Actual resident bytes of this adapter's mounted lora_a/lora_b arrays
(sum of MlxArray.nbytes — the real RAM cost while mounted, not the
on-disk safetensors size). Web chat's adapter routing table shows this
per loaded adapter instead of guessing
from file size.
rank
rank:number|null
scale
scale:number
sizeBytes
sizeBytes:number
skippedTensors
skippedTensors:number
Adapter tensors for modules outside the 7 target suffixes
(per-layer-input projections etc.) — skipped, like the reference.
Actual resident bytes of this adapter's mounted lora_a/lora_b arrays (sum of MlxArray.nbytes — the real RAM cost while mounted, not the on-disk safetensors size). Web chat's adapter routing table shows this per loaded adapter instead of guessing from file size.