Expand description
2026-10-02: Load-time derived weight copies (kernels/circuits/COPIES.toml): every
device allocation the loader makes for a weight beyond the checkpoint tensors it stores
(requantized serving copies, transposed prefill twins, repacks, widened scales, and the
leaked intermediates nothing reads). A rule matches nodes by op, by the checkpoint’s declared
weight format, by the served format and by serve settings, and states the byte law of its copy.
Owner: metrale-circuit (memory). Invariants:
- A copy’s bytes come from the node’s own shape (
out x k, per expert for routed experts) and the rule’s law; nothing is sized from a measured total. - Every rule cites the loader site that allocates it; a rule that matches no node of a circuit is not an error (it belongs to another checkpoint), but an unknown op, format, law or setting spelling is refused at parse.
Structs§
Enums§
- CopyLaw
- 2026-10-02: How a copy’s bytes follow from the weight’s
out x kshape.
Constants§
- COPY_
SETTINGS - 2026-10-02: The settings a rule’s
whenmay read: the serve facts the loader branches on.latent_moeisonwhen the checkpoint’s MoE runs through a latent projection (moe_latent_size > 0, Nemotron-3-Super),offotherwise.moe_expert_tablesis the serve’s plan-based decision on the MoE’s transposed prefill tables (buildorskip).
Functions§
- parse_
copies - 2026-10-02: Parse the rules text.