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

Mojo module

fp6_quantization

MXFP6 activation quantization.

Converts bfloat16 activations to packed MXFP6 (four 6-bit codes per three bytes) with one E8M0 scale per 32 elements, the layout the CDNA4 f8f6f4 block-scaled MFMA consumes and dequant_mxfp6 reverses.

The FP4/FP8 sibling (block_scaled_quantization.quantize_mx_amd) gives each scale group to four cooperating threads, since 32 FP4 elements are only 16 bytes. FP6 keeps one whole MX block per thread instead: 32 elements are exactly 24 bytes and one scale, so the group reduction stays in registers and the store is three aligned 8-byte writes with no cross-lane traffic and no partial-byte ownership.

comptime values

BYTES_PER_THREAD

comptime BYTES_PER_THREAD = 24

ELEMENTS_PER_THREAD

comptime ELEMENTS_PER_THREAD = MXFP6_SF_VECTOR_SIZE

Functions

• ​quantize_mxfp6_amd: Quantizes bfloat16 activations to MXFP6 with E8M0 block scales.