DeepSeek has launched FlashMLA, an innovative Multi-head Latent Attention decoding kernel optimized for NVIDIA’s Hopper GPUs, achieving 3000 GB/s memory bandwidth and 580 TFLOPS. It reduces memory overhead by 40-60% and enables efficient processing of variable-length sequences. FlashMLA demonstrates significant performance improvements and is open-sourced to enhance AI infrastructure, receiving rapid community support.
Chinese hackers evade ESET with MAVInject.exe
Chinese hacking group Earth Preta has been found using a novel technique to bypass antivirus software using a valid Microsoft tool, MAVInject.exe. The group’s malware
