Skip to main content

Module mxfp

Module mxfp 

Source
Expand description

MXFP (OCP Microscaling) quantization descriptor (scheme tag 0x05).

A block quantization format standardised by the Open Compute Project Microscaling specification (OCP MX v1.0). Each contiguous block of elements along a chosen axis shares a single float8_e8m0 exponent-only scale stored in a dedicated buffer. Unlike per-block affine, MXFP requires exact divisibility — partial trailing blocks are not permitted.

See docs/spec/quantization/mxfp.md for the normative definition.

Structs§

Mxfp
Quantization parameters for MXFP (OCP Microscaling) block quantization.

Constants§

MXFP_CANONICAL_BLOCK_SIZE
OCP MX v1.0 canonical block size.
MXFP_MAX_BLOCK_SIZE
Maximum block size for MXFP (inclusive).
MXFP_MIN_BLOCK_SIZE
Minimum block size for MXFP (inclusive).