pub const MIN_CTAS_PER_SM: u32 = 2;Expand description
Tokens per SM the twin insists on before it will take the launch.
Two, not one: at exactly one CTA per SM a trailing partial wave is the whole
kernel, and the parent’s ceil(N/4)-wide grid beats that comfortably. This
is the only number here that is a judgement rather than a contract — it is
deliberately conservative, because being wrong in this direction costs the
prefill lever nothing (a 1168- or 4576-token chunk clears it by 4-17x) and
being wrong in the other direction costs the decode step.