GLM 5.1 runner.
GLM ships as general.architecture = glm4 / glm5 / chatglm /
glm4moe in its GGUF converters. Llama-shaped with a GLM-specific
RoPE layout and RMSNorm placement. This crate is a thin wrapper
over [rlx_llama32::Llama32Runner] with arch validation.
Caveat: GLM-specific RoPE (rope_ratio) and the
post-LayerNorm-on-FFN-input placement aren't yet wired in
rlx-llama32 — runs produce some tokens but won't match the
upstream reference for the first few prompts until those deltas
land (PLAN.md M5 follow-up).