Class DflashDrafter

Our trained DSpark module: block drafting over the target's multi-layer hidden context.

Implements

DsparkDrafterModel

Index

Properties

cfg dims m names targetId

Methods

dispose flatParams forwardRows forwardTrain get installParam projectContextRows save useParams dimsOf init initFromDims load

Properties

`Readonly`cfg

cfg: DflashConfig

`Readonly`dims

dims: TargetDims

`Readonly`m

m: number

`Readonly`names

names: string[]

`Readonly`targetId

targetId: string

Methods

dispose

dispose(): void
Returns void
Implementation of DsparkDrafterModel.dispose
- Defined in packages/inference/src/models/speculative/dflash.ts:540

flatParams

flatParams(): MlxArray[]
Returns MlxArray[]
- Defined in packages/inference/src/models/speculative/dflash.ts:189

forwardRows

forwardRows(
    model: DraftProjection,
    hCtx: MlxArray | null,
    anchors: readonly number[],
    gamma: number,
    opts?: DflashDraftOpts,
    context?: DsparkContextAttention,
): DflashDraftRows
Parameters
- model: DraftProjection
- hCtx: MlxArray | null
- anchors: readonly number[]
- gamma: number
- opts: DflashDraftOpts = {}
- Optionalcontext: DsparkContextAttention
Returns DflashDraftRows
Implementation of DsparkDrafterModel.forwardRows
- Defined in packages/inference/src/models/speculative/dflash.ts:437

forwardTrain

forwardTrain(
    model: DraftProjection,
    hCtx: MlxArray,
    ctxMask: MlxArray | null,
    anchorEmb: MlxArray,
    prevToks: MlxArray,
): DflashTrainOut
Parallel training forward (teacher-forced Markov). hCtx [A,Lctx,m*H], ctxMask [A,Lctx]|null, anchorEmb [A,H], prevToks [A,γ].
Parameters
- model: DraftProjection
- hCtx: MlxArray
- ctxMask: MlxArray | null
- anchorEmb: MlxArray
- prevToks: MlxArray
Returns DflashTrainOut
- Defined in packages/inference/src/models/speculative/dflash.ts:415

get

get(name: string): MlxArray
Parameters
- name: string
Returns MlxArray
- Defined in packages/inference/src/models/speculative/dflash.ts:188

installParam

installParam(i: number, p: MlxArray): void
Parameters
- i: number
- p: MlxArray
Returns void
- Defined in packages/inference/src/models/speculative/dflash.ts:195

projectContextRows

projectContextRows(hidden: MlxArray): { k: MlxArray; v: MlxArray }[]
Context projection is independent of the block and can be retained.
Parameters
- hidden: MlxArray
Returns { k: MlxArray; v: MlxArray }[]
Implementation of DsparkDrafterModel.projectContextRows
- Defined in packages/inference/src/models/speculative/dflash.ts:320

save

save(dir: string): void
Parameters
- dir: string
Returns void
- Defined in packages/inference/src/models/speculative/dflash.ts:516

useParams

useParams<T>(primals: MlxArray[], fn: () => T): T
Type Parameters
- T
Parameters
- primals: MlxArray[]
- fn: () => T
Returns T
- Defined in packages/inference/src/models/speculative/dflash.ts:190

`Static`dimsOf

dimsOf(model: Gemma4Model): TargetDims
Parameters
- model: Gemma4Model
Returns TargetDims
- Defined in packages/inference/src/models/speculative/dflash.ts:142

`Static`init

init(
    model: Gemma4Model,
    cfg: DflashConfig,
    targetId: string,
    seed?: number,
): DflashDrafter
Parameters
- model: Gemma4Model
- cfg: DflashConfig
- targetId: string
- seed: number = 0
Returns DflashDrafter
- Defined in packages/inference/src/models/speculative/dflash.ts:146

`Static`initFromDims

initFromDims(
    dims: TargetDims,
    cfg: DflashConfig,
    targetId: string,
    seed?: number,
): DflashDrafter
Parameters
- dims: TargetDims
- cfg: DflashConfig
- targetId: string
- seed: number = 0
Returns DflashDrafter
- Defined in packages/inference/src/models/speculative/dflash.ts:149

`Static`load

load(dir: string): DflashDrafter
Parameters
- dir: string
Returns DflashDrafter
- Defined in packages/inference/src/models/speculative/dflash.ts:527