Convolved [B,S,2HkDk+Hv*Dv] to owned Q, K, V heads, Dk/Dv <=128. Scale vectors have Dk entries in the input dtype, as MLX's RMS weights do.
Convolved [B,S,2HkDk+Hv*Dv] to owned Q, K, V heads, Dk/Dv <=128. Scale vectors have Dk entries in the input dtype, as MLX's RMS weights do.