Skip to main content

Module link

Module link 

Source
Expand description

The spatial link: each op of the one IR that the shared optimizer leaves the spatial chips (super::shim::SpatialShim), as the section templates it composes from.

The link reads the op and the graph’s tensor facts: operand names, the fills the program declares, the shapes it gives. The decisions only the spatial chips make about a whole program are taken once, before any op is linked, as the Prelude: which tables are laid out as matrices, the interval each product’s operands are proven in, and the Q8_0 view each product over a stored weight reads its activation through. Prelude::agree_layouts bridges a section’s layout to the one the prelude chose, where they differ.

Structs§

Prelude
The decisions the spatial chips make about a whole program before any op of it is linked.

Enums§

Form
How an op is linked: alone, as the one program a single-op source is, or as one section of a composition, its float-ness read off the whole program’s walk.

Functions§

dense_tables
The tables the ops linked of g lay out DENSE, with agreed, the tables a deployment’s other programs lay out dense: every K and V an attention reads in the flash form (and a block of rows’ queries and result, which the flash form reads and writes a tile of rows at a time), and every multi-row tensor an op reads or writes as a matrix (a stacked batch through a matmul, a pointwise op over a matrix). The layout is decided here, once per table, so a row write or a row read on a table and the ops that read it whole can never disagree on it.
link
Op k of g as the template sections it composes from — one per program the composer chains. Every op is one section except the carried-state blocks, whose read half and write half are separate sections precisely so the order between them is a checkable barrier rather than an assumption (from_qwen_ops::from_gated_delta).