Class TrellisEncoder

One codec per distinct k (the L=12 1MAD codebook is shared; k only changes the branching), created lazily so a uniform run builds exactly one.

Index

Constructors

Properties

L T

Methods

codec dispose encode encodeLdlq

Constructors

constructor

new TrellisEncoder(opts?: TrellisEncoderOptions): TrellisEncoder
Parameters
- opts: TrellisEncoderOptions = {}
Returns TrellisEncoder
- Defined in packages/quantize/src/trellis-encoder.ts:135

Properties

`Readonly`L

L: 12 = TRELLIS_L

`Readonly`T

T: 256 = TRELLIS_T

Methods

codec

codec(k: number): Trellis
Parameters
- k: number
Returns Trellis
- Defined in packages/quantize/src/trellis-encoder.ts:140

dispose

dispose(): void
Returns void
- Defined in packages/quantize/src/trellis-encoder.ts:146

encode

encode(folded: MlxArray, axis: 0 | 1, k: number): PackedTrellisTensor
Unweighted encoding of a folded [out, in] weight along axis.
Parameters
- folded: MlxArray
- axis: 0 | 1
- k: number
Returns PackedTrellisTensor
- Defined in packages/quantize/src/trellis-encoder.ts:152

encodeLdlq

encodeLdlq(folded: MlxArray, axis: 0 | 1, Lf: MlxArray, k: number): LdlqEncoding
BlockLDLQ encoding of a folded weight against Hessian factor Lf ([n, n], n = the LDLQ dim). Falls back to the unweighted encoding, with tripped set, when the feedback grows past the guard ceiling.
Parameters
- folded: MlxArray
- axis: 0 | 1
- Lf: MlxArray
- k: number
Returns LdlqEncoding
- Defined in packages/quantize/src/trellis-encoder.ts:193