Class KvScheme

constructor

• new KvScheme(
      kind: KvSchemeKind,
      options: KvSchemeOptions,
      kernels?: AffineKernelChoice,
  ): KvScheme

  Parameters

  • kind: KvSchemeKind
  • options: KvSchemeOptions
  • kernels: AffineKernelChoice = ...

  Returns KvScheme

Readonlykind

Readonlyoptions

Readonlystorage

cacheKey

• get cacheKey(): string

  Returns string

fitOptions

• get fitOptions(): KvSchemeOptions

  Owned copies of the options for memory fit; TurboQuant fits as bf16.

  Returns KvSchemeOptions

label

• get label(): string

  Returns string

quantized

• get quantized(): boolean

  Returns boolean

bytesAt

• bytesAt(config: ModelConfig, tokens: number): number

  Parameters

  • config: ModelConfig
  • tokens: number

  Returns number

describe

• describe(
      config: ModelConfig,
  ): {
      attention: { global: number; sliding_window: number };
      layers: Record<string, number>;
      mode: string;
      recurrent_layers: number;
  }

  Configured attention storage, not live allocator occupancy. Recurrent layers retain their own state and never acquire an attention KV codec.

  Parameters

  • config: ModelConfig

  Returns {
      attention: { global: number; sliding_window: number };
      layers: Record<string, number>;
      mode: string;
      recurrent_layers: number;
  }