Training-only LoRA-input dropout. rate is the drop probability; seed is
set per micro-step by the trainer and is CONSTANT across that step's forward
and any recompute (segmented / gradient-checkpoint), so each layer's mask —
keyed by (seed, the linear's dropoutId) — is reproduced exactly in the
backward. seed === null (the default, and for inference) disables it.
Active-adapter state, shared by every mounted linear of one model. A plain field, NOT a ContextVar port: under continuous batching requests interleave, so the field is only valid for the forward being built. Each execution group binds its rows' adapter list (
bindAdapterContext) and sets/restores it around every forward;adapterScopeddoes the same for a single generation.