First Attentions Last: Better Exploiting First Attentions for Efficient Transformer Training

• URL: http://arxiv.org/abs/2510.14614v1
• Date: Thu, 16 Oct 2025 12:21:46 GMT
• Title: First Attentions Last: Better Exploiting First Attentions for Efficient Transformer Training
• Authors: Gyudong Kim, Hyukju Na, Jin Hyeon Kim, Hyunsung Jang, Jaemin Park, Jaegi Hwang, Namkoo Ha, Seungryong Kim, Young Geun Kim,
• Abstract summary: Existing transformer designs suffer from significant communication overhead, especially in Parallelism (TP)<n>We propose FAL (First Attentions Last), an efficient transformer architecture that redirects the first MHA output to the inputs of the following layers, eliminating the per-block MHA-MLP connections.<n>FAL reduces multi-GPU training time by up to 44%, improves single-GPU throughput by up to 1.18x, and achieves better perplexity compared to the baseline GPT.
• Score: 32.292925864710156
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: As training billion-scale transformers becomes increasingly common, employing multiple distributed GPUs along with parallel training methods has become a standard practice. However, existing transformer designs suffer from significant communication overhead, especially in Tensor Parallelism (TP), where each block's MHA-MLP connection requires an all-reduce communication. Through our investigation, we show that the MHA-MLP connections can be bypassed for efficiency, while the attention output of the first layer can serve as an alternative signal for the bypassed connection. Motivated by the observations, we propose FAL (First Attentions Last), an efficient transformer architecture that redirects the first MHA output to the MLP inputs of the following layers, eliminating the per-block MHA-MLP connections. This removes the all-reduce communication and enables parallel execution of MHA and MLP on a single GPU. We also introduce FAL+, which adds the normalized first attention output to the MHA outputs of the following layers to augment the MLP input for the model quality. Our evaluation shows that FAL reduces multi-GPU training time by up to 44%, improves single-GPU throughput by up to 1.18x, and achieves better perplexity compared to the baseline GPT. FAL+ achieves even lower perplexity without increasing the training time than the baseline.

Related papers

• MLP Can Be A Good Transformer Learner [73.01739251050076]
  Self-attention mechanism is the key of the Transformer but often criticized for its computation demands. This paper introduces a novel strategy that simplifies vision transformers and reduces computational load through the selective removal of non-essential attention layers.
  arXiv  Detail & Related papers  (2024-04-08T16:40:15Z)
• MLP Fusion: Towards Efficient Fine-tuning of Dense and Mixture-of-Experts Language Models [33.86069537521178]
  Fine-tuning a pre-trained language model (PLM) emerges as the predominant strategy in many natural language processing applications.<n>General approaches (e.g. quantization and distillation) have been widely studied to reduce the compute/memory of PLM fine-tuning.<n>We propose one-shot compression techniques specifically designed for fine-tuning.
  arXiv  Detail & Related papers  (2023-07-18T03:12:51Z)
• The Lazy Neuron Phenomenon: On Emergence of Activation Sparsity in Transformers [59.87030906486969]
  This paper studies the curious phenomenon for machine learning models with Transformer architectures that their activation maps are sparse. We show that sparsity is a prevalent phenomenon that occurs for both natural language processing and vision tasks. We discuss how sparsity immediately implies a way to significantly reduce the FLOP count and improve efficiency for Transformers.
  arXiv  Detail & Related papers  (2022-10-12T15:25:19Z)
• Efficient Language Modeling with Sparse all-MLP [53.81435968051093]
  All-MLPs can match Transformers in language modeling, but still lag behind in downstream tasks. We propose sparse all-MLPs with mixture-of-experts (MoEs) in both feature and input (tokens) We evaluate its zero-shot in-context learning performance on six downstream tasks, and find that it surpasses Transformer-based MoEs and dense Transformers.
  arXiv  Detail & Related papers  (2022-03-14T04:32:19Z)
• MLP Architectures for Vision-and-Language Modeling: An Empirical Study [91.6393550858739]
  We initiate the first empirical study on the use of architectures for vision-and-featured (VL) fusion. We find that without pre-training, usings for multimodal fusion has a noticeable performance gap compared to transformers. Instead of heavy multi-head attention, adding tiny one-head attention to encoders is sufficient to achieve comparable performance to transformers.
  arXiv  Detail & Related papers  (2021-12-08T18:26:19Z)
• MLP-Mixer: An all-MLP Architecture for Vision [93.16118698071993]
  We present-Mixer, an architecture based exclusively on multi-layer perceptrons (MLPs). Mixer attains competitive scores on image classification benchmarks, with pre-training and inference comparable to state-of-the-art models.
  arXiv  Detail & Related papers  (2021-05-04T16:17:21Z)

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.