Skip to main content

Mojo function

split_kv_buffer

split_kv_buffer[dtype: DType, target: StringSlice[StaticConstantOrigin] = "cpu"](kv: TileTensor[dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], k: TileTensor[dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], v: TileTensor[dtype, origin, linear_idx_type=linear_idx_type, element_shape_types=element_shape_types], ctx: DeviceContext)

Split a packed KV buffer into separate K and V tensors.

Expects kv shape [S, H * (Dk + Dv)] and writes k/v shapes [S, H, Dk/Dv].

Was this page helpful?