decoder

type Decoder struct {
	// contains filtered or unexported fields
}

type DecodingOptions struct {
	// MaxLen is the maximum number of tokens to generate.
	MaxLen int `json:"max_len" yaml:"max_len"`
	// MinLen is the minimum number of tokens to generate.
	MinLen int `json:"min_len" yaml:"min_len"`
	// StopSequencesIDs is a list of token ids that if generated, the generation process will stop.
	StopSequencesIDs [][]int `json:"stop_sequences_ids" yaml:"stop_sequences_ids"`
	// EndTokenID is the end-of-sequence token (default: 0).
	EndTokenID int `json:"end_token_id" yaml:"end_token_id"`
	// SkipEndTokenID when true, the end token is not added to the generated sequence.
	SkipEndTokenID bool `json:"skip_end_token_id" yaml:"skip_end_token_id"`
	// Temperature is the temperature used to control the randomness of the generated text.
	Temp float64 `json:"temp" yaml:"temp"`
	// TopK is the number of tokens to consider when sampling the next token.
	TopK int `json:"top_k" yaml:"top_k"`
	// TopP is the cumulative probability of the tokens to consider when sampling the next token.
	TopP float64 `json:"top_p" yaml:"top_p"`
	// UseSampling uses sampling to generate the next token.
	UseSampling bool `json:"use_sampling" yaml:"use_sampling"`
}

type GeneratedToken struct {
	// TokenID is the ID of the token predicted by the decoder at the current step.
	TokenID int
	// SumNegLogProbs is the sum of the negative log probabilities up to the current step.
	SumNegLogProbs float64
}

type OutputDiversityControlFunc func(logits mat.Matrix) (mat.Matrix, error)

type OutputSelectionFunc func(logits mat.Matrix) (int, float64, error)