pub const W8A8_PREFILL_KERNELS: [(&str, &str); 2];Expand description
The two (module, function) pairs the W8A8 block-scaled prefill arm needs
(prefill/paged_qkv.rs:220, prefill/paged_oproj.rs:94).
SSOT for three readers that must not drift (#915): init.rs resolves the
handles, Qwen3AttentionLayer::has_w8a8_prefill_kernels tests them, and
w8a8_prefill_kernels_loaded asks the BACKEND the same question before
any layer exists — which is what lets preflight predict, pre-load, whether
the Q and O FP8 prefill twins will be built. A name typo’d in one of the
three would mis-predict ~1.5 GB of residency on the 27B in silence.