QKV Projections Require a Fraction of Their Memory

• URL: http://arxiv.org/abs/2506.02939v1
• Date: Tue, 03 Jun 2025 14:37:17 GMT
• Title: QKV Projections Require a Fraction of Their Memory
• Authors: Malik Khalf, Yara Shamshoum, Nitzan Hodos, Yuval Sieradzki, Assaf Schuster,
• Abstract summary: We propose a novel tensor compression technique that reduces memory consumption of the $Q,K,V$ projections in attention layers by a factor of up to $times 512$.<n>PAMM is fully composable with efficient attention techniques such as FlashAttention, making it a practical and complementary method for memory-efficient LLM training.
• Score: 7.305065320738301
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: The Multi-Head Attention mechanism is central to LLM operation, and multiple works target its compute and memory efficiency during training. While most works focus on approximating the scaled dot product, the memory consumption of the linear projections that compute the $Q$, $K$, and $V$ tensors from the input $x$ is often overlooked. To address this, we propose Point-Approximate Matrix Multiplication (PAMM), a novel tensor compression technique that reduces memory consumption of the $Q,K,V$ projections in attention layers by a factor of up to $\times 512$, effectively erasing their memory footprint, while achieving similar or better final perplexity. PAMM is fully composable with efficient attention techniques such as FlashAttention, making it a practical and complementary method for memory-efficient LLM training.

Related papers

• Memory Is Not the Bottleneck: Cost-Efficient Continual Learning via Weight Space Consolidation [55.77835198580209]
  Continual learning (CL) has traditionally emphasized minimizing exemplar memory usage, assuming that memory is the primary bottleneck.<n>This paper re-examines CL under a more realistic setting with sufficient exemplar memory, where the system can retain a representative portion of past data.<n>We find that, under this regime, stability improves due to reduced forgetting, but plasticity diminishes as the model becomes biased toward prior tasks and struggles to adapt to new ones.
  arXiv  Detail & Related papers  (2025-02-11T05:40:52Z)
• Memory Layers at Scale [67.00854080570979]
  This work takes memory layers beyond proof-of-concept, proving their utility at contemporary scale.<n>On downstream tasks, language models augmented with our improved memory layer outperform dense models with more than twice the budget, as well as mixture-of-expert models when matched for both compute and parameters.<n>We provide a fully parallelizable memory layer implementation, demonstrating scaling laws with up to 128B memory parameters, pretrained to 1 trillion tokens, comparing to base models with up to 8B parameters.
  arXiv  Detail & Related papers  (2024-12-12T23:56:57Z)
• Cut Your Losses in Large-Vocabulary Language Models [102.6981011879656]
  We propose Cut Cross-Entropy (CCE), a method that computes the cross-entropy loss without materializing the logits for all tokens into global memory.<n>CCE reduces the memory footprint of the loss from 24 GB to 1 MB, and the total training-time memory consumption of the head from 28 GB to 1 GB.
  arXiv  Detail & Related papers  (2024-11-13T20:30:15Z)
• CompAct: Compressed Activations for Memory-Efficient LLM Training [7.837209773889032]
  CompAct is a technique that reduces peak memory utilization on GPU by 25-30% for pretraining and 50% for fine-tuning of LLMs.<n>By storing low-rank, compressed activations to be used in the backward pass we greatly reduce the required memory.<n>We expect CompAct's savings to scale even higher for larger models.
  arXiv  Detail & Related papers  (2024-10-20T10:24:38Z)
• LoRC: Low-Rank Compression for LLMs KV Cache with a Progressive Compression Strategy [59.1298692559785]
  Key-Value ( KV) cache is crucial component in serving transformer-based autoregressive large language models (LLMs) Existing approaches to mitigate this issue include: (1) efficient attention variants integrated in upcycling stages; (2) KV cache compression at test time; and (3) KV cache compression at test time. We propose a low-rank approximation of KV weight matrices, allowing plug-in integration with existing transformer-based LLMs without model retraining. Our method is designed to function without model tuning in upcycling stages or task-specific profiling in test stages.
  arXiv  Detail & Related papers  (2024-10-04T03:10:53Z)
• ThinK: Thinner Key Cache by Query-Driven Pruning [63.13363917871414]
  Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications.<n>This paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference.<n>We propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels.
  arXiv  Detail & Related papers  (2024-07-30T17:59:08Z)
• MEMO: Fine-grained Tensor Management For Ultra-long Context LLM Training [24.066283519769968]
  Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications.<n>We propose MEMO, a novel framework for fine-grained activation memory management.<n>MeMO achieves an average of 1.97x and 1.80x MFU compared to Megatron-LM and DeepSpeed.
  arXiv  Detail & Related papers  (2024-07-16T18:59:49Z)
• $\text{Memory}^3$: Language Modeling with Explicit Memory [22.572376536612015]
  We equip large language models (LLMs) with explicit memory, a memory format cheaper than model parameters and text retrieval-augmented generation (RAG) As a preliminary proof of concept, we train from scratch a 2.4B LLM, which achieves better performance than much larger LLMs and RAG models. We introduce a memory circuitry theory to support the externalization of knowledge, and present novel techniques including a memory sparsification mechanism that makes storage tractable.
  arXiv  Detail & Related papers  (2024-07-01T11:07:23Z)
• HiRE: High Recall Approximate Top-$k$ Estimation for Efficient LLM Inference [68.59839755875252]
  HiRE comprises of two novel components: (i) a compression scheme to cheaply predict top-$k$ rows/columns with high recall, followed by full computation restricted to the predicted subset, and (ii) DA-TOP-$k$: an efficient multi-device approximate top-$k$ operator. We demonstrate that on a one billion parameter model, HiRE applied to both the softmax as well as feedforward layers, achieves almost matching pretraining and downstream accuracy, and speeds up inference latency by $1.47times$ on a single TPUv5e device.
  arXiv  Detail & Related papers  (2024-02-14T18:04:36Z)
• UniPT: Universal Parallel Tuning for Transfer Learning with Efficient Parameter and Memory [69.33445217944029]
  PETL is an effective strategy for adapting pre-trained models to downstream domains. Recent PETL works focus on the more valuable memory-efficient characteristic. We propose a new memory-efficient PETL strategy, Universal Parallel Tuning (UniPT)
  arXiv  Detail & Related papers  (2023-08-28T05:38:43Z)
• Sketchy: Memory-efficient Adaptive Regularization with Frequent Directions [22.09320263962004]
  We find the spectra of the Kronecker-factored gradient covariance matrix in deep learning (DL) training tasks are concentrated on a small leading eigenspace. We describe a generic method for reducing memory and compute requirements of maintaining a matrix preconditioner. We show extensions of our work to Shampoo, resulting in a method competitive in quality with Shampoo and Adam, yet requiring only sub-linear memory for tracking second moments.
  arXiv  Detail & Related papers  (2023-02-07T21:50:06Z)
• Mesa: A Memory-saving Training Framework for Transformers [58.78933015299703]
  We present Mesa, a memory-saving training framework for Transformers. Mesa uses exact activations during forward pass while storing a low-precision version of activations to reduce memory consumption during training. Experiments on ImageNet, CIFAR-100 and ADE20K demonstrate that Mesa can reduce half of the memory footprints during training.
  arXiv  Detail & Related papers  (2021-11-22T11:23:01Z)
• Semantically Constrained Memory Allocation (SCMA) for Embedding in Efficient Recommendation Systems [27.419109620575313]
  A key challenge for deep learning models is to work with millions of categorical classes or tokens. We propose a novel formulation of memory shared embedding, where memory is shared in proportion to the overlap in semantic information. We demonstrate a significant reduction in the memory footprint while maintaining performance.
  arXiv  Detail & Related papers  (2021-02-24T19:55:49Z)
• Sub-Linear Memory: How to Make Performers SLiM [38.068090269482425]
  vanilla Transformers require $O(L2)$ in serial time and memory as functions of input length $L$. Recent works proposed various linear self-attention mechanisms, scaling only as $O(L)$ for serial computation. We observe a remarkable computational flexibility: forward and backward propagation can be performed with no approximations using sublinear memory.
  arXiv  Detail & Related papers  (2020-12-21T13:56:04Z)

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.