Skip to main content

/

DocsCode

Mojo struct

KVBufferImpl

struct KVBufferImpl[dtype: DType, layout: Layout, address_space: AddressSpace, alignment: Int, origin: Origin[mut=origin.mut], masked: Bool, layout_int_type: DType, linear_idx_type: DType, //, config: KVBufferConfig, tensor_core_mma: TiledTensorCore[tensor_core_mma.out_type, tensor_core_mma.in_type, tensor_core_mma.shape, tensor_core_mma.group_size, tensor_core_mma.transpose_b], swizzle: Optional[Swizzle], BN: Int, WN: Int, BK: Int, depth: Int, num_threads: Int, num_stages: Int = 1, token_gen: Bool = False]

Fields​

  • ​load_tile (KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].LoadTileType):
  • ​mma_tile (KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].MMATileType):
  • ​smem_ptr (UnsafePointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED]):
  • ​bounds (Int):
  • ​load_tile_id (Int):
  • ​global_iterator (KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].GlobalTiledIteratorType):

Implemented traits​

AnyType, ImplicitlyDestructible, KVBuffer

comptime members​

GlobalTensorType​

comptime GlobalTensorType = LayoutTensor[dtype, layout, origin, address_space=address_space, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment]

GlobalTiledIteratorType​

comptime GlobalTiledIteratorType = LayoutTensorIter[dtype, LayoutTensor._compute_tile_layout[config.btile_dim0, config.btile_dim1]()[0], origin, address_space=address_space, axis=config.iterator_axis, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked or _tile_is_masked[layout, config.btile_dim0, config.btile_dim1]()]

LoadTileType​

comptime LoadTileType = LayoutTensor[dtype, Layout.row_major(VariadicList(((num_stages * KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].num_mmas) * KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].num_k_tiles), KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].simd_width)), MutAnyOrigin, address_space=AddressSpace.LOCAL]

MMA_K​

comptime MMA_K = shape.__getitem__[Int](2)

MMA_N​

comptime MMA_N = shape.__getitem__[Int](1)

mma_tile_layout​

comptime mma_tile_layout = Layout.row_major(VariadicList(KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].num_mmas, KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].simd_width))

MMATileType​

comptime MMATileType = LayoutTensor[dtype, KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].mma_tile_layout, MutAnyOrigin, address_space=AddressSpace.LOCAL]

num_k_tiles​

comptime num_k_tiles = ceildiv(BK, (KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].MMA_K * group_size))

num_mmas​

comptime num_mmas = ceildiv(config.wsize, KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].MMA_N)

num_warps_n​

comptime num_warps_n = (BN // WN)

SharedTileType​

comptime SharedTileType = LayoutTensor[dtype, KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].smem_layout, MutAnyOrigin, address_space=AddressSpace.SHARED]

SharedWarpTileType​

comptime SharedWarpTileType = LayoutTensor[dtype, LayoutTensor._compute_tile_layout[KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].wtile_dim0, KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].wtile_dim1]()[0], MutAnyOrigin, address_space=AddressSpace.SHARED, layout_int_type=_get_layout_type(KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].smem_layout, AddressSpace.SHARED), linear_idx_type=_get_index_type(KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].smem_layout, AddressSpace.SHARED), masked=_tile_is_masked[KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].smem_layout, KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].wtile_dim0, KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].wtile_dim1]()]

simd_width​

comptime simd_width = simd_width_of[dtype]()

smem_layout​

comptime smem_layout = Layout.row_major(VariadicList(config.btile_dim0, config.btile_dim1))

thread_layout​

comptime thread_layout = Layout.row_major(VariadicList(((min(num_threads, ((config.btile_dim0 * config.btile_dim1) // KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].simd_width)) * KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].simd_width) // KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].smem_layout.stride[0].value()), (KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].smem_layout.stride[0].value() // KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].simd_width))) if token_gen else Layout.row_major(VariadicList((num_threads // 4), 4))

wtile_dim0​

comptime wtile_dim0 = config.wtile_dim0

wtile_dim1​

comptime wtile_dim1 = config.wtile_dim1

Methods​

__init__​

__init__(out self, global_tile: LayoutTensor[dtype, layout, origin, address_space=address_space, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], num_b_rows: Optional[Int], shared_ptr: UnsafePointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED])

get_dtype​

static get_dtype() -> DType

Returns:

DType

load_from_dram​

load_from_dram(mut self)

get_mma_tile​

get_mma_tile(self) -> KVBufferImpl[config, tensor_core_mma, swizzle, BN, WN, BK, depth, num_threads, num_stages, token_gen].MMATileType

Returns:

KVBufferImpl

copy_to_shared​

copy_to_shared[tile_id: Int = 0](self)

load_from_shared​

load_from_shared[k_mma: Int](self)

View source

View source

Was this page helpful?

Models

Cookbook

Puzzles

Blog

Community

Contact