Function ffn_w8a16_only 

pub fn ffn_w8a16_only() -> bool

ATLAS_FFN_W8A16_ONLY kill switch: PRESENCE (any value, including empty) keeps the dense-FFN prefill on today’s W8A16 kernels. Presence rather than =1 because this is an escape hatch an operator reaches for while a serve is misbehaving, and ATLAS_FFN_W8A16_ONLY=0 meaning “on” is a trap.

OnceLock-cached: the selector runs per projection per layer per prefill (3 x num_layers times), and std::env::var_os walks the environment block on every call. Cached process-wide is correct here — the variable is read once at first prefill and a serve never rewrites its own environment.