Event-driven spiking neural networks(SNNs) are promising neural networks that reduce the energy consumption of continuously growing AI models. Recently, keeping pace with the development of transformers, transformer-based SNNs were presented. In this work, we propose a novel transformer-to-SNN conversion method that outputs an end-to-end spike-based transformer, named SpikedAttention. Our method directly converts the well-trained transformer without modifying its attention architecture. For the vision task, the proposed method converts Swin Transformer into an SNN without post-training or conversion-aware training, achieving state-of-the-art SNN accuracy on ImageNet dataset, i.e., 80.0\% with 28.7M parameters.

Post-training is essential for adapting Large Language Models (LLMs) to real-world applications. Deploying post-trained models faces significant challenges due to substantial memory overhead and noticeable inference latency. Existing work has identified significant redundancies in LLMs and proposed efficient architectures, namely intra-layer KV sharing and cross-layer KV sharing. However, intra-layer KV sharing still results in high inference costs, while cross-layer KV sharing leads to significant performance degradation. As a result, both methods remain suboptimal for post-training pre-trained LLMs. In this paper, we identify that the \texttt{Softmax} operation is a primary bottleneck for LLM inference and discover that it is actually highly redundant during post-training. We propose Softmax \textbf{Uni}fication in \textbf{Att}e\textbf{n}tion (\textbf{UniAttn}), a novel post-training method that unifies Softmax activations across transformer blocks to reduce LLM inference costs. Additionally, UniAttn adopts a linear projection to compensate for the errors induced by Softmax unification. Experiments show that UniAttn matches the performance of standard post-training while significantly reducing inference costs, outperforming existing efficient architectures during post-training. Our code will be available at \url{https://github.com/Bostoncake/UniAttn}.

Large language models have achieved remarkable success in recent years, primarily due to the implementation of self-attention mechanisms. However, traditional Softmax attention suffers from numerical instability and reduced performance as the length of inference tokens increases. This paper addresses these issues by decomposing the Softmax operation into a non-linear transformation and the $l_1$-norm. We identify the latter as essential for maintaining model performance. By replacing the non-linear transformation with the Softplus activation function and introducing a dynamic scale factor for different token lengths based on invariance entropy, we create a novel attention mechanism with performance better than conventional Softmax attention across various inference lengths. To further improve the length extrapolation ability of the proposed attention mechanism, we introduce a fine-tuning-free re-weighting mechanism that amplifies significant attention weights while diminishing weaker ones, enabling the model to concentrate more effectively on relevant tokens without requiring retraining. When combined with our proposed attention mechanism, this approach demonstrates significant promise in managing longer sequences, maintaining nearly constant validation loss even at 16$\times$ the training token length while ensuring numerical stability. Our code is available at: https://github.com/iminfine/freeatten.

Quantization has established itself as the primary approach for decreasing the computational and storage expenses associated with Large Language Models (LLMs) inference. The majority of current research emphasizes quantizing weights and activations to enable low-bit general-matrix-multiply (GEMM) operations, with the remaining non-linear operations executed at higher precision. In our study, we discovered that following the application of these techniques, the primary bottleneck in LLMs inference lies in the softmax layer. The softmax operation comprises three phases: exponent calculation, accumulation, and normalization, Our work focuses on optimizing the first two phases. We propose an analytical approach to determine the optimal clipping value for the input to the softmax function, enabling sub-4-bit quantization for LLMs inference. This method accelerates the calculations of both $e^x$ and $\sum(e^x)$ with minimal to no accuracy degradation. For example, in LLaMA1-30B, we achieve baseline performance with 2-bit quantization on the well-known "Physical Interaction: Question Answering" (PIQA) dataset evaluation. This ultra-low bit quantization allows, for the first time, an acceleration of approximately 4x in the accumulation phase. The combination of accelerating both $e^x$ and $\sum(e^x)$ results in a 36.9% acceleration in the softmax operation.

When implementations of the Transformer's self-attention layer utilize SRAM instead of DRAM, they can achieve significant speedups. The Tenstorrent Grayskull architecture provides a large SRAM, distributed across a grid of cores. This work presents a fused kernel for Grayskull, that exclusively utilizes its large SRAM by combining matrix multiplication, attention score scaling and Softmax operations. Additionally, a dedicated Softmax kernel utilizing the SRAM and a CPU implementation serving as a baseline are presented. The Softmax operation consumes most of the runtime in the computation of attention weights from queries and keys on Grayskull. The speedup of the dedicated Softmax kernel compared to the CPU implementation is up to $10 \times$, and the Softmax implementation inside the fused kernel is approximately $1.8 \times$ faster than the dedicated Softmax kernel. The time and memory complexity of all implementations is quadratic in sequence length. Currently, the Grayskull e150 is approximately $30 \times$ cheaper for the general public than an Nvidia H100 PCIe (a state-of-the-art GPU) and offers approximately $1.5 \times$ more SRAM.

The Gumbel-Softmax probability distribution allows learning discrete tokens in generative learning, while the Gumbel-Argmax probability distribution is useful in learning discrete structures in discriminative learning. Despite the efforts invested in optimizing these probability models, their statistical properties are under-explored. In this work, we investigate their representation properties and determine for which families of parameters these probability distributions are complete, i.e., can represent any probability distribution, and minimal, i.e., can represent a probability distribution uniquely. We rely on convexity and differentiability to determine these statistical conditions and extend this framework to general probability models, such as Gaussian-Softmax and Gaussian-Argmax. We experimentally validate the qualities of these extensions, which enjoy a faster convergence rate. We conclude the analysis by identifying two sets of parameters that satisfy these assumptions and thus admit a complete and minimal representation. Our contribution is theoretical with supporting practical evaluation.

New models of the attention-based random forests called LARF (Leaf Attention-based Random Forest) are proposed. The first idea behind the models is to introduce a two-level attention, where one of the levels is the "leaf" attention and the attention mechanism is applied to every leaf of trees. The second level is the tree attention depending on the "leaf" attention. The second idea is to replace the softmax operation in the attention with the weighted sum of the softmax operations with different parameters. It is implemented by applying a mixture of the Huber's contamination models and can be regarded as an analog of the multi-head attention with "heads" defined by selecting a value of the softmax parameter. Attention parameters are simply trained by solving the quadratic optimization problem. To simplify the tuning process of the models, it is proposed to make the tuning contamination parameters to be training and to compute them by solving the quadratic optimization problem. Many numerical experiments with real datasets are performed for studying LARFs. The code of proposed algorithms can be found in https://github.com/andruekonst/leaf-attention-forest.

Non-local (NL) block is a popular module that demonstrates the capability to model global contexts. However, NL block generally has heavy computation and memory costs, so it is impractical to apply the block to high-resolution feature maps. In this paper, to investigate the efficacy of NL block, we empirically analyze if the magnitude and direction of input feature vectors properly affect the attention between vectors. The results show the inefficacy of softmax operation which is generally used to normalize the attention map of the NL block. Attention maps normalized with softmax operation highly rely upon magnitude of key vectors, and performance is degenerated if the magnitude information is removed. By replacing softmax operation with the scaling factor, we demonstrate improved performance on CIFAR-10, CIFAR-100, and Tiny-ImageNet. In Addition, our method shows robustness to embedding channel reduction and embedding weight initialization. Notably, our method makes multi-head attention employable without additional computational cost.