Module quantization 

Quantization descriptor types for the Hurray tensor format.

Every quantized tensor carries a quantization descriptor immediately following the tensor descriptor header. The descriptor starts with a 4-byte header (scheme tag, scheme version, flags) followed by a scheme-specific payload.

Supported schemes

Scheme tag	Tier	Type
0x01	1	PerTensorAffine
0x02	1	PerChannelAffine
0x03	1	PerBlockAffine
0x04	1	Nf4
0x05	1	Mxfp

Encoding overview

The canonical encoding path is:

1. Call QuantizationDescriptor::encode_into to write into a caller-owned buffer (zero-alloc, required by Layer 5 streaming writers).
2. Or call QuantizationDescriptor::encode_to_vec for a convenience Vec<u8>.

The canonical decoding path is:

• Call QuantizationDescriptor::decode which returns the descriptor and the number of bytes consumed, allowing streaming readers to advance their cursor without re-scanning.

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

Re-exports

pub use mxfp::Mxfp;

pub use mxfp::MXFP_CANONICAL_BLOCK_SIZE;

pub use mxfp::MXFP_MAX_BLOCK_SIZE;

pub use mxfp::MXFP_MIN_BLOCK_SIZE;

pub use nf4::Nf4;

pub use nf4::NF4_LUT;

pub use nf4::NF4_MIN_BLOCK_SIZE;

pub use per_block_affine::PerBlockAffine;

pub use per_block_affine::PER_BLOCK_AFFINE_MIN_BLOCK_SIZE;

pub use per_channel_affine::PerChannelAffine;

pub use per_tensor_affine::PerTensorAffine;

Modules

mxfp

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

nf4

NF4 (NormalFloat4) quantization descriptor (scheme tag 0x04).

per_block_affine

Per-block affine quantization descriptor (scheme tag 0x03).

per_channel_affine

Per-channel affine quantization descriptor (scheme tag 0x02).

per_tensor_affine

Per-tensor affine quantization descriptor (scheme tag 0x01).

Enums

QuantizationDescriptor

A typed quantization descriptor carrying the parameters for one of the supported quantization schemes.

QuantizationSchemeTag

The wire scheme tag that identifies which quantization scheme a descriptor uses.

Functions

validate_axis

Validates that axis is a valid axis index for a tensor of the given rank.

validate_buffer_placement

Validates that all quantization-parameter buffer indices in desc are valid within the buffers table and do not alias the tensor data buffer.