pub fn gqa_pack_enabled() -> boolExpand description
The armed/disarmed answer for this process, resolved ONCE.
OnceLock, not a per-call env::var: this is read on the decode path, per
layer per token, and spark_model’s hot_path_env_guards exists because
that read takes the process-wide environment lock and costs 5.76 us at 16
threads. Resolving once is also what CUDA graph capture requires — the
kernel a captured graph holds cannot depend on an environment read that
might answer differently on replay.