Expand description
2026-10-02: The REQUIRED pipeline of a node: what the circuit’s formats, under the serving
policy and the plan’s stated routing, ask each step to compute at. This is the “what we
need” view (met circuit precision) and the side every kernel declaration is checked
against (super::check).
Where each requirement comes from:
- inputs and outputs: the formats the plan stores each edge in (the circuit’s, or a rule’s
writes); inside a group, the format a rule states the edge is handed on in (holds); act: the input format, unlessactivation_quantizationroutes the projection’s family to a fixed format at this row count, or, underadaptive(today’s routing, which the rules encode), the rule states the format its kernels run the activation at;weight: the node’s weight format, its values widened to the activation’s precision for a 16/32-bit activation, consumed natively (E4M3, E2M1) under an 8- or 4-bit one (an NVFP4 weight under FP8 activations by the exact E2M1 to E4M3 conversion);cache:kv_cache_dtype;state: the state’s declared format, a keyed one from its setting (ssm_h_storagefromssm_h_dtype);- every accumulation, scale application, element-wise compute, softmax, routing score and blend sum: FP32, the circuit’s reference definition of those ops;
- any step a plan’s rule states (
stepson a pattern element): the stated value. A rule states where today’s routing departs from the reference; the plan prints it, and the kernel must declare the same.
Owner: metrale-circuit (pipeline). Invariants:
- Pure; a setting the requirement reads and the policy does not state is an error.
- The base requirement (no plan) and a plan’s requirement differ only where the plan states something: a stored, held or stated format.
Structs§
- Need
- 2026-10-02: Everything a requirement reads besides the node.
Functions§
- keyed_
state - 2026-10-02: The precision a keyed state format holds under
settings. The recurrent h-state (ssm_h_storage) holds the precision--ssm-h-dtypenames: FP16 underf16andf16-pool(which differ only in the pool’s size), FP32 underf32. - required
- 2026-10-02: The pipeline
needrequires of noden.