Mojo struct
KVCacheMHAOperand
@register_passable(trivial)
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,
ExplicitlyCopyable,
MHAOperand,
Movable,
UnknownDestructibility
Aliases
dtype
alias dtype = get_witness(cache_t, kv_cache::types::KVCacheT, dtype)
Methods
__init__
__init__(cache: cache_t) -> Self
block_paged_ptr
block_paged_ptr[tile_size: Int](self, batch_idx: SIMD[uint32, 1], start_tok_idx: SIMD[uint32, 1], head_idx: SIMD[uint32, 1], head_dim_idx: SIMD[uint32, 1] = 0) -> UnsafePointer[SIMD[get_witness(cache_t, kv_cache::types::KVCacheT, dtype), 1]]
Returns:
cache_length
max_context_length
row_idx
row_idx(self, batch_idx: SIMD[uint32, 1], start_tok_idx: SIMD[uint32, 1]) -> SIMD[uint32, 1]
Returns the row idx when viewing the memory as a matrix.
Returns:
col_idx
col_idx(self, head_idx: SIMD[uint32, 1]) -> SIMD[uint32, 1]
Returns the col idx when viewing the memory as a matrix.
Returns:
create_tma_tile
create_tma_tile[tile_m: Int, tile_n: Int, swizzle_mode: TensorMapSwizzle, *, is_k_major: Bool](self, ctx: DeviceContext) -> TMATensorTile[get_witness(cache_t, kv_cache::types::KVCacheT, dtype), tile_layout_k_major[::DType,::Int,::Int,::TensorMapSwizzle]() if is_k_major else tile_layout_mn_major[::DType,::Int,::Int,::TensorMapSwizzle](), _tma_desc_tile_layout[::DType,::Int,::IndexList[$1, ::DType(), is_k_major]
Creates a TMA tile for efficient GPU memory transfers.
Returns:
TMATensorTile
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!