For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo struct

BlockSizing

struct BlockSizing

Target op count per MMA block for latency hiding.

Heavy blocks (introducing a new M-tile row) get more ops to fill the MMA latency window. Light blocks (continuation of same M-tile) get fewer ops since only one new fragment is needed.

Fields

• ​heavy (Int):
• ​light (Int):
• ​max_globals (Int):

Implemented traits

AnyType, Copyable, Deinitable, ImplicitlyCopyable, Movable

Methods

default_2x2

static def default_2x2() -> Self

2x2 MMA grid default: 4 ops for new M-tile, 2 for continuation.

uniform

static def uniform(n: Int) -> Self

All blocks get the same target size.