Hugging Face's logo

Transformers documentation

Utilities for Generation

You are viewing v4.43.0 version. A newer version v4.46.0 is available.
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Utilities for Generation

This page lists all the utility functions used by generate().

Generate Outputs

The output of generate() is an instance of a subclass of ModelOutput. This output is a data structure containing all the information returned by generate(), but that can also be used as tuple or dictionary.

Here’s an example:

from transformers import GPT2Tokenizer, GPT2LMHeadModel

tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2")
model = GPT2LMHeadModel.from_pretrained("openai-community/gpt2")

inputs = tokenizer("Hello, my dog is cute and ", return_tensors="pt")
generation_output = model.generate(**inputs, return_dict_in_generate=True, output_scores=True)

The generation_output object is a GenerateDecoderOnlyOutput, as we can see in the documentation of that class below, it means it has the following attributes:

  • sequences: the generated sequences of tokens
  • scores (optional): the prediction scores of the language modelling head, for each generation step
  • hidden_states (optional): the hidden states of the model, for each generation step
  • attentions (optional): the attention weights of the model, for each generation step

Here we have the scores since we passed along output_scores=True, but we don’t have hidden_states and attentions because we didn’t pass output_hidden_states=True or output_attentions=True.

You can access each attribute as you would usually do, and if that attribute has not been returned by the model, you will get None. Here for instance generation_output.scores are all the generated prediction scores of the language modeling head, and generation_output.attentions is None.

When using our generation_output object as a tuple, it only keeps the attributes that don’t have None values. Here, for instance, it has two elements, loss then logits, so

generation_output[:2]

will return the tuple (generation_output.sequences, generation_output.scores) for instance.

When using our generation_output object as a dictionary, it only keeps the attributes that don’t have None values. Here, for instance, it has two keys that are sequences and scores.

We document here all output types.

PyTorch

class transformers.generation.GenerateDecoderOnlyOutput

< >

( sequences: LongTensor = None scores: Optional = None logits: Optional = None attentions: Optional = None hidden_states: Optional = None past_key_values: Optional = None )

Parameters

  • sequences (torch.LongTensor of shape (batch_size, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • scores (tuple(torch.FloatTensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of torch.FloatTensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • logits (tuple(torch.FloatTensor) optional, returned when output_logits=True is passed or when config.output_logits=True) — Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of torch.FloatTensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • attentions (tuple(tuple(torch.FloatTensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • hidden_states (tuple(tuple(torch.FloatTensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size, generated_length, hidden_size).
  • past_key_values (tuple(tuple(torch.FloatTensor))), optional, returned when use_cache=True is passed or when config.use_cache=True) — NOTE: some models have a different past_key_values format, confirm with the model’s documentation. Usually a Tuple (one element for each layer of the decoder) of tuples (two elements, key tensor and value tensor). The first Tuple is of length config.n_layers, with each tuple having 2 tensors of shape (batch_size, num_heads, sequence_length, embed_size_per_head)) and optionally if config.is_encoder_decoder=True 2 additional tensors of shape (batch_size, num_heads, encoder_sequence_length, embed_size_per_head).

Outputs of decoder-only generation models, when using non-beam methods.

class transformers.generation.GenerateEncoderDecoderOutput

< >

( sequences: LongTensor = None scores: Optional = None logits: Optional = None encoder_attentions: Optional = None encoder_hidden_states: Optional = None decoder_attentions: Optional = None cross_attentions: Optional = None decoder_hidden_states: Optional = None past_key_values: Optional = None )

Parameters

  • sequences (torch.LongTensor of shape (batch_size*num_return_sequences, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • scores (tuple(torch.FloatTensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of torch.FloatTensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • logits (tuple(torch.FloatTensor) optional, returned when output_logits=True is passed or when config.output_logits=True) — Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of torch.FloatTensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer of the decoder) of shape (batch_size, num_heads, sequence_length, sequence_length).
  • encoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).
  • decoder_attentions (tuple(tuple(torch.FloatTensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • cross_attentions (tuple(tuple(torch.FloatTensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • decoder_hidden_states (tuple(tuple(torch.FloatTensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size, generated_length, hidden_size).
  • past_key_values (tuple(tuple(torch.FloatTensor))), optional, returned when use_cache=True is passed or when config.use_cache=True) — NOTE: some models have a different past_key_values format, confirm with the model’s documentation. Usually a Tuple (one element for each layer of the decoder) of tuples (two elements, key tensor and value tensor). The first Tuple is of length config.n_layers, with each tuple having 2 tensors of shape (batch_size, num_heads, sequence_length, embed_size_per_head)) and optionally if config.is_encoder_decoder=True 2 additional tensors of shape (batch_size, num_heads, encoder_sequence_length, embed_size_per_head).

Outputs of encoder-decoder generation models, when using non-beam methods.

class transformers.generation.GenerateBeamDecoderOnlyOutput

< >

( sequences: LongTensor = None sequences_scores: Optional = None scores: Optional = None logits: Optional = None beam_indices: Optional = None attentions: Optional = None hidden_states: Optional = None past_key_values: Optional = None )

Parameters

  • sequences (torch.LongTensor of shape (batch_size*num_return_sequences, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • sequences_scores (torch.FloatTensor of shape (batch_size*num_return_sequences), optional, returned when output_scores=True is passed or when config.output_scores=True) — Final beam scores of the generated sequences.
  • scores (tuple(torch.FloatTensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam. Tuple of torch.FloatTensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size*num_beams, config.vocab_size).
  • logits (tuple(torch.FloatTensor) optional, returned when output_logits=True is passed or when config.output_logits=True) — Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of torch.FloatTensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • beam_indices (torch.LongTensor, optional, returned when output_scores=True is passed or when config.output_scores=True) — Beam indices of generated token id at each generation step. torch.LongTensor of shape (batch_size*num_return_sequences, sequence_length).
  • attentions (tuple(tuple(torch.FloatTensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size*num_beams, num_heads, generated_length, sequence_length).
  • hidden_states (tuple(tuple(torch.FloatTensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size*num_beams*num_return_sequences, generated_length, hidden_size).
  • past_key_values (tuple(tuple(torch.FloatTensor))), optional, returned when use_cache=True is passed or when config.use_cache=True) — NOTE: some models have a different past_key_values format, confirm with the model’s documentation. Usually a Tuple (one element for each layer of the decoder) of tuples (two elements, key tensor and value tensor). The first Tuple is of length config.n_layers, with each tuple having 2 tensors of shape (batch_size, num_heads, sequence_length, embed_size_per_head)) and optionally if config.is_encoder_decoder=True 2 additional tensors of shape (batch_size, num_heads, encoder_sequence_length, embed_size_per_head).

Outputs of decoder-only generation models, when using beam methods.

class transformers.generation.GenerateBeamEncoderDecoderOutput

< >

( sequences: LongTensor = None sequences_scores: Optional = None scores: Optional = None logits: Optional = None beam_indices: Optional = None encoder_attentions: Optional = None encoder_hidden_states: Optional = None decoder_attentions: Optional = None cross_attentions: Optional = None decoder_hidden_states: Optional = None past_key_values: Optional = None )

Parameters

  • sequences (torch.LongTensor of shape (batch_size*num_return_sequences, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • sequences_scores (torch.FloatTensor of shape (batch_size*num_return_sequences), optional, returned when output_scores=True is passed or when config.output_scores=True) — Final beam scores of the generated sequences.
  • scores (tuple(torch.FloatTensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam. Tuple of torch.FloatTensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size*num_beams, config.vocab_size).
  • logits (tuple(torch.FloatTensor) optional, returned when output_logits=True is passed or when config.output_logits=True) — Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of torch.FloatTensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • beam_indices (torch.LongTensor, optional, returned when output_scores=True is passed or when config.output_scores=True) — Beam indices of generated token id at each generation step. torch.LongTensor of shape (batch_size*num_return_sequences, sequence_length).
  • encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer of the decoder) of shape (batch_size, num_heads, sequence_length, sequence_length).
  • encoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size*num_beams*num_return_sequences, sequence_length, hidden_size).
  • decoder_attentions (tuple(tuple(torch.FloatTensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size*num_beams*num_return_sequences, num_heads, generated_length, sequence_length).
  • cross_attentions (tuple(tuple(torch.FloatTensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • decoder_hidden_states (tuple(tuple(torch.FloatTensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of torch.FloatTensor of shape (batch_size*num_beams*num_return_sequences, generated_length, hidden_size).
  • past_key_values (tuple(tuple(torch.FloatTensor))), optional, returned when use_cache=True is passed or when config.use_cache=True) — NOTE: some models have a different past_key_values format, confirm with the model’s documentation. Usually a Tuple (one element for each layer of the decoder) of tuples (two elements, key tensor and value tensor). The first Tuple is of length config.n_layers, with each tuple having 2 tensors of shape (batch_size, num_heads, sequence_length, embed_size_per_head)) and optionally if config.is_encoder_decoder=True 2 additional tensors of shape (batch_size, num_heads, encoder_sequence_length, embed_size_per_head).

Outputs of encoder-decoder generation models, when using beam methods.

TensorFlow

class transformers.generation.TFGreedySearchEncoderDecoderOutput

< >

( sequences: Tensor = None scores: Optional = None encoder_attentions: Optional = None encoder_hidden_states: Optional = None decoder_attentions: Optional = None cross_attentions: Optional = None decoder_hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of tf.Tensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple of tf.Tensor (one for each layer of the decoder) of shape (batch_size, num_heads, sequence_length, sequence_length).
  • encoder_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).
  • decoder_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • cross_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • decoder_hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, generated_length, hidden_size).

Base class for outputs of encoder-decoder generation models using greedy search. Hidden states and attention weights of the decoder (respectively the encoder) can be accessed via the encoder_attentions and the encoder_hidden_states attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)

class transformers.generation.TFGreedySearchDecoderOnlyOutput

< >

( sequences: Tensor = None scores: Optional = None attentions: Optional = None hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of tf.Tensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, generated_length, hidden_size).

Base class for outputs of decoder-only generation models using greedy search.

class transformers.generation.TFSampleEncoderDecoderOutput

< >

( sequences: Tensor = None scores: Optional = None encoder_attentions: Optional = None encoder_hidden_states: Optional = None decoder_attentions: Optional = None cross_attentions: Optional = None decoder_hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size*num_return_sequences, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of tf.Tensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size*num_return_sequences, config.vocab_size).
  • encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple of tf.Tensor (one for each layer of the decoder) of shape (batch_size*num_return_sequences, num_heads, sequence_length, sequence_length).
  • encoder_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size*num_return_sequences, sequence_length, hidden_size).
  • decoder_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_return_sequences, num_heads, generated_length, sequence_length).
  • cross_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • decoder_hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_return_sequences, generated_length, hidden_size).

Base class for outputs of encoder-decoder generation models using sampling. Hidden states and attention weights of the decoder (respectively the encoder) can be accessed via the encoder_attentions and the encoder_hidden_states attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)

class transformers.generation.TFSampleDecoderOnlyOutput

< >

( sequences: Tensor = None scores: Optional = None attentions: Optional = None hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size*num_return_sequences, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of tf.Tensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size*num_return_sequences, config.vocab_size).
  • attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (num_return_sequences*batch_size, num_heads, generated_length, sequence_length).
  • hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (num_return_sequences*batch_size, generated_length, hidden_size).

Base class for outputs of decoder-only generation models using sampling.

class transformers.generation.TFBeamSearchEncoderDecoderOutput

< >

( sequences: Tensor = None sequences_scores: Optional = None scores: Optional = None beam_indices: Optional = None encoder_attentions: Optional = None encoder_hidden_states: Optional = None decoder_attentions: Optional = None cross_attentions: Optional = None decoder_hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size*num_return_sequences, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • sequences_scores (tf.Tensor of shape (batch_size*num_return_sequences), optional, returned when output_scores=True is passed or when config.output_scores=True) — Final beam scores of the generated sequences.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam. Tuple of tf.Tensorwith up tomax_new_tokenselements (one element for each generated token), with each tensor of shape(batch_size*num_beams, config.vocab_size)`.
  • beam_indices (tf.Tensor, optional, returned when output_scores=True is passed or when config.output_scores=True) — Beam indices of generated token id at each generation step. tf.Tensor of shape (batch_size*num_return_sequences, sequence_length).
  • encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple of tf.Tensor (one for each layer of the decoder) of shape (batch_size, num_heads, sequence_length, sequence_length).
  • encoder_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size*num_beams*num_return_sequences, sequence_length, hidden_size).
  • decoder_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_beams*num_return_sequences, num_heads, generated_length, sequence_length).
  • cross_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • decoder_hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_beams*num_return_sequences, generated_length, hidden_size).

Base class for outputs of encoder-decoder generation models using beam search. Hidden states and attention weights of the decoder (respectively the encoder) can be accessed via the encoder_attentions and the encoder_hidden_states attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)

class transformers.generation.TFBeamSearchDecoderOnlyOutput

< >

( sequences: Tensor = None sequences_scores: Optional = None scores: Optional = None beam_indices: Optional = None attentions: Optional = None hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size*num_return_sequences, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • sequences_scores (tf.Tensor of shape (batch_size*num_return_sequences), optional, returned when output_scores=True is passed or when config.output_scores=True) — Final beam scores of the generated sequences.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam. Tuple of tf.Tensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size*num_beams*num_return_sequences, config.vocab_size).
  • beam_indices (tf.Tensor, optional, returned when output_scores=True is passed or when config.output_scores=True) — Beam indices of generated token id at each generation step. tf.Tensor of shape (batch_size*num_return_sequences, sequence_length).
  • attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_beams, num_heads, generated_length, sequence_length).
  • hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_beams*num_return_sequences, generated_length, hidden_size).

Base class for outputs of decoder-only generation models using beam search.

class transformers.generation.TFBeamSampleEncoderDecoderOutput

< >

( sequences: Tensor = None sequences_scores: Optional = None scores: Optional = None beam_indices: Optional = None encoder_attentions: Optional = None encoder_hidden_states: Optional = None decoder_attentions: Optional = None cross_attentions: Optional = None decoder_hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size*num_beams, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • sequences_scores (tf.Tensor of shape (batch_size * num_return_sequence), optional, returned when output_scores=True is passed or when config.output_scores=True) — Final beam scores of the generated sequences.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam. Tuple of tf.Tensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size*num_beams, config.vocab_size).
  • beam_indices (tf.Tensor, optional, returned when output_scores=True is passed or when config.output_scores=True) — Beam indices of generated token id at each generation step. tf.Tensor of shape (batch_size*num_return_sequences, sequence_length).
  • encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple of tf.Tensor (one for each layer of the decoder) of shape (batch_size, num_heads, sequence_length, sequence_length).
  • encoder_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size*num_beams, sequence_length, hidden_size).
  • decoder_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_beams, num_heads, generated_length, sequence_length).
  • cross_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • decoder_hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_beams, generated_length, hidden_size).

Base class for outputs of encoder-decoder generation models using beam sampling. Hidden states and attention weights of the decoder (respectively the encoder) can be accessed via the encoder_attentions and the encoder_hidden_states attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)

class transformers.generation.TFBeamSampleDecoderOnlyOutput

< >

( sequences: Tensor = None sequences_scores: Optional = None scores: Optional = None beam_indices: Optional = None attentions: Optional = None hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size*num_return_sequences, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • sequences_scores (tf.Tensor of shape (batch_size * num_return_sequence), optional, returned when output_scores=True is passed or when config.output_scores=True) — Final beam scores of the generated sequences.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam. Tuple of tf.Tensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size*num_beams*num_return_sequences, config.vocab_size).
  • beam_indices (tf.Tensor, optional, returned when output_scores=True is passed or when config.output_scores=True) — Beam indices of generated token id at each generation step. tf.Tensor of shape (batch_size*num_return_sequences, sequence_length).
  • attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_beams, num_heads, generated_length, sequence_length).
  • hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size*num_beams, generated_length, hidden_size).

Base class for outputs of decoder-only generation models using beam sample.

class transformers.generation.TFContrastiveSearchEncoderDecoderOutput

< >

( sequences: Tensor = None scores: Optional = None encoder_attentions: Optional = None encoder_hidden_states: Optional = None decoder_attentions: Optional = None cross_attentions: Optional = None decoder_hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of tf.Tensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple of tf.Tensor (one for each layer of the decoder) of shape (batch_size, num_heads, sequence_length, sequence_length).
  • encoder_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).
  • decoder_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • cross_attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • decoder_hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, generated_length, hidden_size).

Base class for outputs of encoder-decoder generation models using contrastive search. Hidden states and attention weights of the decoder (respectively the encoder) can be accessed via the encoder_attentions and the encoder_hidden_states attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)

class transformers.generation.TFContrastiveSearchDecoderOnlyOutput

< >

( sequences: Tensor = None scores: Optional = None attentions: Optional = None hidden_states: Optional = None )

Parameters

  • sequences (tf.Tensor of shape (batch_size, sequence_length)) — The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.
  • scores (tuple(tf.Tensor) optional, returned when output_scores=True is passed or when config.output_scores=True) — Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of tf.Tensor with up to max_new_tokens elements (one element for each generated token), with each tensor of shape (batch_size, config.vocab_size).
  • attentions (tuple(tuple(tf.Tensor)), optional, returned when output_attentions=True is passed or config.output_attentions=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, num_heads, generated_length, sequence_length).
  • hidden_states (tuple(tuple(tf.Tensor)), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of tf.Tensor of shape (batch_size, generated_length, hidden_size).

Base class for outputs of decoder-only generation models using contrastive search.

FLAX

class transformers.generation.FlaxSampleOutput

< >

( sequences: Array = None )

Parameters

  • sequences (jnp.ndarray of shape (batch_size, max_length)) — The generated sequences.

Flax Base class for outputs of decoder-only generation models using sampling.

replace

< >

( **updates )

“Returns a new object replacing the specified fields with new values.

class transformers.generation.FlaxGreedySearchOutput

< >

( sequences: Array = None )

Parameters

  • sequences (jnp.ndarray of shape (batch_size, max_length)) — The generated sequences.

Flax Base class for outputs of decoder-only generation models using greedy search.

replace

< >

( **updates )

“Returns a new object replacing the specified fields with new values.

class transformers.generation.FlaxBeamSearchOutput

< >

( sequences: Array = None scores: Array = None )

Parameters

  • sequences (jnp.ndarray of shape (batch_size, max_length)) — The generated sequences.
  • scores (jnp.ndarray of shape (batch_size,)) — The scores (log probabilities) of the generated sequences.

Flax Base class for outputs of decoder-only generation models using greedy search.

replace

< >

( **updates )

“Returns a new object replacing the specified fields with new values.

LogitsProcessor

A LogitsProcessor can be used to modify the prediction scores of a language model head for generation.

PyTorch

class transformers.AlternatingCodebooksLogitsProcessor

< >

( input_start_len: int semantic_vocab_size: int codebook_size: int )

Parameters

  • input_start_len (int) — The length of the initial input sequence.
  • semantic_vocab_size (int) — Vocabulary size of the semantic part, i.e number of tokens associated to the semantic vocabulary.
  • codebook_size (int) — Number of tokens associated to the codebook.

LogitsProcessor enforcing alternated generation between the two codebooks of Bark.

This logits processor is exclusively compatible with Bark’s fine submodel. See the model documentation for examples.

__call__

< >

( input_ids: LongTensor scores: FloatTensor )

class transformers.ClassifierFreeGuidanceLogitsProcessor

< >

( guidance_scale )

Parameters

  • guidance_scale (float) — The guidance scale for classifier free guidance (CFG). CFG is enabled by setting guidance_scale > 1. Higher guidance scale encourages the model to generate samples that are more closely linked to the input prompt, usually at the expense of poorer quality.

LogitsProcessor for classifier free guidance (CFG). The scores are split over the batch dimension, where the first half correspond to the conditional logits (predicted from the input prompt) and the second half correspond to the unconditional logits (predicted from an empty or ‘null’ prompt). The processor computes a weighted average across the conditional and unconditional logits, parameterised by the guidance_scale.

See the paper for more information.

This logits processor is exclusively compatible with MusicGen

Examples:

>>> from transformers import AutoProcessor, MusicgenForConditionalGeneration

>>> processor = AutoProcessor.from_pretrained("facebook/musicgen-small")
>>> model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")

>>> inputs = processor(
...     text=["80s pop track with bassy drums and synth", "90s rock song with loud guitars and heavy drums"],
...     padding=True,
...     return_tensors="pt",
... )
>>> audio_values = model.generate(**inputs, do_sample=True, guidance_scale=3, max_new_tokens=256)

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.EncoderNoRepeatNGramLogitsProcessor

< >

( encoder_ngram_size: int encoder_input_ids: LongTensor )

Parameters

  • encoder_ngram_size (int) — All ngrams of size ngram_size can only occur within the encoder input ids.
  • encoder_input_ids (int) — The encoder_input_ids that should not be repeated within the decoder ids.

LogitsProcessor that works similarly to NoRepeatNGramLogitsProcessor, but applied exclusively to prevent the repetition of n-grams present in the prompt.

It was designed to promote chattiness in a language model, by preventing the generation of n-grams present in previous conversation rounds.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> model = AutoModelForCausalLM.from_pretrained("bigscience/bloomz-560m")
>>> tokenizer = AutoTokenizer.from_pretrained("bigscience/bloomz-560m")

>>> inputs = tokenizer("Alice: I love cats. What do you love?\nBob:", return_tensors="pt")

>>> # With greedy decoding, we see Bob repeating Alice's opinion. If Bob was a chatbot, it would be a poor one.
>>> outputs = model.generate(**inputs)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
Alice: I love cats. What do you love?
Bob: I love cats. What do you

>>> # With this logits processor, we can prevent Bob from repeating Alice's opinion.
>>> outputs = model.generate(**inputs, encoder_no_repeat_ngram_size=2)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
Alice: I love cats. What do you love?
Bob: My cats are very cute.

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.EncoderRepetitionPenaltyLogitsProcessor

< >

( penalty: float encoder_input_ids: LongTensor )

Parameters

  • penalty (float) — The parameter for repetition penalty. 1.0 means no penalty. Above 1.0 rewards prompt tokens. Between 0.0 and 1.0 penalizes prompt tokens.
  • encoder_input_ids (torch.LongTensor) — The encoder_input_ids that should be repeated within the decoder ids.

LogitsProcessor that works similarly to RepetitionPenaltyLogitsProcessor, but with an inverse penalty that is applied to the tokens present in the prompt. In other words, a penalty above 1.0 increases the odds of selecting tokens that were present in the prompt.

It was designed to avoid hallucination in input-grounded tasks, like summarization. Although originally intended for encoder-decoder models, it can also be used with decoder-only models like LLMs.

Examples:

>>> from transformers import AutoModelForCausalLM, AutoTokenizer

>>> tokenizer = AutoTokenizer.from_pretrained("bigscience/bloomz-560m")
>>> model = AutoModelForCausalLM.from_pretrained("bigscience/bloomz-560m")

>>> inputs = tokenizer(["Alice and Bob. The third member's name was"], return_tensors="pt")
>>> gen_out = model.generate(**inputs)
>>> print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
Alice and Bob. The third member's name was not mentioned.

>>> # With the `encoder_repetition_penalty` argument we can trigger this logits processor in `generate`, which can
>>> # promote the use of prompt tokens ("Bob" in this example)
>>> gen_out = model.generate(**inputs, encoder_repetition_penalty=1.2)
>>> print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
Alice and Bob. The third member's name was Bob. The third member's name was Bob.

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.EpsilonLogitsWarper

< >

( epsilon: float filter_value: float = -inf min_tokens_to_keep: int = 1 )

Parameters

  • epsilon (float) — If set to > 0, only the most tokens with probabilities epsilon or higher are kept for generation.
  • filter_value (float, optional, defaults to -inf) — All filtered values will be set to this float value.
  • min_tokens_to_keep (int, optional, defaults to 1) — Minimum number of tokens that cannot be filtered.

LogitsWarper that performs epsilon-sampling, i.e. restricting to tokens with prob >= epsilon. Takes the largest min_tokens_to_keep tokens if no tokens satisfy this constraint. See Truncation Sampling as Language Model Desmoothing for more information.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed

>>> set_seed(1)
>>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")

>>> inputs = tokenizer("A sequence: 1, 2", return_tensors="pt")

>>> # With sampling, the output is unexpected -- sometimes too unexpected.
>>> outputs = model.generate(**inputs, do_sample=True)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: 1, 2, 3 | < 4 (left-hand pointer) ;
<BLANKLINE>
<BLANKLINE>

>>> # With epsilon sampling, the output gets restricted to high-probability tokens. Note that this is similar to
>>> # Top P sampling, which restricts tokens based on their cumulative probability.
>>> # Pro tip: The paper recomends using `epsilon_cutoff` values between 3e-4 and 9e-4
>>> outputs = model.generate(**inputs, do_sample=True, epsilon_cutoff=0.1)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: 1, 2, 3, 4, 5, 6, 7, 8, 9

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.EtaLogitsWarper

< >

( epsilon: float filter_value: float = -inf min_tokens_to_keep: int = 1 device: str = 'cpu' )

Parameters

  • epsilon (float) — A float value in the range (0, 1). Hyperparameter used to calculate the dynamic cutoff value, eta. The suggested values from the paper ranges from 3e-4 to 4e-3 depending on the size of the model.
  • filter_value (float, optional, defaults to -inf) — All values that are found to be below the dynamic cutoff value, eta, are set to this float value. This parameter is useful when logits need to be modified for very low probability tokens that should be excluded from generation entirely.
  • min_tokens_to_keep (int, optional, defaults to 1) — Specifies the minimum number of tokens that must be kept for generation, regardless of their probabilities. For example, if min_tokens_to_keep is set to 1, at least one token will always be kept for generation, even if all tokens have probabilities below the cutoff eta.
  • device (str, optional, defaults to "cpu") — The device to allocate the tensors.

LogitsWarper that performs eta-sampling, a technique to filter out tokens with probabilities below a dynamic cutoff value, eta, which is calculated based on a combination of the hyperparameter epsilon and the entropy of the token probabilities, i.e. eta := min(epsilon, sqrt(epsilon * e^-entropy(probabilities))). Takes the largest min_tokens_to_keep tokens if no tokens satisfy this constraint. It addresses the issue of poor quality in long samples of text generated by neural language models leading to more coherent and fluent text. See Truncation Sampling as Language Model Desmoothing for more information. Note: do_sample must be set to True for this LogitsWarper to work.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed

>>> set_seed(1)
>>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")

>>> inputs = tokenizer("A sequence: 1, 2", return_tensors="pt")

>>> # With sampling, the output is unexpected -- sometimes too unexpected.
>>> outputs = model.generate(**inputs, do_sample=True)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: 1, 2, 3 | < 4 (left-hand pointer) ;
<BLANKLINE>
<BLANKLINE>

>>> # With eta sampling, the output gets restricted to high-probability tokens. You can see it as a dynamic form of
>>> # epsilon sampling that adapts its cutoff probability based on the entropy (high entropy = lower cutoff).
>>> # Pro tip: The paper recomends using `eta_cutoff` values between 3e-4 to 4e-3
>>> outputs = model.generate(**inputs, do_sample=True, eta_cutoff=0.1)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: 1, 2, 3, 4, 5, 6, 7, 8, 9

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.ExponentialDecayLengthPenalty

< >

( exponential_decay_length_penalty: Tuple eos_token_id: Union input_ids_seq_length: int )

Parameters

  • exponential_decay_length_penalty (tuple(int, float)) — This tuple shall consist of: (start_index, decay_factor) where start_index indicates where penalty starts and decay_factor represents the factor of exponential decay
  • eos_token_id (Union[int, List[int], torch.Tensor]) — The id(s) of the end-of-sequence token.
  • input_ids_seq_length (int) — The length of the input sequence.

LogitsProcessor that exponentially increases the score of the eos_token_id after start_index has been reached. This allows generating shorter sequences without having a hard cutoff, allowing the eos_token to be predicted in a meaningful position.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed

>>> model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")

>>> text = "Just wanted to let you know, I"
>>> inputs = tokenizer(text, return_tensors="pt")

>>> # Let's consider that we want short sentences, so we limit `max_length=30`. However, we observe that the answer
>>> # tends to end abruptly.
>>> set_seed(1)
>>> outputs = model.generate(**inputs, do_sample=True, temperature=0.9, max_length=30, pad_token_id=50256)
>>> print(tokenizer.batch_decode(outputs)[0])
Just wanted to let you know, I received a link to an ebook, the book How To Start A Social Network which was
published in 2010. Although

>>> # To promote the appearance of the EOS token at the right time, we add the `exponential_decay_length_penalty =
>>> # (start_index, decay_factor)`. Instead of cutting at max_tokens, the output comes to an end before and usually
>>> # with more meaning. What happens is that starting from `start_index` the EOS token score will be increased
>>> # by `decay_factor` exponentially. However, if you set a high decay factor, you may also end up with abruptly
>>> # ending sequences.
>>> set_seed(1)
>>> outputs = model.generate(
...     **inputs,
...     do_sample=True,
...     temperature=0.9,
...     max_length=30,
...     pad_token_id=50256,
...     exponential_decay_length_penalty=(15, 1.6),
... )
>>> print(tokenizer.batch_decode(outputs)[0])
Just wanted to let you know, I received a link to an ebook, the book How To Start A Social Network
which<|endoftext|>

>>> # With a small decay factor, you will have a higher chance of getting a meaningful sequence.
>>> set_seed(1)
>>> outputs = model.generate(
...     **inputs,
...     do_sample=True,
...     temperature=0.9,
...     max_length=30,
...     pad_token_id=50256,
...     exponential_decay_length_penalty=(15, 1.01),
... )
>>> print(tokenizer.batch_decode(outputs)[0])
Just wanted to let you know, I received a link to an ebook, the book How To Start A Social Network which was
published in 2010.<|endoftext|>

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.ForcedBOSTokenLogitsProcessor

< >

( bos_token_id: int )

Parameters

  • bos_token_id (int) — The id of the token to force as the first generated token.

LogitsProcessor that enforces the specified token as the first generated token. Used with encoder-decoder models.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

>>> model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-small")
>>> tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-small")

>>> inputs = tokenizer("Translate from English to German: I love cats.", return_tensors="pt")

>>> # By default, it continues generating according to the model's logits
>>> outputs = model.generate(**inputs, max_new_tokens=10)
>>> print(tokenizer.batch_decode(outputs)[0])
<pad> Ich liebe Kitty.</s>

>>> # We can use `forced_bos_token_id` to force the start of generation with an encoder-decoder model
>>> # (including forcing it to end straight away with an EOS token)
>>> outputs = model.generate(**inputs, max_new_tokens=10, forced_bos_token_id=tokenizer.eos_token_id)
>>> print(tokenizer.batch_decode(outputs)[0])
<pad></s>

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.ForcedEOSTokenLogitsProcessor

< >

( max_length: int eos_token_id: Union device: str = 'cpu' )

Parameters

  • max_length (int) — The maximum length of the sequence to be generated.
  • eos_token_id (Union[int, List[int], torch.Tensor]) — The id(s) of the end-of-sequence token.
  • device (str, optional, defaults to "cpu") — The device to allocate the tensors.

LogitsProcessor that enforces the specified token as the last generated token when max_length is reached.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")

>>> inputs = tokenizer("A sequence: 1, 2, 3", return_tensors="pt")

>>> # By default, it continues generating according to the model's logits
>>> outputs = model.generate(**inputs, max_new_tokens=10)
>>> print(tokenizer.batch_decode(outputs)[0])
A sequence: 1, 2, 3, 4, 5, 6, 7, 8

>>> # `forced_eos_token_id` ensures the generation ends with a EOS token
>>> outputs = model.generate(**inputs, max_new_tokens=10, forced_eos_token_id=tokenizer.eos_token_id)
>>> print(tokenizer.batch_decode(outputs)[0])
A sequence: 1, 2, 3, 4, 5, 6, 7,<|endoftext|>

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.ForceTokensLogitsProcessor

< >

( force_token_map: List _has_warned: Optional = False )

This processor takes a list of pairs of integers which indicates a mapping from generation indices to token indices that will be forced before generation. The processor will set their log probs to inf so that they are sampled at their corresponding index. Originally created for Whisper.

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.HammingDiversityLogitsProcessor

< >

( diversity_penalty: float num_beams: int num_beam_groups: int )

Parameters

  • diversity_penalty (float) — 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. A higher diversity_penalty will enforce greater diversity among the beams. Adjusting this value can help strike a balance between diversity and natural likelihood.
  • num_beams (int) — Number of beams for beam search. 1 means no beam search.
  • num_beam_groups (int) — Number of groups to divide num_beams into in order to ensure diversity among different groups of beams. this paper for more details.

LogitsProcessor that enforces diverse beam search.

Note that this logits processor is only effective for PreTrainedModel.group_beam_search. See Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence Models for more details.

Traditional beam search often generates very similar sequences across different beams. HammingDiversityLogitsProcessor addresses this by penalizing beams that generate tokens already chosen by other beams in the same time step.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
>>> import torch

>>> # Initialize the model and tokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-base")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("google-t5/t5-base")

>>> # A long text about the solar system
>>> text = (
...     "The Solar System is a gravitationally bound system comprising the Sun and the objects that orbit it, "
...     "either directly or indirectly. Of the objects that orbit the Sun directly, the largest are the eight "
...     "planets, with the remainder being smaller objects, such as the five dwarf planets and small Solar System "
...     "bodies. The Solar System formed 4.6 billion years ago from the gravitational collapse of a giant "
...     "interstellar molecular cloud."
... )
>>> inputs = tokenizer("summarize: " + text, return_tensors="pt")

>>> # Generate diverse summary
>>> outputs_diverse = model.generate(
...     **inputs,
...     num_beam_groups=2,
...     diversity_penalty=10.0,
...     max_length=100,
...     num_beams=4,
...     num_return_sequences=2,
... )
>>> summaries_diverse = tokenizer.batch_decode(outputs_diverse, skip_special_tokens=True)

>>> # Generate non-diverse summary
>>> outputs_non_diverse = model.generate(
...     **inputs,
...     max_length=100,
...     num_beams=4,
...     num_return_sequences=2,
... )
>>> summary_non_diverse = tokenizer.batch_decode(outputs_non_diverse, skip_special_tokens=True)

>>> # With `diversity_penalty`, the resulting beams are much more diverse
>>> print(summary_non_diverse)
['the solar system formed 4.6 billion years ago from the collapse of a giant interstellar molecular cloud. of the objects that orbit the Sun directly, the largest are the eight planets.',
'the Solar System formed 4.6 billion years ago from the collapse of a giant interstellar molecular cloud. of the objects that orbit the Sun directly, the largest are the eight planets.']

>>> print(summaries_diverse)
['the solar system formed 4.6 billion years ago from the collapse of a giant interstellar molecular cloud. of the objects that orbit the Sun directly, the largest are the eight planets.',
'the solar system formed 4.6 billion years ago from the collapse of a giant interstellar molecular cloud. of the objects that orbit the Sun directly, the largest are the eight planets. the rest of the objects are smaller objects, such as the five dwarf planets and small solar system bodies.']

__call__

< >

( input_ids: LongTensor scores: FloatTensor current_tokens: LongTensor beam_group_idx: int ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search
  • current_tokens (torch.LongTensor of shape (batch_size)) — Indices of input sequence tokens in the vocabulary, corresponding to the tokens selected by the other beam groups in the current generation step.
  • beam_group_idx (int) — The index of the beam group currently being processed.

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.InfNanRemoveLogitsProcessor

< >

( )

LogitsProcessor that removes all nan and inf values to avoid the generation method to fail. Note that using the logits processor should only be used if necessary since it can slow down the generation method.

This logits processor has no generate example, as there shouldn’t be a correct combination of flags that warrants its use.

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.LogitNormalization

< >

( )

LogitsWarper and LogitsProcessor for normalizing the scores using log-softmax. It’s important to normalize the scores during beam search, after applying the logits processors or warpers, since the search algorithm used in this library doesn’t do it (it only does it before, but they may need re-normalization) but it still supposes that the scores are normalized when comparing the hypotheses.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM
>>> import torch

>>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")

>>> inputs = tokenizer("A sequence: 1, 2, 3", return_tensors="pt")

>>> # By default, the scores are not normalized -- the sum of their exponentials is NOT a normalized probability
>>> # distribution, summing to 1
>>> outputs = model.generate(**inputs, return_dict_in_generate=True, output_scores=True)
>>> print(torch.allclose(torch.sum(torch.exp(outputs.scores[-1])), torch.Tensor((1.000,)), rtol=1e-4))
False

>>> # Normalizing them may have a positive impact on beam methods, or when using the scores on your application
>>> outputs = model.generate(**inputs, renormalize_logits=True, return_dict_in_generate=True, output_scores=True)
>>> print(torch.allclose(torch.sum(torch.exp(outputs.scores[-1])), torch.Tensor((1.000,)), rtol=1e-4))
True

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.LogitsProcessor

< >

( )

Abstract base class for all logit processors that can be applied during generation.

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.LogitsProcessorList

< >

( iterable = () )

This class can be used to create a list of LogitsProcessor or LogitsWarper to subsequently process a scores input tensor. This class inherits from list and adds a specific call method to apply each LogitsProcessor or LogitsWarper to the inputs.

__call__

< >

( input_ids: LongTensor scores: FloatTensor **kwargs ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search
  • kwargs (Dict[str, Any], optional) — Additional kwargs that are specific to a logits processor.

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.LogitsWarper

< >

( )

Abstract base class for all logit warpers that can be applied during generation with multinomial sampling.

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.MinLengthLogitsProcessor

< >

( min_length: int eos_token_id: Union device: str = 'cpu' )

Parameters

  • min_length (int) — The minimum length below which the score of eos_token_id is set to -float("Inf").
  • eos_token_id (Union[int, List[int], torch.Tensor]) — The id(s) of the end-of-sequence token.
  • device (str, optional, defaults to "cpu") — The device to allocate the tensors.

LogitsProcessor enforcing a min-length by setting EOS probability to 0. Note that, for decoder-only models like most LLMs, the length includes the prompt.

Examples:

>>> from transformers import AutoModelForCausalLM, AutoTokenizer

>>> tokenizer = AutoTokenizer.from_pretrained("bigscience/bloomz-560m")
>>> model = AutoModelForCausalLM.from_pretrained("bigscience/bloomz-560m")

>>> inputs = tokenizer("A number:", return_tensors="pt")
>>> gen_out = model.generate(**inputs)
>>> print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
A number: one

>>> # setting `min_length` to a value smaller than the uncontrolled output length has no impact
>>> gen_out = model.generate(**inputs, min_length=3)
>>> print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
A number: one

>>> # setting a larger `min_length` will force the model to generate beyond its natural ending point, which is not
>>> # necessarily incorrect
>>> gen_out = model.generate(**inputs, min_length=10)
>>> print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
A number: one thousand, nine hundred and ninety-four

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.MinNewTokensLengthLogitsProcessor

< >

( prompt_length_to_skip: int min_new_tokens: int eos_token_id: Union device: str = 'cpu' )

Parameters

  • prompt_length_to_skip (int) — The input tokens length. Not a valid argument when used with generate as it will automatically assign the input length.
  • min_new_tokens (int) — The minimum new tokens length below which the score of eos_token_id is set to -float("Inf").
  • eos_token_id (Union[int, List[int], torch.Tensor]) — The id(s) of the end-of-sequence token.
  • device (str, optional, defaults to "cpu") — The device to allocate the tensors.

LogitsProcessor enforcing a min-length of new tokens by setting EOS (End-Of-Sequence) token probability to 0. Contrarily to MinLengthLogitsProcessor, this processor ignores the prompt.

Examples:

>>> from transformers import AutoModelForCausalLM, AutoTokenizer

>>> tokenizer = AutoTokenizer.from_pretrained("bigscience/bloomz-560m")
>>> model = AutoModelForCausalLM.from_pretrained("bigscience/bloomz-560m")

>>> inputs = tokenizer(["A number:"], return_tensors="pt")
>>> gen_out = model.generate(**inputs)
>>> print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
A number: one

>>> # setting `min_new_tokens` will force the model to generate beyond its natural ending point, which is not
>>> # necessarily incorrect
>>> gen_out = model.generate(**inputs, min_new_tokens=2)
>>> print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
A number: one thousand

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.MinPLogitsWarper

< >

( min_p: float filter_value: float = -inf min_tokens_to_keep: int = 1 )

Parameters

  • min_p (float) — Minimum token probability, which will be scaled by the probability of the most likely token. It must be a value between 0 and 1. Typical values are in the 0.01-0.2 range, comparably selective as setting top_p in the 0.99-0.8 range (use the opposite of normal top_p values).
  • filter_value (float, optional, defaults to -inf) — All filtered values will be set to this float value.
  • min_tokens_to_keep (int, optional, defaults to 1) — Minimum number of tokens that cannot be filtered.

LogitsWarper that performs min-p, i.e. keeps all tokens that are above a minimum probability, scaled by the probability of the most likely token. As a result, the filter becomes more agressive in the presence of high-probability tokens, which is a sign of a confident output that we shouldn’t deviate from.

Often used together with TemperatureLogitsWarper. Used as an alternative to TopPLogitsWarper and TopKLogitsWarper.

Created by @menhguin and @kalomaze (github handles). Code adapted from this external PR

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed

>>> set_seed(1)
>>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")

>>> inputs = tokenizer("A sequence: 1, 2", return_tensors="pt")

>>> # With sampling, the output is unexpected -- sometimes too unexpected.
>>> outputs = model.generate(**inputs, do_sample=True)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: 1, 2, 3 | < 4 (left-hand pointer) ;
<BLANKLINE>
<BLANKLINE>

>>> # With `min_p` sampling, the output gets restricted to high-probability tokens.
>>> # Pro tip: In practice, LLMs use `min_p` in the 0.01-0.2 range.
>>> outputs = model.generate(**inputs, do_sample=True, min_p=0.1)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: 1, 2, 3, 4, 5, 6, 7, 8, 9

__call__

< >

( input_ids: LongTensor scores: FloatTensor )

class transformers.NoBadWordsLogitsProcessor

< >

( bad_words_ids: List eos_token_id: Union = None )

Parameters

  • bad_words_ids (List[List[int]]) — List of list of token ids that are not allowed to be generated.
  • eos_token_id (Union[int, List[int], torch.Tensor], optional) — The id(s) of the end-of-sequence token.

LogitsProcessor that enforces that specified sequences will never be selected.

In order to get the token ids of the words that should not appear in the generated text, make sure to set add_prefix_space=True when initializing the tokenizer, and use tokenizer(bad_words, add_special_tokens=False).input_ids. The add_prefix_space argument is only supported for some slow tokenizers, as fast tokenizers’ prefixing behaviours come from pre tokenizers. Read more here.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
>>> inputs = tokenizer(["In a word, the cake is a"], return_tensors="pt")

>>> output_ids = model.generate(inputs["input_ids"], max_new_tokens=5, pad_token_id=tokenizer.eos_token_id)
>>> print(tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0])
In a word, the cake is a bit of a mess.

>>> # Now let's take the bad words out. Please note that the tokenizer is initialized differently
>>> tokenizer_with_prefix_space = AutoTokenizer.from_pretrained("openai-community/gpt2", add_prefix_space=True)


>>> def get_tokens_as_list(word_list):
...     "Converts a sequence of words into a list of tokens"
...     tokens_list = []
...     for word in word_list:
...         tokenized_word = tokenizer_with_prefix_space([word], add_special_tokens=False).input_ids[0]
...         tokens_list.append(tokenized_word)
...     return tokens_list


>>> bad_words_ids = get_tokens_as_list(word_list=["mess"])
>>> output_ids = model.generate(
...     inputs["input_ids"], max_new_tokens=5, bad_words_ids=bad_words_ids, pad_token_id=tokenizer.eos_token_id
... )
>>> print(tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0])
In a word, the cake is a bit of a surprise.

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.NoRepeatNGramLogitsProcessor

< >

( ngram_size: int )

Parameters

  • ngram_size (int) — All ngrams of size ngram_size can only occur once.

N-grams are groups of “n” consecutive words, characters, or tokens taken from a sequence of text. Given the sentence: “She runs fast”, the bi-grams (n=2) would be (“she”, “runs”) and (“runs”, “fast”). In text generation, avoiding repetitions of word sequences provides a more diverse output. This LogitsProcessor enforces no repetition of n-grams by setting the scores of banned tokens to negative infinity which eliminates those tokens from consideration when further processing the scores. Note that, for decoder-only models like most LLMs, the prompt is also considered to obtain the n-grams. Fairseq.

Use n-gram penalties with care. For instance, penalizing 2-grams (bigrams) in an article about the city of New York might lead to undesirable outcomes where the city’s name appears only once in the entire text. Reference

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")
>>> inputs = tokenizer(["Today I"], return_tensors="pt")

>>> output = model.generate(**inputs)
>>> print(tokenizer.decode(output[0], skip_special_tokens=True))
Today I’m not sure if I’m going to be able to do it.

>>> # Now let's add ngram size using `no_repeat_ngram_size`. This stops the repetitions ("I’m") in the output.
>>> output = model.generate(**inputs, no_repeat_ngram_size=2)
>>> print(tokenizer.decode(output[0], skip_special_tokens=True))
Today I’m not sure if I can get a better understanding of the nature of this issue

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.PrefixConstrainedLogitsProcessor

< >

( prefix_allowed_tokens_fn: Callable num_beams: int )

Parameters

  • prefix_allowed_tokens_fn (Callable[[int, torch.Tensor], List[int]]) — This function constraints the beam search to allowed tokens only at each step. This function takes 2 arguments inputs_ids and the batch ID batch_id. It has to return a list with the allowed tokens for the next generation step conditioned on the previously generated tokens inputs_ids and the batch ID batch_id.

LogitsProcessor that enforces constrained generation and is useful for prefix-conditioned constrained generation. See Autoregressive Entity Retrieval for more information.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> model = AutoModelForCausalLM.from_pretrained("bigscience/bloomz-560m")
>>> tokenizer = AutoTokenizer.from_pretrained("bigscience/bloomz-560m")

>>> inputs = tokenizer("Alice and Bob", return_tensors="pt")

>>> # By default, it continues generating according to the model's logits
>>> outputs = model.generate(**inputs, max_new_tokens=5)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
Alice and Bob are friends

>>> # We can contrain it with `prefix_allowed_tokens_fn` to force a certain behavior based on a prefix.
>>> # For instance, we can force an entire entity to be generated when its beginning is detected.
>>> entity = tokenizer(" Bob Marley", return_tensors="pt").input_ids[0]  # 3 tokens
>>> def prefix_allowed_tokens_fn(batch_id, input_ids):
...     '''
...     Attempts to generate 'Bob Marley' when 'Bob' is detected.
...     In this case, `batch_id` is not used, but you can set rules for each batch member.
...     '''
...     if input_ids[-1] == entity[0]:
...         return [entity[1].item()]
...     elif input_ids[-2] == entity[0] and input_ids[-1] == entity[1]:
...         return [entity[2].item()]
...     return list(range(tokenizer.vocab_size))  # If no match, allow all tokens

>>> outputs = model.generate(**inputs, max_new_tokens=5, prefix_allowed_tokens_fn=prefix_allowed_tokens_fn)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
Alice and Bob Marley

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.RepetitionPenaltyLogitsProcessor

< >

( penalty: float )

Parameters

  • penalty (float) — The parameter for repetition penalty. 1.0 means no penalty. Above 1.0 penalizes previously generated tokens. Between 0.0 and 1.0 rewards previously generated tokens.

LogitsProcessor that prevents the repetition of previous tokens through a penalty. This penalty is applied at most once per token. Note that, for decoder-only models like most LLMs, the considered tokens include the prompt.

In the original paper, the authors suggest the use of a penalty of around 1.2 to achieve a good balance between truthful generation and lack of repetition. To penalize and reduce repetition, use penalty values above 1.0, where a higher value penalizes more strongly. To reward and encourage repetition, use penalty values between 0.0 and 1.0, where a lower value rewards more strongly.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> # Initializing the model and tokenizer for it
>>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")
>>> inputs = tokenizer(["I'm not going to"], return_tensors="pt")

>>> # This shows a normal generate without any specific parameters
>>> summary_ids = model.generate(**inputs)
>>> print(tokenizer.batch_decode(summary_ids, skip_special_tokens=True)[0])
I'm not going to be able to do that. I'm going to be able to do that

>>> # This generates a penalty for repeated tokens
>>> penalized_ids = model.generate(**inputs, repetition_penalty=1.1)
>>> print(tokenizer.batch_decode(penalized_ids, skip_special_tokens=True)[0])
I'm not going to be able to do that. I'll just have to go out and play

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.SequenceBiasLogitsProcessor

< >

( sequence_bias: Dict )

Parameters

  • sequence_bias (Dict[Tuple[int], float]) — Dictionary that maps a sequence of tokens to its bias term. Positive biases increase the odds of the sequence being selected, while negative biases do the opposite. If a sequence has a length of 1, its bias will always be applied. Otherwise, the bias will only be applied if the sequence in question is about to be completed (in the token selection step after this processor is applied).

LogitsProcessor that applies an additive bias on sequences. The bias is applied to the last token of a sequence when the next generated token can complete it. Consequently, to take the most of biasing sequences with more than one token, consider using beam methods (to gracefully work around partially completed sequences that have a negative bias) and applying the bias to their prefixes (to ensure the bias is applied earlier).

In order to get the token ids of the sequences that you want to bias, make sure to set add_prefix_space=True when initializing the tokenizer, and use tokenizer(bad_words, add_special_tokens=False).input_ids. The add_prefix_space argument is only supported for some slow tokenizers, as fast tokenizers’ prefixing behaviours come from pre tokenizers. Read more here.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
>>> inputs = tokenizer(["The full name of Donald is Donald"], return_tensors="pt")

>>> summary_ids = model.generate(inputs["input_ids"], max_new_tokens=4)
>>> print(tokenizer.batch_decode(summary_ids, skip_special_tokens=True)[0])
The full name of Donald is Donald J. Trump Jr

>>> # Now let's control generation through a bias. Please note that the tokenizer is initialized differently!
>>> tokenizer_with_prefix_space = AutoTokenizer.from_pretrained("openai-community/gpt2", add_prefix_space=True)


>>> def get_tokens_as_tuple(word):
...     return tuple(tokenizer_with_prefix_space([word], add_special_tokens=False).input_ids[0])


>>> # If we add a negative bias without beam search, it may become "stuck" in a prefix without good continuations
>>> sequence_bias = {get_tokens_as_tuple("Trump"): -10.0}
>>> biased_ids = model.generate(inputs["input_ids"], max_new_tokens=4, sequence_bias=sequence_bias)
>>> print(tokenizer.batch_decode(biased_ids, skip_special_tokens=True)[0])
The full name of Donald is Donald J. Donald,

>>> biased_ids = model.generate(inputs["input_ids"], max_new_tokens=4, num_beams=4, sequence_bias=sequence_bias)
>>> print(tokenizer.batch_decode(biased_ids, skip_special_tokens=True)[0])
The full name of Donald is Donald Rumsfeld,

>>> # We can also add a positive bias to nudge the model towards specific tokens or continuations
>>> sequence_bias = {get_tokens_as_tuple("Donald Duck"): 10.0}
>>> biased_ids = model.generate(inputs["input_ids"], max_new_tokens=4, num_beams=4, sequence_bias=sequence_bias)
>>> print(tokenizer.batch_decode(biased_ids, skip_special_tokens=True)[0])
The full name of Donald is Donald Duck.

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.SuppressTokensAtBeginLogitsProcessor

< >

( begin_suppress_tokens begin_index device: str = 'cpu' )

SuppressTokensAtBeginLogitsProcessor supresses 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 are not generated at the begining. Originally created for Whisper.

Examples:

>>> from transformers import AutoProcessor, WhisperForConditionalGeneration
>>> from datasets import load_dataset

>>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
>>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")

>>> # Whisper has `begin_suppress_tokens` set by default (= `[220, 50256]`). 50256 is the EOS token, so this means
>>> # it can't generate and EOS token in the first iteration, but it can in the others.
>>> outputs = model.generate(**inputs, return_dict_in_generate=True, output_scores=True)
>>> print(outputs.scores[0][0, 50256])
tensor(-inf)
>>> print(outputs.scores[-1][0, 50256])  # in other places we can see some probability mass for EOS
tensor(29.9010)

>>> # If we disable `begin_suppress_tokens`, we can generate EOS in the first iteration.
>>> outputs = model.generate(
...     **inputs, return_dict_in_generate=True, output_scores=True, begin_suppress_tokens=None
... )
>>> print(outputs.scores[0][0, 50256])
tensor(11.2027)

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.SuppressTokensLogitsProcessor

< >

( suppress_tokens device: str = 'cpu' )

This processor can be used to suppress a list of tokens. The processor will set their log probs to -inf so that they are not generated. Originally created for Whisper.

Examples:

>>> from transformers import AutoProcessor, WhisperForConditionalGeneration
>>> from datasets import load_dataset

>>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
>>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")

>>> # Whisper has a long list of suppressed tokens. For instance, in this case, the token 1 is suppressed by default.
>>> outputs = model.generate(**inputs, return_dict_in_generate=True, output_scores=True)
>>> print(outputs.scores[1][0, 1])  # 1 (and not 0) is the first freely generated token
tensor(-inf)

>>> # If we disable `suppress_tokens`, we can generate it.
>>> outputs = model.generate(**inputs, return_dict_in_generate=True, output_scores=True, suppress_tokens=None)
>>> print(outputs.scores[1][0, 1])
tensor(6.0678)

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.TemperatureLogitsWarper

< >

( temperature: float )

Parameters

  • temperature (float) — Strictly positive float value used to modulate the logits distribution. A value smaller than 1 decreases randomness (and vice versa), with 0 being equivalent to shifting all probability mass to the most likely token.

LogitsWarper for temperature (exponential scaling output probability distribution), which effectively means that it can control the randomness of the predicted tokens. Often used together with TopPLogitsWarper and TopKLogitsWarper.

Make sure that do_sample=True is included in the generate arguments otherwise the temperature value won’t have any effect.

Examples:

>>> import torch
>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed

>>> set_seed(0)  # for reproducibility

>>> tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
>>> model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2")
>>> model.config.pad_token_id = model.config.eos_token_id
>>> inputs = tokenizer(["Hugging Face Company is"], return_tensors="pt")

>>> # With temperature=1.0, the default, we consistently get random outputs due to random sampling.
>>> generate_kwargs = {"max_new_tokens": 10, "do_sample": True, "temperature": 1.0, "num_return_sequences": 2}
>>> outputs = model.generate(**inputs, **generate_kwargs)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
['Hugging Face Company is one of these companies that is going to take a',
"Hugging Face Company is a brand created by Brian A. O'Neil"]

>>> # However, with temperature close to 0, it approximates greedy decoding strategies (invariant)
>>> generate_kwargs["temperature"] = 0.0001
>>> outputs = model.generate(**inputs, **generate_kwargs)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
['Hugging Face Company is a company that has been around for over 20 years',
'Hugging Face Company is a company that has been around for over 20 years']

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.TopKLogitsWarper

< >

( top_k: int filter_value: float = -inf min_tokens_to_keep: int = 1 )

Parameters

  • top_k (int) — The number of highest probability vocabulary tokens to keep for top-k-filtering.
  • filter_value (float, optional, defaults to -inf) — All filtered values will be set to this float value.
  • min_tokens_to_keep (int, optional, defaults to 1) — Minimum number of tokens that cannot be filtered.

LogitsWarper that performs top-k, i.e. restricting to the k highest probability elements. Often used together with TemperatureLogitsWarper and TopPLogitsWarper.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed

>>> set_seed(1)
>>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")

>>> inputs = tokenizer("A sequence: A, B, C, D", return_tensors="pt")

>>> # With sampling, the output is unexpected -- sometimes too unexpected.
>>> outputs = model.generate(**inputs, do_sample=True)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: A, B, C, D, E — S — O, P — R

>>> # With `top_k` sampling, the output gets restricted the k most likely tokens.
>>> # Pro tip: In practice, LLMs use `top_k` in the 5-50 range.
>>> outputs = model.generate(**inputs, do_sample=True, top_k=2)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: A, B, C, D, E, F, G, H, I

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.TopPLogitsWarper

< >

( top_p: float filter_value: float = -inf min_tokens_to_keep: int = 1 )

Parameters

  • top_p (float) — If set to < 1, only the smallest set of most probable tokens with probabilities that add up to top_p or higher are kept for generation.
  • filter_value (float, optional, defaults to -inf) — All filtered values will be set to this float value.
  • min_tokens_to_keep (int, optional, defaults to 1) — Minimum number of tokens that cannot be filtered.

LogitsWarper that performs top-p, i.e. restricting to top tokens summing to prob_cut_off <= prob_cut_off. Often used together with TemperatureLogitsWarper and TopKLogitsWarper.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed

>>> set_seed(1)
>>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")

>>> inputs = tokenizer("A sequence: 1, 2", return_tensors="pt")

>>> # With sampling, the output is unexpected -- sometimes too unexpected.
>>> outputs = model.generate(**inputs, do_sample=True)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: 1, 2, 3 | < 4 (left-hand pointer) ;
<BLANKLINE>
<BLANKLINE>

>>> # With `top_p` sampling, the output gets restricted to high-probability tokens.
>>> # Pro tip: In practice, LLMs use `top_p` in the 0.9-0.95 range.
>>> outputs = model.generate(**inputs, do_sample=True, top_p=0.1)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
A sequence: 1, 2, 3, 4, 5, 6, 7, 8, 9

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.TypicalLogitsWarper

< >

( mass: float = 0.9 filter_value: float = -inf min_tokens_to_keep: int = 1 )

Parameters

  • mass (float, optional, defaults to 0.9) — Value of typical_p between 0 and 1 inclusive, defaults to 0.9.
  • filter_value (float, optional, defaults to -inf) — All filtered values will be set to this float value.
  • min_tokens_to_keep (int, optional, defaults to 1) — Minimum number of tokens that cannot be filtered.

LogitsWarper that performs typical decoding. Inspired on how humans use language, it prioritizes tokens whose log probability is close to the entropy of the token probability distribution. This means that the most likely tokens may be discarded in the process.

See Typical Decoding for Natural Language Generation for more information.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed

>>> model = AutoModelForCausalLM.from_pretrained("bigscience/bloomz-560m")
>>> tokenizer = AutoTokenizer.from_pretrained("bigscience/bloomz-560m")

>>> inputs = tokenizer("1, 2, 3", return_tensors="pt")

>>> # We can see that greedy decoding produces a sequence of numbers
>>> outputs = model.generate(**inputs)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
1, 2, 3, 4, 5, 6, 7, 8, 9, 10,

>>> # For this particular seed, we can see that sampling produces nearly the same low-information (= low entropy)
>>> # sequence
>>> set_seed(18)
>>> outputs = model.generate(**inputs, do_sample=True)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
1, 2, 3, 4, 5, 6, 7, 8, 9 and 10

>>> # With `typical_p` set, the most obvious sequence is no longer produced, which may be good for your problem
>>> set_seed(18)
>>> outputs = model.generate(
...     **inputs, do_sample=True, typical_p=0.1, return_dict_in_generate=True, output_scores=True
... )
>>> print(tokenizer.batch_decode(outputs.sequences, skip_special_tokens=True)[0])
1, 2, 3 and 5

>>> # We can see that the token corresponding to "4" (token 934) in the second position, the most likely token
>>> # as seen with greedy decoding, was entirely blocked out
>>> print(outputs.scores[1][0, 934])
tensor(-inf)

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.UnbatchedClassifierFreeGuidanceLogitsProcessor

< >

( guidance_scale: float model unconditional_ids: Optional = None unconditional_attention_mask: Optional = None use_cache: Optional = True )

Parameters

  • guidance_scale (float) — The guidance scale for classifier free guidance (CFG). CFG is enabled by setting guidance_scale != 1. Higher guidance scale encourages the model to generate samples that are more closely linked to the input prompt, usually at the expense of poorer quality. A value smaller than 1 has the opposite effect, while making the negative prompt provided with negative_prompt_ids (if any) act as a positive prompt.
  • model (PreTrainedModel) — The model computing the unconditional scores. Supposedly the same as the one computing the conditional scores. Both models must use the same tokenizer.
  • unconditional_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of input sequence tokens in the vocabulary for the unconditional branch. If unset, will default to the last token of the prompt.
  • unconditional_attention_mask (torch.LongTensor of shape (batch_size, sequence_length), optional) — Attention mask for unconditional_ids.
  • use_cache (bool, optional, defaults to True) — Whether to cache key/values during the negative prompt forward pass.

Logits processor for Classifier-Free Guidance (CFG). The processors computes a weighted average across scores from prompt conditional and prompt unconditional (or negative) logits, parameterized by the guidance_scale. The unconditional scores are computed internally by prompting model with the unconditional_ids branch.

See the paper for more information.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
>>> inputs = tokenizer(["Today, a dragon flew over Paris, France,"], return_tensors="pt")
>>> out = model.generate(inputs["input_ids"], guidance_scale=1.5)
>>> tokenizer.batch_decode(out, skip_special_tokens=True)[0]
'Today, a dragon flew over Paris, France, killing at least 50 people and injuring more than 100'

>>> # with a negative prompt
>>> neg_inputs = tokenizer(["A very happy event happened,"], return_tensors="pt")
>>> out = model.generate(inputs["input_ids"], guidance_scale=2, negative_prompt_ids=neg_inputs["input_ids"])
>>> tokenizer.batch_decode(out, skip_special_tokens=True)[0]
'Today, a dragon flew over Paris, France, killing at least 130 people. French media reported that'

>>> # with a positive prompt
>>> neg_inputs = tokenizer(["A very happy event happened,"], return_tensors="pt")
>>> out = model.generate(inputs["input_ids"], guidance_scale=0, negative_prompt_ids=neg_inputs["input_ids"])
>>> tokenizer.batch_decode(out, skip_special_tokens=True)[0]
"Today, a dragon flew over Paris, France, and I'm very happy to be here. I"

__call__

< >

( input_ids scores )

class transformers.WhisperTimeStampLogitsProcessor

< >

( generate_config begin_index: Optional = None _detect_timestamp_from_logprob: Optional = None )

Parameters

  • generate_config (GenerateConfig) — The generate config used to generate the output. The following parameters are required: eos_token_id (int, optional, defaults to 50257): The id of the end-of-sequence token. no_timestamps_token_id (int, optional, defaults to 50363): The id of the "<|notimestamps|>" token. max_initial_timestamp_index (int, optional, defaults to 1): Used to set the maximum value of the initial timestamp. This is used to prevent the model from predicting timestamps that are too far in the future.
  • begin_index (Optional, optional) — Token index of the first token that is generated by the model.
  • _detect_timestamp_from_logprob (bool, optional) — Whether timestamps can be predicted from logprobs over all timestamps.

LogitsProcessor that modifies the logits for the generation of timestamps in the transcription. When the input tokens are at a specific threshold, the processor sets the scores to negative infinity. The processor makes sure that timestamp tokens appear in pairs, by masking out the logits that would break this pairing pattern. This is done to maintain the consistency and structure of generated timestamps. It also ensures that when the predicted probability of sampling any of the timestamp token is greater than any individual non-timestamp token, those non-timestamp logits are set to negative infinity. This is done to ensure the generation of timestamps over other potential tokens.

See the paper for more information.

Examples:

>>> import torch
>>> from transformers import AutoProcessor, WhisperForConditionalGeneration, GenerationConfig
>>> from datasets import load_dataset

>>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
>>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> inputs = processor(ds[3]["audio"]["array"], return_tensors="pt")
>>> input_features = inputs.input_features

>>> #Displaying timestamps
>>> generated_ids = model.generate(inputs=input_features, return_timestamps=True)
>>> transcription = processor.batch_decode(generated_ids, decode_with_timestamps=True)[0]
>>> print("Transcription:", transcription)
Transcription: <|startoftranscript|><|0.00|> He has grave doubts whether Sir Frederick Layton's work is really Greek after all, and can<|6.44|><|6.44|> discover in it but little of rocky Ithaca.<|9.44|><|endoftext|>


>>> #No timestamps & change EOS:
>>> #This allows the user to select a specific token to terminate the sequence on, in this case it's the word "can"(460)
>>> model.generation_config.eos_token_id = 460
>>> generated_ids = model.generate(inputs=input_features,return_timestamps=False)
>>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
>>> print("Transcription:", transcription)
Transcription:  He has grave doubts whether Sir Frederick Layton's work is really Greek after all and can

__call__

< >

( input_ids: LongTensor scores: FloatTensor ) → torch.FloatTensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. What are input IDs?
  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search

Returns

torch.FloatTensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.WatermarkLogitsProcessor

< >

( vocab_size device greenlist_ratio: float = 0.25 bias: float = 2.0 hashing_key: int = 15485863 seeding_scheme: str = 'lefthash' context_width: int = 1 )

Parameters

  • vocab_size (int) — The model tokenizer’s vocab_size. Used to calculate “green” tokens ratio.
  • device (str) — The device where model is allocated.
  • greenlist_ratio (float, optional, optional, defaults to 0.25) — The ratio of “green” tokens used to the vocabulary size. Defaults to 0.25.
  • bias (float, optional, optional, defaults to 2.0) — The bias added to the selected “green” tokens’ logits. Consider lowering the bias if the text generation quality degrades. Recommended values are in the range of [0.5, 2.0]. Defaults to 2.0.
  • hashing_key (int, optional, optional, defaults to 15485863) — Key used for hashing. If you deploy this watermark, we advise using another private key. Defaults to 15485863 (the millionth prime).
  • seeding_scheme (str, optional, optional, defaults to "lefthash") — The seeding scheme used for selecting “green” tokens. Accepts values:
    • “lefthash” (default): “green” tokens selection depend on the last token (Algorithm 2 from paper)
    • “selfhash”: “green” tokens selection depends on the current token itself (Algorithm 3 from paper) The downside of this scheme is that it considers all possible next tokens and can be slower than “lefthash”. The context length of previous tokens to use in seeding. Higher context length makes watermarking more robust.
  • context_width (int, optional, defaults to 1) — The number of previous tokens to use when setting the seed.

Logits processor for watermarking generated text. The processor modifies model output scores by adding a small bias to randomized set of “green” tokens before generating the next token. “Green” tokens selection process depends on the seeding_scheme used. The code was based on the original repo.

The text generated by this LogitsProcessor can be detected using WatermarkDetector. See call() for details,

See the paper for more information.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM, WatermarkingConfig

>>> model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2")
>>> tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
>>> inputs = tokenizer(["Alice and Bob are"], return_tensors="pt")

>>> # normal generation
>>> out = model.generate(inputs["input_ids"], max_length=20, do_sample=False)
>>> tokenizer.batch_decode(out, skip_special_tokens=True)[0]
'Alice and Bob are both in the same room.\n\n"I\'m not sure if you\'re'

>>> # watermarked generation
>>> watermarking_config = WatermarkingConfig(bias=2.5, context_width=2, seeding_scheme="selfhash")
>>> out = model.generate(inputs["input_ids"], watermarking_config=watermarking_config, max_length=20, do_sample=False)
>>> tokenizer.batch_decode(out, skip_special_tokens=True)[0]
'Alice and Bob are both still alive and well and the story is pretty much a one-hour adventure'

>>> # to detect watermarked text use the WatermarkDetector class
>>> from transformers import WatermarkDetector
>>> detector = WatermarkDetector(model_config=model.config, device="cpu", watermarking_config= watermarking_config)
>>> detection_preds = detector(out)
>>> detection_preds
array([ True])

__call__

< >

( input_ids: LongTensor scores: FloatTensor )

TensorFlow

class transformers.TFForcedBOSTokenLogitsProcessor

< >

( bos_token_id: int )

Parameters

  • bos_token_id (int) — The id of the token to force as the first generated token.

TFLogitsProcessor that enforces the specified token as the first generated token.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFForcedEOSTokenLogitsProcessor

< >

( max_length: int eos_token_id: int )

Parameters

  • max_length (int) — The maximum length of the sequence to be generated.
  • eos_token_id (int) — The id of the token to force as the last generated token when max_length is reached.

TFLogitsProcessor that enforces the specified token as the last generated token when max_length is reached.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFForceTokensLogitsProcessor

< >

( force_token_map: List )

This processor takes a list of pairs of integers which indicates a mapping from generation indices to token indices that will be forced before sampling. The processor will set their log probs to 0 and all other tokens to -inf so that they are sampled at their corresponding index.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFLogitsProcessor

< >

( )

Abstract base class for all logit processors that can be applied during generation.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int ) → tf.Tensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (tf.Tensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (tf.Tensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search.
  • cur_len (int) — The current length of valid input sequence tokens. In the TF implementation, the input_ids’ sequence length is the maximum length generate can produce, and we need to know which of its tokens are valid.
  • kwargs (Dict[str, Any], optional) — Additional logits processor specific kwargs.

Returns

tf.Tensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

TF method for processing logits.

class transformers.TFLogitsProcessorList

< >

( iterable = () )

This class can be used to create a list of TFLogitsProcessor to subsequently process a scores input tensor. This class inherits from list and adds a specific call method to apply each TFLogitsProcessor to the inputs.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int **kwargs ) → tf.Tensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (tf.Tensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (tf.Tensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search.
  • cur_len (int) — The current length of valid input sequence tokens. In the TF implementation, the input_ids’ sequence length is the maximum length generate can produce, and we need to know which of its tokens are valid.
  • kwargs (Dict[str, Any], optional) — Additional logits processor specific kwargs.

Returns

tf.Tensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.TFLogitsWarper

< >

( )

Abstract base class for all logit warpers that can be applied during generation with multinomial sampling.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int ) → tf.Tensor of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (tf.Tensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (tf.Tensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search.
  • cur_len (int) — The current length of valid input sequence tokens. In the TF implementation, the input_ids’ sequence length is the maximum length generate can produce, and we need to know which of its tokens are valid.
  • kwargs (Dict[str, Any], optional) — Additional logits processor specific kwargs.

Returns

tf.Tensor of shape (batch_size, config.vocab_size)

The processed prediction scores.

TF method for warping logits.

class transformers.TFMinLengthLogitsProcessor

< >

( min_length: int eos_token_id: int )

Parameters

  • min_length (int) — The minimum length below which the score of eos_token_id is set to -float("Inf").
  • eos_token_id (int) — The id of the end-of-sequence token.

TFLogitsProcessor enforcing a min-length by setting EOS probability to 0.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFNoBadWordsLogitsProcessor

< >

( bad_words_ids: List eos_token_id: int )

Parameters

  • bad_words_ids (List[List[int]]) — List of list of token ids that are not allowed to be generated. In order to get the tokens of the words that should not appear in the generated text, make sure to set add_prefix_space=True when initializing the tokenizer, and use tokenizer(bad_words, add_special_tokens=False).input_ids. The add_prefix_space argument is only supported for some slow tokenizers, as fast tokenizers’ prefixing behaviours come from pre tokenizers. Read more here.
  • eos_token_id (int) — The id of the end-of-sequence token.

TFLogitsProcessor that enforces that specified sequences will never be sampled.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFNoRepeatNGramLogitsProcessor

< >

( ngram_size: int )

Parameters

  • ngram_size (int) — All ngrams of size ngram_size can only occur once.

TFLogitsProcessor that enforces no repetition of n-grams. See Fairseq.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFRepetitionPenaltyLogitsProcessor

< >

( penalty: float )

Parameters

  • repetition_penalty (float) — The parameter for repetition penalty. 1.0 means no penalty. See this paper for more details.

TFLogitsProcessor enforcing an exponential penalty on repeated sequences.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFSuppressTokensAtBeginLogitsProcessor

< >

( begin_suppress_tokens begin_index )

TFSuppressTokensAtBeginLogitsProcessor 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.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFSuppressTokensLogitsProcessor

< >

( suppress_tokens )

This processor can be used to suppress a list of tokens. The processor will set their log probs to -inf so that they are not sampled.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFTemperatureLogitsWarper

< >

( temperature: float )

Parameters

  • temperature (float) — The value used to module the logits distribution.

TFLogitsWarper for temperature (exponential scaling output probability distribution).

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFTopKLogitsWarper

< >

( top_k: int filter_value: float = -inf min_tokens_to_keep: int = 1 )

Parameters

  • top_k (int) — The number of highest probability vocabulary tokens to keep for top-k-filtering.
  • filter_value (float, optional, defaults to -inf) — All filtered values will be set to this float value.
  • min_tokens_to_keep (int, optional, defaults to 1) — Minimum number of tokens that cannot be filtered.

TFLogitsWarper that performs top-k, i.e. restricting to the k highest probability elements.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

class transformers.TFTopPLogitsWarper

< >

( top_p: float filter_value: float = -inf min_tokens_to_keep: int = 1 )

Parameters

  • top_p (float) — If set to < 1, only the smallest set of most probable tokens with probabilities that add up to top_p or higher are kept for generation.
  • filter_value (float, optional, defaults to -inf) — All filtered values will be set to this float value.
  • min_tokens_to_keep (int, optional, defaults to 1) — Minimum number of tokens that cannot be filtered.

TFLogitsWarper that performs top-p, i.e. restricting to top tokens summing to <= prob_cut_off.

__call__

< >

( input_ids: Tensor scores: Tensor cur_len: int )

FLAX

class transformers.FlaxForcedBOSTokenLogitsProcessor

< >

( bos_token_id: int )

Parameters

  • bos_token_id (int) — The id of the token to force as the first generated token.

FlaxLogitsProcessor that enforces the specified token as the first generated token.

__call__

< >

( input_ids: Array scores: Array cur_len: int )

class transformers.FlaxForcedEOSTokenLogitsProcessor

< >

( max_length: int eos_token_id: int )

Parameters

  • max_length (int) — The maximum length of the sequence to be generated.
  • eos_token_id (int) — The id of the token to force as the last generated token when max_length is reached.

FlaxLogitsProcessor that enforces the specified token as the last generated token when max_length is reached.

__call__

< >

( input_ids: Array scores: Array cur_len: int )

class transformers.FlaxForceTokensLogitsProcessor

< >

( force_token_map )

Parameters

  • force_token_map (list) — Map giving token ids and indices where they will be forced to be sampled.

FlaxLogitsProcessor that takes a list of pairs of integers which indicates a mapping from generation indices to token indices that will be forced before sampling. The processor will set their log probs to 0 and all other tokens to -inf so that they are sampled at their corresponding index.

__call__

< >

( input_ids: Array scores: Array cur_len: int )

class transformers.FlaxLogitsProcessor

< >

( )

Abstract base class for all logit processors that can be applied during generation.

__call__

< >

( input_ids: Array scores: Array ) → jnp.ndarray of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (jnp.ndarray of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (jnp.ndarray of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search
  • kwargs (Dict[str, Any], optional) — Additional logits processor specific kwargs.

Returns

jnp.ndarray of shape (batch_size, config.vocab_size)

The processed prediction scores.

Flax method for processing logits.

class transformers.FlaxLogitsProcessorList

< >

( iterable = () )

This class can be used to create a list of FlaxLogitsProcessor or FlaxLogitsWarper to subsequently process a scores input tensor. This class inherits from list and adds a specific call method to apply each FlaxLogitsProcessor or FlaxLogitsWarper to the inputs.

__call__

< >

( input_ids: Array scores: Array cur_len: int **kwargs ) → jnp.ndarray of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (jnp.ndarray of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (jnp.ndarray of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search
  • kwargs (Dict[str, Any], optional) — Additional logits processor specific kwargs.

Returns

jnp.ndarray of shape (batch_size, config.vocab_size)

The processed prediction scores.

class transformers.FlaxLogitsWarper

< >

( )

Abstract base class for all logit warpers that can be applied during generation with multinomial sampling.

__call__

< >

( input_ids: Array scores: Array ) → jnp.ndarray of shape (batch_size, config.vocab_size)

Parameters

  • input_ids (jnp.ndarray of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (jnp.ndarray of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search
  • kwargs (Dict[str, Any], optional) — Additional logits processor specific kwargs.

Returns

jnp.ndarray of shape (batch_size, config.vocab_size)

The processed prediction scores.

Flax method for warping logits.

class transformers.FlaxMinLengthLogitsProcessor

< >

( min_length: int eos_token_id: int )

Parameters

  • min_length (int) — The minimum length below which the score of eos_token_id is set to -float("Inf").
  • eos_token_id (int) — The id of the end-of-sequence token.

FlaxLogitsProcessor enforcing a min-length by setting EOS probability to 0.

__call__

< >

( input_ids: Array scores: Array cur_len: int )

class transformers.FlaxSuppressTokensAtBeginLogitsProcessor

< >

( begin_suppress_tokens begin_index )

Parameters

  • begin_suppress_tokens (List[int]) — Tokens to not sample.
  • begin_index (int) — Index where the tokens are suppressed.

FlaxLogitsProcessor supressing 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 are not sampled at the begining of the generation.

__call__

< >

( input_ids scores cur_len: int )

class transformers.FlaxSuppressTokensLogitsProcessor

< >

( suppress_tokens: list )

Parameters

  • suppress_tokens (list) — Tokens to not sample.

FlaxLogitsProcessor suppressing a list of tokens at each decoding step. The processor will set their log probs to be -inf so they are not sampled.

__call__

< >

( input_ids: Array scores: Array cur_len: int )

class transformers.FlaxTemperatureLogitsWarper

< >

( temperature: float )

Parameters

  • temperature (float) — The value used to module the logits distribution.

FlaxLogitsWarper for temperature (exponential scaling output probability distribution).

__call__

< >

( input_ids: Array scores: Array cur_len: int )

class transformers.FlaxTopKLogitsWarper

< >

( top_k: int filter_value: float = -inf min_tokens_to_keep: int = 1 )

Parameters

  • top_k (int) — The number of highest probability vocabulary tokens to keep for top-k-filtering.
  • filter_value (float, optional, defaults to -inf) — All filtered values will be set to this float value.
  • min_tokens_to_keep (int, optional, defaults to 1) — Minimum number of tokens that cannot be filtered.

FlaxLogitsWarper that performs top-k, i.e. restricting to the k highest probability elements.

__call__

< >

( input_ids: Array scores: Array cur_len: int )

class transformers.FlaxTopPLogitsWarper

< >

( top_p: float filter_value: float = -inf min_tokens_to_keep: int = 1 )

Parameters

  • top_p (float) — If set to < 1, only the smallest set of most probable tokens with probabilities that add up to top_p or higher are kept for generation.
  • filter_value (float, optional, defaults to -inf) — All filtered values will be set to this float value.
  • min_tokens_to_keep (int, optional, defaults to 1) — Minimum number of tokens that cannot be filtered.

FlaxLogitsWarper that performs top-p, i.e. restricting to top tokens summing to prob_cut_off <= prob_cut_off.

__call__

< >

( input_ids: Array scores: Array cur_len: int )

class transformers.FlaxWhisperTimeStampLogitsProcessor

< >

( generate_config model_config decoder_input_length )

Parameters

  • generate_config (GenerateConfig) — The generate config used to generate the output. The following parameters are required: eos_token_id (int, optional, defaults to 50257): The id of the end-of-sequence token. no_timestamps_token_id (int, optional, defaults to 50363): The id of the "<|notimestamps|>" token. max_initial_timestamp_index (int, optional, defaults to 1): Used to set the maximum value of the initial timestamp. This is used to prevent the model from predicting timestamps that are too far in the future.

Whisper specific Processor. This processor can be used to force a list of tokens. The processor will set their log probs to inf so that they are sampled at their corresponding index.

__call__

< >

( input_ids scores cur_len )

StoppingCriteria

A StoppingCriteria can be used to change when to stop generation (other than EOS token). Please note that this is exclusively available to our PyTorch implementations.

class transformers.StoppingCriteria

< >

( )

Abstract base class for all stopping criteria that can be applied during generation.

If your stopping criteria depends on the scores input, make sure you pass return_dict_in_generate=True, output_scores=True to generate.

__call__

< >

( input_ids: LongTensor scores: FloatTensor **kwargs )

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax or scores for each vocabulary token after SoftMax. If this stopping criteria depends on the scores input, make sure you pass return_dict_in_generate=True, output_scores=True to generate.
  • kwargs (Dict[str, Any], optional) — Additional stopping criteria specific kwargs.

class transformers.StoppingCriteriaList

< >

( iterable = () )

__call__

< >

( input_ids: LongTensor scores: FloatTensor **kwargs )

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax or scores for each vocabulary token after SoftMax. If this stopping criteria depends on the scores input, make sure you pass return_dict_in_generate=True, output_scores=True to generate.
  • kwargs (Dict[str, Any], optional) — Additional stopping criteria specific kwargs.

class transformers.MaxLengthCriteria

< >

( max_length: int max_position_embeddings: Optional = None )

Parameters

  • max_length (int) — The maximum length that the output sequence can have in number of tokens.
  • max_position_embeddings (int, optional) — The maximum model length, as defined by the model’s config.max_position_embeddings attribute.

This class can be used to stop generation whenever the full generated number of tokens exceeds max_length. Keep in mind for decoder-only type of transformers, this will include the initial prompted tokens.

__call__

< >

( input_ids: LongTensor scores: FloatTensor **kwargs )

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax or scores for each vocabulary token after SoftMax. If this stopping criteria depends on the scores input, make sure you pass return_dict_in_generate=True, output_scores=True to generate.
  • kwargs (Dict[str, Any], optional) — Additional stopping criteria specific kwargs.

class transformers.MaxTimeCriteria

< >

( max_time: float initial_timestamp: Optional = None )

Parameters

  • max_time (float) — The maximum allowed time in seconds for the generation.
  • initial_time (float, optional, defaults to time.time()) — The start of the generation allowed time.

This class can be used to stop generation whenever the full generation exceeds some amount of time. By default, the time will start being counted when you initialize this function. You can override this by passing an initial_time.

__call__

< >

( input_ids: LongTensor scores: FloatTensor **kwargs )

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax or scores for each vocabulary token after SoftMax. If this stopping criteria depends on the scores input, make sure you pass return_dict_in_generate=True, output_scores=True to generate.
  • kwargs (Dict[str, Any], optional) — Additional stopping criteria specific kwargs.

class transformers.StopStringCriteria

< >

( tokenizer: PreTrainedTokenizerBase stop_strings: Union )

Parameters

  • tokenizer (PreTrainedTokenizer) — The model’s associated tokenizer (necessary to extract vocab and tokenize the termination sequences)
  • stop_strings (Union[str, List[str]]) — A list of strings that should end generation. If a string is passed, it will be treated like a list with a single element.

This class can be used to stop generation whenever specific string sequences are generated. It preprocesses the strings together with the tokenizer vocab to find positions where tokens can validly complete the stop strings.

Generation is stopped as soon as a token is generated that completes any of the stop strings. We want to catch any instance in which the stop string would be present in the decoded output, which means we must also catch cases with “overhangs” off one or both ends. To make this more concrete, for the stop string “stop”, any of the following token sequences would trigger the match:

  • [“st”, “op”]
  • [“stop”]
  • [“st”, “opera”]
  • [“sto”, “pper”]
  • [“las”, “topper”]
  • [“s”, “to”, “pped”]

Note that a match will only be triggered if the stop string is at the end of the generated sequence. In other words, these sequences will not trigger a match:

  • [“stop”, “at”]
  • [“st”, “op”, “at”]
  • [“st”, “opera”, “tion”]

The reason these are not a match is that the stop string does not overlap with the final token. If you can remove one or more tokens from the end of the sequence without destroying the stop string, then this criterion will not match that stop string. This is by design; because this check is run after each token is generated, we can’t miss a valid stop string if one is generated, but we don’t want to halt generation just because the stop string exists somewhere in the past input_ids.

How is the match actually performed, though? We do it in quite a confusing way, because we want the entire match process to be compilable with Torch or XLA, which means we cannot use standard string methods. However, it is possible, with some work, to do string matching with pure tensor operations. We’ll begin by describing the algorithm we use with standard string operations, and then at the end we’ll explain how this is converted to pure tensor operations.

The key to the algorithm is an observation: Because the stop string must overlap with the end of the token sequence, we can start at the end of the sequence and work backwards. Specifically, we check that there is an overlap between the start of the final token and the end of the stop_string, or to put it another way, stop_string[-i:] == token[:i] for some i > 0. If you look at the positive examples above, you’ll see the last token in all of them fulfills this property:

  • [“st”, “op”] (overlap is “op”, overlap length == 2)
  • [“stop”] (overlap is “stop”, overlap length == 4)
  • [“st”, “opera”] (overlap is “op”, overlap length == 2)
  • [“sto”, “pper”] (overlap is “p”, overlap length == 1)
  • [“las”, “topper”] (overlap is “top”, overlap length == 3)
  • [“s”, “to”, “pped”] (overlap is “p”, overlap length == 1)

It’s impossible to construct a matching sequence that does not have this property (feel free to verify this yourself). However, although this overlap between the start of the final token and the end of the stop string is necessary for a match, it is not sufficient. We also need to check that the rest of the token sequence is consistent with the stop string.

How do we do that? Let’s use [“s”, “to”, “pped”] as an example. We know that the final token, “pped”, has an overlap of 1 with the stop string, “stop”. We then go back to the previous token, “to”. Since we have already matched 1 character from the stop string, the remainder to check is “sto”. We check that the next token “to” matches the end of the remainder, which it does. We have now matched 3 characters from the stop string, and the remainder to match is “s”. We go back to the previous token again, which is also “s”. This is a match, and so we have matched the entire stop string.

How does it work when the tokens run off the start of the stop string, though? Let’s consider the example of [“las”, “topper”]. The final token, “topper”, has an overlap of 3 with the stop string, “stop”. Therefore, the remaining stop string to match is “s”. We go back to the previous token, “las”. Because the remainder to match is just “s”, with length 1, we consider only the final 1 character from the token, which is “s”. This matches the stop string, and so the entire string is matched.

How do we compute these matches with tensor operations, though? Simply: we efficiently precompute the necessary information for all tokens! For every token, we compute:

  • Its overlap with the end of the stop string, if any
  • The positions inside the stop string where the token matches, including matches that run off the start.
  • The total length of the token

For example, for the token “pped”, we would compute an end overlap of 1, no internal matching positions, and a length of 4. For the token “to”, we would compute no end overlap, a single internal matching position of 1 (counting from the end), and a length of 2. For the token “s”, we would compute no end overlap, a single internal matching position of 3 (again counting from the end) and a length of 1.

As long as we have this information, we can execute the algorithm above without any string comparison operations. We simply perform the following steps:

  • Check if the final token has an end-overlap with the start string
  • Continue backwards, keeping track of how much of the stop string we’ve matched so far
  • At each point, check if the next token has the current position as one of its valid positions
  • Continue until either a match fails, or we completely match the whole stop string

Again, consider [“s”, “to”, “pped”] as an example. “pped” has an end overlap of 1, so we can begin a match. We have matched 1 character so far, so we check that the next token “to”, has 1 as a valid position (again, counting from the end). It does, so we add the length of “to” to our position tracker. We have now matched 3 characters, so we check that the next token “s” has 3 as a valid position. It does, so we add its length to the position tracker. The position tracker is now 4, which is the length of the stop string. We have matched the entire stop string.

In the second case, [“las”, “topper”], “topper” has an end overlap of 3, so we can begin a match. We have matched 3 characters so far, so we check that the next token “las” has 3 as a valid position. It does, because we allow tokens to match positions that run off the start of the stop string. We add its length to the position tracker. The position tracker is now 6, which is greater than the length of the stop string! Don’t panic, though - this also counts as a match of the stop string. We have matched the entire stop string.

Examples:

>>> from transformers import AutoModelForCausalLM, AutoTokenizer

>>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2")
>>> model = AutoModelForCausalLM.from_pretrained("microsoft/phi-2")
>>> inputs = tokenizer("The biggest states in the USA by land area:", return_tensors="pt")

>>> gen_out = model.generate(**inputs)
>>> print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
The biggest states in the USA by land area:
- Alaska
- Texas
- California

>>> # Passing one or more stop strings will halt generation after those strings are emitted
>>> # Note that generating with stop strings requires you to pass the tokenizer too
>>> gen_out = model.generate(**inputs, stop_strings=["Texas"], tokenizer=tokenizer)
>>> print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
The biggest states in the USA by land area:
- Alaska
- Texas

__call__

< >

( input_ids: LongTensor scores: FloatTensor **kwargs )

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax or scores for each vocabulary token after SoftMax. If this stopping criteria depends on the scores input, make sure you pass return_dict_in_generate=True, output_scores=True to generate.
  • kwargs (Dict[str, Any], optional) — Additional stopping criteria specific kwargs.

class transformers.EosTokenCriteria

< >

( eos_token_id: Union )

Parameters

  • eos_token_id (Union[int, List[int], torch.Tensor]) — The id(s) of the end-of-sequence token.

This class can be used to stop generation whenever the “end-of-sequence” token is generated. By default, it uses the model.generation_config.eos_token_id.

__call__

< >

( input_ids: LongTensor scores: FloatTensor **kwargs )

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • scores (torch.FloatTensor of shape (batch_size, config.vocab_size)) — Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax or scores for each vocabulary token after SoftMax. If this stopping criteria depends on the scores input, make sure you pass return_dict_in_generate=True, output_scores=True to generate.
  • kwargs (Dict[str, Any], optional) — Additional stopping criteria specific kwargs.

Constraints

A Constraint can be used to force the generation to include specific tokens or sequences in the output. Please note that this is exclusively available to our PyTorch implementations.

class transformers.Constraint

< >

( )

Abstract base class for all constraints that can be applied during generation. It must define how the constraint can be satisfied.

All classes that inherit Constraint must follow the requirement that

completed = False
while not completed:
    _, completed = constraint.update(constraint.advance())

will always terminate (halt).

advance

< >

( ) → token_ids(torch.tensor)

Returns

token_ids(torch.tensor)

Must be a tensor of a list of indexable tokens, not some integer.

When called, returns the token that would take this constraint one step closer to being fulfilled.

copy

< >

( stateful = False ) → constraint(Constraint)

Returns

constraint(Constraint)

The same constraint as the one being called from.

Creates a new instance of this constraint.

does_advance

< >

( token_id: int )

Reads in a token and returns whether it creates progress.

remaining

< >

( )

Returns the number of remaining steps of advance() in order to complete this constraint.

reset

< >

( )

Resets the state of this constraint to its initialization. We would call this in cases where the fulfillment of a constraint is abrupted by an unwanted token.

test

< >

( )

Tests whether this constraint has been properly defined.

update

< >

( token_id: int ) → stepped(bool)

Returns

stepped(bool)

Whether this constraint has become one step closer to being fulfuilled. completed(bool): Whether this constraint has been completely fulfilled by this token being generated. reset (bool): Whether this constraint has reset its progress by this token being generated.

Reads in a token and returns booleans that indicate the progress made by it. This function will update the state of this object unlikes does_advance(self, token_id: int).

This isn’t to test whether a certain token will advance the progress; it’s to update its state as if it has been generated. This becomes important if token_id != desired token (refer to else statement in PhrasalConstraint)

class transformers.PhrasalConstraint

< >

( token_ids: List )

Parameters

  • token_ids (List[int]) — The id of the token that must be generated by the output.

Constraint enforcing that an ordered sequence of tokens is included in the output.

class transformers.DisjunctiveConstraint

< >

( nested_token_ids: List )

Parameters

  • nested_token_ids (List[List[int]]) — A list of words, where each word is a list of ids. This constraint is fulfilled by generating just one from the list of words.

A special Constraint that is fulfilled by fulfilling just one of several constraints.

class transformers.ConstraintListState

< >

( constraints: List )

Parameters

  • constraints (List[Constraint]) — A list of Constraint objects that must be fulfilled by the beam scorer.

A class for beam scorers to track its progress through a list of constraints.

advance

< >

( )

The list of tokens to generate such that we can make progress. By “list” we don’t mean the list of token that will fully fulfill a constraint.

Given constraints c_i = {t_ij | j == # of tokens}, If we’re not in the middle of progressing through a specific constraint c_i, we return:

[t_k1 for k in indices of unfulfilled constraints]

If we are in the middle of a constraint, then we return: [t_ij], where i is the index of the inprogress constraint, j is the next step for the constraint.

Though we don’t care which constraint is fulfilled first, if we are in the progress of fulfilling a constraint, that’s the only one we’ll return.

reset

< >

( token_ids: Optional )

token_ids: the tokens generated thus far to reset the state of the progress through constraints.

BeamSearch

class transformers.BeamScorer

< >

( )

Abstract base class for all beam scorers that are used for ~PreTrainedModel.beam_search and ~PreTrainedModel.beam_sample.

process

< >

( input_ids: LongTensor next_scores: FloatTensor next_tokens: LongTensor next_indices: LongTensor **kwargs ) → UserDict

Parameters

  • input_ids (torch.LongTensor of shape (batch_size * num_beams, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using any class inheriting from PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • next_scores (torch.FloatTensor of shape (batch_size, 2 * num_beams)) — Current scores of the top 2 * num_beams non-finished beam hypotheses.
  • next_tokens (torch.LongTensor of shape (batch_size, 2 * num_beams)) — input_ids of the tokens corresponding to the top 2 * num_beams non-finished beam hypotheses.
  • next_indices (torch.LongTensor of shape (batch_size, 2 * num_beams)) — Beam indices indicating to which beam hypothesis the next_tokens correspond.
  • pad_token_id (int, optional) — The id of the padding token.
  • eos_token_id (Union[int, List[int]], optional) — The id of the end-of-sequence token. Optionally, use a list to set multiple end-of-sequence tokens.
  • beam_indices (torch.LongTensor, optional) — Beam indices indicating to which beam hypothesis each token correspond.
  • group_index (int, optional) — The index of the group of beams. Used with ~PreTrainedModel.group_beam_search.

Returns

UserDict

A dictionary composed of the fields as defined above:

  • next_beam_scores (torch.FloatTensor of shape (batch_size * num_beams)) — Updated scores of all non-finished beams.
  • next_beam_tokens (torch.FloatTensor of shape (batch_size * num_beams)) — Next tokens to be added to the non-finished beam_hypotheses.
  • next_beam_indices (torch.FloatTensor of shape (batch_size * num_beams)) — Beam indices indicating to which beam the next tokens shall be added.

finalize

< >

( input_ids: LongTensor next_scores: FloatTensor next_tokens: LongTensor next_indices: LongTensor max_length: int **kwargs ) → torch.LongTensor of shape (batch_size * num_return_sequences, sequence_length)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size * num_beams, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using any class inheriting from PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • final_beam_scores (torch.FloatTensor of shape (batch_size * num_beams)) — The final scores of all non-finished beams.
  • final_beam_tokens (torch.FloatTensor of shape (batch_size * num_beams)) — The last tokens to be added to the non-finished beam_hypotheses.
  • final_beam_indices (torch.FloatTensor of shape (batch_size * num_beams)) — The beam indices indicating to which beam the final_beam_tokens shall be added.
  • pad_token_id (int, optional) — The id of the padding token.
  • eos_token_id (Union[int, List[int]], optional) — The id of the end-of-sequence token. Optionally, use a list to set multiple end-of-sequence tokens.

Returns

torch.LongTensor of shape (batch_size * num_return_sequences, sequence_length)

The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.

class transformers.BeamSearchScorer

< >

( batch_size: int num_beams: int device: device length_penalty: Optional = 1.0 do_early_stopping: Union = False num_beam_hyps_to_keep: Optional = 1 num_beam_groups: Optional = 1 max_length: Optional = None )

Parameters

  • batch_size (int) — Batch Size of input_ids for which standard beam search decoding is run in parallel.
  • num_beams (int) — Number of beams for beam search.
  • device (torch.device) — Defines the device type (e.g., "cpu" or "cuda") on which this instance of BeamSearchScorer will be allocated.
  • length_penalty (float, optional, defaults to 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.
  • do_early_stopping (bool or str, optional, defaults to 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).
  • num_beam_hyps_to_keep (int, optional, defaults to 1) — The number of beam hypotheses that shall be returned upon calling finalize().
  • num_beam_groups (int, optional, defaults to 1) — Number of groups to divide num_beams into in order to ensure diversity among different groups of beams. See this paper for more details.
  • max_length (int, optional) — The maximum length of the sequence to be generated.

BeamScorer implementing standard beam search decoding.

Adapted in part from Facebook’s XLM beam search code.

Reference for the diverse beam search algorithm and implementation Ashwin Kalyan’s DBS implementation

process

< >

( input_ids: LongTensor next_scores: FloatTensor next_tokens: LongTensor next_indices: LongTensor pad_token_id: Union = None eos_token_id: Union = None beam_indices: Optional = None group_index: Optional = 0 decoder_prompt_len: Optional = 0 )

finalize

< >

( input_ids: LongTensor final_beam_scores: FloatTensor final_beam_tokens: LongTensor final_beam_indices: LongTensor max_length: int pad_token_id: Union = None eos_token_id: Union = None beam_indices: Optional = None decoder_prompt_len: Optional = 0 )

class transformers.ConstrainedBeamSearchScorer

< >

( batch_size: int num_beams: int constraints: List device: device length_penalty: Optional = 1.0 do_early_stopping: Union = False num_beam_hyps_to_keep: Optional = 1 num_beam_groups: Optional = 1 max_length: Optional = None )

Parameters

  • batch_size (int) — Batch Size of input_ids for which standard beam search decoding is run in parallel.
  • num_beams (int) — Number of beams for beam search.
  • constraints (List[Constraint]) — A list of positive constraints represented as Constraint objects that must be fulfilled in the generation output. For more information, the documentation of Constraint should be read.
  • device (torch.device) — Defines the device type (e.g., "cpu" or "cuda") on which this instance of BeamSearchScorer will be allocated.
  • length_penalty (float, optional, defaults to 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.
  • do_early_stopping (bool or str, optional, defaults to 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).
  • num_beam_hyps_to_keep (int, optional, defaults to 1) — The number of beam hypotheses that shall be returned upon calling finalize().
  • num_beam_groups (int, optional, defaults to 1) — Number of groups to divide num_beams into in order to ensure diversity among different groups of beams. See this paper for more details.
  • max_length (int, optional) — The maximum length of the sequence to be generated.

BeamScorer implementing constrained beam search decoding.

process

< >

( input_ids: LongTensor next_scores: FloatTensor next_tokens: LongTensor next_indices: LongTensor scores_for_all_vocab: FloatTensor pad_token_id: Union = None eos_token_id: Union = None beam_indices: Optional = None decoder_prompt_len: Optional = 0 ) → UserDict

Parameters

  • input_ids (torch.LongTensor of shape (batch_size * num_beams, sequence_length)) — Indices of input sequence tokens in the vocabulary.

    Indices can be obtained using any class inheriting from PreTrainedTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • next_scores (torch.FloatTensor of shape (batch_size, 2 * num_beams)) — Current scores of the top 2 * num_beams non-finished beam hypotheses.
  • next_tokens (torch.LongTensor of shape (batch_size, 2 * num_beams)) — input_ids of the tokens corresponding to the top 2 * num_beams non-finished beam hypotheses.
  • next_indices (torch.LongTensor of shape (batch_size, 2 * num_beams)) — Beam indices indicating to which beam hypothesis the next_tokens correspond.
  • scores_for_all_vocab (torch.FloatTensor of shape (batch_size * num_beams, sequence_length)) — The scores of all tokens in the vocabulary for each of the beam hypotheses.
  • pad_token_id (int, optional) — The id of the padding token.
  • eos_token_id (Union[int, List[int]], optional) — The id of the end-of-sequence token. Optionally, use a list to set multiple end-of-sequence tokens.
  • beam_indices (torch.LongTensor, optional) — Beam indices indicating to which beam hypothesis each token correspond.
  • decoder_prompt_len (int, optional) — The length of prompt that is included in the input to decoder.

Returns

UserDict

A dictionary composed of the fields as defined above:

  • next_beam_scores (torch.FloatTensor of shape (batch_size * num_beams)) — Updated scores of all non-finished beams.

  • next_beam_tokens (torch.FloatTensor of shape (batch_size * num_beams)) — Next tokens to be added to the non-finished beam_hypotheses.

  • next_beam_indices (torch.FloatTensor of shape (batch_size * num_beams)) — Beam indices indicating to which beam the next tokens shall be added.

finalize

< >

( input_ids: LongTensor final_beam_scores: FloatTensor final_beam_tokens: LongTensor final_beam_indices: LongTensor max_length: int pad_token_id: Union = None eos_token_id: Union = None beam_indices: Optional = None decoder_prompt_len: Optional = 0 )

Streamers

class transformers.TextStreamer

< >

( tokenizer: AutoTokenizer skip_prompt: bool = False **decode_kwargs )

Parameters

  • tokenizer (AutoTokenizer) — The tokenized used to decode the tokens.
  • skip_prompt (bool, optional, defaults to False) — Whether to skip the prompt to .generate() or not. Useful e.g. for chatbots.
  • decode_kwargs (dict, optional) — Additional keyword arguments to pass to the tokenizer’s decode method.

Simple text streamer that prints the token(s) to stdout as soon as entire words are formed.

The API for the streamer classes is still under development and may change in the future.

Examples:

>>> from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

>>> tok = AutoTokenizer.from_pretrained("openai-community/gpt2")
>>> model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2")
>>> inputs = tok(["An increasing sequence: one,"], return_tensors="pt")
>>> streamer = TextStreamer(tok)

>>> # Despite returning the usual output, the streamer will also print the generated text to stdout.
>>> _ = model.generate(**inputs, streamer=streamer, max_new_tokens=20)
An increasing sequence: one, two, three, four, five, six, seven, eight, nine, ten, eleven,

end

< >

( )

Flushes any remaining cache and prints a newline to stdout.

on_finalized_text

< >

( text: str stream_end: bool = False )

Prints the new text to stdout. If the stream is ending, also prints a newline.

put

< >

( value )

Receives tokens, decodes them, and prints them to stdout as soon as they form entire words.

class transformers.TextIteratorStreamer

< >

( tokenizer: AutoTokenizer skip_prompt: bool = False timeout: Optional = None **decode_kwargs )

Parameters

  • tokenizer (AutoTokenizer) — The tokenized used to decode the tokens.
  • skip_prompt (bool, optional, defaults to False) — Whether to skip the prompt to .generate() or not. Useful e.g. for chatbots.
  • timeout (float, optional) — The timeout for the text queue. If None, the queue will block indefinitely. Useful to handle exceptions in .generate(), when it is called in a separate thread.
  • decode_kwargs (dict, optional) — Additional keyword arguments to pass to the tokenizer’s decode method.

Streamer that stores print-ready text in a queue, to be used by a downstream application as an iterator. This is useful for applications that benefit from acessing the generated text in a non-blocking way (e.g. in an interactive Gradio demo).

The API for the streamer classes is still under development and may change in the future.

Examples:

>>> from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
>>> from threading import Thread

>>> tok = AutoTokenizer.from_pretrained("openai-community/gpt2")
>>> model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2")
>>> inputs = tok(["An increasing sequence: one,"], return_tensors="pt")
>>> streamer = TextIteratorStreamer(tok)

>>> # Run the generation in a separate thread, so that we can fetch the generated text in a non-blocking way.
>>> generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=20)
>>> thread = Thread(target=model.generate, kwargs=generation_kwargs)
>>> thread.start()
>>> generated_text = ""
>>> for new_text in streamer:
...     generated_text += new_text
>>> generated_text
'An increasing sequence: one, two, three, four, five, six, seven, eight, nine, ten, eleven,'

on_finalized_text

< >

( text: str stream_end: bool = False )

Put the new text in the queue. If the stream is ending, also put a stop signal in the queue.

Caches

class transformers.Cache

< >

( )

Base, abstract class for all caches. The actual data structure is specific to each subclass.

update

< >

( key_states: Tensor value_states: Tensor layer_idx: int cache_kwargs: Optional = None )

Parameters

  • key_states (torch.Tensor) — The new key states to cache.
  • value_states (torch.Tensor) — The new value states to cache.
  • layer_idx (int) — The index of the layer to cache the states for.
  • cache_kwargs (Dict[str, Any], optional) — Additional arguments for the cache subclass. These are specific to each subclass and allow new types of cache to be created.

Updates the cache with the new key_states and value_states for the layer layer_idx.

class transformers.CacheConfig

< >

( cache_implementation: None )

Base class for cache configs

update

< >

( **kwargs ) → Dict[str, Any]

Parameters

  • kwargs (Dict[str, Any]) — Dictionary of attributes to tentatively update this class.

Returns

Dict[str, Any]

Dictionary containing all the key-value pairs that were not used to update the instance.

Updates attributes of this class instance with attributes from kwargs if they match existing attributes, returning all the unused kwargs.

class transformers.QuantizedCacheConfig

< >

( backend: str = 'quanto' nbits: Optional = 4 axis_key: Optional = 0 axis_value: Optional = 0 q_group_size: Optional = 64 residual_length: Optional = 128 compute_dtype: Optional = torch.float16 device: Optional = 'cpu' )

Parameters

  • backend (str, optional, defaults to "quanto") — Backend to use when performing quantization, Can be one of [quanto, HQQ]
  • nbits (Optional[int], optional, defaults to 4) — Number of bits, can be 2 or 4 for the quanto backend and one of [1, 2, 3, 4, 8] for the HQQ backend. Defaults to 2.
  • axis_key (int, optional, defaults to 0) — Axis over which to perform grouping for the key tensors. Can be [0, -1] for quanto backend and [0, 1] for HQQ backend.
  • axis_value (int, optional, defaults to 0) — Axis over which to perform grouping for the value tensors. Can be [0, -1] for quanto backend and [0, 1] for HQQ backend.
  • q_group_size (Optional[int], optional, defaults to 64) — Size of the quantization group, should be a divisor of the model’s hidden dimension. Defaults to 64.
  • residual_length (Optional[int], optional, defaults to 128) — Length of the residual cache which will always be stored in original presicion. Defaults to 128.
  • compute_dtype (torch.dtype, optional, defaults to torch.float16) — The defualt dtype used for computations in the model. Keys and Values will be cast to this dtype after dequantization.
  • device (str, optional, defaults to "cpu") — Device on which to perform computations, should be same as the model’s device.

Configuration class for quantized cache settings.

validate

< >

( )

Validates if the arguments passed are correct

class transformers.DynamicCache

< >

( )

A cache that grows dynamically as more tokens are generated. This is the default for generative models.

It stores the Key and Value states as a list of tensors, one for each layer. The expected shape for each tensor is [batch_size, num_heads, seq_len, head_dim].

update

< >

( key_states: Tensor value_states: Tensor layer_idx: int cache_kwargs: Optional = None )

Parameters

  • key_states (torch.Tensor) — The new key states to cache.
  • value_states (torch.Tensor) — The new value states to cache.
  • layer_idx (int) — The index of the layer to cache the states for.
  • cache_kwargs (Dict[str, Any], optional) — Additional arguments for the cache subclass. No additional arguments are used in DynamicCache.

Updates the cache with the new key_states and value_states for the layer layer_idx.

get_seq_length

< >

( layer_idx: Optional = 0 )

Returns the sequence length of the cached states. A layer index can be optionally passed.

reorder_cache

< >

( beam_idx: LongTensor )

Reorders the cache for beam search, given the selected beam indices.

to_legacy_cache

< >

( )

Converts the DynamicCache instance into the its equivalent in the legacy cache format. Used for backward compatibility.

from_legacy_cache

< >

( past_key_values: Optional = None )

Converts a cache in the legacy cache format into an equivalent DynamicCache. Used for backward compatibility.

class transformers.QuantizedCache

< >

( cache_config: QuantizedCacheConfig )

A quantizer cache similar to what is described in the KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache paper. It allows the model to generate longer sequence length without allocating too much memory for Key and Value cache by applying quantization.

The cache has two types of storage, one for original precision and one for the quantized cache. A residual length is set as a maximum capacity for the original precision cache. When the length goes beyond maximum capacity, the original precision cache is discarded and moved into the quantized cache. The quantization is done per-channel with a set q_group_size for both Keys and Values, in contrast to what was described in the paper.

It stores Keys and Values a list of quantized tensors (tuples in case we need to store metadata), one for each layer. Additionally, it stores the Key and Value in original precision states as a list of tensors, one for each layer. The size of each tensor is [batch_size, num_heads, seq_len - residual_length, head_dim]

update

< >

( key_states: Tensor value_states: Tensor layer_idx: int cache_kwargs: Optional = None )

get_seq_length

< >

( layer_idx: Optional = 0 )

Returns the sequence length of the cached states. A layer index can be optionally passed.

class transformers.QuantoQuantizedCache

< >

( cache_config: CacheConfig )

Parameters

  • cache_config (QuantizedCacheConfig,) — A configuration containing all the arguments to be used by the quantizer, including axis, qtype and group size.

Quantized Cache class that uses quanto as a backend to perform quantization. Current implementation supports int2 and int4 dtypes only.

class transformers.HQQQuantizedCache

< >

( cache_config: CacheConfig )

Parameters

  • cache_config (QuantizedCacheConfig,) — A configuration containing all the arguments to be used by the quantizer, including axis, qtype and group size.

Quantized Cache class that uses HQQ as a backend to perform quantization. Current implementation supports int2, int4, int8 dtypes.

class transformers.SinkCache

< >

( window_length: int num_sink_tokens: int )

Parameters

  • window_length (int) — The length of the context window.
  • num_sink_tokens (int) — The number of sink tokens. See the original paper for more information.

A cache that as described in the Attention Sinks paper. It allows the model to generate beyond the length of its context window, without losing fluency in the conversation. As it discards past tokens, the model will lose the ability to generate tokens that depend on the context that was discarded.

It stores the Key and Value states as a list of tensors, one for each layer. The expected shape for each tensor is [batch_size, num_heads, seq_len, head_dim].

update

< >

( key_states: Tensor value_states: Tensor layer_idx: int cache_kwargs: Optional = None )

Parameters

  • key_states (torch.Tensor) — The new key states to cache.
  • value_states (torch.Tensor) — The new value states to cache.
  • layer_idx (int) — The index of the layer to cache the states for.
  • cache_kwargs (Dict[str, Any], optional) — Additional arguments for the cache subclass. The following arguments can be used in SinkCache: sin, cos and partial_rotation_size. These arguments are used with models using RoPE, to recompute the rotation as the tokens are shifted.

Updates the cache with the new key_states and value_states for the layer layer_idx.

get_seq_length

< >

( layer_idx: Optional = 0 )

Returns the sequence length of the cached states. A layer index can be optionally passed.

reorder_cache

< >

( beam_idx: LongTensor )

Reorders the cache for beam search, given the selected beam indices.

class transformers.StaticCache

< >

( config: PretrainedConfig max_batch_size: int max_cache_len: int device dtype = None )

Parameters

  • config (`PretrainedConfig) — The configuration file defining the shape-related attributes required to initialize the static cache.
  • max_batch_size (int) — The maximum batch size with which the model will be used.
  • max_cache_len (int) — The maximum sequence length with which the model will be used.
  • device (torch.device) — The device on which the cache should be initialized. Should be the same as the layer.
  • dtype (optional, defaults to torch.float32) — The default dtype to use when initializing the layer.

Static Cache class to be used with torch.compile(model).

update

< >

( key_states: Tensor value_states: Tensor layer_idx: int cache_kwargs: Optional = None )

Parameters

  • key_states (torch.Tensor) — The new key states to cache.
  • value_states (torch.Tensor) — The new value states to cache.
  • layer_idx (int) — The index of the layer to cache the states for.
  • cache_kwargs (Dict[str, Any], optional) — Additional arguments for the cache subclass. The StaticCache needs the cache_position input to know how where to write in the cache.

Updates the cache with the new key_states and value_states for the layer layer_idx. It is VERY important to index using a tensor, otherwise you introduce a copy to the device.

get_seq_length

< >

( layer_idx: Optional = 0 )

Returns the sequence length of the cached states that were seen by the model.

reset

< >

( )

Resets the cache values while preserving the objects

class transformers.EncoderDecoderCache

< >

( self_attention_cache: Cache cross_attention_cache: Cache )

Base, abstract class for all encoder-decoder caches. Can be used to hold combinations of self-attention and cross-attention caches.

get_seq_length

< >

( layer_idx: Optional = 0 )

Returns the sequence length of the cached states. A layer index can be optionally passed.

to_legacy_cache

< >

( )

Converts the EncoderDecoderCache instance into its equivalent in the legacy cache format.

from_legacy_cache

< >

( past_key_values: Optional = None )

Converts a cache in the legacy cache format into an equivalent EncoderDecoderCache.

reset

< >

( )

reorder_cache

< >

( beam_idx: LongTensor )

Reorders the cache for beam search, given the selected beam indices.

Watermark Utils

class transformers.WatermarkDetector

< >

( model_config: PretrainedConfig device: str watermarking_config: Union ignore_repeated_ngrams: bool = False max_cache_size: int = 128 )

Parameters

  • model_config (PretrainedConfig) — The model config that will be used to get model specific arguments used when generating.
  • device (str) — The device which was used during watermarked text generation.
  • watermarking_config (Union[WatermarkingConfig, Dict]) — The exact same watermarking config and arguments used when generating text.
  • ignore_repeated_ngrams (bool, optional, defaults to False) — Whether to count every unique ngram only once or not.
  • max_cache_size (int, optional, defaults to 128) — The max size to be used for LRU caching of seeding/sampling algorithms called for every token.

Detector for detection of watermark generated text. The detector needs to be given the exact same settings that were given during text generation to replicate the watermark greenlist generation and so detect the watermark. This includes the correct device that was used during text generation, the correct watermarking arguments and the correct tokenizer vocab size. The code was based on the original repo.

See the paper for more information.

Examples:

>>> from transformers import AutoTokenizer, AutoModelForCausalLM, WatermarkDetector, WatermarkingConfig

>>> model_id = "openai-community/gpt2"
>>> model = AutoModelForCausalLM.from_pretrained(model_id)
>>> tok = AutoTokenizer.from_pretrained(model_id)
>>> tok.pad_token_id = tok.eos_token_id
>>> tok.padding_side = "left"

>>> inputs = tok(["This is the beginning of a long story", "Alice and Bob are"], padding=True, return_tensors="pt")
>>> input_len = inputs["input_ids"].shape[-1]

>>> # first generate text with watermark and without
>>> watermarking_config = WatermarkingConfig(bias=2.5, seeding_scheme="selfhash")
>>> out_watermarked = model.generate(**inputs, watermarking_config=watermarking_config, do_sample=False, max_length=20)
>>> out = model.generate(**inputs, do_sample=False, max_length=20)

>>> # now we can instantiate the detector and check the generated text
>>> detector = WatermarkDetector(model_config=model.config, device="cpu", watermarking_config=watermarking_config)
>>> detection_out_watermarked = detector(out_watermarked, return_dict=True)
>>> detection_out = detector(out, return_dict=True)
>>> detection_out_watermarked.prediction
array([ True,  True])

>>> detection_out.prediction
array([False,  False])

__call__

< >

( input_ids: LongTensor z_threshold: float = 3.0 return_dict: bool = False ) → WatermarkDetectorOutput or np.array

Parameters

  • input_ids (torch.LongTensor) — The watermark generated text. It is advised to remove the prompt, which can affect the detection.
  • z_threshold (Dict, optional, defaults to 3.0) — Changing this threshold will change the sensitivity of the detector. Higher z threshold gives less sensitivity and vice versa for lower z threshold.
  • return_dict (bool, optional, defaults to False) — Whether to return ~generation.WatermarkDetectorOutput or not. If not it will return boolean predictions,

Returns

WatermarkDetectorOutput or np.array

A WatermarkDetectorOutput if return_dict=True otherwise a np.array.

ma

< > Update on GitHub

←Utilities for Trainer Utilities for Image Processors→