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 package
layout
Provides layout and tensor types, which abstract memory layout for multidimensional data.
-
The
Layouttype represents a mapping between a set of logical coordinates and a linear index. It can be used, for example, to map logical tensor coordinates to a memory address, or to map GPU threads to tiles of data. -
The
TileTensortype is the tensor with explicit memory layout, andTileTensorItersteps through a buffer of equally shaped tiles.
Packages
-
atoms: Defines atoms: wrappers for single hardware intrinsics as typed operations onTileTensoroperands. -
tileops: Defines TileOps: tile-level operations built fromlayout.atoms.
Modules
-
bit_layout: DefinesBitLayout, a layout that is linear over F2. -
coord: Subpackage exposingstd.utils.coordaslayout.coordfor kernel imports. -
layout: Provides a mixed compile-time/runtime layout system for tensor memory mapping. -
math: Implements tensor math operations. -
numpy: NumPy interoperability forTileTensor. -
owned_tensor: Defines tensors that own their storage, with the storage supplied by an engine. -
swizzle: Defines swizzle layouts for optimizing memory access patterns. -
tensor_core: Tensor Core Module for High-Performance Matrix Operations. -
tensor_core_async: Tensor Core Async Module. -
tensor_engine: Defines the engines that operate on tile-backed tensor storage. -
tile_async_copy: Tile copies into shared memory laid out forldmatrixreaders. -
tile_io: Trait and utilities for copying data betweenTileTensors. -
tile_iterator: ATileTensoriterator that walks a buffer one tile at a time. -
tile_tensor: TileTensor type for structured memory access with compile-time layout information. -
tma_async: Tensor Memory Accelerator (TMA) Asynchronous Operations Module. -
tmem_engine: Defines theTensorEnginethat views Blackwell Tensor Memory (TMEM). -
two_level_layout: Provides a two-levelTensorLayoutconformer.