Affected by 
Documentation
¶
Overview ¶
Package quant holds the quantization format facts shared by the model importer, the runtime loader, and `ollama show`. It deliberately has no dependency on the MLX C library, so any package can use it without pulling in cgo — which is what keeps these facts from drifting between separate hand-maintained copies.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func Bits ¶
Bits returns the bit width of a recognized quantization type, or 0 if the type is empty or unrecognized. Unlike Params it applies no fallback, so callers that size or display tensors never act on an unknown type.
func Canonical ¶
Canonical returns the canonical name for a quantization type, resolving aliases (for example "FP8" and "Q8" both map to "int8"). It returns "" for the empty string and for any type it does not recognize.
func PackFactor ¶
PackFactor returns how many quantized values are packed into one 32-bit word, or 0 for an empty or unrecognized type. MLX stores quantized weights packed into U32 words, so a tensor's logical last dimension is its stored last dimension times this factor.
Types ¶
This section is empty.
Source Files