pub const CAUSAL_ARCHS: [&str; 9];Expand description
The architectures this stage’s generator can run: llama.cpp’s causal text models.
A guard, not a preference. An encoder or encoder-decoder gguf loads
and then aborts the process mid-decode (ggml_abort on a cross-attention
input a causal decode never sets), and an abort cannot be caught, so the
check is by name and it happens before the load. The cleanup feed reads
this same list to decide what it offers, so a candidate and the file the
load accepts cannot disagree.