Quantizations¶

The Quantizations repository provides MyeLiN Codec implementations for weight storage. Each codec is a Compiler-time package export: encoded parameter bytes and graph decode behavior go into the deployment result, not the codec plugin library itself.

Exports in the supplied source¶

Package	Export	Description from the package contract
f32	f32	32-bit floating-point storage.
f16	f16	IEEE half precision.
bf16	bf16	Brain floating-point 16-bit storage.
int8	int8	Symmetric 8-bit storage with per-output-row scaling.
q4k	q4k	Q4 affine G32 representation.
q5k	q5k	GGML-style Q5_K, 256-value superblocks.
q6k	q6k	GGML-style Q6_K, 256-value superblocks.
q8	q8	Grouped 8-bit representation with graph decode.
kblock	kblock	Facade over Q5_K/Q6_K layouts.

The umbrella quantizations manifest exports the same codecs as separate build packages; each codec directory also has its own standalone synps.toml.

Use a codec¶

[dependencies]
f16 = { git = "https://github.com/useplx/Quantizations.git", version = "v1.0.0", subdir = "F16" }
q4k = { git = "https://github.com/useplx/Quantizations.git", version = "v1.0.0", subdir = "Q4K" }

[policies."model"]
dtype = "f16"

[policies."model".backend]
device = "local"

[[policies."model".schemes]]
format = "f16"

[[policies."model.layers.*.mlp.down_proj.weight".schemes]]
format = "q4k"

Design contract¶

PLeX's generic scheme specifies scope, lifetime, update behavior, and the codec export. Storage-specific representation belongs to each codec implementation. The codec's graph decode() remains visible to compiler optimization so decode and the consuming operation can be optimized together when the backend supports it.

This repository provides formats, not a universal recommendation that one format is fastest or most accurate on every model/device. Measure the finished artifact in the target application and hardware environment.