Inverted LoRA-input dropout, keyed by (seed, id) so it is deterministic —
the same (seed, id, shape) reproduces the mask, which is what makes the
segmented / gradient-checkpoint recompute correct. kept ⇒ x/(1-p),
dropped ⇒ 0 (preserves the expectation). Caller owns the result.
Inverted LoRA-input dropout, keyed by (seed, id) so it is deterministic — the same (seed, id, shape) reproduces the mask, which is what makes the segmented / gradient-checkpoint recompute correct. kept ⇒ x/(1-p), dropped ⇒ 0 (preserves the expectation). Caller owns the result.