pub fn w8a8_decode_proj_disabled() -> boolExpand description
ATLAS_NO_W8A8_DECODE_PROJ kill switch: PRESENCE (any value, including
empty) keeps every 5..16-row decode projection on today’s w8a16_gemv_batch16
tiers. Presence rather than =1 for the same reason ATLAS_FFN_W8A16_ONLY
is presence-checked — an operator reaches for it while a serve is
misbehaving, and ...=0 meaning “on” is a trap.
OnceLock-cached and therefore constant for the life of the process, which
is what makes it safe to branch on under CUDA-graph capture.