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

fp4_materialize_kernel

def fp4_materialize_kernel[out_type: DType, w_layout: TensorLayout, s_layout: TensorLayout, out_layout: TensorLayout](out_w: TileTensor[out_type, out_layout, MutAnyOrigin], packed: TileTensor[DType.uint8, w_layout, ImmutAnyOrigin], scales: TileTensor[DType.float8_e4m3fn, s_layout, ImmutAnyOrigin])

Materializes the packed-FP4 weight into a dense [N, K] out_type buffer.

One thread per output element (n, k). packed is [N, K//2] (lo-nibble first), scales is [N, K//16]. Used by the Stage-1 oracle: it dequants the weight to bf16 so the EXISTING AppleM5MatMul can consume it, proving the dequant math against a host reference before the fused loader is written.

Parameters:

out_type (DType): Output element dtype (bf16 for the Apple W4A16 path).
w_layout (TensorLayout): TileTensor layout of packed, a plain rank-2 [N, K//2] uint8 buffer.
s_layout (TensorLayout): TileTensor layout of scales, a plain rank-2 [N, K//16] float8_e4m3fn buffer.
out_layout (TensorLayout): TileTensor layout of out_w, a plain rank-2 [N, K] buffer of out_type.

Args:

out_w (TileTensor[out_type, out_layout, MutAnyOrigin]): Output weight buffer of shape [N, K] and dtype out_type, written one element per thread.
packed (TileTensor[DType.uint8, w_layout, ImmutAnyOrigin]): 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, s_layout, ImmutAnyOrigin]): Per-block FP8-E4M3 scales of shape [N, K//16], applied as abs(scale) over each 16-element K block.