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

NoPartition

struct NoPartition[dtype: DType]

A single-partition (non-split-K) scheme for MHA decoding.

Uses the standard single-pass flash-attention decode without any inter-CTA reduction. do_partition is False, so the split-K reduction kernel is compiled away entirely.

Parameters

• ​dtype (DType): The accumulator element type (same as the output type).

Implemented traits

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

comptime members

accum_dtype

comptime accum_dtype = dtype

do_partition

comptime do_partition = False

Methods

__init__

def __init__() -> Self

num_partitions

def num_partitions(self) -> UInt32

Returns:

UInt32

max_num_partitions

def max_num_partitions(self) -> UInt32

Returns:

UInt32

get_exp_sum_qk_max_pointer

def get_exp_sum_qk_max_pointer(self) -> Pointer[Scalar[Self.accum_dtype], MutAnyOrigin]

Returns:

Pointer[Scalar[Self.accum_dtype], MutAnyOrigin]