utils.py

# coding=utf-8
# Copyright 2020 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import copy
import inspect
import warnings
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union

import torch
import torch.distributed as dist
from torch import nn

from ..deepspeed import is_deepspeed_zero3_enabled
from ..modeling_outputs import CausalLMOutputWithPast, Seq2SeqLMOutput
from ..models.auto import (
    MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
    MODEL_FOR_CAUSAL_LM_MAPPING,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
    MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
    MODEL_FOR_VISION_2_SEQ_MAPPING,
)
from ..utils import ModelOutput, logging
from .beam_constraints import DisjunctiveConstraint, PhrasalConstraint
from .beam_search import BeamScorer, BeamSearchScorer, ConstrainedBeamSearchScorer
from .configuration_utils import GenerationConfig
from .logits_process import (
    ClassifierFreeGuidanceLogitsProcessor,
    EncoderNoRepeatNGramLogitsProcessor,
    EncoderRepetitionPenaltyLogitsProcessor,
    EpsilonLogitsWarper,
    EtaLogitsWarper,
    ExponentialDecayLengthPenalty,
    ForcedBOSTokenLogitsProcessor,
    ForcedEOSTokenLogitsProcessor,
    ForceTokensLogitsProcessor,
    HammingDiversityLogitsProcessor,
    InfNanRemoveLogitsProcessor,
    LogitNormalization,
    LogitsProcessorList,
    MinLengthLogitsProcessor,
    MinNewTokensLengthLogitsProcessor,
    NoBadWordsLogitsProcessor,
    NoRepeatNGramLogitsProcessor,
    PrefixConstrainedLogitsProcessor,
    RepetitionPenaltyLogitsProcessor,
    SequenceBiasLogitsProcessor,
    SuppressTokensAtBeginLogitsProcessor,
    SuppressTokensLogitsProcessor,
    TemperatureLogitsWarper,
    TopKLogitsWarper,
    TopPLogitsWarper,
    TypicalLogitsWarper,
)
from .stopping_criteria import (
    MaxLengthCriteria,
    MaxTimeCriteria,
    StoppingCriteria,
    StoppingCriteriaList,
    validate_stopping_criteria,
)


if TYPE_CHECKING:
    from ..modeling_utils import PreTrainedModel
    from .streamers import BaseStreamer

logger = logging.get_logger(__name__)


@dataclass
class GreedySearchDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using greedy search.


    Args:
        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)`.
        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)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class ContrastiveSearchEncoderDecoderOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using contrastive search.

    Args:
        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)`.
        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)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class ContrastiveSearchDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using contrastive search.

    Args:
        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)`.
        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)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class GreedySearchEncoderDecoderOutput(ModelOutput):
    """
    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)


    Args:
        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)`.
        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)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class SampleDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using sampling.


    Args:
        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*num_return_sequences, 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 `(num_return_sequences*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 `(num_return_sequences*batch_size, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class SampleEncoderDecoderOutput(ModelOutput):
    """
    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)


    Args:
        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*num_return_sequences, 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_return_sequences, 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_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_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_return_sequences, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class BeamSearchDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using beam search.

    Args:
        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*num_return_sequences, 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)`.
    """

    sequences: torch.LongTensor = None
    sequences_scores: Optional[torch.FloatTensor] = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    beam_indices: Optional[torch.LongTensor] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class BeamSearchEncoderDecoderOutput(ModelOutput):
    """
    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)

    Args:
        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)`.
        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)`.
    """

    sequences: torch.LongTensor = None
    sequences_scores: Optional[torch.FloatTensor] = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    beam_indices: Optional[torch.LongTensor] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class BeamSampleDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using beam sample.

    Args:
        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_sequence)`, *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*num_return_sequences, 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, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    sequences_scores: Optional[torch.FloatTensor] = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    beam_indices: Optional[torch.LongTensor] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class BeamSampleEncoderDecoderOutput(ModelOutput):
    """
    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)

    Args:
        sequences (`torch.LongTensor` 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 (`torch.FloatTensor` 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(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)`).
        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, 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_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, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    sequences_scores: Optional[torch.FloatTensor] = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    beam_indices: Optional[torch.LongTensor] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


GreedySearchOutput = Union[GreedySearchEncoderDecoderOutput, GreedySearchDecoderOnlyOutput]
SampleOutput = Union[SampleEncoderDecoderOutput, SampleDecoderOnlyOutput]
BeamSearchOutput = Union[BeamSearchEncoderDecoderOutput, BeamSearchDecoderOnlyOutput]
BeamSampleOutput = Union[BeamSampleEncoderDecoderOutput, BeamSampleDecoderOnlyOutput]
ContrastiveSearchOutput = Union[ContrastiveSearchEncoderDecoderOutput, ContrastiveSearchDecoderOnlyOutput]
GenerateOutput = Union[GreedySearchOutput, SampleOutput, BeamSearchOutput, BeamSampleOutput, ContrastiveSearchOutput]


class GenerationMixin:
    """
    A class containing all functions for auto-regressive text generation, to be used as a mixin in [`PreTrainedModel`].

    The class exposes [`~generation.GenerationMixin.generate`], which can be used for:
        - *greedy decoding* by calling [`~generation.GenerationMixin.greedy_search`] if `num_beams=1` and
          `do_sample=False`
        - *contrastive search* by calling [`~generation.GenerationMixin.contrastive_search`] if `penalty_alpha>0` and
          `top_k>1`
        - *multinomial sampling* by calling [`~generation.GenerationMixin.sample`] if `num_beams=1` and
          `do_sample=True`
        - *beam-search decoding* by calling [`~generation.GenerationMixin.beam_search`] if `num_beams>1` and
          `do_sample=False`
        - *beam-search multinomial sampling* by calling [`~generation.GenerationMixin.beam_sample`] if `num_beams>1`
          and `do_sample=True`
        - *diverse beam-search decoding* by calling [`~generation.GenerationMixin.group_beam_search`], if `num_beams>1`
          and `num_beam_groups>1`
        - *constrained beam-search decoding* by calling [`~generation.GenerationMixin.constrained_beam_search`], if
          `constraints!=None` or `force_words_ids!=None`

    You do not need to call any of the above methods directly. Pass custom parameter values to 'generate' instead. To
    learn more about decoding strategies refer to the [text generation strategies guide](../generation_strategies).
    """

    def prepare_inputs_for_generation(self, *args, **kwargs):
        raise NotImplementedError(
            "A model class needs to define a `prepare_inputs_for_generation` method in order to use `.generate()`."
        )

    def _prepare_model_inputs(
        self,
        inputs: Optional[torch.Tensor] = None,
        bos_token_id: Optional[int] = None,
        model_kwargs: Optional[Dict[str, torch.Tensor]] = None,
    ) -> Tuple[torch.Tensor, Optional[str], Dict[str, torch.Tensor]]:
        """
        This function extracts the model-specific `inputs` for generation.
        """
        # 1. retrieve all kwargs that are non-None or non-model input related.
        # some encoder-decoder models have different names for model and encoder
        if (
            self.config.is_encoder_decoder
            and hasattr(self, "encoder")
            and self.encoder.main_input_name != self.main_input_name
        ):
            input_name = self.encoder.main_input_name
        else:
            input_name = self.main_input_name

        model_kwargs = {k: v for k, v in model_kwargs.items() if v is not None or k != input_name}

        # 2. check whether model_input_name is passed as kwarg
        # if yes and `inputs` is None use kwarg inputs
        inputs_kwarg = model_kwargs.pop(input_name, None)
        if inputs_kwarg is not None and inputs is not None:
            raise ValueError(
                f"`inputs`: {inputs}` were passed alongside {input_name} which is not allowed."
                f"Make sure to either pass {inputs} or {input_name}=..."
            )
        elif inputs_kwarg is not None:
            inputs = inputs_kwarg

        # 3. In the presence of `inputs_embeds` for text models:
        # - decoder-only models should complain if the user attempts to pass `inputs_embeds`, but the model
        # doesn't have its forwarding implemented. `inputs_embeds` is kept in `model_kwargs` and can coexist with
        # input_ids (`inputs_embeds` will be used in the 1st generation step, as opposed to `input_ids`)
        # - encoder-decoder models should complain if the user attempts to pass `inputs_embeds` and `input_ids`, and
        # pull the former to inputs. It will be used in place of `input_ids` to get the encoder hidden states.
        if input_name == "input_ids" and "inputs_embeds" in model_kwargs:
            if not self.config.is_encoder_decoder:
                has_inputs_embeds_forwarding = "inputs_embeds" in set(
                    inspect.signature(self.prepare_inputs_for_generation).parameters.keys()
                )
                if not has_inputs_embeds_forwarding:
                    raise ValueError(
                        f"You passed `inputs_embeds` to `.generate()`, but the model class {self.__class__.__name__} "
                        "doesn't have its forwarding implemented. See the GPT2 implementation for an example "
                        "(https://github.com/huggingface/transformers/pull/21405), and feel free to open a PR with it!"
                    )
                # In this case, `input_ids` is moved to the `model_kwargs`, so a few automations (like the creation of
                # the attention mask) can rely on the actual model input.
                model_kwargs["input_ids"] = self._maybe_initialize_input_ids_for_generation(
                    inputs, bos_token_id, model_kwargs=model_kwargs
                )
            else:
                if inputs is not None:
                    raise ValueError("You passed `inputs_embeds` and `input_ids` to `.generate()`. Please pick one.")
            inputs, input_name = model_kwargs["inputs_embeds"], "inputs_embeds"

        # 4. if `inputs` is still None, try to create `input_ids` from BOS token
        inputs = self._maybe_initialize_input_ids_for_generation(inputs, bos_token_id, model_kwargs)
        return inputs, input_name, model_kwargs

    def adjust_logits_during_generation(self, logits: torch.FloatTensor, **kwargs) -> torch.FloatTensor:
        """
        Implement in subclasses of [`PreTrainedModel`] for custom behavior to adjust the logits in the generate method.
        """
        return logits

    def _maybe_initialize_input_ids_for_generation(
        self,
        inputs: Optional[torch.Tensor] = None,
        bos_token_id: Optional[int] = None,
        model_kwargs: Optional[Dict[str, torch.Tensor]] = None,
    ) -> torch.LongTensor:
        """Initializes input ids for generation, if necessary."""
        if inputs is not None:
            return inputs

        encoder_outputs = model_kwargs.get("encoder_outputs")
        if self.config.is_encoder_decoder and encoder_outputs is not None:
            # make dummy input_ids with value -100, as a sanity check ensuring that they won't be used for encoding
            shape = encoder_outputs.last_hidden_state.size()[:-1]
            return torch.ones(shape, dtype=torch.long, device=self.device) * -100

        if bos_token_id is None:
            raise ValueError("`bos_token_id` has to be defined when no `input_ids` are provided.")

        # If there is some tensor in `model_kwargs`, we can infer the batch size from it. This is helpful with
        # soft-prompting or in multimodal implementations built on top of decoder-only language models.
        batch_size = 1
        for value in model_kwargs.values():
            if isinstance(value, torch.Tensor):
                batch_size = value.shape[0]
                break
        return torch.ones((batch_size, 1), dtype=torch.long, device=self.device) * bos_token_id

    def _prepare_attention_mask_for_generation(
        self,
        inputs: torch.Tensor,
        pad_token_id: Optional[int],
        eos_token_id: Optional[Union[int, List[int]]],
    ) -> torch.LongTensor:
        is_input_ids = len(inputs.shape) == 2 and inputs.dtype in [torch.int, torch.long]
        is_pad_token_in_inputs = (pad_token_id is not None) and (pad_token_id in inputs)
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        is_pad_token_not_equal_to_eos_token_id = (eos_token_id is None) or (pad_token_id not in eos_token_id)

        # Check if input is input_ids and padded -> only then is attention_mask defined
        if is_input_ids and is_pad_token_in_inputs and is_pad_token_not_equal_to_eos_token_id:
            return inputs.ne(pad_token_id).long()
        else:
            return torch.ones(inputs.shape[:2], dtype=torch.long, device=inputs.device)

    def _prepare_encoder_decoder_kwargs_for_generation(
        self, inputs_tensor: torch.Tensor, model_kwargs, model_input_name: Optional[str] = None
    ) -> Dict[str, Any]:
        # 1. get encoder
        encoder = self.get_encoder()
        # Compatibility with Accelerate big model inference: we need the encoder to outputs stuff on the same device
        # as the inputs.
        if hasattr(encoder, "_hf_hook"):
            encoder._hf_hook.io_same_device = True

        # 2. Prepare encoder args and encoder kwargs from model kwargs.
        irrelevant_prefix = ["decoder_", "cross_attn", "use_cache"]
        encoder_kwargs = {
            argument: value
            for argument, value in model_kwargs.items()
            if not any(argument.startswith(p) for p in irrelevant_prefix)
        }
        encoder_signature = set(inspect.signature(encoder.forward).parameters)
        encoder_accepts_wildcard = "kwargs" in encoder_signature or "model_kwargs" in encoder_signature
        if not encoder_accepts_wildcard:
            encoder_kwargs = {
                argument: value for argument, value in encoder_kwargs.items() if argument in encoder_signature
            }

        # 3. make sure that encoder returns `ModelOutput`
        model_input_name = model_input_name if model_input_name is not None else self.main_input_name
        encoder_kwargs["return_dict"] = True
        encoder_kwargs[model_input_name] = inputs_tensor
        model_kwargs["encoder_outputs"]: ModelOutput = encoder(**encoder_kwargs)

        return model_kwargs

    def _prepare_decoder_input_ids_for_generation(
        self,
        batch_size: int,
        model_input_name: str,
        model_kwargs: Dict[str, torch.Tensor],
        decoder_start_token_id: int = None,
        bos_token_id: int = None,
        device: torch.device = None,
    ) -> Tuple[torch.LongTensor, Dict[str, torch.Tensor]]:
        """Prepares `decoder_input_ids` for generation with encoder-decoder models"""
        # 1. Check whether the user has defined `decoder_input_ids` manually. To facilitate in terms of input naming,
        # we also allow the user to pass it under `input_ids`, if the encoder does not use it as the main input.
        if model_kwargs is not None and "decoder_input_ids" in model_kwargs:
            decoder_input_ids = model_kwargs.pop("decoder_input_ids")
        elif "input_ids" in model_kwargs and model_input_name != "input_ids":
            decoder_input_ids = model_kwargs.pop("input_ids")
        else:
            decoder_input_ids = None

        # 2. Encoder-decoder models expect the `decoder_input_ids` to start with a special token. Let's ensure that.
        decoder_start_token_id = self._get_decoder_start_token_id(decoder_start_token_id, bos_token_id)
        if device is None:
            device = self.device
        decoder_input_ids_start = torch.ones((batch_size, 1), dtype=torch.long, device=device) * decoder_start_token_id

        # no user input -> use decoder_start_token_id as decoder_input_ids
        if decoder_input_ids is None:
            decoder_input_ids = decoder_input_ids_start
        # exception: Donut checkpoints have task-specific decoder starts and don't expect a BOS token
        elif self.config.model_type == "vision-encoder-decoder" and "donut" in self.name_or_path.lower():
            pass
        # user input but doesn't start with decoder_start_token_id -> prepend decoder_start_token_id (and adjust
        # decoder_attention_mask if provided)
        elif (decoder_input_ids[:, 0] != decoder_start_token_id).all().item():
            decoder_input_ids = torch.cat([decoder_input_ids_start, decoder_input_ids], dim=-1)
            if "decoder_attention_mask" in model_kwargs:
                decoder_attention_mask = model_kwargs["decoder_attention_mask"]
                decoder_attention_mask = torch.cat(
                    (torch.ones_like(decoder_attention_mask)[:, :1], decoder_attention_mask),
                    dim=-1,
                )
                model_kwargs["decoder_attention_mask"] = decoder_attention_mask

        return decoder_input_ids, model_kwargs

    def _get_decoder_start_token_id(self, decoder_start_token_id: int = None, bos_token_id: int = None) -> int:
        decoder_start_token_id = (
            decoder_start_token_id
            if decoder_start_token_id is not None
            else self.generation_config.decoder_start_token_id
        )
        bos_token_id = bos_token_id if bos_token_id is not None else self.generation_config.bos_token_id

        if decoder_start_token_id is not None:
            return decoder_start_token_id
        elif bos_token_id is not None:
            return bos_token_id
        raise ValueError(
            "`decoder_start_token_id` or `bos_token_id` has to be defined for encoder-decoder generation."
        )

    @staticmethod
    def _expand_inputs_for_generation(
        expand_size: int = 1,
        is_encoder_decoder: bool = False,
        input_ids: Optional[torch.LongTensor] = None,
        **model_kwargs,
    ) -> Tuple[torch.LongTensor, Dict[str, Any]]:
        """Expands tensors from [batch_size, ...] to [batch_size * expand_size, ...]"""

        def _expand_dict_for_generation(dict_to_expand):
            for key in dict_to_expand:
                if dict_to_expand[key] is not None and isinstance(dict_to_expand[key], torch.Tensor):
                    dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=0)
            return dict_to_expand

        if input_ids is not None:
            input_ids = input_ids.repeat_interleave(expand_size, dim=0)

        model_kwargs = _expand_dict_for_generation(model_kwargs)

        if is_encoder_decoder:
            if model_kwargs.get("encoder_outputs") is None:
                raise ValueError("If `is_encoder_decoder` is True, make sure that `encoder_outputs` is defined.")
            model_kwargs["encoder_outputs"] = _expand_dict_for_generation(model_kwargs["encoder_outputs"])

        return input_ids, model_kwargs

    def _extract_past_from_model_output(self, outputs: ModelOutput, standardize_cache_format: bool = False):
        past_key_values = None
        if "past_key_values" in outputs:
            past_key_values = outputs.past_key_values
        elif "mems" in outputs:
            past_key_values = outputs.mems
        elif "past_buckets_states" in outputs:
            past_key_values = outputs.past_buckets_states

        # Bloom fix: standardizes the cache format when requested
        if standardize_cache_format and hasattr(self, "_convert_to_standard_cache"):
            batch_size = outputs.logits.shape[0]
            past_key_values = self._convert_to_standard_cache(past_key_values, batch_size=batch_size)
        return past_key_values

    def _update_model_kwargs_for_generation(
        self,
        outputs: ModelOutput,
        model_kwargs: Dict[str, Any],
        is_encoder_decoder: bool = False,
        standardize_cache_format: bool = False,
    ) -> Dict[str, Any]:
        # update past_key_values
        model_kwargs["past_key_values"] = self._extract_past_from_model_output(
            outputs, standardize_cache_format=standardize_cache_format
        )
        if getattr(outputs, "state", None) is not None:
            model_kwargs["state"] = outputs.state

        # update token_type_ids with last value
        if "token_type_ids" in model_kwargs:
            token_type_ids = model_kwargs["token_type_ids"]
            model_kwargs["token_type_ids"] = torch.cat([token_type_ids, token_type_ids[:, -1].unsqueeze(-1)], dim=-1)

        if not is_encoder_decoder:
            # update attention mask
            if "attention_mask" in model_kwargs:
                attention_mask = model_kwargs["attention_mask"]
                model_kwargs["attention_mask"] = torch.cat(
                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
                )
        else:
            # update decoder attention mask
            if "decoder_attention_mask" in model_kwargs:
                decoder_attention_mask = model_kwargs["decoder_attention_mask"]
                model_kwargs["decoder_attention_mask"] = torch.cat(
                    [decoder_attention_mask, decoder_attention_mask.new_ones((decoder_attention_mask.shape[0], 1))],
                    dim=-1,
                )

        return model_kwargs

    def _reorder_cache(self, past_key_values, beam_idx):
        raise NotImplementedError(
            f"Make sure that a `_reorder_cache` function is correctly implemented in {self.__class__.__module__} to"
            f" enable beam search for {self.__class__}"
        )

    def _get_logits_warper(
        self,
        generation_config: GenerationConfig,
    ) -> LogitsProcessorList:
        """
        This class returns a [`LogitsProcessorList`] list object that contains all relevant [`LogitsWarper`] instances
        used for multinomial sampling.
        """

        # instantiate warpers list
        warpers = LogitsProcessorList()

        # the following idea is largely copied from this PR: https://github.com/huggingface/transformers/pull/5420/files
        # all samplers can be found in `generation_utils_samplers.py`
        if generation_config.temperature is not None and generation_config.temperature != 1.0:
            warpers.append(TemperatureLogitsWarper(generation_config.temperature))
        min_tokens_to_keep = 2 if generation_config.num_beams > 1 else 1
        if generation_config.top_k is not None and generation_config.top_k != 0:
            warpers.append(TopKLogitsWarper(top_k=generation_config.top_k, min_tokens_to_keep=min_tokens_to_keep))
        if generation_config.top_p is not None and generation_config.top_p < 1.0:
            warpers.append(TopPLogitsWarper(top_p=generation_config.top_p, min_tokens_to_keep=min_tokens_to_keep))
        if generation_config.typical_p is not None and generation_config.typical_p < 1.0:
            warpers.append(
                TypicalLogitsWarper(mass=generation_config.typical_p, min_tokens_to_keep=min_tokens_to_keep)
            )
        if generation_config.epsilon_cutoff is not None and 0.0 < generation_config.epsilon_cutoff < 1.0:
            warpers.append(
                EpsilonLogitsWarper(epsilon=generation_config.epsilon_cutoff, min_tokens_to_keep=min_tokens_to_keep)
            )
        if generation_config.eta_cutoff is not None and 0.0 < generation_config.eta_cutoff < 1.0:
            warpers.append(
                EtaLogitsWarper(epsilon=generation_config.eta_cutoff, min_tokens_to_keep=min_tokens_to_keep)
            )
        # `LogitNormalization` should always be the last logit processor, when present
        if generation_config.renormalize_logits is True:
            warpers.append(LogitNormalization())
        return warpers

    def _get_logits_processor(
        self,
        generation_config: GenerationConfig,
        input_ids_seq_length: int,
        encoder_input_ids: torch.LongTensor,
        prefix_allowed_tokens_fn: Callable[[int, torch.Tensor], List[int]],
        logits_processor: Optional[LogitsProcessorList],
    ) -> LogitsProcessorList:
        """
        This class returns a [`LogitsProcessorList`] list object that contains all relevant [`LogitsProcessor`]
        instances used to modify the scores of the language model head.
        """
        # instantiate processors list
        processors = LogitsProcessorList()

        if generation_config.sequence_bias is not None:
            processors.append(SequenceBiasLogitsProcessor(sequence_bias=generation_config.sequence_bias))

        if generation_config.diversity_penalty is not None and generation_config.diversity_penalty > 0.0:
            processors.append(
                HammingDiversityLogitsProcessor(
                    diversity_penalty=generation_config.diversity_penalty,
                    num_beams=generation_config.num_beams,
                    num_beam_groups=generation_config.num_beam_groups,
                )
            )
        if (
            generation_config.encoder_repetition_penalty is not None
            and generation_config.encoder_repetition_penalty != 1.0
        ):
            processors.append(
                EncoderRepetitionPenaltyLogitsProcessor(
                    penalty=generation_config.encoder_repetition_penalty, encoder_input_ids=encoder_input_ids
                )
            )
        if generation_config.repetition_penalty is not None and generation_config.repetition_penalty != 1.0:
            processors.append(RepetitionPenaltyLogitsProcessor(penalty=generation_config.repetition_penalty))
        if generation_config.no_repeat_ngram_size is not None and generation_config.no_repeat_ngram_size > 0:
            processors.append(NoRepeatNGramLogitsProcessor(generation_config.no_repeat_ngram_size))
        if (
            generation_config.encoder_no_repeat_ngram_size is not None
            and generation_config.encoder_no_repeat_ngram_size > 0
        ):
            if self.config.is_encoder_decoder:
                processors.append(
                    EncoderNoRepeatNGramLogitsProcessor(
                        generation_config.encoder_no_repeat_ngram_size, encoder_input_ids
                    )
                )
            else:
                raise ValueError(
                    "It's impossible to use `encoder_no_repeat_ngram_size` with decoder-only architecture"
                )
        if generation_config.bad_words_ids is not None:
            processors.append(
                NoBadWordsLogitsProcessor(generation_config.bad_words_ids, generation_config.eos_token_id)
            )
        if (
            generation_config.min_length is not None
            and generation_config.eos_token_id is not None
            and generation_config.min_length > 0
        ):
            processors.append(MinLengthLogitsProcessor(generation_config.min_length, generation_config.eos_token_id))
        if (
            generation_config.min_new_tokens is not None
            and generation_config.eos_token_id is not None
            and generation_config.min_new_tokens > 0
        ):
            processors.append(
                MinNewTokensLengthLogitsProcessor(
                    input_ids_seq_length, generation_config.min_new_tokens, generation_config.eos_token_id
                )
            )
        if prefix_allowed_tokens_fn is not None:
            processors.append(
                PrefixConstrainedLogitsProcessor(
                    prefix_allowed_tokens_fn, generation_config.num_beams // generation_config.num_beam_groups
                )
            )
        if generation_config.forced_bos_token_id is not None:
            processors.append(ForcedBOSTokenLogitsProcessor(generation_config.forced_bos_token_id))
        if generation_config.forced_eos_token_id is not None:
            processors.append(
                ForcedEOSTokenLogitsProcessor(generation_config.max_length, generation_config.forced_eos_token_id)
            )
        if generation_config.remove_invalid_values is True:
            processors.append(InfNanRemoveLogitsProcessor())
        if generation_config.exponential_decay_length_penalty is not None:
            processors.append(
                ExponentialDecayLengthPenalty(
                    generation_config.exponential_decay_length_penalty,
                    generation_config.eos_token_id,
                    input_ids_seq_length,
                )
            )
        if generation_config.suppress_tokens is not None:
            processors.append(SuppressTokensLogitsProcessor(generation_config.suppress_tokens))
        if generation_config.begin_suppress_tokens is not None:
            begin_index = input_ids_seq_length
            begin_index = (
                begin_index
                if (input_ids_seq_length > 1 or generation_config.forced_bos_token_id is None)
                else begin_index + 1
            )
            if generation_config.forced_decoder_ids is not None:
                # generation starts after the last token that is forced
                begin_index += generation_config.forced_decoder_ids[-1][0]
            processors.append(
                SuppressTokensAtBeginLogitsProcessor(generation_config.begin_suppress_tokens, begin_index)
            )
        if generation_config.forced_decoder_ids is not None:
            processors.append(ForceTokensLogitsProcessor(generation_config.forced_decoder_ids))
        if generation_config.guidance_scale is not None and generation_config.guidance_scale > 1:
            processors.append(ClassifierFreeGuidanceLogitsProcessor(generation_config.guidance_scale))
        processors = self._merge_criteria_processor_list(processors, logits_processor)
        # `LogitNormalization` should always be the last logit processor, when present
        if generation_config.renormalize_logits is True:
            processors.append(LogitNormalization())
        return processors

    def _get_stopping_criteria(
        self, generation_config: GenerationConfig, stopping_criteria: Optional[StoppingCriteriaList]
    ) -> StoppingCriteriaList:
        criteria = StoppingCriteriaList()
        if generation_config.max_length is not None:
            max_position_embeddings = getattr(self.config, "max_position_embeddings", None)
            criteria.append(
                MaxLengthCriteria(
                    max_length=generation_config.max_length,
                    max_position_embeddings=max_position_embeddings,
                )
            )
        if generation_config.max_time is not None:
            criteria.append(MaxTimeCriteria(max_time=generation_config.max_time))
        criteria = self._merge_criteria_processor_list(criteria, stopping_criteria)
        return criteria

    def _merge_criteria_processor_list(
        self,
        default_list: Union[LogitsProcessorList, StoppingCriteriaList],
        custom_list: Union[LogitsProcessorList, StoppingCriteriaList],
    ) -> Union[LogitsProcessorList, StoppingCriteriaList]:
        if len(custom_list) == 0:
            return default_list
        for default in default_list:
            for custom in custom_list:
                if type(custom) is type(default):
                    object_type = "stopping criteria" if isinstance(custom, StoppingCriteria) else "logits processor"
                    raise ValueError(
                        f"A custom {object_type} of type {type(custom)} with values {custom} has been passed to"
                        f" `.generate()`, but it has already been created with the values {default}. {default} has been"
                        " created by passing the corresponding arguments to generate or by the model's config default"
                        f" values. If you just want to change the default values of {object_type} consider passing"
                        f" them as arguments to `.generate()` instead of using a custom {object_type}."
                    )
        default_list.extend(custom_list)
        return default_list

    def compute_transition_scores(
        self,
        sequences: torch.Tensor,
        scores: Tuple[torch.Tensor],
        beam_indices: Optional[torch.Tensor] = None,
        normalize_logits: bool = False,
    ) -> torch.Tensor:
        """
        Computes the transition scores of sequences given the generation scores (and beam indices, if beam search was
        used). This is a convenient method to quicky obtain the scores of the selected tokens at generation time.

        Parameters:
            sequences (`torch.LongTensor`):
                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)`):
                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 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)`.
            beam_indices (`torch.LongTensor`, *optional*):
                Beam indices of generated token id at each generation step. `torch.LongTensor` of shape
                `(batch_size*num_return_sequences, sequence_length)`. Only required if a `num_beams>1` at
                generate-time.
            normalize_logits (`bool`, *optional*, defaults to `False`):
                Whether to normalize the logits (which, for legacy reasons, may be unnormalized).

        Return:
            `torch.Tensor`: A `torch.Tensor` of shape `(batch_size*num_return_sequences, sequence_length)` containing
                the transition scores (logits)

        Examples:

        ```python
        >>> from transformers import GPT2Tokenizer, AutoModelForCausalLM
        >>> import numpy as np

        >>> tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
        >>> model = AutoModelForCausalLM.from_pretrained("gpt2")
        >>> tokenizer.pad_token_id = tokenizer.eos_token_id
        >>> inputs = tokenizer(["Today is"], return_tensors="pt")

        >>> # Example 1: Print the scores for each token generated with Greedy Search
        >>> outputs = model.generate(**inputs, max_new_tokens=5, return_dict_in_generate=True, output_scores=True)
        >>> transition_scores = model.compute_transition_scores(
        ...     outputs.sequences, outputs.scores, normalize_logits=True
        ... )
        >>> # input_length is the length of the input prompt for decoder-only models, like the GPT family, and 1 for
        >>> # encoder-decoder models, like BART or T5.
        >>> input_length = 1 if model.config.is_encoder_decoder else inputs.input_ids.shape[1]
        >>> generated_tokens = outputs.sequences[:, input_length:]
        >>> for tok, score in zip(generated_tokens[0], transition_scores[0]):
        ...     # | token | token string | logits | probability
        ...     print(f"| {tok:5d} | {tokenizer.decode(tok):8s} | {score.numpy():.3f} | {np.exp(score.numpy()):.2%}")
        |   262 |  the     | -1.414 | 24.33%
        |  1110 |  day     | -2.609 | 7.36%
        |   618 |  when    | -2.010 | 13.40%
        |   356 |  we      | -1.859 | 15.58%
        |   460 |  can     | -2.508 | 8.14%

        >>> # Example 2: Reconstruct the sequence scores from Beam Search
        >>> outputs = model.generate(
        ...     **inputs,
        ...     max_new_tokens=5,
        ...     num_beams=4,
        ...     num_return_sequences=4,
        ...     return_dict_in_generate=True,
        ...     output_scores=True,
        ... )
        >>> transition_scores = model.compute_transition_scores(
        ...     outputs.sequences, outputs.scores, outputs.beam_indices, normalize_logits=False
        ... )
        >>> # If you sum the generated tokens' scores and apply the length penalty, you'll get the sequence scores.
        >>> # Tip: recomputing the scores is only guaranteed to match with `normalize_logits=False`. Depending on the
        >>> # use case, you might want to recompute it with `normalize_logits=True`.
        >>> output_length = input_length + np.sum(transition_scores.numpy() < 0, axis=1)
        >>> length_penalty = model.generation_config.length_penalty
        >>> reconstructed_scores = transition_scores.sum(axis=1) / (output_length**length_penalty)
        >>> print(np.allclose(outputs.sequences_scores, reconstructed_scores))
        True
        ```"""
        # 1. In absence of `beam_indices`, we can assume that we come from e.g. greedy search, which is equivalent
        # to a beam search approach were the first (and only) beam is always selected
        if beam_indices is None:
            beam_indices = torch.arange(scores[0].shape[0]).view(-1, 1).to(sequences.device)
            beam_indices = beam_indices.expand(-1, len(scores))

        # 2. reshape scores as [batch_size*vocab_size, # generation steps] with # generation steps being
        # seq_len - input_length
        scores = torch.stack(scores).reshape(len(scores), -1).transpose(0, 1)

        # 3. Optionally normalize the logits (across the vocab dimension)
        if normalize_logits:
            scores = scores.reshape(-1, self.config.vocab_size, scores.shape[-1])
            scores = torch.nn.functional.log_softmax(scores, dim=1)
            scores = scores.reshape(-1, scores.shape[-1])

        # 4. cut beam_indices to longest beam length
        beam_indices_mask = beam_indices < 0
        max_beam_length = (1 - beam_indices_mask.long()).sum(-1).max()
        beam_indices = beam_indices.clone()[:, :max_beam_length]
        beam_indices_mask = beam_indices_mask[:, :max_beam_length]

        # 5. Set indices of beams that finished early to 0; such indices will be masked correctly afterwards
        beam_indices[beam_indices_mask] = 0

        # 6. multiply beam_indices with vocab size to gather correctly from scores
        beam_sequence_indices = beam_indices * self.config.vocab_size

        # 7. Define which indices contributed to scores
        cut_idx = sequences.shape[-1] - max_beam_length
        indices = sequences[:, cut_idx:] + beam_sequence_indices

        # 8. Compute scores
        transition_scores = scores.gather(0, indices)

        # 9. Mask out transition_scores of beams that stopped early
        transition_scores[beam_indices_mask] = 0

        return transition_scores

    def _validate_model_class(self):
        """
        Confirms that the model class is compatible with generation. If not, raises an exception that points to the
        right class to use.
        """
        if not self.can_generate():
            generate_compatible_mappings = [
                MODEL_FOR_CAUSAL_LM_MAPPING,
                MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
                MODEL_FOR_VISION_2_SEQ_MAPPING,
                MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
                MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
            ]
            generate_compatible_classes = set()
            for model_mapping in generate_compatible_mappings:
                supported_models = model_mapping.get(type(self.config), default=None)
                if supported_models is not None:
                    generate_compatible_classes.add(supported_models.__name__)
            exception_message = (
                f"The current model class ({self.__class__.__name__}) is not compatible with `.generate()`, as "
                "it doesn't have a language model head."
            )
            if generate_compatible_classes:
                exception_message += f" Please use one of the following classes instead: {generate_compatible_classes}"
            raise TypeError(exception_message)

    def _validate_model_kwargs(self, model_kwargs: Dict[str, Any]):
        """Validates model kwargs for generation. Generate argument typos will also be caught here."""
        # Excludes arguments that are handled before calling any model function
        if self.config.is_encoder_decoder:
            for key in ["decoder_input_ids"]:
                model_kwargs.pop(key, None)

        unused_model_args = []
        model_args = set(inspect.signature(self.prepare_inputs_for_generation).parameters)
        # `kwargs`/`model_kwargs` is often used to handle optional forward pass inputs like `attention_mask`. If
        # `prepare_inputs_for_generation` doesn't accept them, then a stricter check can be made ;)
        if "kwargs" in model_args or "model_kwargs" in model_args:
            model_args |= set(inspect.signature(self.forward).parameters)

        # Encoder-Decoder models may also need Encoder arguments from `model_kwargs`
        if self.config.is_encoder_decoder:
            base_model = getattr(self, self.base_model_prefix, None)

            # allow encoder kwargs
            encoder = getattr(self, "encoder", None)
            # `MusicgenForConditionalGeneration` has `text_encoder` and `audio_encoder`.
            # Also, it has `base_model_prefix = "encoder_decoder"` but there is no `self.encoder_decoder`
            # TODO: A better way to handle this.
            if encoder is None and base_model is not None:
                encoder = getattr(base_model, "encoder", None)

            if encoder is not None:
                encoder_model_args = set(inspect.signature(encoder.forward).parameters)
                model_args |= encoder_model_args

            # allow decoder kwargs
            decoder = getattr(self, "decoder", None)
            if decoder is None and base_model is not None:
                decoder = getattr(base_model, "decoder", None)

            if decoder is not None:
                decoder_model_args = set(inspect.signature(decoder.forward).parameters)
                model_args |= {f"decoder_{x}" for x in decoder_model_args}

        for key, value in model_kwargs.items():
            if value is not None and key not in model_args:
                unused_model_args.append(key)

        if unused_model_args:
            raise ValueError(
                f"The following `model_kwargs` are not used by the model: {unused_model_args} (note: typos in the"
                " generate arguments will also show up in this list)"
            )

    @torch.no_grad()
    def generate(
        self,
        inputs: Optional[torch.Tensor] = None,
        generation_config: Optional[GenerationConfig] = None,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
        synced_gpus: Optional[bool] = None,
        assistant_model: Optional["PreTrainedModel"] = None,
        streamer: Optional["BaseStreamer"] = None,
        **kwargs,
    ) -> Union[GenerateOutput, torch.LongTensor]:
        r"""

        Generates sequences of token ids for models with a language modeling head.

        <Tip warning={true}>

        Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
        model's default generation configuration. You can override any `generation_config` by passing the corresponding
        parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`.

        For an overview of generation strategies and code examples, check out the [following
        guide](../generation_strategies).

        </Tip>

        Parameters:
            inputs (`torch.Tensor` of varying shape depending on the modality, *optional*):
                The sequence used as a prompt for the generation or as model inputs to the encoder. If `None` the
                method initializes it with `bos_token_id` and a batch size of 1. For decoder-only models `inputs`
                should of in the format of `input_ids`. For encoder-decoder models *inputs* can represent any of
                `input_ids`, `input_values`, `input_features`, or `pixel_values`.
            generation_config (`~generation.GenerationConfig`, *optional*):
                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
                passed to generate matching the attributes of `generation_config` will override them. If
                `generation_config` is not provided, the default will be used, which had the following loading
                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
                default values, whose documentation should be checked to parameterize generation.
            logits_processor (`LogitsProcessorList`, *optional*):
                Custom logits processors that complement the default logits processors built from arguments and
                generation config. If a logit processor is passed that is already created with the arguments or a
                generation config an error is thrown. This feature is intended for advanced users.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                Custom stopping criteria that complement the default stopping criteria built from arguments and a
                generation config. If a stopping criteria is passed that is already created with the arguments or a
                generation config an error is thrown. This feature is intended for advanced users.
            prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], List[int]]`, *optional*):
                If provided, this function constraints the beam search to allowed tokens only at each step. If not
                provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and
                `input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned
                on the batch ID `batch_id` and the previously generated tokens `inputs_ids`. This argument is useful
                for constrained generation conditioned on the prefix, as described in [Autoregressive Entity
                Retrieval](https://arxiv.org/abs/2010.00904).
            synced_gpus (`bool`, *optional*):
                Whether to continue running the while loop until max_length. Unless overridden this flag will be set to
                `True` under DeepSpeed ZeRO Stage 3 multiple GPUs environment to avoid hanging if one GPU finished
                generating before other GPUs. Otherwise it'll be set to `False`.
            assistant_model (`PreTrainedModel`, *optional*):
                An assistant model that can be used to accelerate generation. The assistant model must have the exact
                same tokenizer. The acceleration is achieved when forecasting candidate tokens with the assistent model
                is much faster than running generation with the model you're calling generate from. As such, the
                assistant model should be much smaller.
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
            kwargs (`Dict[str, Any]`, *optional*):
                Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
                forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
                specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.

        Return:
            [`~utils.ModelOutput`] or `torch.LongTensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True`
            or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor`.

                If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible
                [`~utils.ModelOutput`] types are:

                    - [`~generation.GreedySearchDecoderOnlyOutput`],
                    - [`~generation.SampleDecoderOnlyOutput`],
                    - [`~generation.BeamSearchDecoderOnlyOutput`],
                    - [`~generation.BeamSampleDecoderOnlyOutput`]

                If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible
                [`~utils.ModelOutput`] types are:

                    - [`~generation.GreedySearchEncoderDecoderOutput`],
                    - [`~generation.SampleEncoderDecoderOutput`],
                    - [`~generation.BeamSearchEncoderDecoderOutput`],
                    - [`~generation.BeamSampleEncoderDecoderOutput`]
        """

        if synced_gpus is None:
            if is_deepspeed_zero3_enabled() and dist.get_world_size() > 1:
                synced_gpus = True
            else:
                synced_gpus = False

        # 1. Handle `generation_config` and kwargs that might update it, and validate the `.generate()` call
        self._validate_model_class()

        # priority: `generation_config` argument > `model.generation_config` (the default generation config)
        if generation_config is None:
            # legacy: users may modify the model configuration to control generation -- update the generation config
            # model attribute accordingly, if it was created from the model config
            if self.generation_config._from_model_config:
                new_generation_config = GenerationConfig.from_model_config(self.config)
                if new_generation_config != self.generation_config:
                    warnings.warn(
                        "You have modified the pretrained model configuration to control generation. This is a"
                        " deprecated strategy to control generation and will be removed soon, in a future version."
                        " Please use a generation configuration file (see"
                        " https://huggingface.co/docs/transformers/main_classes/text_generation )"
                    )
                    self.generation_config = new_generation_config
            generation_config = self.generation_config

        generation_config = copy.deepcopy(generation_config)
        model_kwargs = generation_config.update(**kwargs)  # All unused kwargs must be model kwargs
        generation_config.validate()
        self._validate_model_kwargs(model_kwargs.copy())

        # 2. Set generation parameters if not already defined
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()

        if generation_config.pad_token_id is None and generation_config.eos_token_id is not None:
            if model_kwargs.get("attention_mask", None) is None:
                logger.warning(
                    "The attention mask and the pad token id were not set. As a consequence, you may observe "
                    "unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results."
                )
            eos_token_id = generation_config.eos_token_id
            if isinstance(eos_token_id, list):
                eos_token_id = eos_token_id[0]
            logger.warning(f"Setting `pad_token_id` to `eos_token_id`:{eos_token_id} for open-end generation.")
            generation_config.pad_token_id = eos_token_id

        # 3. Define model inputs
        # inputs_tensor has to be defined
        # model_input_name is defined if model-specific keyword input is passed
        # otherwise model_input_name is None
        # all model-specific keyword inputs are removed from `model_kwargs`
        inputs_tensor, model_input_name, model_kwargs = self._prepare_model_inputs(
            inputs, generation_config.bos_token_id, model_kwargs
        )
        batch_size = inputs_tensor.shape[0]

        # 4. Define other model kwargs
        model_kwargs["output_attentions"] = generation_config.output_attentions
        model_kwargs["output_hidden_states"] = generation_config.output_hidden_states
        # decoder-only models with inputs_embeds forwarding must use caching (otherwise we can't detect whether we are
        # generating the first new token or not, and we only want to use the embeddings for the first new token)
        if not self.config.is_encoder_decoder and model_input_name == "inputs_embeds":
            model_kwargs["use_cache"] = True
        else:
            model_kwargs["use_cache"] = generation_config.use_cache

        accepts_attention_mask = "attention_mask" in set(inspect.signature(self.forward).parameters.keys())
        requires_attention_mask = "encoder_outputs" not in model_kwargs

        if model_kwargs.get("attention_mask", None) is None and requires_attention_mask and accepts_attention_mask:
            model_kwargs["attention_mask"] = self._prepare_attention_mask_for_generation(
                inputs_tensor, generation_config.pad_token_id, generation_config.eos_token_id
            )

        # decoder-only models should use left-padding for generation
        if not self.config.is_encoder_decoder:
            # If `input_ids` was given, check if the last id in any sequence is `pad_token_id`
            # Note: If using, `inputs_embeds` this check does not work, because we want to be more hands-off.
            if (
                generation_config.pad_token_id is not None
                and len(inputs_tensor.shape) == 2
                and torch.sum(inputs_tensor[:, -1] == generation_config.pad_token_id) > 0
            ):
                logger.warning(
                    "A decoder-only architecture is being used, but right-padding was detected! For correct "
                    "generation results, please set `padding_side='left'` when initializing the tokenizer."
                )

        if self.config.is_encoder_decoder and "encoder_outputs" not in model_kwargs:
            # if model is encoder decoder encoder_outputs are created
            # and added to `model_kwargs`
            model_kwargs = self._prepare_encoder_decoder_kwargs_for_generation(
                inputs_tensor, model_kwargs, model_input_name
            )

        # 5. Prepare `input_ids` which will be used for auto-regressive generation
        if self.config.is_encoder_decoder:
            input_ids, model_kwargs = self._prepare_decoder_input_ids_for_generation(
                batch_size=batch_size,
                model_input_name=model_input_name,
                model_kwargs=model_kwargs,
                decoder_start_token_id=generation_config.decoder_start_token_id,
                bos_token_id=generation_config.bos_token_id,
                device=inputs_tensor.device,
            )
        else:
            input_ids = inputs_tensor if model_input_name == "input_ids" else model_kwargs.pop("input_ids")

        if streamer is not None:
            streamer.put(input_ids.cpu())

        # 6. Prepare `max_length` depending on other stopping criteria.
        input_ids_seq_length = input_ids.shape[-1]
        has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
        if has_default_max_length and generation_config.max_new_tokens is None and generation_config.max_length != 20:
            # 20 is the default max_length of the generation config
            warnings.warn(
                f"Using the model-agnostic default `max_length` (={generation_config.max_length}) "
                "to control the generation length.  recommend setting `max_new_tokens` to control the maximum length of the generation.",
                UserWarning,
            )
        elif generation_config.max_new_tokens is not None:
            if not has_default_max_length:
                logger.warning(
                    f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
                    f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
                    "Please refer to the documentation for more information. "
                    "(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)"
                )
            generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length

        if generation_config.min_length is not None and generation_config.min_length > generation_config.max_length:
            raise ValueError(
                f"Unfeasible length constraints: the minimum length ({generation_config.min_length}) is larger than"
                f" the maximum length ({generation_config.max_length})"
            )
        if input_ids_seq_length >= generation_config.max_length:
            input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
            logger.warning(
                f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
                f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
                " increasing `max_new_tokens`."
            )

        # 7. determine generation mode
        is_constraint_gen_mode = (
            generation_config.constraints is not None or generation_config.force_words_ids is not None
        )

        is_contrastive_search_gen_mode = (
            (generation_config.num_beams == 1)
            and generation_config.top_k is not None
            and generation_config.top_k > 1
            and generation_config.do_sample is False
            and generation_config.penalty_alpha is not None
            and generation_config.penalty_alpha > 0
        )

        is_greedy_gen_mode = (
            (generation_config.num_beams == 1)
            and (generation_config.num_beam_groups == 1)
            and generation_config.do_sample is False
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
        is_sample_gen_mode = (
            (generation_config.num_beams == 1)
            and (generation_config.num_beam_groups == 1)
            and generation_config.do_sample is True
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
        is_beam_gen_mode = (
            (generation_config.num_beams > 1)
            and (generation_config.num_beam_groups == 1)
            and generation_config.do_sample is False
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
        is_beam_sample_gen_mode = (
            (generation_config.num_beams > 1)
            and (generation_config.num_beam_groups == 1)
            and generation_config.do_sample is True
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
        is_group_beam_gen_mode = (
            (generation_config.num_beams > 1)
            and (generation_config.num_beam_groups > 1)
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
        is_assisted_gen_mode = False
        if assistant_model is not None:
            if not (is_greedy_gen_mode or is_sample_gen_mode):
                raise ValueError(
                    "You've set `assistant_model`, which triggers assisted generate. Currently, assisted generate "
                    "is only supported with Greedy Search and Sample."
                )
            is_assisted_gen_mode = True

        if generation_config.num_beam_groups > generation_config.num_beams:
            raise ValueError("`num_beam_groups` has to be smaller or equal to `num_beams`")
        if is_group_beam_gen_mode and generation_config.do_sample is True:
            raise ValueError(
                "Diverse beam search cannot be used in sampling mode. Make sure that `do_sample` is set to `False`."
            )

        if streamer is not None and (generation_config.num_beams > 1):
            raise ValueError(
                "`streamer` cannot be used with beam search (yet!). Make sure that `num_beams` is set to 1."
            )

        if self.device.type != input_ids.device.type:
            warnings.warn(
                "You are calling .generate() with the `input_ids` being on a device type different"
                f" than your model's device. `input_ids` is on {input_ids.device.type}, whereas the model"
                f" is on {self.device.type}. You may experience unexpected behaviors or slower generation."
                " Please make sure that you have put `input_ids` to the"
                f" correct device by calling for example input_ids = input_ids.to('{self.device.type}') before"
                " running `.generate()`.",
                UserWarning,
            )

        # 8. prepare distribution pre_processing samplers
        logits_processor = self._get_logits_processor(
            generation_config=generation_config,
            input_ids_seq_length=input_ids_seq_length,
            encoder_input_ids=inputs_tensor,
            prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
            logits_processor=logits_processor,
        )

        # 9. prepare stopping criteria
        stopping_criteria = self._get_stopping_criteria(
            generation_config=generation_config, stopping_criteria=stopping_criteria
        )
        # 10. go into different generation modes
        if is_assisted_gen_mode:
            if generation_config.num_return_sequences > 1:
                raise ValueError(
                    "num_return_sequences has to be 1 when doing assisted generate, "
                    f"but is {generation_config.num_return_sequences}."
                )
            if batch_size > 1:
                raise ValueError("assisted generate is only supported for batch_size = 1")
            if not model_kwargs["use_cache"]:
                raise ValueError("assisted generate requires `use_cache=True`")

            # 11. If the assistant model is an encoder-decoder, prepare its encoder outputs
            if assistant_model.config.is_encoder_decoder:
                assistant_model_kwargs = copy.deepcopy(model_kwargs)
                inputs_tensor, model_input_name, assistant_model_kwargs = assistant_model._prepare_model_inputs(
                    inputs_tensor, assistant_model.generation_config.bos_token_id, assistant_model_kwargs
                )
                assistant_model_kwargs = assistant_model._prepare_encoder_decoder_kwargs_for_generation(
                    inputs_tensor, assistant_model_kwargs, model_input_name
                )
                model_kwargs["assistant_encoder_outputs"] = assistant_model_kwargs["encoder_outputs"]

            # 12. run assisted generate
            return self.assisted_decoding(
                input_ids,
                assistant_model=assistant_model,
                do_sample=generation_config.do_sample,
                logits_processor=logits_processor,
                logits_warper=self._get_logits_warper(generation_config) if generation_config.do_sample else None,
                stopping_criteria=stopping_criteria,
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
                synced_gpus=synced_gpus,
                streamer=streamer,
                **model_kwargs,
            )
        if is_greedy_gen_mode:
            if generation_config.num_return_sequences > 1:
                raise ValueError(
                    "num_return_sequences has to be 1 when doing greedy search, "
                    f"but is {generation_config.num_return_sequences}."
                )

            # 11. run greedy search
            return self.greedy_search(
                input_ids,
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
                synced_gpus=synced_gpus,
                streamer=streamer,
                **model_kwargs,
            )

        elif is_contrastive_search_gen_mode:
            if generation_config.num_return_sequences > 1:
                raise ValueError(
                    "num_return_sequences has to be 1 when doing contrastive search, "
                    f"but is {generation_config.num_return_sequences}."
                )
            if not model_kwargs["use_cache"]:
                raise ValueError("Contrastive search requires `use_cache=True`")

            return self.contrastive_search(
                input_ids,
                top_k=generation_config.top_k,
                penalty_alpha=generation_config.penalty_alpha,
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
                synced_gpus=synced_gpus,
                streamer=streamer,
                sequential=generation_config.low_memory,
                **model_kwargs,
            )

        elif is_sample_gen_mode:
            # 11. prepare logits warper
            logits_warper = self._get_logits_warper(generation_config)

            # 12. expand input_ids with `num_return_sequences` additional sequences per batch
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
                expand_size=generation_config.num_return_sequences,
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )

            # 13. run sample
            return self.sample(
                input_ids,
                logits_processor=logits_processor,
                logits_warper=logits_warper,
                stopping_criteria=stopping_criteria,
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
                synced_gpus=synced_gpus,
                streamer=streamer,
                **model_kwargs,
            )

        elif is_beam_gen_mode:
            if generation_config.num_return_sequences > generation_config.num_beams:
                raise ValueError("`num_return_sequences` has to be smaller or equal to `num_beams`.")

            if stopping_criteria.max_length is None:
                raise ValueError("`max_length` needs to be a stopping_criteria for now.")

            # 11. prepare beam search scorer
            beam_scorer = BeamSearchScorer(
                batch_size=batch_size,
                num_beams=generation_config.num_beams,
                device=inputs_tensor.device,
                length_penalty=generation_config.length_penalty,
                do_early_stopping=generation_config.early_stopping,
                num_beam_hyps_to_keep=generation_config.num_return_sequences,
                max_length=generation_config.max_length,
            )
            # 12. interleave input_ids with `num_beams` additional sequences per batch
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
                expand_size=generation_config.num_beams,
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )
            # 13. run beam search
            return self.beam_search(
                input_ids,
                beam_scorer,
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
                synced_gpus=synced_gpus,
                **model_kwargs,
            )

        elif is_beam_sample_gen_mode:
            # 11. prepare logits warper
            logits_warper = self._get_logits_warper(generation_config)

            if stopping_criteria.max_length is None:
                raise ValueError("`max_length` needs to be a stopping_criteria for now.")
            # 12. prepare beam search scorer
            beam_scorer = BeamSearchScorer(
                batch_size=batch_size * generation_config.num_return_sequences,
                num_beams=generation_config.num_beams,
                device=inputs_tensor.device,
                length_penalty=generation_config.length_penalty,
                do_early_stopping=generation_config.early_stopping,
                max_length=generation_config.max_length,
            )

            # 13. interleave input_ids with `num_beams` additional sequences per batch
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
                expand_size=generation_config.num_beams * generation_config.num_return_sequences,
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )

            # 14. run beam sample
            return self.beam_sample(
                input_ids,
                beam_scorer,
                logits_processor=logits_processor,
                logits_warper=logits_warper,
                stopping_criteria=stopping_criteria,
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
                synced_gpus=synced_gpus,
                **model_kwargs,
            )

        elif is_group_beam_gen_mode:
            if generation_config.num_return_sequences > generation_config.num_beams:
                raise ValueError("`num_return_sequences` has to be smaller or equal to `num_beams`.")

            if generation_config.num_beams % generation_config.num_beam_groups != 0:
                raise ValueError("`num_beams` should be divisible by `num_beam_groups` for group beam search.")

            if generation_config.diversity_penalty == 0.0:
                raise ValueError(
                    "`diversity_penalty` should be greater than `0.0`, otherwise your beam groups will be identical."
                )

            if stopping_criteria.max_length is None:
                raise ValueError("`max_length` needs to be a stopping_criteria for now.")

            has_default_typical_p = kwargs.get("typical_p") is None and generation_config.typical_p == 1.0
            if not has_default_typical_p:
                raise ValueError("Decoder argument `typical_p` is not supported with beam groups.")

            # 11. prepare beam search scorer
            beam_scorer = BeamSearchScorer(
                batch_size=batch_size,
                num_beams=generation_config.num_beams,
                device=inputs_tensor.device,
                length_penalty=generation_config.length_penalty,
                do_early_stopping=generation_config.early_stopping,
                num_beam_hyps_to_keep=generation_config.num_return_sequences,
                num_beam_groups=generation_config.num_beam_groups,
                max_length=generation_config.max_length,
            )
            # 12. interleave input_ids with `num_beams` additional sequences per batch
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
                expand_size=generation_config.num_beams,
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )
            # 13. run beam search
            return self.group_beam_search(
                input_ids,
                beam_scorer,
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
                synced_gpus=synced_gpus,
                **model_kwargs,
            )

        elif is_constraint_gen_mode:
            if generation_config.num_return_sequences > generation_config.num_beams:
                raise ValueError("`num_return_sequences` has to be smaller or equal to `num_beams`.")

            if stopping_criteria.max_length is None:
                raise ValueError("`max_length` needs to be a stopping_criteria for now.")

            if generation_config.num_beams <= 1:
                raise ValueError("`num_beams` needs to be greater than 1 for constrained generation.")

            if generation_config.do_sample:
                raise ValueError("`do_sample` needs to be false for constrained generation.")

            if generation_config.num_beam_groups is not None and generation_config.num_beam_groups > 1:
                raise ValueError("`num_beam_groups` not supported yet for constrained generation.")

            final_constraints = []
            if generation_config.constraints is not None:
                final_constraints = generation_config.constraints

            if generation_config.force_words_ids is not None:

                def typeerror():
                    raise ValueError(
                        "`force_words_ids` has to either be a `List[List[List[int]]]` or `List[List[int]]`"
                        f"of positive integers, but is {generation_config.force_words_ids}."
                    )

                if (
                    not isinstance(generation_config.force_words_ids, list)
                    or len(generation_config.force_words_ids) == 0
                ):
                    typeerror()

                for word_ids in generation_config.force_words_ids:
                    if isinstance(word_ids[0], list):
                        if not isinstance(word_ids, list) or len(word_ids) == 0:
                            typeerror()
                        if any(not isinstance(token_ids, list) for token_ids in word_ids):
                            typeerror()
                        if any(
                            any((not isinstance(token_id, int) or token_id < 0) for token_id in token_ids)
                            for token_ids in word_ids
                        ):
                            typeerror()

                        constraint = DisjunctiveConstraint(word_ids)
                    else:
                        if not isinstance(word_ids, list) or len(word_ids) == 0:
                            typeerror()
                        if any((not isinstance(token_id, int) or token_id < 0) for token_id in word_ids):
                            typeerror()

                        constraint = PhrasalConstraint(word_ids)
                    final_constraints.append(constraint)

            # 11. prepare beam search scorer
            constrained_beam_scorer = ConstrainedBeamSearchScorer(
                constraints=final_constraints,
                batch_size=batch_size,
                num_beams=generation_config.num_beams,
                device=inputs_tensor.device,
                length_penalty=generation_config.length_penalty,
                do_early_stopping=generation_config.early_stopping,
                num_beam_hyps_to_keep=generation_config.num_return_sequences,
                max_length=generation_config.max_length,
            )
            # 12. interleave input_ids with `num_beams` additional sequences per batch
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
                expand_size=generation_config.num_beams,
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )
            # 13. run beam search
            return self.constrained_beam_search(
                input_ids,
                constrained_beam_scorer=constrained_beam_scorer,
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
                synced_gpus=synced_gpus,
                **model_kwargs,
            )

    @torch.no_grad()
    def contrastive_search(
        self,
        input_ids: torch.LongTensor,
        top_k: Optional[int] = 1,
        penalty_alpha: Optional[float] = 0,
        logits_processor: Optional[LogitsProcessorList] = None,
        logits_warper: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[Union[int, List[int]]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: bool = False,
        streamer: Optional["BaseStreamer"] = None,
        sequential: Optional[bool] = None,
        **model_kwargs,
    ) -> Union[ContrastiveSearchOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **contrastive search** and can
        be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.contrastive_search`] directly. Use
        generate() instead. For an overview of generation strategies and code examples, check the [following
        guide](../generation_strategies).

        </Tip>

        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            top_k (`int`, *optional*, defaults to 1):
                The size of the candidate set that is used to re-rank for contrastive search
            penalty_alpha (`float`, *optional*, defaults to 0):
                The degeneration penalty for contrastive search; activate when it is larger than 0
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            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.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
            sequential (`bool`, *optional*):
                Switches topk hidden state computation from parallel to sequential to reduce memory if True.
            model_kwargs:
                Additional model specific keyword arguments will be forwarded to the `forward` function of the model.
                If model is an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.ContrastiveSearchDecoderOnlyOutput`], [`~generation.ContrastiveSearchEncoderDecoderOutput`]
            or `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.ContrastiveSearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.ContrastiveSearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:
        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForCausalLM,
        ...     StoppingCriteriaList,
        ...     MaxLengthCriteria,
        ... )

        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-125m")
        >>> model = AutoModelForCausalLM.from_pretrained("facebook/opt-125m")
        >>> # set pad_token_id to eos_token_id because OPT does not have a PAD token
        >>> model.config.pad_token_id = model.config.eos_token_id
        >>> input_prompt = "DeepMind Company is"
        >>> input_ids = tokenizer(input_prompt, return_tensors="pt")
        >>> stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=64)])
        >>> outputs = model.contrastive_search(
        ...     **input_ids, penalty_alpha=0.6, top_k=4, stopping_criteria=stopping_criteria
        ... )
        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['DeepMind Company is a company that focuses on the development and commercialization of artificial intelligence (AI). DeepMind’s mission is to help people understand and solve problems that are difficult to solve in the world today.\n\nIn this post, we talk about the benefits of deep learning in business and how it']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        logits_warper = logits_warper if logits_warper is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
        sequential = sequential if sequential is not None else self.generation_config.low_memory
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
        )
        return_dict_in_generate = (
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
        )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # keep track of which sequences are already finished
        unfinished_sequences = torch.ones(input_ids.shape[0], dtype=torch.long, device=input_ids.device)

        this_peer_finished = False  # used by synced_gpus only
        batch_size = input_ids.shape[0]

        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # if the first step in the loop, encode all the prefix and obtain: (1) past_key_values;
            # (2) last_hidden_states; (3) logit_for_next_step; (4) update model kwargs for the next step
            if model_kwargs.get("past_key_values") is None:
                # prepare inputs
                model_kwargs["use_cache"] = True
                model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

                # encode the given prefix and prepare model inputs; encoder-decoder model process the prefix and save
                # the `encoder_outputs`
                outputs = self(
                    **model_inputs, return_dict=True, output_hidden_states=True, output_attentions=output_attentions
                )

                # last decoder hidden states will be used to compute the degeneration penalty (cosine similarity with
                # previous tokens)
                if self.config.is_encoder_decoder:
                    last_hidden_states = outputs.decoder_hidden_states[-1]
                else:
                    last_hidden_states = outputs.hidden_states[-1]

                # next logit for contrastive search to select top-k candidate tokens
                logit_for_next_step = outputs.logits[:, -1, :]

                model_kwargs = self._update_model_kwargs_for_generation(
                    outputs,
                    model_kwargs,
                    is_encoder_decoder=self.config.is_encoder_decoder,
                    standardize_cache_format=True,
                )
                if not sequential:
                    # Expands model inputs top_k times, for batched forward passes (akin to beam search).
                    _, model_kwargs = self._expand_inputs_for_generation(
                        expand_size=top_k, is_encoder_decoder=self.config.is_encoder_decoder, **model_kwargs
                    )

                past_key_values = model_kwargs.get("past_key_values")
                if past_key_values is None:
                    raise ValueError(
                        f"{self.__class__.__name__} does not support caching and therefore **can't** be used "
                        "for contrastive search."
                    )
                elif (
                    not isinstance(past_key_values[0], (tuple, torch.Tensor))
                    or past_key_values[0][0].shape[0] != batch_size
                ):
                    raise ValueError(
                        f"{self.__class__.__name__} does not have a standard cache format and therefore **can't** be "
                        "used for contrastive search without further modifications."
                    )

            # contrastive_search main logic start:
            # contrastive search decoding consists of two steps: (1) candidate tokens recall; (2) candidate re-rank by
            # degeneration penalty
            logit_for_next_step = logits_processor(input_ids, logit_for_next_step)
            logit_for_next_step = logits_warper(input_ids, logit_for_next_step)
            next_probs = nn.functional.softmax(logit_for_next_step, dim=-1)
            top_k_probs, top_k_ids = torch.topk(next_probs, dim=-1, k=top_k)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (logit_for_next_step,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # Replicates the new past_key_values to match the `top_k` candidates
            new_key_values = []
            for layer in model_kwargs["past_key_values"]:
                items = []
                # item is either the key or the value matrix
                for item in layer:
                    if sequential:
                        items.append(item.repeat_interleave(1, dim=0))
                    else:
                        items.append(item.repeat_interleave(top_k, dim=0))
                new_key_values.append(items)
            model_kwargs["past_key_values"] = new_key_values

            if sequential:
                all_outputs = {key: [] for key in outputs}  # defined in first loop iteration
                all_last_hstates, all_hstates, all_logits = [], [], []
                for i in range(top_k):
                    # compute the candidate tokens by the language model and collect their hidden_states
                    next_model_inputs = self.prepare_inputs_for_generation(top_k_ids[:, i].view(-1, 1), **model_kwargs)

                    outputs = self(
                        **next_model_inputs,
                        return_dict=True,
                        output_hidden_states=True,
                        output_attentions=output_attentions,
                    )
                    for key in all_outputs:
                        all_outputs[key].append(outputs[key])

                    if self.config.is_encoder_decoder:
                        next_hidden = outputs.decoder_hidden_states[-1]
                        full_hidden_states = outputs.decoder_hidden_states

                    else:
                        next_hidden = outputs.hidden_states[-1]
                        full_hidden_states = outputs.hidden_states

                    all_last_hstates.append(torch.squeeze(next_hidden, 0))
                    all_hstates.append(full_hidden_states)
                    all_logits.append(outputs.logits[:, -1, :])

                # stack hidden states
                next_hidden = torch.stack([all_last_hstates[i] for i in range(top_k)], dim=0)
                final_full_hstates = [0 for i in range(len(full_hidden_states))]
                for layer in range(len(full_hidden_states)):
                    final_full_hstates[layer] = torch.stack(
                        [torch.squeeze(all_hstates[i][layer], 0) for i in range(top_k)], dim=0
                    )
                full_hidden_states = tuple(final_full_hstates)

                # stack logits
                logits = torch.cat(all_logits, dim=0)

            else:
                # compute the candidate tokens by the language model and collect their hidden_states
                # assembles top_k_ids into batch of size k
                next_model_inputs = self.prepare_inputs_for_generation(top_k_ids.view(-1, 1), **model_kwargs)

                outputs = self(
                    **next_model_inputs,
                    return_dict=True,
                    output_hidden_states=True,
                    output_attentions=output_attentions,
                )
                # name is different for encoder-decoder and decoder-only models
                if self.config.is_encoder_decoder:
                    next_hidden = outputs.decoder_hidden_states[-1]
                    full_hidden_states = outputs.decoder_hidden_states
                else:
                    next_hidden = outputs.hidden_states[-1]
                    full_hidden_states = outputs.hidden_states

                logits = outputs.logits[:, -1, :]

            context_hidden = last_hidden_states.repeat_interleave(top_k, dim=0)

            # compute the degeneration penalty and re-rank the candidates based on the degeneration penalty and the
            # model confidence. Keeping `selected_idx` on CPU enables multi-device contrastive search and doesn't
            # introduce (noticeable) slowdowns on single-device runs.
            selected_idx = _ranking_fast(context_hidden, next_hidden, top_k_probs, penalty_alpha, top_k)
            selected_idx = selected_idx.to("cpu")

            # prepare for the next step: (1) next token_id; (2) past_key_values; (3) last_hidden_states for computing
            # the degeneration penalty; (4) logits for selecting next top-k candidates; (5) selected tokens scores
            # (model confidence minus degeneration penalty); (6) decoder hidden_states
            next_tokens = top_k_ids[range(len(top_k_ids)), selected_idx]
            next_hidden = torch.stack(torch.split(next_hidden.squeeze(dim=1), top_k))
            next_hidden = next_hidden[range(batch_size), selected_idx, :]
            last_hidden_states = torch.cat([last_hidden_states, next_hidden.unsqueeze(1)], dim=1)

            next_decoder_hidden_states = ()
            for layer in full_hidden_states:
                layer = torch.stack(torch.split(layer, top_k))[range(batch_size), selected_idx, :]
                next_decoder_hidden_states += (layer,)

            # generate past_key_values cache of only the selected token
            if sequential:
                next_model_input = self.prepare_inputs_for_generation(
                    top_k_ids[:, selected_idx].view(-1, 1), **model_kwargs
                )

                selected_outputs = self(
                    **next_model_input,
                    return_dict=True,
                    output_hidden_states=False,
                    output_attentions=False,
                )
                next_past_key_values = selected_outputs["past_key_values"]

            else:
                next_past_key_values = self._extract_past_from_model_output(outputs, standardize_cache_format=True)
                new_key_values = ()
                for layer in next_past_key_values:
                    items = ()
                    # item is either the key or the value matrix
                    for item in layer:
                        item = torch.stack(torch.split(item, top_k, dim=0))  # [B, K, num_head, seq_len, esz]
                        item = item[range(batch_size), selected_idx, ...]  # [B, num_head, seq_len, esz]
                        items += (item,)
                    new_key_values += (items,)
                next_past_key_values = new_key_values

            logit_for_next_step = torch.stack(torch.split(logits, top_k))[range(batch_size), selected_idx, :]

            # Rebuilds the relevant parts of the model output for the selected token, for use in the next iteration
            if self.config.is_encoder_decoder:
                next_step_cross_attentions = ()
                next_step_decoder_attentions = ()
                if output_attentions:
                    for layer in outputs.cross_attentions:
                        layer = torch.stack(torch.split(layer, top_k, dim=0))[range(batch_size), selected_idx, ...]
                        next_step_cross_attentions += (layer,)
                    for layer in outputs.decoder_attentions:
                        layer = torch.stack(torch.split(layer, top_k, dim=0))[range(batch_size), selected_idx, ...]
                        next_step_decoder_attentions += (layer,)
                outputs = Seq2SeqLMOutput(
                    past_key_values=next_past_key_values,
                    decoder_hidden_states=next_decoder_hidden_states,
                    decoder_attentions=next_step_decoder_attentions or None,
                    cross_attentions=next_step_cross_attentions or None,
                )
            else:
                next_step_attentions = ()
                if output_attentions:
                    for layer in outputs.attentions:
                        layer = torch.stack(torch.split(layer, top_k, dim=0))[range(batch_size), selected_idx, ...]
                        next_step_attentions += (layer,)
                outputs = CausalLMOutputWithPast(
                    past_key_values=next_past_key_values,
                    hidden_states=next_decoder_hidden_states,
                    attentions=next_step_attentions or None,
                )
            # contrastive_search main logic end

            if synced_gpus and this_peer_finished:
                continue  # don't waste resources running the code we don't need

            # finished sentences should have their next token be a padding token
            if eos_token_id is not None:
                if pad_token_id is None:
                    raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.")
                next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)

            # update generated ids, model inputs, and length for next step
            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
            if streamer is not None:
                streamer.put(next_tokens.cpu())
            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )

            # if eos_token was found in one sentence, set sentence to finished
            if eos_token_id_tensor is not None:
                unfinished_sequences = unfinished_sequences.mul(
                    next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
                )

                # stop when each sentence is finished
                if unfinished_sequences.max() == 0:
                    this_peer_finished = True

            # stop if we exceed the maximum length
            if stopping_criteria(input_ids, scores):
                this_peer_finished = True

            if this_peer_finished and not synced_gpus:
                break

        if streamer is not None:
            streamer.end()

        if return_dict_in_generate:
            if self.config.is_encoder_decoder:
                return ContrastiveSearchEncoderDecoderOutput(
                    sequences=input_ids,
                    scores=scores,
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return ContrastiveSearchDecoderOnlyOutput(
                    sequences=input_ids,
                    scores=scores,
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return input_ids

    def greedy_search(
        self,
        input_ids: torch.LongTensor,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[Union[int, List[int]]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: bool = False,
        streamer: Optional["BaseStreamer"] = None,
        **model_kwargs,
    ) -> Union[GreedySearchOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **greedy decoding** and can be
        used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.greedy_search`] directly. Use generate()
        instead. For an overview of generation strategies and code examples, check the [following
        guide](../generation_strategies).

        </Tip>


        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.

            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            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.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
            model_kwargs:
                Additional model specific keyword arguments will be forwarded to the `forward` function of the model.
                If model is an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.GreedySearchDecoderOnlyOutput`], [`~generation.GreedySearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.GreedySearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.GreedySearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForCausalLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     StoppingCriteriaList,
        ...     MaxLengthCriteria,
        ... )

        >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
        >>> model = AutoModelForCausalLM.from_pretrained("gpt2")

        >>> # set pad_token_id to eos_token_id because GPT2 does not have a PAD token
        >>> model.generation_config.pad_token_id = model.generation_config.eos_token_id

        >>> input_prompt = "It might be possible to"
        >>> input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         MinLengthLogitsProcessor(10, eos_token_id=model.generation_config.eos_token_id),
        ...     ]
        ... )
        >>> stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=20)])

        >>> outputs = model.greedy_search(
        ...     input_ids, logits_processor=logits_processor, stopping_criteria=stopping_criteria
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ["It might be possible to get a better understanding of the nature of the problem, but it's not"]
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList([MaxLengthCriteria(max_length=max_length)])` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
        )
        return_dict_in_generate = (
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
        )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # keep track of which sequences are already finished
        unfinished_sequences = torch.ones(input_ids.shape[0], dtype=torch.long, device=input_ids.device)

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # prepare model inputs
            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            # forward pass to get next token
            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]

            # pre-process distribution
            next_tokens_scores = logits_processor(input_ids, next_token_logits)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (next_tokens_scores,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # argmax
            next_tokens = torch.argmax(next_tokens_scores, dim=-1)

            # finished sentences should have their next token be a padding token
            if eos_token_id is not None:
                if pad_token_id is None:
                    raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.")
                next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)

            # update generated ids, model inputs, and length for next step
            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
            if streamer is not None:
                streamer.put(next_tokens.cpu())
            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )

            # if eos_token was found in one sentence, set sentence to finished
            if eos_token_id_tensor is not None:
                unfinished_sequences = unfinished_sequences.mul(
                    next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
                )

                # stop when each sentence is finished
                if unfinished_sequences.max() == 0:
                    this_peer_finished = True

            # stop if we exceed the maximum length
            if stopping_criteria(input_ids, scores):
                this_peer_finished = True

            if this_peer_finished and not synced_gpus:
                break

        if streamer is not None:
            streamer.end()

        if return_dict_in_generate:
            if self.config.is_encoder_decoder:
                return GreedySearchEncoderDecoderOutput(
                    sequences=input_ids,
                    scores=scores,
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return GreedySearchDecoderOnlyOutput(
                    sequences=input_ids,
                    scores=scores,
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return input_ids

    def sample(
        self,
        input_ids: torch.LongTensor,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        logits_warper: Optional[LogitsProcessorList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[Union[int, List[int]]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: bool = False,
        streamer: Optional["BaseStreamer"] = None,
        **model_kwargs,
    ) -> Union[SampleOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **multinomial sampling** and
        can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.sample`] directly. Use generate() instead.
        For an overview of generation strategies and code examples, check the [following
        guide](../generation_strategies).

        </Tip>

        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            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.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
            model_kwargs:
                Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is
                an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.SampleDecoderOnlyOutput`], [`~generation.SampleEncoderDecoderOutput`] or `torch.LongTensor`:
            A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.SampleDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.SampleEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForCausalLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     TopKLogitsWarper,
        ...     TemperatureLogitsWarper,
        ...     StoppingCriteriaList,
        ...     MaxLengthCriteria,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
        >>> model = AutoModelForCausalLM.from_pretrained("gpt2")

        >>> # set pad_token_id to eos_token_id because GPT2 does not have a EOS token
        >>> model.config.pad_token_id = model.config.eos_token_id
        >>> model.generation_config.pad_token_id = model.config.eos_token_id

        >>> input_prompt = "Today is a beautiful day, and"
        >>> input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         MinLengthLogitsProcessor(15, eos_token_id=model.generation_config.eos_token_id),
        ...     ]
        ... )
        >>> # instantiate logits processors
        >>> logits_warper = LogitsProcessorList(
        ...     [
        ...         TopKLogitsWarper(50),
        ...         TemperatureLogitsWarper(0.7),
        ...     ]
        ... )

        >>> stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=20)])

        >>> torch.manual_seed(0)  # doctest: +IGNORE_RESULT
        >>> outputs = model.sample(
        ...     input_ids,
        ...     logits_processor=logits_processor,
        ...     logits_warper=logits_warper,
        ...     stopping_criteria=stopping_criteria,
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['Today is a beautiful day, and we must do everything possible to make it a day of celebration.']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
        logits_warper = logits_warper if logits_warper is not None else LogitsProcessorList()
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
        )
        return_dict_in_generate = (
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
        )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # keep track of which sequences are already finished
        unfinished_sequences = torch.ones(input_ids.shape[0], dtype=torch.long, device=input_ids.device)

        this_peer_finished = False  # used by synced_gpus only
        # auto-regressive generation
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # prepare model inputs
            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            # forward pass to get next token
            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]

            # pre-process distribution
            next_token_scores = logits_processor(input_ids, next_token_logits)
            next_token_scores = logits_warper(input_ids, next_token_scores)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (next_token_scores,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # sample
            probs = nn.functional.softmax(next_token_scores, dim=-1)
            next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)

            # finished sentences should have their next token be a padding token
            if eos_token_id is not None:
                if pad_token_id is None:
                    raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.")
                next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)

            # update generated ids, model inputs, and length for next step
            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
            if streamer is not None:
                streamer.put(next_tokens.cpu())
            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )

            # if eos_token was found in one sentence, set sentence to finished
            if eos_token_id_tensor is not None:
                unfinished_sequences = unfinished_sequences.mul(
                    next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
                )

                # stop when each sentence is finished
                if unfinished_sequences.max() == 0:
                    this_peer_finished = True

            # stop if we exceed the maximum length
            if stopping_criteria(input_ids, scores):
                this_peer_finished = True

            if this_peer_finished and not synced_gpus:
                break

        if streamer is not None:
            streamer.end()

        if return_dict_in_generate:
            if self.config.is_encoder_decoder:
                return SampleEncoderDecoderOutput(
                    sequences=input_ids,
                    scores=scores,
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return SampleDecoderOnlyOutput(
                    sequences=input_ids,
                    scores=scores,
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return input_ids

    def beam_search(
        self,
        input_ids: torch.LongTensor,
        beam_scorer: BeamScorer,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[Union[int, List[int]]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: bool = False,
        **model_kwargs,
    ) -> Union[BeamSearchOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **beam search decoding** and
        can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.beam_search`] directly. Use generate()
        instead. For an overview of generation strategies and code examples, check the [following
        guide](../generation_strategies).

        </Tip>

        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            beam_scorer (`BeamScorer`):
                An derived instance of [`BeamScorer`] that defines how beam hypotheses are constructed, stored and
                sorted during generation. For more information, the documentation of [`BeamScorer`] should be read.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            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.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            model_kwargs:
                Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is
                an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`generation.BeamSearchDecoderOnlyOutput`], [`~generation.BeamSearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.BeamSearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.BeamSearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.


        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForSeq2SeqLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     BeamSearchScorer,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
        >>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

        >>> encoder_input_str = "translate English to German: How old are you?"
        >>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids


        >>> # lets run beam search using 3 beams
        >>> num_beams = 3
        >>> # define decoder start token ids
        >>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
        >>> input_ids = input_ids * model.config.decoder_start_token_id

        >>> # add encoder_outputs to model keyword arguments
        >>> model_kwargs = {
        ...     "encoder_outputs": model.get_encoder()(
        ...         encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
        ...     )
        ... }

        >>> # instantiate beam scorer
        >>> beam_scorer = BeamSearchScorer(
        ...     batch_size=1,
        ...     num_beams=num_beams,
        ...     device=model.device,
        ... )

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
        ...     ]
        ... )

        >>> outputs = model.beam_search(input_ids, beam_scorer, logits_processor=logits_processor, **model_kwargs)

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['Wie alt bist du?']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
        if len(stopping_criteria) == 0:
            warnings.warn("You don't have defined any stopping_criteria, this will likely loop forever", UserWarning)
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
        )
        return_dict_in_generate = (
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
        )

        batch_size = len(beam_scorer._beam_hyps)
        num_beams = beam_scorer.num_beams

        batch_beam_size, cur_len = input_ids.shape

        if num_beams * batch_size != batch_beam_size:
            raise ValueError(
                f"Batch dimension of `input_ids` should be {num_beams * batch_size}, but is {batch_beam_size}."
            )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        beam_indices = (
            tuple(() for _ in range(batch_beam_size)) if (return_dict_in_generate and output_scores) else None
        )
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # initialise score of first beam with 0 and the rest with -1e9. This makes sure that only tokens
        # of the first beam are considered to avoid sampling the exact same tokens across all beams.
        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
        beam_scores[:, 1:] = -1e9
        beam_scores = beam_scores.view((batch_size * num_beams,))

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                cur_len = cur_len + 1
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]
            # hack: adjust tokens for Marian. For Marian we have to make sure that the `pad_token_id`
            # cannot be generated both before and after the `nn.functional.log_softmax` operation.
            next_token_logits = self.adjust_logits_during_generation(next_token_logits, cur_len=cur_len)
            next_token_scores = nn.functional.log_softmax(
                next_token_logits, dim=-1
            )  # (batch_size * num_beams, vocab_size)

            next_token_scores_processed = logits_processor(input_ids, next_token_scores)
            next_token_scores = next_token_scores_processed + beam_scores[:, None].expand_as(next_token_scores)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (next_token_scores_processed,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # reshape for beam search
            vocab_size = next_token_scores.shape[-1]
            next_token_scores = next_token_scores.view(batch_size, num_beams * vocab_size)

            # Sample 2 next tokens for each beam (so we have some spare tokens and match output of beam search)
            next_token_scores, next_tokens = torch.topk(
                next_token_scores, 2 * num_beams, dim=1, largest=True, sorted=True
            )

            next_indices = torch.div(next_tokens, vocab_size, rounding_mode="floor")
            next_tokens = next_tokens % vocab_size

            # stateless
            beam_outputs = beam_scorer.process(
                input_ids,
                next_token_scores,
                next_tokens,
                next_indices,
                pad_token_id=pad_token_id,
                eos_token_id=eos_token_id,
                beam_indices=beam_indices,
            )

            beam_scores = beam_outputs["next_beam_scores"]
            beam_next_tokens = beam_outputs["next_beam_tokens"]
            beam_idx = beam_outputs["next_beam_indices"]

            input_ids = torch.cat([input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1)

            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )
            if model_kwargs["past_key_values"] is not None:
                model_kwargs["past_key_values"] = self._reorder_cache(model_kwargs["past_key_values"], beam_idx)

            if return_dict_in_generate and output_scores:
                beam_indices = tuple((beam_indices[beam_idx[i]] + (beam_idx[i],) for i in range(len(beam_indices))))

            # increase cur_len
            cur_len = cur_len + 1

            if beam_scorer.is_done or stopping_criteria(input_ids, scores):
                if not synced_gpus:
                    break
                else:
                    this_peer_finished = True

        sequence_outputs = beam_scorer.finalize(
            input_ids,
            beam_scores,
            next_tokens,
            next_indices,
            pad_token_id=pad_token_id,
            eos_token_id=eos_token_id,
            max_length=stopping_criteria.max_length,
            beam_indices=beam_indices,
        )

        if return_dict_in_generate:
            if not output_scores:
                sequence_outputs["sequence_scores"] = None

            if self.config.is_encoder_decoder:
                return BeamSearchEncoderDecoderOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return BeamSearchDecoderOnlyOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return sequence_outputs["sequences"]

    def beam_sample(
        self,
        input_ids: torch.LongTensor,
        beam_scorer: BeamScorer,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        logits_warper: Optional[LogitsProcessorList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[Union[int, List[int]]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: bool = False,
        **model_kwargs,
    ) -> Union[BeamSampleOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **beam search multinomial
        sampling** and can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.beam_sample`] directly. Use generate()
        instead. For an overview of generation strategies and code examples, check the [following
        guide](../generation_strategies).

        </Tip>

        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            beam_scorer (`BeamScorer`):
                A derived instance of [`BeamScorer`] that defines how beam hypotheses are constructed, stored and
                sorted during generation. For more information, the documentation of [`BeamScorer`] should be read.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            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.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            model_kwargs:
                Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is
                an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.BeamSampleDecoderOnlyOutput`], [`~generation.BeamSampleEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.BeamSampleDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.BeamSampleEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForSeq2SeqLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     TopKLogitsWarper,
        ...     TemperatureLogitsWarper,
        ...     BeamSearchScorer,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
        >>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

        >>> encoder_input_str = "translate English to German: How old are you?"
        >>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids

        >>> # lets run beam search using 3 beams
        >>> num_beams = 3
        >>> # define decoder start token ids
        >>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
        >>> input_ids = input_ids * model.config.decoder_start_token_id

        >>> # add encoder_outputs to model keyword arguments
        >>> model_kwargs = {
        ...     "encoder_outputs": model.get_encoder()(
        ...         encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
        ...     )
        ... }

        >>> # instantiate beam scorer
        >>> beam_scorer = BeamSearchScorer(
        ...     batch_size=1,
        ...     max_length=model.config.max_length,
        ...     num_beams=num_beams,
        ...     device=model.device,
        ... )

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id)]
        ... )
        >>> # instantiate logits processors
        >>> logits_warper = LogitsProcessorList(
        ...     [
        ...         TopKLogitsWarper(50),
        ...         TemperatureLogitsWarper(0.7),
        ...     ]
        ... )

        >>> outputs = model.beam_sample(
        ...     input_ids, beam_scorer, logits_processor=logits_processor, logits_warper=logits_warper, **model_kwargs
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['Wie alt bist du?']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
        )
        return_dict_in_generate = (
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
        )

        batch_size = len(beam_scorer._beam_hyps)
        num_beams = beam_scorer.num_beams

        batch_beam_size, cur_len = input_ids.shape

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        beam_indices = (
            tuple(() for _ in range(batch_beam_size)) if (return_dict_in_generate and output_scores) else None
        )
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
        beam_scores = beam_scores.view((batch_size * num_beams,))

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                cur_len = cur_len + 1
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]

            # hack: adjust tokens for Marian. For Marian we have to make sure that the `pad_token_id`
            # cannot be generated both before and after the `nn.functional.log_softmax` operation.
            next_token_logits = self.adjust_logits_during_generation(next_token_logits, cur_len=cur_len)
            next_token_scores = nn.functional.log_softmax(
                next_token_logits, dim=-1
            )  # (batch_size * num_beams, vocab_size)

            next_token_scores_processed = logits_processor(input_ids, next_token_scores)
            next_token_scores = next_token_scores_processed + beam_scores[:, None].expand_as(next_token_scores)
            # Note: logits warpers are intentionally applied after adding running beam scores. On some logits warpers
            # (like top_p) this is indiferent, but on others (like temperature) it is not. For reference, see
            # https://github.com/huggingface/transformers/pull/5420#discussion_r449779867
            next_token_scores = logits_warper(input_ids, next_token_scores)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (logits_warper(input_ids, next_token_scores_processed),)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # reshape for beam search
            vocab_size = next_token_scores.shape[-1]
            next_token_scores = next_token_scores.view(batch_size, num_beams * vocab_size)

            probs = nn.functional.softmax(next_token_scores, dim=-1)

            next_tokens = torch.multinomial(probs, num_samples=2 * num_beams)
            next_token_scores = torch.gather(next_token_scores, -1, next_tokens)

            next_token_scores, _indices = torch.sort(next_token_scores, descending=True, dim=1)
            next_tokens = torch.gather(next_tokens, -1, _indices)

            next_indices = torch.div(next_tokens, vocab_size, rounding_mode="floor")
            next_tokens = next_tokens % vocab_size

            # stateless
            beam_outputs = beam_scorer.process(
                input_ids,
                next_token_scores,
                next_tokens,
                next_indices,
                pad_token_id=pad_token_id,
                eos_token_id=eos_token_id,
                beam_indices=beam_indices,
            )
            beam_scores = beam_outputs["next_beam_scores"]
            beam_next_tokens = beam_outputs["next_beam_tokens"]
            beam_idx = beam_outputs["next_beam_indices"]

            input_ids = torch.cat([input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1)

            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )
            if model_kwargs["past_key_values"] is not None:
                model_kwargs["past_key_values"] = self._reorder_cache(model_kwargs["past_key_values"], beam_idx)

            if return_dict_in_generate and output_scores:
                beam_indices = tuple((beam_indices[beam_idx[i]] + (beam_idx[i],) for i in range(len(beam_indices))))

            # increase cur_len
            cur_len = cur_len + 1

            if beam_scorer.is_done or stopping_criteria(input_ids, scores):
                if not synced_gpus:
                    break
                else:
                    this_peer_finished = True

        sequence_outputs = beam_scorer.finalize(
            input_ids,
            beam_scores,
            next_tokens,
            next_indices,
            pad_token_id=pad_token_id,
            eos_token_id=eos_token_id,
            max_length=stopping_criteria.max_length,
            beam_indices=beam_indices,
        )

        if return_dict_in_generate:
            if not output_scores:
                sequence_outputs["sequence_scores"] = None

            if self.config.is_encoder_decoder:
                return BeamSampleEncoderDecoderOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return BeamSampleDecoderOnlyOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return sequence_outputs["sequences"]

    def group_beam_search(
        self,
        input_ids: torch.LongTensor,
        beam_scorer: BeamScorer,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[Union[int, List[int]]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: bool = False,
        **model_kwargs,
    ):
        r"""
        Generates sequences of token ids for models with a language modeling head using **diverse beam search
        decoding** and can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.group_beam_search`] directly. Use
        generate() instead. For an overview of generation strategies and code examples, check the [following
        guide](../generation_strategies).

        </Tip>

        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            beam_scorer (`BeamScorer`):
                An derived instance of [`BeamScorer`] that defines how beam hypotheses are constructed, stored and
                sorted during generation. For more information, the documentation of [`BeamScorer`] should be read.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            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.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)

            model_kwargs:
                Additional model specific kwargs that will be forwarded to the `forward` function of the model. If
                model is an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.BeamSearchDecoderOnlyOutput`], [`~generation.BeamSearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.BeamSearchDecoderOnlyOutput`] if [`~generation.BeamSearchDecoderOnlyOutput`] if
            `model.config.is_encoder_decoder=False` and `return_dict_in_generate=True` or a
            [`~generation.BeamSearchEncoderDecoderOutput`] if `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForSeq2SeqLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     HammingDiversityLogitsProcessor,
        ...     BeamSearchScorer,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
        >>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

        >>> encoder_input_str = "translate English to German: How old are you?"
        >>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids


        >>> # lets run diverse beam search using 6 beams
        >>> num_beams = 6
        >>> # define decoder start token ids
        >>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
        >>> input_ids = input_ids * model.config.decoder_start_token_id

        >>> # add encoder_outputs to model keyword arguments
        >>> model_kwargs = {
        ...     "encoder_outputs": model.get_encoder()(
        ...         encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
        ...     )
        ... }

        >>> # instantiate beam scorer
        >>> beam_scorer = BeamSearchScorer(
        ...     batch_size=1,
        ...     max_length=model.config.max_length,
        ...     num_beams=num_beams,
        ...     device=model.device,
        ...     num_beam_groups=3,
        ... )

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         HammingDiversityLogitsProcessor(5.5, num_beams=6, num_beam_groups=3),
        ...         MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
        ...     ]
        ... )

        >>> outputs = model.group_beam_search(
        ...     input_ids, beam_scorer, logits_processor=logits_processor, **model_kwargs
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['Wie alt bist du?']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
        )
        return_dict_in_generate = (
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
        )

        num_beams = beam_scorer.num_beams
        num_beam_groups = beam_scorer.num_beam_groups
        num_sub_beams = num_beams // num_beam_groups
        batch_size = len(beam_scorer._beam_hyps) // num_beam_groups
        device = input_ids.device

        batch_beam_size, cur_len = input_ids.shape

        if return_dict_in_generate and output_scores:
            beam_indices = [tuple(() for _ in range(num_sub_beams * batch_size)) for _ in range(num_beam_groups)]
        else:
            beam_indices = None

        if num_beams * batch_size != batch_beam_size:
            raise ValueError(
                f"Batch dimension of `input_ids` should be {num_beams * batch_size}, but is {batch_beam_size}."
            )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # initialise score of first beam of each group with 0 and the rest with -1e9. This ensures that the beams in
        # the same group don't produce same tokens everytime.
        beam_scores = torch.full((batch_size, num_beams), -1e9, dtype=torch.float, device=device)
        beam_scores[:, ::num_sub_beams] = 0
        beam_scores = beam_scores.view((batch_size * num_beams,))

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # predicted tokens in cur_len step
            current_tokens = torch.zeros(batch_size * num_beams, dtype=input_ids.dtype, device=device)

            # indices which will form the beams in the next time step
            reordering_indices = torch.zeros(batch_size * num_beams, dtype=torch.long, device=device)

            # do one decoder step on all beams of all sentences in batch
            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                cur_len = cur_len + 1
                continue  # don't waste resources running the code we don't need

            if output_scores:
                processed_score = torch.zeros_like(outputs.logits[:, -1, :])

            for beam_group_idx in range(num_beam_groups):
                group_start_idx = beam_group_idx * num_sub_beams
                group_end_idx = min(group_start_idx + num_sub_beams, num_beams)
                group_size = group_end_idx - group_start_idx

                # indices of beams of current group among all sentences in batch
                batch_group_indices = []

                for batch_idx in range(batch_size):
                    batch_group_indices.extend(
                        [batch_idx * num_beams + idx for idx in range(group_start_idx, group_end_idx)]
                    )
                group_input_ids = input_ids[batch_group_indices]

                # select outputs of beams of current group only
                next_token_logits = outputs.logits[batch_group_indices, -1, :]

                # hack: adjust tokens for Marian. For Marian we have to make sure that the `pad_token_id`
                # cannot be generated both before and after the `nn.functional.log_softmax` operation.
                next_token_logits = self.adjust_logits_during_generation(next_token_logits, cur_len=cur_len)
                next_token_scores = nn.functional.log_softmax(
                    next_token_logits, dim=-1
                )  # (batch_size * group_size, vocab_size)
                vocab_size = next_token_scores.shape[-1]

                next_token_scores_processed = logits_processor(
                    group_input_ids, next_token_scores, current_tokens=current_tokens, beam_group_idx=beam_group_idx
                )
                next_token_scores = next_token_scores_processed + beam_scores[batch_group_indices].unsqueeze(-1)
                next_token_scores = next_token_scores.expand_as(next_token_scores_processed)

                if output_scores:
                    processed_score[batch_group_indices] = next_token_scores_processed

                # reshape for beam search
                next_token_scores = next_token_scores.view(batch_size, group_size * vocab_size)

                # Sample 2 next tokens for each beam (so we have some spare tokens and match output of beam search)
                next_token_scores, next_tokens = torch.topk(
                    next_token_scores, 2 * group_size, dim=1, largest=True, sorted=True
                )

                next_indices = torch.div(next_tokens, vocab_size, rounding_mode="floor")
                next_tokens = next_tokens % vocab_size

                # stateless
                process_beam_indices = sum(beam_indices, ()) if beam_indices is not None else None
                beam_outputs = beam_scorer.process(
                    group_input_ids,
                    next_token_scores,
                    next_tokens,
                    next_indices,
                    pad_token_id=pad_token_id,
                    eos_token_id=eos_token_id,
                    beam_indices=process_beam_indices,
                    group_index=beam_group_idx,
                )
                beam_scores[batch_group_indices] = beam_outputs["next_beam_scores"]
                beam_next_tokens = beam_outputs["next_beam_tokens"]
                beam_idx = beam_outputs["next_beam_indices"]

                if return_dict_in_generate and output_scores:
                    beam_indices[beam_group_idx] = tuple(
                        beam_indices[beam_group_idx][beam_idx[i]] + (beam_idx[i],) for i in range(len(beam_indices[0]))
                    )

                input_ids[batch_group_indices] = group_input_ids[beam_idx]
                group_input_ids = torch.cat([group_input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1)
                current_tokens[batch_group_indices] = group_input_ids[:, -1]

                # (beam_idx // group_size) -> batch_idx
                # (beam_idx % group_size) -> offset of idx inside the group
                reordering_indices[batch_group_indices] = (
                    num_beams * torch.div(beam_idx, group_size, rounding_mode="floor")
                    + group_start_idx
                    + (beam_idx % group_size)
                )

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (processed_score,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            input_ids = torch.cat([input_ids, current_tokens.unsqueeze(-1)], dim=-1)

            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )
            if model_kwargs["past_key_values"] is not None:
                model_kwargs["past_key_values"] = self._reorder_cache(
                    model_kwargs["past_key_values"], reordering_indices
                )

            # increase cur_len
            cur_len = cur_len + 1

            if beam_scorer.is_done or stopping_criteria(input_ids, scores):
                if not synced_gpus:
                    break
                else:
                    this_peer_finished = True

        final_beam_indices = sum(beam_indices, ()) if beam_indices is not None else None
        sequence_outputs = beam_scorer.finalize(
            input_ids,
            beam_scores,
            next_tokens,
            next_indices,
            pad_token_id=pad_token_id,
            eos_token_id=eos_token_id,
            max_length=stopping_criteria.max_length,
            beam_indices=final_beam_indices,
        )

        if return_dict_in_generate:
            if not output_scores:
                sequence_outputs["sequence_scores"] = None

            if self.config.is_encoder_decoder:
                return BeamSearchEncoderDecoderOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return BeamSearchDecoderOnlyOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return sequence_outputs["sequences"]

    def constrained_beam_search(
        self,
        input_ids: torch.LongTensor,
        constrained_beam_scorer: ConstrainedBeamSearchScorer,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[Union[int, List[int]]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: Optional[bool] = None,
        **model_kwargs,
    ) -> Union[BeamSearchOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **constrained beam search
        decoding** and can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.constrained_beam_search`] directly. Use
        generate() instead. For an overview of generation strategies and code examples, check the [following
        guide](../generation_strategies).

        </Tip>

        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            constrained_beam_scorer (`ConstrainedBeamSearchScorer`):
                A derived instance of [`BeamScorer`] that defines how beam hypotheses are constructed, stored and
                sorted during generation, while satisfying a list of positive constraints. For more information, the
                documentation of [`ConstrainedBeamSearchScorer`] should be read.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            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.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            model_kwargs:
                Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is
                an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`generation.BeamSearchDecoderOnlyOutput`], [`~generation.BeamSearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.BeamSearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.BeamSearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.


        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForSeq2SeqLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     ConstrainedBeamSearchScorer,
        ...     PhrasalConstraint,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
        >>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

        >>> encoder_input_str = "translate English to German: How old are you?"
        >>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids


        >>> # lets run beam search using 3 beams
        >>> num_beams = 3
        >>> # define decoder start token ids
        >>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
        >>> input_ids = input_ids * model.config.decoder_start_token_id

        >>> # add encoder_outputs to model keyword arguments
        >>> model_kwargs = {
        ...     "encoder_outputs": model.get_encoder()(
        ...         encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
        ...     )
        ... }

        >>> constraint_str = "Sie"
        >>> constraint_token_ids = tokenizer.encode(constraint_str)[:-1]  # slice to remove eos token
        >>> constraints = [PhrasalConstraint(token_ids=constraint_token_ids)]


        >>> # instantiate beam scorer
        >>> beam_scorer = ConstrainedBeamSearchScorer(
        ...     batch_size=1, num_beams=num_beams, device=model.device, constraints=constraints
        ... )

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
        ...     ]
        ... )

        >>> outputs = model.constrained_beam_search(
        ...     input_ids, beam_scorer, constraints=constraints, logits_processor=logits_processor, **model_kwargs
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['Wie alt sind Sie?']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
        if len(stopping_criteria) == 0:
            warnings.warn("You don't have defined any stopping_criteria, this will likely loop forever", UserWarning)
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
        )
        return_dict_in_generate = (
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
        )

        batch_size = len(constrained_beam_scorer._beam_hyps)
        num_beams = constrained_beam_scorer.num_beams

        batch_beam_size, cur_len = input_ids.shape

        if num_beams * batch_size != batch_beam_size:
            raise ValueError(
                f"Batch dimension of `input_ids` should be {num_beams * batch_size}, but is {batch_beam_size}."
            )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        beam_indices = (
            tuple(() for _ in range(batch_beam_size)) if (return_dict_in_generate and output_scores) else None
        )
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # initialise score of first beam with 0 and the rest with -1e9. This makes sure that only tokens
        # of the first beam are considered to avoid sampling the exact same tokens across all beams.
        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
        beam_scores[:, 1:] = -1e9
        beam_scores = beam_scores.view((batch_size * num_beams,))

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                cur_len = cur_len + 1
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]
            # hack: adjust tokens for Marian. For Marian we have to make sure that the `pad_token_id`
            # cannot be generated both before and after the `nn.functional.log_softmax` operation.
            next_token_logits = self.adjust_logits_during_generation(next_token_logits, cur_len=cur_len)
            next_token_scores = nn.functional.log_softmax(
                next_token_logits, dim=-1
            )  # (batch_size * num_beams, vocab_size)

            next_token_scores_processed = logits_processor(input_ids, next_token_scores)

            next_token_scores = next_token_scores_processed + beam_scores[:, None].expand_as(next_token_scores)

            scores_for_all_vocab = next_token_scores.clone()

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (next_token_scores,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # reshape for beam search
            vocab_size = next_token_scores.shape[-1]
            next_token_scores = next_token_scores.view(batch_size, num_beams * vocab_size)

            # Sample 2 next tokens for each beam (so we have some spare tokens and match output of beam search)
            next_token_scores, next_tokens = torch.topk(
                next_token_scores, 2 * num_beams, dim=1, largest=True, sorted=True
            )

            next_indices = (next_tokens / vocab_size).long()
            next_tokens = next_tokens % vocab_size

            # stateless
            beam_outputs = constrained_beam_scorer.process(
                input_ids,
                next_token_scores,
                next_tokens,
                next_indices,
                scores_for_all_vocab,
                pad_token_id=pad_token_id,
                eos_token_id=eos_token_id,
                beam_indices=beam_indices,
            )
            beam_scores = beam_outputs["next_beam_scores"]
            beam_next_tokens = beam_outputs["next_beam_tokens"]
            beam_idx = beam_outputs["next_beam_indices"]

            input_ids = torch.cat([input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1)
            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )
            if model_kwargs["past_key_values"] is not None:
                model_kwargs["past_key_values"] = self._reorder_cache(model_kwargs["past_key_values"], beam_idx)

            if return_dict_in_generate and output_scores:
                beam_indices = tuple((beam_indices[beam_idx[i]] + (beam_idx[i],) for i in range(len(beam_indices))))

            # increase cur_len
            cur_len = cur_len + 1

            if constrained_beam_scorer.is_done or stopping_criteria(input_ids, scores):
                if not synced_gpus:
                    break
                else:
                    this_peer_finished = True

        sequence_outputs = constrained_beam_scorer.finalize(
            input_ids,
            beam_scores,
            next_tokens,
            next_indices,
            pad_token_id=pad_token_id,
            eos_token_id=eos_token_id,
            max_length=stopping_criteria.max_length,
            beam_indices=beam_indices,
        )

        if return_dict_in_generate:
            if not output_scores:
                sequence_outputs["sequence_scores"] = None
            if self.config.is_encoder_decoder:
                return BeamSearchEncoderDecoderOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return BeamSearchDecoderOnlyOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return sequence_outputs["sequences"]

    def assisted_decoding(
        self,
        input_ids: torch.LongTensor,
        assistant_model: "PreTrainedModel",
        do_sample: bool = False,
        logits_processor: Optional[LogitsProcessorList] = None,
        logits_warper: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[Union[int, List[int]]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: bool = False,
        streamer: Optional["BaseStreamer"] = None,
        **model_kwargs,
    ):
        r"""
        Generates sequences of token ids for models with a language modeling head using **greedy decoding** or
        **sample** (depending on `do_sample`), assisted by a smaller model. Can be used for text-decoder, text-to-text,
        speech-to-text, and vision-to-text models.

        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.assisted_decoding`] directly. Use
        generate() instead. For an overview of generation strategies and code examples, check the [following
        guide](../generation_strategies).

        </Tip>

        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            assistant_model (`PreTrainedModel`, *optional*):
                An assistant model that can be used to accelerate generation. The assistant model must have the exact
                same tokenizer. The acceleration is achieved when forecasting candidate tokens with the assistent model
                is much faster than running generation with the model you're calling generate from. As such, the
                assistant model should be much smaller.
            do_sample (`bool`, *optional*, defaults to `False`):
                Whether or not to use sampling ; use greedy decoding otherwise.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            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.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
            model_kwargs:
                Additional model specific keyword arguments will be forwarded to the `forward` function of the model.
                If model is an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.GreedySearchDecoderOnlyOutput`], [`~generation.GreedySearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.GreedySearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.GreedySearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForCausalLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     StoppingCriteriaList,
        ...     MaxLengthCriteria,
        ... )

        >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
        >>> model = AutoModelForCausalLM.from_pretrained("gpt2")
        >>> assistant_model = AutoModelForCausalLM.from_pretrained("distilgpt2")
        >>> # set pad_token_id to eos_token_id because GPT2 does not have a PAD token
        >>> model.generation_config.pad_token_id = model.generation_config.eos_token_id
        >>> input_prompt = "It might be possible to"
        >>> input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids
        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         MinLengthLogitsProcessor(10, eos_token_id=model.generation_config.eos_token_id),
        ...     ]
        ... )
        >>> stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=20)])
        >>> outputs = model.assisted_decoding(
        ...     input_ids,
        ...     assistant_model=assistant_model,
        ...     logits_processor=logits_processor,
        ...     stopping_criteria=stopping_criteria,
        ... )
        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ["It might be possible to get a better understanding of the nature of the problem, but it's not"]
        ```"""
        # Assistant: initialize assistant-related variables
        if not hasattr(assistant_model, "max_assistant_tokens"):
            assistant_model.max_assistant_tokens = 5  # this value, which will be updated, persists across calls

        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        logits_warper = logits_warper if logits_warper is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
        if eos_token_id is not None and pad_token_id is None:
            raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.")
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
        )
        return_dict_in_generate = (
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
        )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # keep track of which sequences are already finished
        unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)

        # other auxiliary variables
        max_len = stopping_criteria[0].max_length
        assistant_kv_indexing = (
            1
            if "bloom" in assistant_model.__class__.__name__.lower()
            or (
                assistant_model.config.architectures is not None
                and "bloom" in assistant_model.config.architectures[0].lower()
            )
            else 0
        )

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # Assistant: main logic start
            cur_len = input_ids.shape[-1]

            #  1. Forecast next N tokens using the assistant model. This `for` block can be replaced with a
            # `.generate()` call if we decide to add `past_key_values` as a possible output of generate, as we
            # need access to the assistant cache to secure strong speedups.
            candidate_input_ids = input_ids
            for _ in range(int(assistant_model.max_assistant_tokens)):
                # 1.1. use the assistant model to obtain the next candidate logits
                if "assistant_past_key_values" in model_kwargs:
                    prev_seq_len = model_kwargs["assistant_past_key_values"][0][assistant_kv_indexing].shape[-2]
                    # `new_token_len` can be 1 or 2 (next token in assistant + last token picked by the larger model)
                    new_token_len = candidate_input_ids.shape[1] - prev_seq_len
                    assist_inputs = candidate_input_ids[:, -new_token_len:]
                    assist_attn = torch.ones_like(candidate_input_ids)
                    # TODO (joao): make it compatible with models that use unconventional fwd pass logic, like blip2
                    if assistant_model.config.is_encoder_decoder:
                        assistant_model_outputs = assistant_model(
                            decoder_input_ids=assist_inputs,
                            decoder_attention_mask=assist_attn,
                            past_key_values=model_kwargs["assistant_past_key_values"],
                            encoder_outputs=model_kwargs["assistant_encoder_outputs"],
                        )
                    else:
                        assistant_model_outputs = assistant_model(
                            assist_inputs,
                            attention_mask=assist_attn,
                            past_key_values=model_kwargs["assistant_past_key_values"],
                        )
                else:
                    if assistant_model.config.is_encoder_decoder:
                        assistant_model_outputs = assistant_model(
                            decoder_input_ids=candidate_input_ids,
                            encoder_outputs=model_kwargs["assistant_encoder_outputs"],
                        )
                    else:
                        assistant_model_outputs = assistant_model(candidate_input_ids)

                # 1.2. greedily select the next candidate token
                model_kwargs["assistant_past_key_values"] = assistant_model_outputs.past_key_values
                if len(logits_processor) > 0:
                    assistant_model_outputs.logits[:, -1, :] = logits_processor(
                        candidate_input_ids, assistant_model_outputs.logits[:, -1, :]
                    )
                new_token = assistant_model_outputs.logits[:, -1, :].argmax(dim=-1)
                candidate_input_ids = torch.cat((candidate_input_ids, new_token[:, None]), dim=-1)

                # 1.3. stop assistant generation on EOS
                if eos_token_id_tensor is not None:
                    last_assistant_token_is_eos = new_token.tile(eos_token_id_tensor.shape[0], 1)
                    last_assistant_token_is_eos = (
                        ~last_assistant_token_is_eos.ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0).bool()
                    )
                    if last_assistant_token_is_eos:
                        break
                else:
                    last_assistant_token_is_eos = False

            candidate_length = candidate_input_ids.shape[1] - input_ids.shape[1]

            # 2. Use the original model to obtain the next token logits given the candidate sequence. We obtain
            # `candidate_length + 1` relevant logits from this process: in the event that all candidates are correct,
            # we use this forward pass to also pick the subsequent logits in the original model.

            # 2.1. Run a forward pass on the candidate sequence
            if "past_key_values" in model_kwargs:
                model_attn = torch.ones_like(candidate_input_ids)
                model_input_ids = candidate_input_ids[:, -candidate_length - 1 :]
                if self.config.is_encoder_decoder:
                    outputs = self(
                        decoder_input_ids=model_input_ids,
                        decoder_attention_mask=model_attn,
                        past_key_values=model_kwargs["past_key_values"],
                        encoder_outputs=model_kwargs["encoder_outputs"],
                        output_attentions=output_attentions,
                        output_hidden_states=output_hidden_states,
                        use_cache=True,
                    )
                else:
                    outputs = self(
                        model_input_ids,
                        attention_mask=model_attn,
                        past_key_values=model_kwargs["past_key_values"],
                        output_attentions=output_attentions,
                        output_hidden_states=output_hidden_states,
                        use_cache=True,
                    )
            else:
                if self.config.is_encoder_decoder:
                    outputs = self(
                        decoder_input_ids=candidate_input_ids,
                        encoder_outputs=model_kwargs["encoder_outputs"],
                        output_attentions=output_attentions,
                        output_hidden_states=output_hidden_states,
                        use_cache=True,
                    )
                else:
                    outputs = self(
                        candidate_input_ids,
                        output_attentions=output_attentions,
                        output_hidden_states=output_hidden_states,
                        use_cache=True,
                    )

            # 2.2. Process the new logits
            new_logits = outputs.logits[:, -candidate_length - 1 :]  # excludes the input prompt if present
            if len(logits_processor) > 0:
                for i in range(candidate_length):
                    new_logits[:, i, :] = logits_processor(candidate_input_ids[:, : cur_len + i], new_logits[:, i, :])
            if len(logits_warper) > 0:
                for i in range(candidate_length):
                    new_logits[:, i, :] = logits_warper(candidate_input_ids[:, : cur_len + i], new_logits[:, i, :])

            # 3. Obtain the next tokens from the original model logits.
            if do_sample:
                probs = new_logits[:, -candidate_length - 1 :, :].softmax(dim=-1)
                selected_tokens = torch.multinomial(probs[0, :, :], num_samples=1).squeeze(1)[None, :]
            else:
                selected_tokens = new_logits[:, -candidate_length - 1 :, :].argmax(dim=-1)

            # 4. Compare the argmax from the original model logits with the assistant forecasted tokens. We can keep
            # the assistant forecasted tokens until the first mismatch, or until the max length is reached.
            candidate_new_tokens = candidate_input_ids[:, -candidate_length:]
            n_matches = ((~(candidate_new_tokens == selected_tokens[:, :-1])).cumsum(dim=-1) < 1).sum()

            # 5. Update variables according to the number of matching assistant tokens. Remember: the token generated
            # by the model after the last candidate match is also valid, as it is generated from a correct sequence.
            # Because of this last token, assisted generation search reduces to a normal greedy search/sample if there
            # is no match.

            # 5.1. Ensure we don't generate beyond max_len or an EOS token
            if last_assistant_token_is_eos and n_matches == candidate_length:
                n_matches -= 1
            n_matches = min(n_matches, max_len - cur_len - 1)

            # 5.2. Get the valid continuation, after the matching tokens
            valid_tokens = selected_tokens[:, : n_matches + 1]
            input_ids = torch.cat((input_ids, valid_tokens), dim=-1)
            if streamer is not None:
                streamer.put(valid_tokens.cpu())
            new_cur_len = input_ids.shape[-1]

            # 5.3. Discard past key values relative to unused assistant tokens
            new_cache_size = new_cur_len - 1
            outputs.past_key_values = _crop_past_key_values(self, outputs.past_key_values, new_cache_size)
            model_kwargs["assistant_past_key_values"] = _crop_past_key_values(
                assistant_model, model_kwargs["assistant_past_key_values"], new_cache_size - 1
            )  # the assistant does not have the token after the last match, hence the -1

            # 6. Adjust the max number of assistant tokens to use in the next iteration. This is a simple heuristic,
            # probably can be improved -- we want to balance the benefits of getting assistant tokens correct with the
            # cost of forecasting incorrect assistant tokens.
            if n_matches == int(assistant_model.max_assistant_tokens):
                assistant_model.max_assistant_tokens += 2.0
            else:
                assistant_model.max_assistant_tokens = max(1.0, assistant_model.max_assistant_tokens - 1.0)

            # Assistant: main logic end

            if synced_gpus and this_peer_finished:
                continue  # don't waste resources running the code we don't need

            # Store scores, attentions and hidden_states when required
            # Assistant: modified to append one tuple element per token, as in the other generation methods.
            if return_dict_in_generate:
                if output_scores:
                    scores += tuple(new_logits[:, i, :] for i in range(n_matches + 1))

                if "past_key_values" not in model_kwargs:
                    added_len = new_cur_len
                else:
                    added_len = n_matches + 1

                if output_attentions:
                    if self.config.is_encoder_decoder:
                        cross_attentions = _split_model_outputs(
                            cross_attentions, outputs.cross_attentions, cur_len, added_len
                        )
                        decoder_attentions = _split_model_outputs(
                            decoder_attentions,
                            outputs.decoder_attentions,
                            cur_len,
                            added_len,
                            is_decoder_attention=True,
                        )
                    else:
                        decoder_attentions = _split_model_outputs(
                            decoder_attentions,
                            outputs.attentions,
                            cur_len,
                            added_len,
                            is_decoder_attention=True,
                        )
                if output_hidden_states:
                    if self.config.is_encoder_decoder:
                        decoder_hidden_states = _split_model_outputs(
                            decoder_hidden_states, outputs.decoder_hidden_states, cur_len, added_len
                        )
                    else:
                        decoder_hidden_states = _split_model_outputs(
                            decoder_hidden_states, outputs.hidden_states, cur_len, added_len
                        )

            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )

            # if eos_token was found in one sentence, set sentence to finished
            if eos_token_id_tensor is not None:
                unfinished_sequences = unfinished_sequences.mul(
                    input_ids[:, -1]
                    .tile(eos_token_id_tensor.shape[0], 1)
                    .ne(eos_token_id_tensor.unsqueeze(1))
                    .prod(dim=0)
                )

                # stop when each sentence is finished
                if unfinished_sequences.max() == 0:
                    this_peer_finished = True

            # stop if we exceed the maximum length
            if stopping_criteria(input_ids, scores):
                this_peer_finished = True

            if this_peer_finished and not synced_gpus:
                break

        if streamer is not None:
            streamer.end()

        if return_dict_in_generate:
            if self.config.is_encoder_decoder:
                return GreedySearchEncoderDecoderOutput(
                    sequences=input_ids,
                    scores=scores,
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return GreedySearchDecoderOnlyOutput(
                    sequences=input_ids,
                    scores=scores,
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return input_ids


def _crop_past_key_values(model, past_key_values, maximum_length):
    """Crops the past key values up to a certain maximum length."""
    new_past = []
    if model.config.is_encoder_decoder:
        for idx in range(len(past_key_values)):
            new_past.append(
                (
                    past_key_values[idx][0][:, :, :maximum_length, :],
                    past_key_values[idx][1][:, :, :maximum_length, :],
                    past_key_values[idx][2],
                    past_key_values[idx][3],
                )
            )
        past_key_values = tuple(new_past)
    # bloom is special
    elif "bloom" in model.__class__.__name__.lower() or (
        model.config.architectures is not None and "bloom" in model.config.architectures[0].lower()
    ):
        for idx in range(len(past_key_values)):
            new_past.append(
                (
                    past_key_values[idx][0][:, :, :maximum_length],
                    past_key_values[idx][1][:, :maximum_length, :],
                )
            )
        past_key_values = tuple(new_past)
    # gptbigcode is too
    elif "gptbigcode" in model.__class__.__name__.lower() or (
        model.config.architectures is not None and "gptbigcode" in model.config.architectures[0].lower()
    ):
        if model.config.multi_query:
            for idx in range(len(past_key_values)):
                past_key_values[idx] = past_key_values[idx][:, :maximum_length, :]
        else:
            for idx in range(len(past_key_values)):
                past_key_values[idx] = past_key_values[idx][:, :, :maximum_length, :]
    else:
        for idx in range(len(past_key_values)):
            new_past.append(
                (
                    past_key_values[idx][0][:, :, :maximum_length, :],
                    past_key_values[idx][1][:, :, :maximum_length, :],
                )
            )
        past_key_values = tuple(new_past)
    return past_key_values


def _split_model_outputs(outputs, new_outputs, cur_len, added_len, is_decoder_attention=False):
    """
    Given the (decoder/cross attentions)/(decoder hidden states) for multiple generated tokens, splits it into a tuple
    where each member corresponds to a single generated token.
    """
    # Retrocompatibility: in our generation functions, the first iteration includes the attention/hidden states for the
    # prompt.
    if len(outputs) == 0:
        new_tuple = ()
        for layer in new_outputs:
            last_dim_size = cur_len if is_decoder_attention else layer.shape[-1]
            new_tuple += (layer[..., :cur_len, :last_dim_size],)
        outputs += (new_tuple,)
        # The first iteration contains the prompt + 1 generated token, let's update the length variables accordingly
        cur_len += 1
        added_len -= cur_len

    for i in range(added_len):
        new_tuple = ()
        for layer in new_outputs:
            last_dim_size = cur_len + i if is_decoder_attention else layer.shape[-1]
            new_tuple += (layer[..., i : i + 1, :last_dim_size],)
        outputs += (new_tuple,)
    return outputs


def top_k_top_p_filtering(
    logits: torch.FloatTensor,
    top_k: int = 0,
    top_p: float = 1.0,
    filter_value: float = -float("Inf"),
    min_tokens_to_keep: int = 1,
) -> torch.FloatTensor:
    """
    Filter a distribution of logits using top-k and/or nucleus (top-p) filtering

    Args:
        logits: logits distribution shape (batch size, vocabulary size)
        top_k (`int`, *optional*, defaults to 0):
            If > 0, only keep the top k tokens with highest probability (top-k filtering)
        top_p (`float`, *optional*, defaults to 1.0):
            If < 1.0, only keep the top tokens with cumulative probability >= top_p (nucleus filtering). Nucleus
            filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
        min_tokens_to_keep (`int`, *optional*, defaults to 1):
            Minimumber of tokens we keep per batch example in the output.

    From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
    """
    if top_k > 0:
        logits = TopKLogitsWarper(top_k=top_k, filter_value=filter_value, min_tokens_to_keep=min_tokens_to_keep)(
            None, logits
        )

    if 0 <= top_p <= 1.0:
        logits = TopPLogitsWarper(top_p=top_p, filter_value=filter_value, min_tokens_to_keep=min_tokens_to_keep)(
            None, logits
        )

    return logits


def _ranking_fast(
    context_hidden: torch.FloatTensor,
    next_hidden: torch.FloatTensor,
    next_top_k_probs: torch.FloatTensor,
    alpha: float,
    beam_width: int,
) -> torch.FloatTensor:
    """
    Reranks the top_k candidates based on a degeneration penalty (cosine similarity with previous tokens), as described
    in the paper "A Contrastive Framework for Neural Text Generation". Returns the index of the best candidate for each
    row in the batch.
    """
    norm_context_hidden = context_hidden / context_hidden.norm(dim=2, keepdim=True)
    norm_next_hidden = next_hidden / next_hidden.norm(dim=2, keepdim=True)
    cosine_matrix = torch.matmul(norm_context_hidden, norm_next_hidden.transpose(1, 2)).squeeze(-1)  # [B*K, S]
    degeneration_penalty, _ = torch.max(cosine_matrix, dim=-1)  # [B*K]
    next_top_k_probs = next_top_k_probs.view(-1)  # [B*K]
    contrastive_score = (1.0 - alpha) * next_top_k_probs - alpha * degeneration_penalty
    contrastive_score = torch.stack(torch.split(contrastive_score, beam_width))  # [B, K]
    _, selected_idx = contrastive_score.max(dim=-1)  # [B]
    return selected_idx