Core Insights - The article discusses the advancements in attention mechanisms for large models, particularly focusing on the introduction of SageAttention3, which offers significant performance improvements over previous versions and competitors [1][2]. Group 1: Introduction and Background - The need for optimizing attention speed has become crucial as the sequence length in large models increases [7]. - Previous versions of SageAttention (V1, V2, V2++) achieved acceleration factors of 2.1, 3, and 3.9 times respectively compared to FlashAttention [2][5]. Group 2: Technical Innovations - SageAttention3 provides a 5x inference acceleration compared to FlashAttention, achieving 1040 TOPS on RTX 5090, outperforming even the more expensive H100 with FlashAttention3 by 1.65 times [2][5]. - The introduction of trainable 8-bit attention (SageBwd) allows for training acceleration while maintaining the same results as full precision attention in various fine-tuning tasks [2][5]. Group 3: Methodology - The research team employed Microscaling FP4 quantization to enhance the precision of FP4 quantization, utilizing NVFP4 format for better accuracy [15][16]. - A two-level quantization approach was proposed to address the narrow range of scaling factors for the P matrix, improving overall precision [15][16]. Group 4: Experimental Results - SageAttention3 demonstrated impressive performance in various models, maintaining end-to-end accuracy in video and image generation tasks [21][22]. - In specific tests, SageAttention3 achieved a 3x acceleration in HunyuanVideo, with significant reductions in processing time across multiple models [33][34].