No-Q Attention Block for Small Language Model Training

A drop-in attention module for training small language models from scratch that removes the query projection (Q = x), giving ~8% fewer parameters and better validation quality (3.18% lower BPC on TinyStories, 2.24% on FineWeb).