One activation row; gate and up at the same width. The threadgroup codebook decode where the gate's GPU family qualifies it (applied by the kernel at the measured artifact shape).
Readonly
x [.., in] (borrowed, one row) to an owned [.., rows].
x
One activation row; gate and up at the same width. The threadgroup codebook decode where the gate's GPU family qualifies it (applied by the kernel at the measured artifact shape).