Breaking the Memory Barrier: Near Infinite Batch Size Scaling for Contrastive Loss

• URL: http://arxiv.org/abs/2410.17243v1
• Date: Tue, 22 Oct 2024 17:59:30 GMT
• Title: Breaking the Memory Barrier: Near Infinite Batch Size Scaling for Contrastive Loss
• Authors: Zesen Cheng, Hang Zhang, Kehan Li, Sicong Leng, Zhiqiang Hu, Fei Wu, Deli Zhao, Xin Li, Lidong Bing,
• Abstract summary: We propose a tile-based strategy that partitions the contrastive loss calculation into arbitrary small blocks. We also introduce a multi-level tiling strategy to leverage the hierarchical structure of distributed systems. Compared to SOTA memory-efficient solutions, it achieves a two-order-of-magnitude reduction in memory while maintaining comparable speed.
• Score: 59.835032408496545
• License:
• Abstract: Contrastive loss is a powerful approach for representation learning, where larger batch sizes enhance performance by providing more negative samples to better distinguish between similar and dissimilar data. However, scaling batch sizes is constrained by the quadratic growth in GPU memory consumption, primarily due to the full instantiation of the similarity matrix. To address this, we propose a tile-based computation strategy that partitions the contrastive loss calculation into arbitrary small blocks, avoiding full materialization of the similarity matrix. Furthermore, we introduce a multi-level tiling strategy to leverage the hierarchical structure of distributed systems, employing ring-based communication at the GPU level to optimize synchronization and fused kernels at the CUDA core level to reduce I/O overhead. Experimental results show that the proposed method scales batch sizes to unprecedented levels. For instance, it enables contrastive training of a CLIP-ViT-L/14 model with a batch size of 4M or 12M using 8 or 32 A800 80GB without sacrificing any accuracy. Compared to SOTA memory-efficient solutions, it achieves a two-order-of-magnitude reduction in memory while maintaining comparable speed. The code will be made publicly available.

Related papers

• 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. 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)
• Distributed bundle adjustment with block-based sparse matrix compression for super large scale datasets [0.0]
  We propose a distributed bundle adjustment (DBA) method using the exact Levenberg-Marquardt (LM) algorithm for super large-scale datasets. For the first time, we conducted parallel bundle adjustment using LM algorithm on a real datasets with 1.18 million images and a synthetic dataset with 10 million images.
  arXiv  Detail & Related papers  (2023-07-17T10:43:54Z)
• SqueezeLLM: Dense-and-Sparse Quantization [80.32162537942138]
  Main bottleneck for generative inference with LLMs is memory bandwidth, rather than compute, for single batch inference. We introduce SqueezeLLM, a post-training quantization framework that enables lossless compression to ultra-low precisions of up to 3-bit. Our framework incorporates two novel ideas: (i) sensitivity-based non-uniform quantization, which searches for the optimal bit precision assignment based on second-order information; and (ii) the Dense-and-Sparse decomposition that stores outliers and sensitive weight values in an efficient sparse format.
  arXiv  Detail & Related papers  (2023-06-13T08:57:54Z)
• DisCo-CLIP: A Distributed Contrastive Loss for Memory Efficient CLIP Training [13.953918004371493]
  DisCo-CLIP is a memory-efficient CLIP training approach. DisCo-CLIP can enable contrastive training of a ViT-B/32 model with a batch size of 32K or 196K.
  arXiv  Detail & Related papers  (2023-04-17T17:58:21Z)
• 8-bit Optimizers via Block-wise Quantization [57.25800395197516]
  Statefuls maintain statistics over time, e.g., the exponentially smoothed sum (SGD with momentum) or squared sum (Adam) of past values. This state can be used to accelerate optimization compared to plain gradient descent but uses memory that might otherwise be allocated to model parameters. In this paper, we develop first gradients that use 8-bit statistics while maintaining the performance levels of using 32-bit gradient states.
  arXiv  Detail & Related papers  (2021-10-06T15:43:20Z)
• Recall@k Surrogate Loss with Large Batches and Similarity Mixup [62.67458021725227]
  Direct optimization, by gradient descent, of an evaluation metric is not possible when it is non-differentiable. In this work, a differentiable surrogate loss for the recall is proposed. The proposed method achieves state-of-the-art results in several image retrieval benchmarks.
  arXiv  Detail & Related papers  (2021-08-25T11:09:11Z)
• Memory and Computation-Efficient Kernel SVM via Binary Embedding and Ternary Model Coefficients [18.52747917850984]
  Kernel approximation is widely used to scale up kernel SVM training and prediction. Memory and computation costs of kernel approximation models are still too high if we want to deploy them on memory-limited devices. We propose a novel memory and computation-efficient kernel SVM model by using both binary embedding and binary model coefficients.
  arXiv  Detail & Related papers  (2020-10-06T09:41:54Z)
• 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.