Generation
Last updated
Last updated
Classes, functions, and utilities for generation.
Todo
Describe how to create a custom GenerationConfig.
static
⇐ Callable
⇐ Callable
⇐ LogitsProcessor
⇒ Tensor
⇐ LogitsProcessor
⇒ Object
⇐ LogitsProcessor
⇐ LogitsProcessor
⇒ Object
⇐ LogitsProcessor
⇒ Tensor
⇐ LogitsProcessor
⇒ Map.<string, Array<number>>
⇒ Array.<number>
⇒ Array.<number>
⇒ Object
⇐ LogitsProcessor
⇒ Object
⇐ LogitsProcessor
⇒ Object
⇐ LogitsProcessor
⇒ Object
instance
⇒ void
⇒ Array
⇒ number
static
⇒ Sampler
inner
⇐ Sampler
⇒ Array
⇐ Sampler
⇒ Array
⇐ Sampler
⇒ Array
A class representing a list of logits processors. A logits processor is a function that modifies the logits output of a language model. This class provides methods for adding new processors and applying all processors to a batch of logits.
Constructs a new instance of LogitsProcessorList.
Adds a new logits processor to the list.
item
LogitsProcessor
The logits processor function to add.
Adds multiple logits processors to the list.
items
Array.<LogitsProcessor>
The logits processor functions to add.
Applies all logits processors in the list to a batch of logits, modifying them in-place.
input_ids
Array.<number>
The input IDs for the language model.
batchedLogits
Array.<Array<number>>
A 2D array of logits, where each row corresponds to a single input sequence in the batch.
Base class for processing logits.
Apply the processor to the input logits.
Error Throws an error if `_call` is not implemented in the subclass.
input_ids
Array
The input ids.
logits
Tensor
The logits to process.
A logits processor that forces a specific token to be generated by the decoder.
Constructs a new instance of ForceTokensLogitsProcessor.
forced_decoder_ids
Array
The ids of tokens that should be forced.
Apply the processor to the input logits.
input_ids
Array
The input ids.
logits
Tensor
The logits to process.
A LogitsProcessor that forces a BOS token at the beginning of the generated sequence.
Create a ForcedBOSTokenLogitsProcessor.
bos_token_id
number
The ID of the beginning-of-sequence token to be forced.
Apply the BOS token forcing to the logits.
input_ids
Array
The input IDs.
logits
Object
The logits.
A logits processor that forces end-of-sequence token probability to 1.
Create a ForcedEOSTokenLogitsProcessor.
max_length
number
Max length of the sequence.
forced_eos_token_id
number | Array<number>
The ID of the end-of-sequence token to be forced.
Apply the processor to input_ids and logits.
input_ids
Array.<number>
The input ids.
logits
Tensor
The logits tensor.
A LogitsProcessor that suppresses a list of tokens as soon as the generate function starts generating using begin_index tokens. This should ensure that the tokens defined by begin_suppress_tokens at not sampled at the begining of the generation.
Create a SuppressTokensAtBeginLogitsProcessor.
begin_suppress_tokens
Array.<number>
The IDs of the tokens to suppress.
begin_index
number
The number of tokens to generate before suppressing tokens.
Apply the BOS token forcing to the logits.
input_ids
Array
The input IDs.
logits
Object
The logits.
A LogitsProcessor that handles adding timestamps to generated text.
Constructs a new WhisperTimeStampLogitsProcessor.
generate_config
Object
The config object passed to the generate() method of a transformer model.
generate_config.eos_token_id
number
The ID of the end-of-sequence token.
generate_config.no_timestamps_token_id
number
The ID of the token used to indicate that a token should not have a timestamp.
[generate_config.forced_decoder_ids]
Array.<Array<number>>
An array of two-element arrays representing decoder IDs that are forced to appear in the output. The second element of each array indicates whether the token is a timestamp.
[generate_config.max_initial_timestamp_index]
number
The maximum index at which an initial timestamp can appear.
Modify the logits to handle timestamp tokens.
input_ids
Array
The input sequence of tokens.
logits
Tensor
The logits output by the model.
A logits processor that disallows ngrams of a certain size to be repeated.
Create a NoRepeatNGramLogitsProcessor.
no_repeat_ngram_size
number
The no-repeat-ngram size. All ngrams of this size can only occur once.
Generate n-grams from a sequence of token ids.
prevInputIds
Array.<number>
List of previous input ids
Generate n-grams from a sequence of token ids.
bannedNgrams
Map.<string, Array<number>>
Map of banned n-grams
prevInputIds
Array.<number>
List of previous input ids
Calculate banned n-gram tokens
prevInputIds
Array.<number>
List of previous input ids
Apply the no-repeat-ngram processor to the logits.
input_ids
Array
The input IDs.
logits
Object
The logits.
A logits processor that penalises repeated output tokens.
Create a RepetitionPenaltyLogitsProcessor.
penalty
number
The penalty to apply for repeated tokens.
Apply the repetition penalty to the logits.
input_ids
Array
The input IDs.
logits
Object
The logits.
A logits processor that enforces a minimum number of tokens.
Create a MinLengthLogitsProcessor.
min_length
number
The minimum length below which the score of eos_token_id is set to negative infinity.
eos_token_id
number | Array<number>
The ID/IDs of the end-of-sequence token.
Apply logit processor.
input_ids
Array
The input IDs.
logits
Object
The logits.
A logits processor that enforces a minimum number of new tokens.
Create a MinNewTokensLengthLogitsProcessor.
prompt_length_to_skip
number
The input tokens length.
min_new_tokens
number
The minimum new tokens length below which the score of eos_token_id is set to negative infinity.
eos_token_id
number | Array<number>
The ID/IDs of the end-of-sequence token.
Apply logit processor.
input_ids
Array
The input IDs.
logits
Object
The logits.
Class that holds a configuration for a generation task.
Create a GenerationConfig object
[kwargs]
Object
{}
The configuration parameters. If not set, the default values are used.
[kwargs.max_length]
number
20
The maximum length the generated tokens can have. Corresponds to the length of the input prompt + max_new_tokens. Its effect is overridden by max_new_tokens, if also set.
[kwargs.max_new_tokens]
number
The maximum numbers of tokens to generate, ignoring the number of tokens in the prompt.
[kwargs.min_length]
number
0
The minimum length of the sequence to be generated. Corresponds to the length of the input prompt + min_new_tokens. Its effect is overridden by min_new_tokens, if also set.
[kwargs.min_new_tokens]
number
The minimum numbers of tokens to generate, ignoring the number of tokens in the prompt.
[kwargs.early_stopping]
boolean | "never"
false
Controls the stopping condition for beam-based methods, like beam-search. It accepts the following values:
true, where the generation stops as soon as there are num_beams complete candidates;
false, where an heuristic is applied and the generation stops when is it very unlikely to find better candidates;
"never", where the beam search procedure only stops when there cannot be better candidates (canonical beam search algorithm).
[kwargs.max_time]
number
The maximum amount of time you allow the computation to run for in seconds. Generation will still finish the current pass after allocated time has been passed.
[kwargs.do_sample]
boolean
false
Whether or not to use sampling; use greedy decoding otherwise.
[kwargs.num_beams]
number
1
Number of beams for beam search. 1 means no beam search.
[kwargs.num_beam_groups]
number
1
[kwargs.penalty_alpha]
number
The values balance the model confidence and the degeneration penalty in contrastive search decoding.
[kwargs.use_cache]
boolean
true
Whether or not the model should use the past last key/values attentions (if applicable to the model) to speed up decoding.
[kwargs.temperature]
number
1.0
The value used to modulate the next token probabilities.
[kwargs.top_k]
number
50
The number of highest probability vocabulary tokens to keep for top-k-filtering.
[kwargs.top_p]
number
1.0
If set to float < 1, only the smallest set of most probable tokens with probabilities that add up to top_p or higher are kept for generation.
[kwargs.typical_p]
number
1.0
[kwargs.epsilon_cutoff]
number
0.0
[kwargs.eta_cutoff]
number
0.0
[kwargs.diversity_penalty]
number
0.0
This value is subtracted from a beam's score if it generates a token same as any beam from other group at a particular time. Note that diversity_penalty is only effective if group beam search is enabled.
[kwargs.repetition_penalty]
number
1.0
[kwargs.encoder_repetition_penalty]
number
1.0
The paramater for encoder_repetition_penalty. An exponential penalty on sequences that are not in the original input. 1.0 means no penalty.
[kwargs.length_penalty]
number
1.0
Exponential penalty to the length that is used with beam-based generation. It is applied as an exponent to the sequence length, which in turn is used to divide the score of the sequence. Since the score is the log likelihood of the sequence (i.e. negative), length_penalty > 0.0 promotes longer sequences, while length_penalty < 0.0 encourages shorter sequences.
[kwargs.no_repeat_ngram_size]
number
0
If set to int > 0, all ngrams of that size can only occur once.
[kwargs.bad_words_ids]
Array.<Array<number>>
List of token ids that are not allowed to be generated. In order to get the token ids of the words that should not appear in the generated text, use (await tokenizer(bad_words, {add_prefix_space: true, add_special_tokens: false})).input_ids.
[kwargs.force_words_ids]
Array<Array<number>> | Array<Array<Array<number>>>
[kwargs.renormalize_logits]
boolean
false
Whether to renormalize the logits after applying all the logits processors or warpers (including the custom ones). It's highly recommended to set this flag to true as the search algorithms suppose the score logits are normalized but some logit processors or warpers break the normalization.
[kwargs.constraints]
Array.<Object>
Custom constraints that can be added to the generation to ensure that the output will contain the use of certain tokens as defined by Constraint objects, in the most sensible way possible.
[kwargs.forced_bos_token_id]
number
The id of the token to force as the first generated token after the decoder_start_token_id. Useful for multilingual models like mBART where the first generated token needs to be the target language token.
[kwargs.forced_eos_token_id]
number | Array<number>
The id of the token to force as the last generated token when max_length is reached. Optionally, use a list to set multiple end-of-sequence tokens.
[kwargs.remove_invalid_values]
boolean
false
Whether to remove possible nan and inf outputs of the model to prevent the generation method to crash. Note that using remove_invalid_values can slow down generation.
[kwargs.exponential_decay_length_penalty]
Array.<number>
This Tuple adds an exponentially increasing length penalty, after a certain amount of tokens have been generated. The tuple shall consist of: (start_index, decay_factor) where start_index indicates where penalty starts and decay_factor represents the factor of exponential decay.
[kwargs.suppress_tokens]
Array.<number>
A list of tokens that will be suppressed at generation. The SupressTokens logit processor will set their log probs to -inf so that they are not sampled.
[kwargs.begin_suppress_tokens]
Array.<number>
A list of tokens that will be suppressed at the beginning of the generation. The SupressBeginTokens logit processor will set their log probs to -inf so that they are not sampled.
[kwargs.forced_decoder_ids]
Array.<Array<number>>
A list of pairs of integers which indicates a mapping from generation indices to token indices that will be forced before sampling. For example, [[1, 123]] means the second generated token will always be a token of index 123.
[kwargs.num_return_sequences]
number
1
The number of independently computed returned sequences for each element in the batch.
[kwargs.output_attentions]
boolean
false
Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more details.
[kwargs.output_hidden_states]
boolean
false
Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more details.
[kwargs.output_scores]
boolean
false
Whether or not to return the prediction scores. See scores under returned tensors for more details.
[kwargs.return_dict_in_generate]
boolean
false
Whether or not to return a ModelOutput instead of a plain tuple.
[kwargs.pad_token_id]
number
The id of the padding token.
[kwargs.bos_token_id]
number
The id of the beginning-of-sequence token.
[kwargs.eos_token_id]
number | Array<number>
The id of the end-of-sequence token. Optionally, use a list to set multiple end-of-sequence tokens.
[kwargs.encoder_no_repeat_ngram_size]
number
0
If set to int > 0, all ngrams of that size that occur in the encoder_input_ids cannot occur in the decoder_input_ids.
[kwargs.decoder_start_token_id]
number
If an encoder-decoder model starts decoding with a different token than bos, the id of that token.
[kwargs.generation_kwargs]
Object
{}
Additional generation kwargs will be forwarded to the generate function of the model. Kwargs that are not present in generate's signature will be used in the model forward pass.
Sampler is a base class for all sampling methods used for text generation.
instance
static
Creates a new Sampler object with the specified generation config.
generation_config
GenerationConfig
The generation config.
Executes the sampler, using the specified logits.
logits
Tensor
index
number
Abstract method for sampling the logits.
Error
logits
Tensor
index
number
Returns the specified logits as an array, with temperature applied.
logits
Tensor
index
number
Selects an item randomly based on the specified probabilities.
probabilities
Array
An array of probabilities to use for selection.
Returns a Sampler object based on the specified options.
generation_config
GenerationConfig
An object containing options for the sampler.
Class representing a Greedy Sampler.
Sample the maximum probability of a given logits tensor.
logits
Tensor
[index]
number
-1
Class representing a MultinomialSampler.
Sample from the logits.
logits
Tensor
index
number
Class representing a BeamSearchSampler.
Sample from the logits.
logits
Tensor
index
number
Kind: static class of
Extends: Callable
⇐ Callable
Kind: instance method of
Kind: instance method of
Kind: instance method of
Kind: static class of
Extends: Callable
Kind: instance abstract method of Throws:
Kind: static class of
Extends: LogitsProcessor
⇐ LogitsProcessor
⇒ Tensor
Kind: instance method of
Returns: Tensor - The processed logits.
Kind: static class of
Extends: LogitsProcessor
⇐ LogitsProcessor
⇒ Object
Kind: instance method of
Returns: Object - The logits with BOS token forcing.
Kind: static class of
Extends: LogitsProcessor
⇐ LogitsProcessor
Kind: instance method of
Kind: static class of
Extends: LogitsProcessor
⇐ LogitsProcessor
⇒ Object
Kind: instance method of
Returns: Object - The logits with BOS token forcing.
Kind: static class of
Extends: LogitsProcessor
⇐ LogitsProcessor
⇒ Tensor
Kind: instance method of
Returns: Tensor - The modified logits.
Kind: static class of
Extends: LogitsProcessor
⇐ LogitsProcessor
⇒ Map.<string, Array<number>>
⇒ Array.<number>
⇒ Array.<number>
⇒ Object
Kind: instance method of
Returns: Map.<string, Array<number>> - Map of generated n-grams
Kind: instance method of
Returns: Array.<number> - Map of generated n-grams
Kind: instance method of
Returns: Array.<number> - Map of generated n-grams
Kind: instance method of
Returns: Object - The logits with no-repeat-ngram processing.
Kind: static class of
Extends: LogitsProcessor
⇐ LogitsProcessor
⇒ Object
Kind: instance method of
Returns: Object - The logits with repetition penalty processing.
Kind: static class of
Extends: LogitsProcessor
⇐ LogitsProcessor
⇒ Object
Kind: instance method of
Returns: Object - The processed logits.
Kind: static class of
Extends: LogitsProcessor
⇐ LogitsProcessor
⇒ Object
Kind: instance method of
Returns: Object - The processed logits.
Kind: static class of
Number of groups to divide num_beams into in order to ensure diversity among different groups of beams. See for more details.
Local typicality measures how similar the conditional probability of predicting a target token next is to the expected conditional probability of predicting a random token next, given the partial text already generated. If set to float < 1, the smallest set of the most locally typical tokens with probabilities that add up to typical_p or higher are kept for generation. See for more details.
If set to float strictly between 0 and 1, only tokens with a conditional probability greater than epsilon_cutoff will be sampled. In the paper, suggested values range from 3e-4 to 9e-4, depending on the size of the model. See for more details.
Eta sampling is a hybrid of locally typical sampling and epsilon sampling. If set to float strictly between 0 and 1, a token is only considered if it is greater than either eta_cutoff or sqrt(eta_cutoff) * exp(-entropy(softmax(next_token_logits))). The latter term is intuitively the expected next token probability, scaled by sqrt(eta_cutoff). In the paper, suggested values range from 3e-4 to 2e-3, depending on the size of the model. See for more details.
The parameter for repetition penalty. 1.0 means no penalty. See for more details.
List of token ids that must be generated. If given a number[][], this is treated as a simple list of words that must be included, the opposite to bad_words_ids. If given number[][][], this triggers a , where one can allow different forms of each word.
Kind: static class of
⇒ void
⇒ Array
⇒ number
⇒ Sampler
Kind: instance method of
Kind: instance method of Throws:
Kind: instance method of
Kind: instance method of
Returns: number - The index of the selected item.
Kind: static method of
Returns: Sampler - A Sampler object.
Kind: inner class of
Extends: Sampler
Kind: instance method of
Returns: Array - An array with a single tuple, containing the index of the maximum value and a meaningless score (since this is a greedy search).
Kind: inner class of
Extends: Sampler
Kind: instance method of
Kind: inner class of
Extends: Sampler
Kind: instance method of