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

KVCvt2MmaProducer

struct KVCvt2MmaProducer[dtype: DType, config: MLA_SM100_Decode_Config]

Produces converted BF16 KV tiles for the MMA consumer pipeline.

Parameters

• ​dtype (DType): Element type of the converted KV tiles stored in SMEM.
• ​config (MLA_SM100_Decode_Config): Decode config supplying num_kv_stages, BN_QK, and q_depth for pipeline stage sizing.

Fields

• ​pipe (KVCvt2MmaProducer[dtype, config].PipeT):
• ​smem (Pointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED]):

Implemented traits

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

comptime members

kv_stage_elems

comptime kv_stage_elems = (config * config)

PipeT

comptime PipeT = KVPipelineGeneric[config.num_kv_stages, Int(1), _resolve_warpgroup_size(), Int(2)]

Methods

__init__

def __init__(pipe: KVPipelineGeneric[config.num_kv_stages, Int(1), _resolve_warpgroup_size(), Int(2)], smem: Pointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED]) -> Self

acquire

def acquire(self)

stage_index

def stage_index(self) -> UInt32

Returns:

UInt32

stage_base_ptr

def stage_base_ptr(self) -> Pointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED]

Returns:

Pointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED]

commit_all

def commit_all(mut self)