ConSmax: Hardware-Friendly Alternative Softmax with Learnable Parameters

arXiv:2402.10930v1 Announce Type: cross
Abstract: The self-attention mechanism sets transformer-based large language model (LLM) apart from the convolutional and recurrent neural networks. Despite the performance improvement, achieving real-time LLM inference on silicon is challenging due to the extensively used Softmax in self-attention. Apart from the non-linearity, the low arithmetic intensity greatly reduces the processing parallelism, which becomes the bottleneck especially when dealing with a longer context. To address this challenge, we propose Constant Softmax (ConSmax), a software-hardware co-design as an …

abstract arxiv attention cs.ai cs.ar cs.lg hardware improvement inference intensity language language model large language large language model llm low networks neural networks parameters performance real-time recurrent neural networks self-attention silicon softmax transformer type