Skip to main content

/

DocsCode

Mojo struct

MLA_SM100_Decode_KV_FP8

@register_passable(trivial) struct MLA_SM100_Decode_KV_FP8[q_type: DType, KVLUTType: MHAOperand, output_type: DType, SplitAccumType: OptionalPointer, MaskType: MHAMask, ScoreModType: ScoreModTrait, config: MLA_SM100_Decode_Config, use_score_mod: Bool, ValidLengthType: OptionalPointer, _is_cache_length_accurate: Bool = False, ragged: Bool = False]

Implemented traits​

AnyType, Copyable, ImplicitlyCopyable, ImplicitlyDestructible, Movable, RegisterPassable, TrivialRegisterPassable

comptime members​

__copyinit__is_trivial​

comptime __copyinit__is_trivial = True

__del__is_trivial​

comptime __del__is_trivial = True

__moveinit__is_trivial​

comptime __moveinit__is_trivial = True

AccumType​

comptime AccumType = get_accum_type[q_type]()

BlockElems​

comptime BlockElems = (config * config)

bytes_per_element​

comptime bytes_per_element = size_of[q_type]()

Common_MLA_Op​

comptime Common_MLA_Op = MLA_SM100_Decode_Common[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged]

kv_type​

comptime kv_type = KVLUTType.dtype

KVStageElems​

comptime KVStageElems = (MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].NumQKBlocks * MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].BlockElems)

NumQKBlocks​

comptime NumQKBlocks = (config // config)

NumVOBlocks​

comptime NumVOBlocks = (config // config)

output_tile_width​

comptime output_tile_width = ((config // 2) * (4 // size_of[output_type]()))

UMMAPVSS​

comptime UMMAPVSS = DecodeSM100PVSS[q_type, MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].AccumType, config=config]

UMMAQKTSS​

comptime UMMAQKTSS = DecodeSM100QKTSS[q_type, MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].AccumType, config=config]

Methods​

kernel​

static kernel(q_tma: TMATensorTile[q_type, tile_layout_k_major[q_type, config.BM, config.BK0, config.swizzle_mode](), _tma_desc_tile_layout[q_type, 2, IndexList(config.BM, config.BK0, Tuple()), config.swizzle_mode]()], k_tma: TMATensorTile[MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].kv_type, _split_last_layout[MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].kv_type](IndexList(config.BK1, 1, config.BK0, Tuple()), config, True), _ragged_desc_layout[MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].kv_type](IndexList(config.BK1, 1, config.BK0, Tuple()), config)], o_tma: TMATensorTile[output_type, tile_layout_k_major[output_type, config.out_rows, config.BN, config.swizzle_mode](), _tma_desc_tile_layout[output_type, 2, IndexList(config.out_rows, config.BN, Tuple()), config.swizzle_mode]()], kv_lut: KVLUTType, scale: Float32, batch_size: Int, q_max_seq_len: Int, num_partitions: Int, max_cache_valid_length: Int, mla_decode_pack: MLA_Decode_Pack[ValidLengthType, MaskType, ScoreModType, SplitAccumType])

load​

static load(q_tma: TMATensorTile[q_type, tile_layout_k_major[q_type, config.BM, config.BK0, config.swizzle_mode](), _tma_desc_tile_layout[q_type, 2, IndexList(config.BM, config.BK0, Tuple()), config.swizzle_mode]()], k_tma_fp8: TMATensorTile[MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].kv_type, _split_last_layout[MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].kv_type](IndexList(config.BK1, 1, config.BK0, Tuple()), config, True), _ragged_desc_layout[MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].kv_type](IndexList(config.BK1, 1, config.BK0, Tuple()), config)], kv_lut: KVLUTType, q_smem: UnsafePointer[Scalar[q_type], MutAnyOrigin, address_space=AddressSpace.SHARED], kv_smem_fp8: UnsafePointer[Scalar[MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].kv_type], MutAnyOrigin, address_space=AddressSpace.SHARED], mbar_q: UnsafePointer[SharedMemBarrier, MutAnyOrigin, address_space=AddressSpace.SHARED], kv_load2cvt_pipe: KVPipelineGeneric[config.num_kv_stages, 1, 1, (WARPGROUP_SIZE + 2)], offset_position: OffsetPosition[config, KVLUTType, ragged, _is_cache_length_accurate, ValidLengthType, config.decoding_warp_split_k])

convertFP8ToBF16​

static convertFP8ToBF16(kv_smem_fp8: UnsafePointer[Scalar[MLA_SM100_Decode_KV_FP8[q_type, KVLUTType, output_type, SplitAccumType, MaskType, ScoreModType, config, use_score_mod, ValidLengthType, _is_cache_length_accurate, ragged].kv_type], MutAnyOrigin, address_space=AddressSpace.SHARED], kv_smem_bf16: UnsafePointer[Scalar[q_type], MutAnyOrigin, address_space=AddressSpace.SHARED], kv_load2cvt_pipe: KVPipelineGeneric[config.num_kv_stages, 1, 1, (WARPGROUP_SIZE + 2)], kv_cvt2mma_pipe: KVPipelineGeneric[config.num_kv_stages, 1, WARPGROUP_SIZE, 2], num_k_tiles: Int)

mmaQK​

static mmaQK(tmem_addr: UInt32, q_smem: UnsafePointer[Scalar[q_type], MutAnyOrigin, address_space=AddressSpace.SHARED], kv_smem: UnsafePointer[Scalar[q_type], MutAnyOrigin, address_space=AddressSpace.SHARED], mbar_q: UnsafePointer[SharedMemBarrier, MutAnyOrigin, address_space=AddressSpace.SHARED], s_bars: DecodeSM100MiscMBars[2, 1, WARPGROUP_SIZE], kv_cvt2mma_pipe: KVPipelineGeneric[config.num_kv_stages, 1, WARPGROUP_SIZE, 2], kv_load2cvt_pipe: KVPipelineGeneric[config.num_kv_stages, 1, 1, (WARPGROUP_SIZE + 2)], offset_position: OffsetPosition[config, KVLUTType, ragged, _is_cache_length_accurate, ValidLengthType, config.decoding_warp_split_k])

mmaPV​

static mmaPV(tmem_addr: UInt32, kv_smem: UnsafePointer[Scalar[q_type], MutAnyOrigin, address_space=AddressSpace.SHARED], p_bars: DecodeSM100MiscMBars[2, WARPGROUP_SIZE, 1], o_bars: DecodeSM100MiscMBars[2, 1, WARPGROUP_SIZE], kv_cvt2mma_pipe: KVPipelineGeneric[config.num_kv_stages, 1, WARPGROUP_SIZE, 2], kv_load2cvt_pipe: KVPipelineGeneric[config.num_kv_stages, 1, 1, (WARPGROUP_SIZE + 2)], offset_position: OffsetPosition[config, KVLUTType, ragged, _is_cache_length_accurate, ValidLengthType, config.decoding_warp_split_k])

View source

View source

Was this page helpful?

Models

Cookbook

Puzzles

Blog

Community

Contact