For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

enqueue_fp4_materialize

def enqueue_fp4_materialize[out_type: DType](out_w: TileTensor[out_type, Storage=out_w.Storage, address_space=out_w.address_space, linear_idx_type=out_w.linear_idx_type], packed: TileTensor[DType.uint8, Storage=packed.Storage, address_space=packed.address_space, linear_idx_type=packed.linear_idx_type], scales: TileTensor[DType.float8_e4m3fn, Storage=scales.Storage, address_space=scales.address_space, linear_idx_type=scales.linear_idx_type], ctx: DeviceContext)

Enqueues fp4_materialize_kernel: packed FP4 + scales -> dense [N, K].

out_w is [N, K], packed is [N, K//2], scales is [N, K//16]. The grid is (ceil(K/16), ceil(N/16)) threadgroups of 16x16 threads; bounds are checked per thread so ragged K/N are fine.

Parameters:

out_type (DType): Output element dtype (bf16 for the Apple W4A16 path).

Args:

out_w (TileTensor[out_type, Storage=out_w.Storage, address_space=out_w.address_space, linear_idx_type=out_w.linear_idx_type]): Output weight buffer of shape [N, K] and dtype out_type, written by the enqueued kernel.
packed (TileTensor[DType.uint8, Storage=packed.Storage, address_space=packed.address_space, linear_idx_type=packed.linear_idx_type]): Input packed FP4 weights as uint8 of shape [N, K//2], two E2M1 nibbles per byte with the low nibble first.
scales (TileTensor[DType.float8_e4m3fn, Storage=scales.Storage, address_space=scales.address_space, linear_idx_type=scales.linear_idx_type]): Per-block FP8-E4M3 scales of shape [N, K//16], applied as abs(scale) over each 16-element K block.
ctx (DeviceContext): Device context used to enqueue the kernel onto the GPU.