The performance of standard Transformer models can be enhanced using high-performance computing strategies that are broadly applicable to many deep learning models, not just LLMs. These strategies generally fall into two categories. The first is the use of low-precision implementations, which involves performing arithmetic operations with 8-bit or 16-bit fixed-point data types instead of the conventional 32-bit or 64-bit floating-point types. This shift increases computational efficiency and memory throughput, enabling the processing of longer sequences. The second category consists of hardware-aware techniques, which optimize model performance for specific hardware, such as using IO-aware self-attention implementations on modern GPUs.