Records every tensor a universal load consumes; finish diffs against the shard index so a descriptor mistake is a LOAD error (unconsumed / missing tensors named), never a silently-wrong model (02d723a:docs/design/generic-model-support.md ยง4.1).
finish
02d723a:docs/design/generic-model-support.md
Readonly
drop = the arch's sanitize rules (tensors mlx-lm discards on load, e.g. rotary_emb.inv_freq, or lm_head.weight under tied embeddings).
drop
Records every tensor a universal load consumes;
finishdiffs against the shard index so a descriptor mistake is a LOAD error (unconsumed / missing tensors named), never a silently-wrong model (02d723a:docs/design/generic-model-support.mdยง4.1).