Mojo struct
KVCacheMHAOperand
struct KVCacheMHAOperand[cache_t: KVCacheT]
An implementation for mo.opaque KVCacheT arguments to MHA kernels.
We can eventually remove this trait and just add it as a sub-trait in the KVCacheT type, but we need to solve some cyclic dependencies first.
Fields
- cache (
cache_t):
Implemented traits
AnyType,
Copyable,
DevicePassable,
ImplicitlyCopyable,
ImplicitlyDestructible,
MHAOperand,
Movable,
RegisterPassable,
TrivialRegisterPassable
comptime members
device_type
comptime device_type = KVCacheMHAOperand[cache_t]
dtype
comptime dtype = cache_t.dtype
page_size
comptime page_size = cache_t.page_size_
quantization_enabled
comptime quantization_enabled = cache_t.quantization_enabled
quantization_granularity
comptime quantization_granularity = cache_t.quantization_granularity
scale_dtype
comptime scale_dtype = cache_t.scale_dtype
Methods
__init__
__init__(cache: cache_t) -> Self
get_type_name
block_paged_ptr
block_paged_ptr[tile_size: Int](self, batch_idx: UInt32, start_tok_idx: UInt32, head_idx: UInt32, head_dim_idx: UInt32 = 0) -> UnsafePointer[Scalar[KVCacheMHAOperand[cache_t].dtype], ImmutAnyOrigin]
Returns:
scales_block_paged_ptr
scales_block_paged_ptr(self, batch_idx: Int, start_tok_idx: Int, head_idx: Int, head_dim_idx: Int = 0) -> UnsafePointer[Scalar[KVCacheMHAOperand[cache_t].scale_dtype], ImmutAnyOrigin]
Returns:
load_scale
load_scale[width: Int](self, batch_idx: Int, start_tok_idx: Int, head_idx: Int, head_dim_idx: Int) -> SIMD[KVCacheMHAOperand[cache_t].scale_dtype, width]
Returns:
cache_length
max_context_length
num_kv_rows
num_kv_rows(self) -> Int
Returns the total number of virtual rows in the KV memory view.
Returns:
row_idx
row_idx(self, batch_idx: UInt32, start_tok_idx: UInt32) -> UInt32
Returns the row idx when viewing the memory as a matrix.
Returns:
create_tma_tile
create_tma_tile[swizzle_mode: TensorMapSwizzle, *, BN: Int, depth: Int, BK: Int = padded_depth[KVCacheMHAOperand[cache_t].dtype, swizzle_mode, depth]()](self, ctx: DeviceContext, out tma: TMATensorTile[KVCacheMHAOperand[cache_t].dtype, 3, _padded_shape[3, KVCacheMHAOperand[cache_t].dtype, IndexList(VariadicList(BN, 1, BK), Tuple()), swizzle_mode](), _ragged_shape[3, KVCacheMHAOperand[cache_t].dtype, IndexList(VariadicList(BN, 1, BK), Tuple()), swizzle_mode]()])
Creates a TMA tile for efficient GPU memory transfers.
Returns:
create_scale_tma_tile
create_scale_tma_tile[BMN: Int](self, ctx: DeviceContext, out tma: TMATensorTile[KVCacheMHAOperand[cache_t].scale_dtype, 2, Index[Int, Int](VariadicPack(1, BMN))])
Creates a TMA tile for efficient GPU memory transfers. This is useful for m-major MMA operations where we don't need to mask any extra rows.
Returns:
create_ragged_tma_tile
create_ragged_tma_tile[swizzle_mode: TensorMapSwizzle, *, BN: Int, depth: Int, BK: Int = padded_depth[KVCacheMHAOperand[cache_t].dtype, swizzle_mode, depth]()](self, ctx: DeviceContext, out tma: RaggedTMA3DTile[KVCacheMHAOperand[cache_t].dtype, swizzle_mode, BN, BK])
Returns:
create_rope_tma_tile
create_rope_tma_tile[swizzle_mode: TensorMapSwizzle, *, BN: Int, BK: Int, padded_depth: Int](self, ctx: DeviceContext, out tma: TMATensorTile[DType.bfloat16, 3, _padded_shape[3, DType.bfloat16, IndexList(VariadicList(BN, 1, BK), Tuple()), swizzle_mode](), _ragged_shape[3, DType.bfloat16, IndexList(VariadicList(BN, 1, BK), Tuple()), swizzle_mode]()])
Delegates to the underlying KVCache to create a BF16 rope TMA tile.
Returns:
create_gather4_tma_tile
create_gather4_tma_tile[row_width: Int, swizzle_mode: TensorMapSwizzle = TensorMapSwizzle.SWIZZLE_NONE](self, ctx: DeviceContext, out tma: TMATensorTile[KVCacheMHAOperand[cache_t].dtype, 2, IndexList(VariadicList(4, row_width), Tuple()), IndexList(VariadicList(1, row_width), Tuple())])
Creates a 2D TMA gather4 descriptor for this KV cache operand.
Returns:
scales_raw_ptr
scales_raw_ptr(self) -> UnsafePointer[Float32, MutAnyOrigin]
Returns the base pointer to the quantization scales tensor.
Returns:
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!