Efficient Long-context Language Model Training by Core Attention Disaggregation

• URL: http://arxiv.org/abs/2510.18121v1
• Date: Mon, 20 Oct 2025 21:40:51 GMT
• Title: Efficient Long-context Language Model Training by Core Attention Disaggregation
• Abstract summary: We present a technique that improves long-context large language model training by decoupling the core attention computation, softmax(QKT)V, from the rest of the model.<n>We implement CAD in a system called DistCA, which uses a ping-pong execution scheme to fully overlap communication with computation and in-place execution on attention servers to reduce memory use.
• Score: 40.14172357304901
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We present core attention disaggregation (CAD), a technique that improves long-context large language model training by decoupling the core attention computation, softmax(QK^T)V, from the rest of the model and executing it on a separate pool of devices. In existing systems, core attention is colocated with other layers; at long context lengths, its quadratic compute growth compared to the near-linear growth of other components causes load imbalance and stragglers across data and pipeline parallel groups. CAD is enabled by two observations. First, core attention is stateless: it has no trainable parameters and only minimal transient data, so balancing reduces to scheduling compute-bound tasks. Second, it is composable: modern attention kernels retain high efficiency when processing fused batches of token-level shards with arbitrary lengths. CAD partitions core attention into token-level tasks and dispatches them to dedicated attention servers, which dynamically rebatch tasks to equalize compute without sacrificing kernel efficiency. We implement CAD in a system called DistCA, which uses a ping-pong execution scheme to fully overlap communication with computation and in-place execution on attention servers to reduce memory use. On 512 H200 GPUs and context lengths up to 512k tokens, DistCA improves end-to-end training throughput by up to 1.35x, eliminates data and pipeline parallel stragglers, and achieves near-perfect compute and memory balance.

Related papers

• Accelerating Unified Multimodal Models with Core-Expansion Routing and Unified Computation Scheduling [61.28599393707943]
  Unified multimodal models jointly support understanding and generation, but incur substantial redundant computation across tokens, layers, and generation timesteps.<n>We propose a task-shared core scorer and progress-conditioned generation expansions, optimized through generation decomposition and cross-task core alignment.<n> Experiments on two representative UMM architectures demonstrate consistent quality--efficiency improvements across both tasks, retaining 98.03% of dense understanding performance with a 1.93times$ end-to-end inference speedup.
  arXiv  Detail & Related papers  (2026-08-29T14:27:55Z)
• Stream-CQSA: Avoiding Out-of-Memory in Attention Computation via Flexible Workload Scheduling [0.0]
  Stream-CQSA is a memory-adaptive scheduling framework that partitions attention into subproblems that fit within arbitrary memory budgets.<n> Experiments demonstrate predictable memory scaling and show that exact attention over billion-token sequences can be executed on a single GPU via streaming.
  arXiv  Detail & Related papers  (2026-04-22T17:46:09Z)
• Out of the Memory Barrier: A Highly Memory Efficient Training System for LLMs with Million-Token Contexts [68.79341332280062]
  Training Large Language Models (LLMs) on long contexts is severely constrained by prohibitive GPU memory overhead, not training time.<n>We introduce OOMB, a highly memory-efficient training system that directly confronts this barrier.<n>Our approach employs a chunk-recurrent training framework with on-the-fly activation recomputation, which maintains a constant activation memory footprint.
  arXiv  Detail & Related papers  (2026-02-02T13:52:40Z)
• Training-free Context-adaptive Attention for Efficient Long Context Modeling [57.703159205740185]
  Training-free Context-adaptive Attention (TCA-Attention) is a training-free sparse attention mechanism that selectively attends to only the informative tokens for efficient long-context inference.<n>TCA-Attention achieves a 2.8$times$ speedup and reduces KV cache by 61% at 128K context length while maintaining performance comparable to full attention.
  arXiv  Detail & Related papers  (2025-12-10T01:54:57Z)
• Attention and Compression is all you need for Controllably Efficient Language Models [16.42720496730602]
  Compress & Attend Transformer (CAT) is a conceptually simple architecture employing dense attention and compression.<n>Cat can be trained with multiple chunk sizes at once, unlocking control of quality-compute trade-offs directly at test-time.<n>A single CAT matches dense transformer in language modeling across model scales while being 1.4-3x faster and requiring 2-9x lower total memory usage.
  arXiv  Detail & Related papers  (2025-11-07T15:13:28Z)
• DELTA: Dynamic Layer-Aware Token Attention for Efficient Long-Context Reasoning [6.468843780300177]
  We present textbfDELTA, a training-free sparse attention mechanism that achieves computational efficiency without sacrificing model accuracy.<n>Our results show that selective reuse of intermediate attention maps offers a robust path toward efficient long-context reasoning.
  arXiv  Detail & Related papers  (2025-10-10T21:37:49Z)
• HGCA: Hybrid GPU-CPU Attention for Long Context LLM Inference [8.826966369389893]
  We present HGCA, a hybrid CPU- GPU attention mechanism for large language models.<n>We show that HGCA achieves superior scalability, supports longer sequences and larger batch sizes, and outperforms existing sparse attention baselines in both performance and accuracy.<n> Experiments across diverse models and workloads show that HGCA achieves superior scalability, supports longer sequences and larger batch sizes, and outperforms existing sparse attention baselines in both performance and accuracy.
  arXiv  Detail & Related papers  (2025-07-03T20:20:33Z)
• ByteScale: Efficient Scaling of LLM Training with a 2048K Context Length on More Than 12,000 GPUs [22.542224045868117]
  We introduce ByteScale, an efficient framework for large-scale mixed training of long and short sequences.<n>ByteScale is based on Hybrid Data Parallelism (HDP), which unifies the inter- and intra-data partitioning with a dynamic mesh design.<n>Experiment results show that ByteScale outperforms the state-of-the-art training system by up to 7.89x.
  arXiv  Detail & Related papers  (2025-02-28T17:01:03Z)
• MAS-Attention: Memory-Aware Stream Processing for Attention Acceleration on Resource-Constrained Edge Devices [24.1144641404561]
  We propose a scheme for exact attention inference acceleration on memory-constrained edge accelerators.<n>We show up to 2.75x speedup and 54% reduction in energy consumption as compared to the state-of-the-art attention fusion method (FLAT) in the edge computing scenario.
  arXiv  Detail & Related papers  (2024-11-20T19:44:26Z)
• Tree Attention: Topology-aware Decoding for Long-Context Attention on GPU clusters [10.403248386029407]
  Our formulation reveals that the reduction across the sequence axis can be efficiently computed in parallel through a tree reduction.<n>Our algorithm, called Tree Attention, for parallelizing exact attention across multiple GPUs enables cross-device decoding.<n>We demonstrate that Tree Attention speeds up decoding up to 4x on Llama 3.1-8B and can be applied to a variety of hardware and networking setups.
  arXiv  Detail & Related papers  (2024-08-07T21:16:55Z)
• Efficient Dataset Distillation Using Random Feature Approximation [109.07737733329019]
  We propose a novel algorithm that uses a random feature approximation (RFA) of the Neural Network Gaussian Process (NNGP) kernel. Our algorithm provides at least a 100-fold speedup over KIP and can run on a single GPU. Our new method, termed an RFA Distillation (RFAD), performs competitively with KIP and other dataset condensation algorithms in accuracy over a range of large-scale datasets.
  arXiv  Detail & Related papers  (2022-10-21T15:56:13Z)
• Dynamic Split Computing for Efficient Deep Edge Intelligence [78.4233915447056]
  We introduce dynamic split computing, where the optimal split location is dynamically selected based on the state of the communication channel. We show that dynamic split computing achieves faster inference in edge computing environments where the data rate and server load vary over time.
  arXiv  Detail & Related papers  (2022-05-23T12:35:18Z)
• DistGNN: Scalable Distributed Training for Large-Scale Graph Neural Networks [58.48833325238537]
  Full-batch training on Graph Neural Networks (GNN) to learn the structure of large graphs is a critical problem that needs to scale to hundreds of compute nodes to be feasible. In this paper, we presentGNN that optimize the well-known Deep Graph Library (DGL) for full-batch training on CPU clusters. Our results on four common GNN benchmark datasets show up to 3.7x speed-up using a single CPU socket and up to 97x speed-up using 128 CPU sockets.
  arXiv  Detail & Related papers  (2021-04-14T08:46:35Z)
• Joint Parameter-and-Bandwidth Allocation for Improving the Efficiency of Partitioned Edge Learning [73.82875010696849]
  Machine learning algorithms are deployed at the network edge for training artificial intelligence (AI) models. This paper focuses on the novel joint design of parameter (computation load) allocation and bandwidth allocation.
  arXiv  Detail & Related papers  (2020-03-10T05:52:15Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.