惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

L
LangChain Blog
B
Blog RSS Feed
阮一峰的网络日志
阮一峰的网络日志
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
H
Help Net Security
MyScale Blog
MyScale Blog
WordPress大学
WordPress大学
Microsoft Azure Blog
Microsoft Azure Blog
GbyAI
GbyAI
小众软件
小众软件
大猫的无限游戏
大猫的无限游戏
Martin Fowler
Martin Fowler
Vercel News
Vercel News
S
SegmentFault 最新的问题
M
MIT News - Artificial intelligence
Microsoft Security Blog
Microsoft Security Blog
G
Google Developers Blog
Last Week in AI
Last Week in AI
Hugging Face - Blog
Hugging Face - Blog
酷 壳 – CoolShell
酷 壳 – CoolShell
博客园 - 【当耐特】
Google DeepMind News
Google DeepMind News
Engineering at Meta
Engineering at Meta
云风的 BLOG
云风的 BLOG

cs.LG updates on arXiv.org

Memory-Guided Trust-Region Bayesian Optimization (MG-TuRBO) for High Dimensions EngageTriBoost: Predictive Modeling of User Engagement in Digital Mental Health Intervention Using Explainable Machine Learning Reservoir observer enhanced with residual calibration and attention mechanism Efficient RL Training for LLMs with Experience Replay Wireless Communication Enhanced Value Decomposition for Multi-Agent Reinforcement Learning Adversarial Sensor Errors for Safe and Robust Wind Turbine Fleet Control IKKA: Inversion Classification via Critical Anomalies for Robust Visual Servoing Adaptive Simulation Experiment for LLM Policy Optimization EvoLen: Evolution-Guided Tokenization for DNA Language Model Smartwatch-Based Sitting Time Estimation in Real-World Office Settings Structural Evaluation Metrics for SVG Generation via Leave-One-Out Analysis Loom: A Scalable Analytical Neural Computer Architecture Spectral Geometry of LoRA Adapters Encodes Training Objective and Predicts Harmful Compliance Finite-Sample Analysis of Nonlinear Independent Component Analysis:Sample Complexity and Identifiability Bounds How does Chain of Thought decompose complex tasks? Uncertainty-Aware Transformers: Conformal Prediction for Language Models Adaptive Candidate Point Thompson Sampling for High-Dimensional Bayesian Optimization Using Synthetic Data for Machine Learning-based Childhood Vaccination Prediction in Narok, Kenya Delve into the Applicability of Advanced Optimizers for Multi-Task Learning Bridging SFT and RL: Dynamic Policy Optimization for Robust Reasoning Multi-Agent Decision-Focused Learning via Value-Aware Sequential Communication Predictive Entropy Links Calibration and Paraphrase Sensitivity in Medical Vision-Language Models Efficient Hierarchical Implicit Flow Q-learning for Offline Goal-conditioned Reinforcement Learning Modality-Aware Zero-Shot Pruning and Sparse Attention for Efficient Multimodal Edge Inference The nextAI Solution to the NeurIPS 2023 LLM Efficiency Challenge Feature-Label Modal Alignment for Robust Partial Multi-Label Learning Integrated electro-optic attention nonlinearities for transformers Toward World Models for Epidemiology Tracing the Chain: Deep Learning for Stepping-Stone Intrusion Detection Batch Distillation Data for Developing Machine Learning Anomaly Detection Methods
StreamIndex: Memory-Bounded Compressed Sparse Attention v...
Jaber Jaber, · 2026-05-05 · via cs.LG updates on arXiv.org

View PDF HTML (experimental)

Abstract:DeepSeek-V3.2 and V4 introduce Compressed Sparse Attention (CSA): a lightning indexer (a learned scoring projection over compressed keys) scores them, the top-k are selected per query, and a sparse attention kernel reads only those. Public CSA implementations materialize a [B, S, H_I, T] FP32 score tensor before the top-k reduction. With H_I=64 indexer heads and the V4-Flash compression ratio m=4, that intermediate is 256 GB at sequence length S=65,536, exceeding any single-GPU high-bandwidth-memory (HBM) budget. We present StreamIndex, a Triton implementation of the CSA pipeline whose central component is a chunked partition-merge top-k driver that never materializes the full intermediate. On synthetic-but-realistic V4-shaped inputs at the indexer-step (layer) level on a single NVIDIA H200, the materialize path runs out of memory (OOMs) at S=65,536 with V4-Flash dimensions; StreamIndex runs the same indexer to S=1,048,576 with 6.21 GB peak HBM, a 32x regime extension. Set-overlap recall against the materialize ground truth is bit-exact at small S where both fit; across three 5-point design-space sweeps (chunk size, key-tile size, top-k), mean recall rounds to 1.0000 with min recall at least 0.9980 in every cell. The chunked driver composes with TileLang's pipelined attention kernel: at S=262,144 with V4-Flash dimensions, the materialize indexer paired with TileLang attention OOMs while the chunked indexer paired with the same attention runs in 1.97 s at 18.56 GB peak. Our contribution targets the indexer step; we make no claim of a faster attention kernel or of real-checkpoint end-to-end behavior. Code: this https URL.
Comments: 11 pages, 3 figures, 7 tables, 2 algorithms, 36 references. Memory-bounded indexer kernel for DeepSeek-V4 CSA via chunked partition-merge top-k. Code: this https URL
Subjects: Machine Learning (cs.LG); Performance (cs.PF)
ACM classes: C.1.2; I.2.7
Cite as: arXiv:2605.02568 [cs.LG]
  (or arXiv:2605.02568v1 [cs.LG] for this version)
  https://doi.org/10.48550/arXiv.2605.02568

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Jaber Jaber [view email]
[v1] Mon, 4 May 2026 13:19:29 UTC (66 KB)