torch.js has not been released yet.

torch.DTypeSerializationConfig

export interface DTypeSerializationConfig<_Components extends number, _ComponentType extends DType> {
  /**
   * Safetensors dtype string for interop.
   * For custom dtypes, use format like 'CUSTOM:quaternion_f32'.
   */
  readonly safetensors_dtype?: string;

  /**
   * Custom serialization function.
   * If not provided, uses raw TypedArray bytes.
   *
   * @param data - The tensor data
   * @param numel - Number of logical elements
   * @returns Serialized data
   */
  serialize?(data: ArrayBufferView, numel: number): ArrayBuffer;

  /**
   * Custom deserialization function.
   * If not provided, interprets bytes as TypedArray.
   *
   * @param buffer - The serialized data
   * @param numel - Number of logical elements
   * @returns Deserialized data
   */
  deserialize?(buffer: ArrayBuffer, numel: number): ArrayBufferView;
}

_Componentsextends number

_ComponentTypeextends DType

readonlysafetensors_dtype(string)optional: – Safetensors dtype string for interop. For custom dtypes, use format like 'CUSTOM:quaternion_f32'.
serialize((data: ArrayBufferView, numel: number) => ArrayBuffer)optional: – Custom serialization function. If not provided, uses raw TypedArray bytes.
deserialize((buffer: ArrayBuffer, numel: number) => ArrayBufferView)optional: – Custom deserialization function. If not provided, interprets bytes as TypedArray.

Serialization configuration for custom dtypes.