at every denoising step, limiting both memory capacity and generation speed. Reducing these costs requires deciding which past tokens to use for denoising the current block (selection) and which to keep in memory for future blocks (eviction). We propose MaskAhead, a training-free method that solves both tasks with a single mask-query-based ranking mechanism. Current-block masks guide selection, while probes of upcoming masked blocks guide eviction. Both rank KV entries by their estimated contribution to the attention output. Our quantized variant, Q-MaskAhead, computes selection and attention directly from low-bit KV, largely preserving the selected entries. Experiments on Fast-dLLM-v2, DreamReasoner, and LLaDA2.0-mini cover long-generation reasoning, long-prompt question answering, and needle-in-a-haystack retrieval. On long-prompt QA, MaskAhead reduces KV memory by 9.5× on average with a 1.2-point mean F1 loss relative to dense inference. Q-MaskAhead increases the reduction to 20.1× with a 2.3-point mean F1 loss. In a batch-32 systems profile, MaskAhead achieves 1.23× end-to-end and 1.68× decode-stage speedups over dense inference.
Diffusion language models decode tokens in parallel, but their bidirectional denoiser rules out the naive key--value (KV) cache behind fast autoregressive inference. Block diffusion restores caching by decoding block-by-block, and the block caches deployed on it so far are tied to attention: O(L)in memory and, if used as training-free retrofits, only an approximation of the model's computation. Both constraints can be overcome: sequence mixers that summarize finalized blocks into a reusable state support block caching, and the corresponding block-causal training objective makes the cache exact. We study this recipe at scale, pretraining three 3B block-diffusion denoisers (attention, mamba, and hybrid) on 300B tokens under one single-frontier objective and decoding all three through a single cached interface. Only the state-space cache is O(1) in sequence length: its memory and per-step latency stay constant at any context length, while an attention cache remains O(L). At 256k tokens (where attention has grown to 82GB and 29 ms/step), the Mamba cache delivers 4.3x lower latency, 11x less memory, and 2.6x higher single-stream throughput; and because that footprint is constant it scales with batch as well, reaching 14x the aggregate throughput, where attention cannot run beyond a single stream. The same linear-state bias lets the Mamba and hybrid backbones keep retrieving out to 8-16x their training length, whereas attention's retrieval collapses at 2x, at no measured quality cost.
The key-value (KV) cache is a major bottleneck in long-context inference, where memory and computation grow with sequence length. Existing KV eviction methods reduce this cost but typically degrade performance relative to full-cache inference. Our key insight is that full-cache attention is not always optimal: in long contexts, irrelevant tokens can dilute attention away from useful evidence, so selective, learnable eviction can improve generation rather than merely approximate the full cache. We introduce a global retention-based KV eviction method that learns each token's future utility under a unified memory budget. Lightweight retention gates assign utility scores to cached KV entries, and a shared final scoring projection calibrates these scores across all layers and heads. This enables a single global eviction policy in which tokens from different layers, heads, and modalities compete directly for cache capacity. We further provide theoretical analysis showing that preferentially retaining useful tokens reduces attention dilution, and we justify geometric retention as a query-agnostic proxy for future utility. Across diverse long-context language and vision-language reasoning, and multi-turn dialogue benchmarks, our method substantially reduces KV memory while matching or surpassing full-cache inference. These results suggest that learned, globally calibrated KV eviction is not only a compression technique, but also a mechanism for improving long-context reasoning.
Ngoc Bui, Hieu Trung Nguyen, Arman Cohan +1
Department of Computer Science, Yale University · The Chinese University of Hong Kong
Diffusion LLMs (dLLMs) improve GPU utilization over autoregressive decoding by generating multiple tokens per forward pass, but their KV cache still grows linearly with context, limiting throughput at long contexts. KV cache offloading to host DRAM alleviates this memory pressure, but the limited PCIe bandwidth necessitates recalling only a sparse subset of KV entries. In block dLLMs, the relevant KV entries remain consistent across denoising steps within a block, enabling high-accuracy selection by identifying the top-k entries once and reusing them throughout all denoising steps. This property appears attractive for offloading as it amortizes the selection overhead across the entire block, but it requires exact attention over the full KV cache, which is too expensive under offloading. We present HERALD, a KV offloading system for block dLLMs that resolves this through two opportunities that reduce the required selection compute by a factor of the block size and enable selection to be overlapped with denoising. Across three block dLLMs and five long context tasks, HERALD achieves near-lossless accuracy at 5-10% KV budget and up to 1.59x lower per block latency and 2.47x higher throughput over GPU-only inference, with speedups growing with context length.
Omin Kwon, Doyeon Kim, Jongseok Park +3
Seoul National University Seoul, Republic of Korea · UC Berkeley Berkeley, CA, USA