analiza aiqbench
Elastic Threshold Attention addresses the memory-bandwidth bottlenecks of long-context decoding by introducing an end-to-end trainable mechanism that dynamically allocates attention resources based on query representations. This approach eliminates rigid heuristics, allowing models to maintain dense-like quality while achieving significant speedups through hardware-accelerated pruning of uninformative tokens. The method effectively balances computational efficiency with reasoning fidelity, offering a practical path forward for scaling context windows beyond current limits.
Elastic Threshold Attention: Learned Contextual Sparsity for Long-Context Decoding
arXiv:2609.20888v1 Announce Type: new Abstract: Massive KV caches can cause severe memory-bandwidth bottlenecks during long-context decoding. Sparse attention methods mitigate this via selective loading, but that comes at a cost: rigid heuristics drop necessary context, leading to quality degradation. We introduce \textbf{Elastic Threshold Attention (ETA)}, an end-to-end trainable architecture that achieves hardware-accelerated decoding speed without sacrificing dense model quality. ETA predicts dynamic, contextual thresholds directly from query representations, allowing the model to allocate dense-like context to difficult retrieval or reasoning steps while pruning routine tokens. To learn this policy from scratch without representation collapse, ETA \emph{multiplicatively suppresses} sub-threshold logits toward zero during training rather than deleting them. Training against this smooth uniform attention floor provides a distributed probability reservoir that \textbf{causes localized attention sinks on initial tokens to disappear}. It also enables the model to hard-prune uninformative KV blocks at inference time and absorb incidental tokens co-admitted by coarse GPU block selection. As a result, a 1.45B pretrained ETA model rivals dense attention across language modeling, commonsense reasoning, and long-context needle retrieval at $\approx 85\%$ training sparsity and $\approx 38\%$ active decode density. At inference time, we implement a custom decode kernel in Triton that screens KV blocks in $O(1)$ time using cached geometric-probabilistic bounds, delivering up to $2.5\times$ wall-clock decode speedups over FlashAttention-2 on sequences up to 512K tokens. Finally, we introduce an offline calibration algorithm for domain-specific deployments that freezes per-head constant thresholds to eliminate predictor overhead, cutting attention compute by an additional $27\%$.
Abstract:Massive KV caches can cause severe memory-bandwidth bottlenecks during long-context decoding. Sparse attention methods mitigate this via selective loading, but that comes at a cost: rigid heuristics drop necessary context, leading to quality degradation. We introduce \textbf{Elastic Threshold Attention (ETA)}, an end-to-end trainable architecture that achieves hardware-accelerated decoding speed without sacrificing dense model quality. ETA predicts dynamic, contextual thresholds directly from query representations, allowing the model to allocate dense-like context to difficult retrieval or reasoning steps while pruning routine tokens. To learn this policy from scratch without representation collapse, ETA \emph{multiplicatively suppresses} sub-threshold logits toward zero during training rather than deleting them. Training against this smooth uniform attention floor provides a distributed probability reservoir that \textbf{causes localized attention sinks on initial tokens to disappear}. It also enables the model to hard-prune uninformative KV blocks at inference time and absorb incidental tokens co-admitted by coarse GPU block selection. As a result, a 1.45B pretrained ETA model rivals dense attention across language modeling, commonsense reasoning, and long-context needle retrieval at $\approx 85\%$ training sparsity and $\approx 38\%$ active decode density. At inference time, we implement a custom decode kernel in Triton that screens KV blocks in $O(1)$ time using cached geometric-probabilistic bounds, delivering up to $2.5\times$ wall-clock decode speedups over FlashAttention-2 on sequences up to 512K tokens. Finally, we introduce an offline calibration algorithm for domain-specific deployments that freezes per-head constant thresholds to eliminate predictor overhead, cutting attention compute by an additional $27\%$.
Machine Learning (cs.LG)
arXiv:2609.20888 [cs.LG]
(or arXiv:2609.20888v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.20888 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 335 cuvinte
Subiecte conexe
Sparrow: Sparse Rollout for Stable and Efficient Long-context RL of Large Language Models
arXiv · 9 iun. 2026
LoGo: Token-Level Dynamic Local-Global Attention
arXiv · 1 sept. 2026
EntmaxKV: Support-Aware Decoding for Entmax Attention
arXiv · 30 sept. 2026
FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving
arXiv · 21 aug. 2026