Mojo function
split_kv_buffer
split_kv_buffer[dtype: DType, target: StringSlice[StaticConstantOrigin] = "cpu"](kv: TileTensor[dtype, kv.LayoutType, kv.origin, linear_idx_type=kv.linear_idx_type, element_shape_types=kv.element_shape_types], k: TileTensor[dtype, k.LayoutType, k.origin, linear_idx_type=k.linear_idx_type, element_shape_types=k.element_shape_types], v: TileTensor[dtype, v.LayoutType, v.origin, linear_idx_type=v.linear_idx_type, element_shape_types=v.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?
Thank you! We'll create more content like this.
Thank you for helping us improve!