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

FlashAttentionAlgorithm

struct FlashAttentionAlgorithm

Identifies which flash-attention algorithm variant to use for a kernel launch.

The four variants range from a naive reference implementation to the latest warp-specialized FA3 pipeline. The default constructed value is FLASH_ATTENTION_3. Use init() to resolve an unspecified (-1) value to the best algorithm for the target dtype and GPU architecture.

Implemented traits

AnyType, Copyable, Defaultable, Deinitable, ImplicitlyCopyable, Movable, RegisterPassable, TrivialRegisterPassable, Writable

comptime members

FLASH_ATTENTION_1

comptime FLASH_ATTENTION_1 = FlashAttentionAlgorithm(Int(1))

FLASH_ATTENTION_2

comptime FLASH_ATTENTION_2 = FlashAttentionAlgorithm(Int(2))

FLASH_ATTENTION_3

comptime FLASH_ATTENTION_3 = FlashAttentionAlgorithm(Int(3))

NAIVE

comptime NAIVE = FlashAttentionAlgorithm(Int(0))

Methods

__init__

def __init__() -> Self

def __init__(value: Int) -> Self

__eq__

def __eq__(self, other: Self) -> Bool

Returns:

Bool

def __eq__(self, version: Int) -> Bool

Returns:

Bool

__ne__

def __ne__(self, other: Self) -> Bool

Returns:

Bool

init

def init(self, dtype: DType) -> Self

write_to

def write_to(self, mut writer: T)