Function cpu_attention_enabled 

pub fn cpu_attention_enabled() -> bool

Whether native attention should use the CPU reference.

Default OFF (use the GPU attention path) — as of #49 the GPU path also honors norm_weight for prefill K/V and produces coherent output, and it is faster. The CPU SDPA reference remains available as a correctness fallback / cross-check via QUALIA_LLM_CPU_ATTENTION=1 or set_cpu_attention.