For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo struct
KVLoad2CvtProducer
struct KVLoad2CvtProducer[dtype: DType, config: MLA_SM100_Decode_Config]
Fieldsβ
- βpipe (
KVLoad2CvtProducer[dtype, config].KVPipeType): - βsmem_upper_fp8 (
UnsafePointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED]):
Implemented traitsβ
AnyType,
Copyable,
ImplicitlyCopyable,
ImplicitlyDeletable,
Movable,
RegisterPassable,
TrivialRegisterPassable
comptime membersβ
bf16_stage_elemsβ
comptime bf16_stage_elems = (config * config)
fp8_stage_stride_elemsβ
comptime fp8_stage_stride_elems = (Int(2) * Int((mul config.BN_QK, config.q_depth)))
KVPipeTypeβ
comptime KVPipeType = KVPipelineGeneric[config.num_kv_stages, Int(1), (Int(1) + Int(32) if (xor (eq (xor (lt config.scale_block_size, 1), True), False), True) else Int(0)), (_resolve_warpgroup_size() + Int(2))]
Methodsβ
__init__β
def __init__(pipe: KVPipelineGeneric[config.num_kv_stages, Int(1), (Int(1) + Int(32) if (xor (eq (xor (lt config.scale_block_size, 1), True), False), True) else Int(0)), (_resolve_warpgroup_size() + Int(2))], smem_upper_fp8: UnsafePointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED]) -> Self
initβ
def init(self)
stage_base_ptrβ
def stage_base_ptr[*, qk_stage: Int = Int(0)](self) -> UnsafePointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED]
Returns:
UnsafePointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED]
producer_mbarβ
def producer_mbar[*, qk_stage: Int = Int(0)](self) -> MBarType
Returns:
MBarType
acquireβ
def acquire[*, qk_stage: Int = Int(0)](self)
commit_stepβ
def commit_step(mut self)
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!