GTA & GLA — Hardware-Efficient Attention

Tri Dao's Grouped-Tied Attention (GTA) and Grouped-Latent Attention (GLA) for fast decoding (ICML 2025): both roughly halve the KV cache versus GQA and run up to 2× faster than FlashMLA, targeting the memory bottleneck in long-context and high-throughput serving. Part of the same Princeton efficient-kernels line as FlashAttention-3.