Class AdamW

Index

Constructors

Accessors

Methods

dispose evalState getParam step

Constructors

constructor

new AdamW(
    params: MlxArray[],
    opts: AdamWOptions,
    write?: (i: number, p: MlxArray) => void,
): AdamW
Parameters
- params: MlxArray[]
  live param arrays (e.g. flatParams(lora)).
- opts: AdamWOptions
- Optionalwrite: (i: number, p: MlxArray) => void
  callback to push a replaced param back into its owner (e.g. lw.a / lw.b). If omitted, only the internal array is updated (the caller must read params via getParam).
Returns AdamW
- Defined in packages/training/src/optimizer.ts:49

Accessors

lr

get lr(): number
Returns number
- Defined in packages/training/src/optimizer.ts:71
set lr(value: number): void
Parameters
- value: number
Returns void
- Defined in packages/training/src/optimizer.ts:68

step_count

get step_count(): number
Returns number
- Defined in packages/training/src/optimizer.ts:75

Methods

dispose

dispose(): void
Returns void
- Defined in packages/training/src/optimizer.ts:150

evalState

evalState(): void
Eval all optimizer state + params (call once per step after step).

Returns void
- Defined in packages/training/src/optimizer.ts:146

getParam

getParam(i: number): MlxArray
Current parameter array at index i (after the latest step).
Parameters
- i: number
Returns MlxArray
- Defined in packages/training/src/optimizer.ts:64

step

step(grads: MlxArray[]): void
Apply one AdamW update. grads[i] is dL/dparams[i]. Takes ownership of the grads (they are disposed here).
Parameters
- grads: MlxArray[]
Returns void
- Defined in packages/training/src/optimizer.ts:81