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

MaskStrategy

struct MaskStrategy

Bit-flag enum that selects the masking strategy for a tile iteration set.

Strategies are combined with bitwise OR. NO_MASK skips masking entirely. COMPUTED calls the mask functor per element. OUT_OF_BOUNDS clips keys at num_keys. BITMASK reads a 32-bit column-visibility mask from MHAMask.mask_bits(), which subsumes the older triangular and OOB strategies for SM100 kernels.

Implemented traits

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

comptime members

BITMASK

comptime BITMASK = MaskStrategy(Int32(16))

Mask provides its own 32-bit visibility bitmask per 32-col batch via MHAMask.mask_bits(). Subsumes the historical LOWER_TRIANGULAR / UPPER_TRIANGULAR (and OUT_OF_BOUNDS for masks that fold it into mask_bits).

COMPUTED

comptime COMPUTED = MaskStrategy(Int32(4))

Mask where the call operator must be used to compute the masked value.

NO_MASK

comptime NO_MASK = MaskStrategy(Int32(0))

No mask is to be applied.

OUT_OF_BOUNDS

comptime OUT_OF_BOUNDS = MaskStrategy(Int32(8))

Check if we are out of bounds, e.g. clip at num_keys after a COMPUTED mask. Used in combination with COMPUTED and as the kernel-hardcoded "no mask, just clip" form (e.g. hot-path optimizations in softmax_warp that bypass BITMASK when the runtime mask status is NO_MASK).

Methods

__init__

def __init__(value: Int32) -> Self

__eq__

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

Returns:

Bool

__ne__

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

Returns:

Bool

__contains__

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

Returns:

Bool

__and__

def __and__(self, other: Self) -> Self

__or__

def __or__(self, other: Self) -> Self