//clika-runtime/io.clika.runtime/EncodeOptions
EncodeOptions
[common]
data class EncodeOptions(val addSpecialTokens: Boolean = true, val returnTensors: Boolean = false, val varlen: Boolean = false, val padId: Int? = null, val device: Placement? = null, val maxLength: Long? = null, val truncationSide: TruncationSide = TruncationSide.RIGHT, val paddingSide: PaddingSide = PaddingSide.RIGHT, val padToMultipleOf: Long? = null)
The per-call controls of the tensor-shaped encode entry points (Tokenizer.encodeText, Tokenizer.encodeBatch).
Constructors
| EncodeOptions | [common] constructor(addSpecialTokens: Boolean = true, returnTensors: Boolean = false, varlen: Boolean = false, padId: Int? = null, device: Placement? = null, maxLength: Long? = null, truncationSide: TruncationSide = TruncationSide.RIGHT, paddingSide: PaddingSide = PaddingSide.RIGHT, padToMultipleOf: Long? = null) |
Properties
| Name | Summary |
|---|---|
| addSpecialTokens | [common] val addSpecialTokens: Boolean = true wrap the ids with the model's special tokens (bos, eos, ...). |
| device | [common] val device: Placement? = null where the tensors land: a Device, or a Stream so they land stream-ordered on it; null takes the tokenizer's default placement. |
| maxLength | [common] val maxLength: Long? = null cap each sequence at this many tokens, trimming the text per truncationSide before the special tokens attach, so the total stays within the cap; null means no truncation. |
| paddingSide | [common] val paddingSide: PaddingSide which end a padded batch fills; the lengths always carry the real counts. |
| padId | [common] val padId: Int? = null the pad id of a padded batch; null takes the tokenizer's own. |
| padToMultipleOf | [common] val padToMultipleOf: Long? = null round the padded row length up to a multiple of this (padded layout only; refused with varlen). |
| returnTensors | [common] val returnTensors: Boolean = false build the tensor columns of Encoded beside the id rows. |
| truncationSide | [common] val truncationSide: TruncationSide |
| varlen | [common] val varlen: Boolean = false lay inputIds out as one flat row with a cuSeqlens offset table instead of a padded [B, S] matrix; varlen never pads. |