# 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 os import warnings from dataclasses import dataclass from typing import TYPE_CHECKING, Any, Callable, Optional, Union import torch import torch.distributed as dist from packaging import version from torch import nn from ..cache_utils import ( Cache, DynamicCache, EncoderDecoderCache, QuantizedCache, StaticCache, ) from ..dynamic_module_utils import ( check_python_requirements, get_cached_module_file, get_class_in_module, resolve_trust_remote_code, ) from ..integrations.deepspeed import is_deepspeed_zero3_enabled from ..integrations.fsdp import is_fsdp_managed_module from ..masking_utils import create_masks_for_generate from ..pytorch_utils import isin_mps_friendly from ..tokenization_utils import ExtensionsTrie from ..utils import ( ModelOutput, TransformersKwargs, is_accelerate_available, is_hqq_available, is_optimum_quanto_available, is_torchdynamo_exporting, logging, ) from .candidate_generator import ( AssistantVocabTranslatorCache, AssistedCandidateGenerator, AssistedCandidateGeneratorDifferentTokenizers, CandidateGenerator, EarlyExitCandidateGenerator, PromptLookupCandidateGenerator, UniversalSpeculativeDecodingGenerator, _prepare_attention_mask, _prepare_token_type_ids, ) from .configuration_utils import ( ALL_STATIC_CACHE_IMPLEMENTATIONS, DEPRECATED_STATIC_CACHE_IMPLEMENTATIONS, STATIC_CACHE_IMPLEMENTATIONS, GenerationConfig, GenerationMode, ) from .continuous_batching import ContinuousMixin from .logits_process import ( EncoderNoRepeatNGramLogitsProcessor, EncoderRepetitionPenaltyLogitsProcessor, EpsilonLogitsWarper, EtaLogitsWarper, ExponentialDecayLengthPenalty, ForcedBOSTokenLogitsProcessor, ForcedEOSTokenLogitsProcessor, InfNanRemoveLogitsProcessor, LogitNormalization, LogitsProcessorList, MinLengthLogitsProcessor, MinNewTokensLengthLogitsProcessor, MinPLogitsWarper, NoBadWordsLogitsProcessor, NoRepeatNGramLogitsProcessor, PrefixConstrainedLogitsProcessor, RepetitionPenaltyLogitsProcessor, SequenceBiasLogitsProcessor, SuppressTokensAtBeginLogitsProcessor, SuppressTokensLogitsProcessor, TemperatureLogitsWarper, TopKLogitsWarper, TopPLogitsWarper, TypicalLogitsWarper, UnbatchedClassifierFreeGuidanceLogitsProcessor, ) from .stopping_criteria import ( ConfidenceCriteria, EosTokenCriteria, MaxLengthCriteria, MaxTimeCriteria, StoppingCriteria, StoppingCriteriaList, StopStringCriteria, ) if TYPE_CHECKING: from ..modeling_utils import PreTrainedModel from ..tokenization_utils_base import PreTrainedTokenizerBase from .streamers import BaseStreamer logger = logging.get_logger(__name__) if is_accelerate_available(): from accelerate.hooks import AlignDevicesHook, add_hook_to_module # Variable names used to hold the cache at generation time ALL_CACHE_NAMES = [ "past_key_values", # default "cache_params", # mamba-based models "state", # rwkv "mems", # xlnet "past_buckets_states", # reformer ] GENERATION_MODES_MAPPING = { GenerationMode.SAMPLE: "_sample", GenerationMode.GREEDY_SEARCH: "_sample", GenerationMode.BEAM_SEARCH: "_beam_search", GenerationMode.BEAM_SAMPLE: "_beam_search", GenerationMode.ASSISTED_GENERATION: "_assisted_decoding", # Deprecated methods GenerationMode.DOLA_GENERATION: "transformers-community/dola", GenerationMode.CONTRASTIVE_SEARCH: "transformers-community/contrastive-search", GenerationMode.GROUP_BEAM_SEARCH: "transformers-community/group-beam-search", GenerationMode.CONSTRAINED_BEAM_SEARCH: "transformers-community/constrained-beam-search", } @dataclass class GenerateDecoderOnlyOutput(ModelOutput): """ Outputs of decoder-only generation models, when using non-beam methods. 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`): Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size, config.vocab_size)`. logits (`tuple(torch.FloatTensor)` *optional*, returned when `output_logits=True`): Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size, config.vocab_size)`. attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True`): 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`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size, generated_length, hidden_size)`. past_key_values (`Cache`, *optional*, returned when `use_cache=True`): Returns the model cache, used to speed up decoding. Different models have a different cache format, check the model's documentation. Usually, a [`~cache_utils.Cache`] instance. """ sequences: torch.LongTensor scores: Optional[tuple[torch.FloatTensor]] = None logits: Optional[tuple[torch.FloatTensor]] = None attentions: Optional[tuple[tuple[torch.FloatTensor]]] = None hidden_states: Optional[tuple[tuple[torch.FloatTensor]]] = None past_key_values: Optional[Cache] = None @dataclass class GenerateEncoderDecoderOutput(ModelOutput): """ Outputs of encoder-decoder generation models, when using non-beam methods. 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`): Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size, config.vocab_size)`. logits (`tuple(torch.FloatTensor)` *optional*, returned when `output_logits=True`): Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size, config.vocab_size)`. encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True`): 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`): 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`): 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`): 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`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size, generated_length, hidden_size)`. past_key_values (`Cache`, *optional*, returned when `use_cache=True`): Returns the model cache, used to speed up decoding. Different models have a different cache format, check the model's documentation. Usually, a [`~cache_utils.Cache`] instance. """ sequences: torch.LongTensor scores: Optional[tuple[torch.FloatTensor]] = None logits: 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 past_key_values: Optional[Cache] = None @dataclass class GenerateBeamDecoderOnlyOutput(ModelOutput): """ Outputs of decoder-only generation models, when using beam methods. 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`): Final beam scores of the generated `sequences`. scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True`): Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size*num_beams, config.vocab_size)`. logits (`tuple(torch.FloatTensor)` *optional*, returned when `output_logits=True`): Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size*num_beams, config.vocab_size)`. beam_indices (`torch.LongTensor`, *optional*, returned when `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`): 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`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size*num_beams*num_return_sequences, generated_length, hidden_size)`. past_key_values (`Cache`, *optional*, returned when `use_cache=True`): Returns the model cache, used to speed up decoding. Different models have a different cache format, check the model's documentation. Usually, a [`~cache_utils.Cache`] instance. """ sequences: torch.LongTensor sequences_scores: Optional[torch.FloatTensor] = None scores: Optional[tuple[torch.FloatTensor]] = None logits: 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 past_key_values: Optional[Cache] = None @dataclass class GenerateBeamEncoderDecoderOutput(ModelOutput): """ Outputs of encoder-decoder generation models, when using beam methods. 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`): Final beam scores of the generated `sequences`. scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True`): Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size*num_beams, config.vocab_size)`. logits (`tuple(torch.FloatTensor)` *optional*, returned when `output_logits=True`): Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size*num_beams, config.vocab_size)`. beam_indices (`torch.LongTensor`, *optional*, returned when `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`): 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`): 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`): 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`): 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`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size*num_beams*num_return_sequences, generated_length, hidden_size)`. past_key_values (`Cache`, *optional*, returned when `use_cache=True`): Returns the model cache, used to speed up decoding. Different models have a different cache format, check the model's documentation. Usually, a [`~cache_utils.Cache`] instance. """ sequences: torch.LongTensor sequences_scores: Optional[torch.FloatTensor] = None scores: Optional[tuple[torch.FloatTensor]] = None logits: 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 past_key_values: Optional[Cache] = None # TODO (joao): remove the equivalent classes and typing shortcuts below in v5 # Equivalent classes (kept for retrocompatibility purposes) GreedySearchDecoderOnlyOutput = GenerateDecoderOnlyOutput ContrastiveSearchDecoderOnlyOutput = GenerateDecoderOnlyOutput SampleDecoderOnlyOutput = GenerateDecoderOnlyOutput ContrastiveSearchEncoderDecoderOutput = GenerateEncoderDecoderOutput GreedySearchEncoderDecoderOutput = GenerateEncoderDecoderOutput SampleEncoderDecoderOutput = GenerateEncoderDecoderOutput BeamSearchDecoderOnlyOutput = GenerateBeamDecoderOnlyOutput BeamSampleDecoderOnlyOutput = GenerateBeamDecoderOnlyOutput BeamSearchEncoderDecoderOutput = GenerateBeamEncoderDecoderOutput BeamSampleEncoderDecoderOutput = GenerateBeamEncoderDecoderOutput GreedySearchOutput = Union[GreedySearchEncoderDecoderOutput, GreedySearchDecoderOnlyOutput] SampleOutput = Union[SampleEncoderDecoderOutput, SampleDecoderOnlyOutput] BeamSearchOutput = Union[BeamSearchEncoderDecoderOutput, BeamSearchDecoderOnlyOutput] BeamSampleOutput = Union[BeamSampleEncoderDecoderOutput, BeamSampleDecoderOnlyOutput] ContrastiveSearchOutput = Union[ContrastiveSearchEncoderDecoderOutput, ContrastiveSearchDecoderOnlyOutput] # Typing shortcuts GenerateNonBeamOutput = Union[GenerateDecoderOnlyOutput, GenerateEncoderDecoderOutput] GenerateBeamOutput = Union[GenerateBeamDecoderOnlyOutput, GenerateBeamEncoderDecoderOutput] GenerateOutput = Union[GenerateNonBeamOutput, GenerateBeamOutput] class GenerationMixin(ContinuousMixin): """ A class containing all functions for auto-regressive text generation, to be used as a mixin in model classes. Inheriting from this class causes the model to have special generation-related behavior, such as loading a `GenerationConfig` at initialization time or ensuring `generate`-related tests are run in `transformers` CI. A model class should inherit from `GenerationMixin` to enable calling methods like `generate`, or when it has defined a custom `generate` method that relies on `GenerationMixin`, directly or indirectly, which approximately shares the same interface to public methods like `generate`. Three examples: - `LlamaForCausalLM` should inherit from `GenerationMixin` to enable calling `generate` and other public methods in the mixin; - `BlipForQuestionAnswering` has a custom `generate` method that approximately shares the same interface as `GenerationMixin.generate` (it has a few extra arguments, and the same output). That function also calls `GenerationMixin.generate` indirectly, through an inner model. As such, `BlipForQuestionAnswering` should inherit from `GenerationMixin` to benefit from all generation-related automation in our codebase; - `BarkModel` has a custom `generate` method and one of its inner models calls `GenerationMixin.generate`. However, its `generate` does not share the same interface as `GenerationMixin.generate`. In this case, `BarkModel` should NOT inherit from `GenerationMixin`, as it breaks the `generate` interface. The class exposes [`~generation.GenerationMixin.generate`], which can be used for: - *greedy decoding* if `num_beams=1` and `do_sample=False` - *multinomial sampling* if `num_beams=1` and `do_sample=True` - *beam-search decoding* if `num_beams>1` and `do_sample=False` - *beam-search multinomial sampling* if `num_beams>1` and `do_sample=True` - *assisted decoding* if `assistant_model` or `prompt_lookup_num_tokens` is passed to `.generate()` To learn more about decoding strategies refer to the [text generation strategies guide](../generation_strategies). """ def load_custom_generate( self, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]] = None, trust_remote_code: Optional[bool] = None, **kwargs, ) -> Callable: """ Loads and returns a custom generate function, given a model repo. Args: pretrained_model_name_or_path (`str` or `os.PathLike`): Can be either: - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co. - A path to a *directory* containing model weights saved using [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`. trust_remote_code (`bool`, *optional*): Whether or not to allow for custom models defined on the Hub in their own modeling files. This option should only be set to `True` for repositories you trust and in which you have read the code, as it will execute code present on the Hub on your local machine. **kwargs: Additional keyword arguments for remote code loading. Raises: OSError: If `pretrained_model_name_or_path` does not contain a `custom_generate` subdirectory. Returns: A callable that can be used to generate text. """ # Fetches the generate.py file from the model repo. If it doesn't exist, a file in `.no_exist` cache directory # is created (preventing future hub requests), and an OSError is raised. try: module = get_cached_module_file( pretrained_model_name_or_path, module_file="custom_generate/generate.py", **kwargs ) except OSError: raise OSError( f"`{pretrained_model_name_or_path}` does not contain a `custom_generate` subdirectory with a " "`generate.py` file, can't load the custom generate function." ) # Handle opt-in `trust_remote_code` and related exceptions is_local_code = os.path.exists(pretrained_model_name_or_path) error_message = ( f"The repository `{pretrained_model_name_or_path}` contains custom generation code that will override " "the default `generate` method." ) resolve_trust_remote_code( trust_remote_code, pretrained_model_name_or_path, has_local_code=is_local_code, has_remote_code=not is_local_code, error_message=error_message, ) # Load the custom generate function check_python_requirements( pretrained_model_name_or_path, requirements_file="custom_generate/requirements.txt", **kwargs ) custom_generate_function = get_class_in_module("generate", module) return custom_generate_function def _cache_dependant_input_preparation( self, input_ids: torch.LongTensor, inputs_embeds: Optional[torch.FloatTensor], cache_position: Optional[torch.LongTensor], ) -> tuple[torch.FloatTensor, torch.LongTensor]: """ Generic cache-dependent input preparation The code is put in a separate function to allow granular unit testing as it needs a different implementation to be exportable. If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens - Exception 1: when passing input_embeds, input_ids may be missing entries - Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here - Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case. - Exception 4: If input_embeds are passed then slice it through `cache_position`, to keep only the unprocessed tokens and generate the first token for each sequence. Later use the generated Input ids for continuation. The current implementation does not rely on ``self`` and could be a class method. It is left as a standard method to be easily rewritten. """ if is_torchdynamo_exporting(): return self._cache_dependant_input_preparation_exporting(input_ids, inputs_embeds, cache_position) if inputs_embeds is not None and input_ids.shape[1] == 0: # Exception 4 inputs_embeds = inputs_embeds[:, -cache_position.shape[0] :] elif ( inputs_embeds is not None # Exception 1 or (cache_position[-1] >= input_ids.shape[1]) # Exception 3 ): input_ids = input_ids[:, -cache_position.shape[0] :] elif input_ids.shape[1] != cache_position.shape[0]: # Default case (the "else", a no op, is Exception 2) input_ids = input_ids[:, cache_position] return inputs_embeds, input_ids def _cache_dependant_input_preparation_exporting( self, input_ids: torch.LongTensor, inputs_embeds: Optional[torch.FloatTensor], cache_position: Optional[torch.LongTensor], ) -> tuple[torch.FloatTensor, torch.LongTensor]: """ This method implements method ``_cache_dependant_input_preparation`` with :func:`torch.cond` to make it exportable with :func:`torch.export.export`. The code is put in a separate function to allow granular unit testing. """ if inputs_embeds is None: input_ids = input_ids[:, cache_position] else: # This is the code we need to implemented with torch.cond. # if input_ids.shape[1] == 0: # inputs_embeds = inputs_embeds[:, -cache_position.shape[0] :] # else: # if cache_position[-1] >= input_ids.shape[1]: # input_ids = input_ids[:, -cache_position.shape[0] :] # else: # if input_ids.shape[1] != cache_position.shape[0]: # input_ids = input_ids[:, cache_position] # We need to clone the outputs to avoid aliasing. def branch_1(inputs_embeds, cache_position): return inputs_embeds[:, -cache_position.shape[0] :].clone() def branch_2(input_ids, cache_position): return input_ids[:, -cache_position.shape[0] :].clone() def branch_3(input_ids, cache_position): return input_ids[:, cache_position].clone() inputs_embeds, input_ids = torch.cond( input_ids.shape[1] == 0, ( lambda input_ids, inputs_embeds, cache_position: ( branch_1(inputs_embeds, cache_position), input_ids.clone(), ) ), ( lambda input_ids, inputs_embeds, cache_position: ( inputs_embeds, torch.cond( cache_position[-1] >= input_ids.shape[1], branch_2, lambda input_ids, cache_position: ( torch.cond( input_ids.shape[1] != cache_position.shape[0], branch_3, (lambda input_ids, cache_position: input_ids.clone()), [input_ids, cache_position], ) ), [input_ids, cache_position], ), ) ), [input_ids, inputs_embeds, cache_position], ) return inputs_embeds, input_ids def prepare_inputs_for_generation( self, input_ids: torch.LongTensor, past_key_values: Optional[Cache] = None, attention_mask: Optional[torch.LongTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, cache_position: Optional[torch.LongTensor] = None, **kwargs, ): """ Prepare the model inputs for generation. Notable steps include selecting the correct input key and cloning when appropriate, creating position_ids from the attention_mask when missing, slicing inputs and converting 2D attention masks to 4D for compilable caches, and finally forwarding all additional keyword arguments unchanged to the model's forward pass. See the forward pass in the model documentation for expected arguments (different models might have different requirements for e.g. `past_key_values`). This function should work as is for most LLMs. """ # 1. Handle BC: model_inputs = {} model_inputs["cache_position"] = cache_position # 2. Generic cache-dependent input preparation if past_key_values is not None: model_inputs["past_key_values"] = past_key_values # TODO (joao): handle the case where cache length == input_ids length. The function below results in an # exception because we get empty input_ids after slicing. In essence, we need to roll back the cache 1 # token to recompute the logits for the first token to be generated (but not all caches support roll backs) inputs_embeds, input_ids = self._cache_dependant_input_preparation( input_ids, inputs_embeds, cache_position ) # 3. Prepare base model inputs input_ids_key = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids" # if `inputs_embeds` are passed, we only want to use them in the 1st generation step for every prompt. if not self.config.is_encoder_decoder: if inputs_embeds is not None and len(cache_position) == inputs_embeds.shape[1]: model_inputs[input_ids_key] = None model_inputs["inputs_embeds"] = inputs_embeds else: # `clone` calls in this function ensure a consistent stride. See #32227 model_inputs[input_ids_key] = input_ids.clone(memory_format=torch.contiguous_format) model_inputs["inputs_embeds"] = None else: model_inputs[input_ids_key] = input_ids.clone(memory_format=torch.contiguous_format) # 4. Create missing `position_ids` on the fly encoder_attention_mask = attention_mask if self.config.is_encoder_decoder else None attention_mask = ( kwargs.pop("decoder_attention_mask", None) if self.config.is_encoder_decoder else attention_mask ) attention_mask_key = "decoder_attention_mask" if self.config.is_encoder_decoder else "attention_mask" position_ids_key = "decoder_position_ids" if self.config.is_encoder_decoder else "position_ids" if ( attention_mask is not None and kwargs.get(position_ids_key) is None and position_ids_key in set(inspect.signature(self.forward).parameters.keys()) ): position_ids = attention_mask.long().cumsum(-1) - 1 position_ids.masked_fill_(attention_mask == 0, 1) kwargs[position_ids_key] = position_ids # placed in kwargs for further processing (see below) # 5. Slice model inputs if it's an input that should have the same length as `input_ids` for model_input_name in ["position_ids", "token_type_ids", "decoder_position_ids"]: model_input = kwargs.get(model_input_name) if model_input is not None: if past_key_values is not None: current_input_length = ( model_inputs["inputs_embeds"].shape[1] if model_inputs.get("inputs_embeds") is not None else model_inputs[input_ids_key].shape[1] ) model_input = model_input[:, -current_input_length:] model_input = model_input.clone(memory_format=torch.contiguous_format) model_inputs[model_input_name] = model_input # 6. Create 4D attention mask is we are using a compilable cache (important for performant compiled forward # pass) if ( isinstance(past_key_values, Cache) and past_key_values.is_compileable and attention_mask is not None and attention_mask.ndim == 2 ): if not self.config.is_encoder_decoder and model_inputs["inputs_embeds"] is not None: batch_size, sequence_length, _ = model_inputs["inputs_embeds"].shape else: batch_size, sequence_length = model_inputs[input_ids_key].shape[:2] # Create the causal mask with fixed shape in advance, to reduce recompilations. If the function to create # the 4D causal mask exists, it should be present in the base model (XXXModel class) or in its decoder. base_model = getattr(self, self.base_model_prefix, self) decoder = base_model.get_decoder() if hasattr(base_model, "get_decoder") else None causal_mask_creation_function = getattr( base_model, "_prepare_4d_causal_attention_mask_with_cache_position", None ) if causal_mask_creation_function is None and decoder is not None: # it may be in the decoder causal_mask_creation_function = getattr( decoder, "_prepare_4d_causal_attention_mask_with_cache_position", None ) # If it's not defined, it means the model uses the new general mask API if causal_mask_creation_function is None: # can't be found token_type_ids = model_inputs.get("token_type_ids") position_ids = model_inputs.get(position_ids_key) # Some models may overwrite the general one causal_mask_creation_function = getattr(self, "create_masks_for_generate", create_masks_for_generate) attention_mask = causal_mask_creation_function( config=self.config, # we only need batch size, seq_length and dtype here - we don't care about the values of the embeddings input_embeds=torch.empty((batch_size, sequence_length), dtype=self.dtype), attention_mask=attention_mask, cache_position=cache_position, past_key_values=past_key_values, position_ids=position_ids, token_type_ids=token_type_ids, ) else: attention_mask = causal_mask_creation_function( attention_mask, sequence_length=sequence_length, target_length=past_key_values.get_max_cache_shape(), dtype=self.dtype, cache_position=cache_position, batch_size=batch_size, config=self.config, past_key_values=past_key_values, ) if attention_mask is not None: model_inputs[attention_mask_key] = attention_mask if encoder_attention_mask is not None: model_inputs["attention_mask"] = encoder_attention_mask # 7. Forward ALL kwargs that are uninitialized (e.g. `use_cache`). for key, value in kwargs.items(): if key not in model_inputs: model_inputs[key] = value # 8. Remove unexpected `generate` inputs (TODO @joao: fix trainer and examples) model_inputs.pop("labels", None) return model_inputs def _prepare_model_inputs( self, inputs: Optional[torch.Tensor] = None, bos_token_id: Optional[torch.Tensor] = 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 model_kwargs["inputs_embeds"] is None: model_kwargs.pop("inputs_embeds") elif 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 ) inputs, input_name = model_kwargs["inputs_embeds"], "inputs_embeds" 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 _maybe_initialize_input_ids_for_generation( self, inputs: Optional[torch.Tensor] = None, bos_token_id: Optional[torch.Tensor] = 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 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 if "inputs_embeds" in model_kwargs: return torch.ones((batch_size, 0), dtype=torch.long, device=self.device) if bos_token_id is None: raise ValueError("`bos_token_id` has to be defined when no `input_ids` are provided.") return torch.ones((batch_size, 1), dtype=torch.long, device=self.device) * bos_token_id def _prepare_attention_mask_for_generation( self, inputs_tensor: torch.Tensor, generation_config: GenerationConfig, model_kwargs: dict[str, Any], ) -> torch.LongTensor: pad_token_id = generation_config._pad_token_tensor eos_token_id = generation_config._eos_token_tensor # `input_ids` may be present in the model kwargs, instead of being the main input (e.g. multimodal model) if "input_ids" in model_kwargs and model_kwargs["input_ids"].shape[1] > 0: inputs_tensor = model_kwargs["input_ids"] # No information for attention mask inference -> return default attention mask default_attention_mask = torch.ones(inputs_tensor.shape[:2], dtype=torch.long, device=inputs_tensor.device) if pad_token_id is None: return default_attention_mask is_input_ids = len(inputs_tensor.shape) == 2 and inputs_tensor.dtype in [torch.int, torch.long] if not is_input_ids: return default_attention_mask is_pad_token_in_inputs = (pad_token_id is not None) and ( isin_mps_friendly(elements=inputs_tensor, test_elements=pad_token_id).any() ) is_pad_token_not_equal_to_eos_token_id = (eos_token_id is None) or ~( isin_mps_friendly(elements=eos_token_id, test_elements=pad_token_id).any() ) can_infer_attention_mask = is_pad_token_in_inputs * is_pad_token_not_equal_to_eos_token_id attention_mask_from_padding = inputs_tensor.ne(pad_token_id).long() attention_mask = ( attention_mask_from_padding * can_infer_attention_mask + default_attention_mask * ~can_infer_attention_mask ) return attention_mask def _prepare_encoder_decoder_kwargs_for_generation( self, inputs_tensor: torch.Tensor, model_kwargs, model_input_name: Optional[str], generation_config: GenerationConfig, ) -> 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(self, "hf_device_map"): if hasattr(encoder, "_hf_hook"): encoder._hf_hook.io_same_device = True else: add_hook_to_module(encoder, AlignDevicesHook(io_same_device=True)) # 2. Prepare encoder args and encoder kwargs from model kwargs and generation config. 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 } encoder_kwargs["output_attentions"] = generation_config.output_attentions encoder_kwargs["output_hidden_states"] = generation_config.output_hidden_states # 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) # type: ignore 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: torch.Tensor, device: Optional[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. `decoder_start_token_id` must have shape (batch_size, 1) if device is None: device = self.device if decoder_start_token_id.ndim == 1: if decoder_start_token_id.shape[0] != batch_size: raise ValueError( f"`decoder_start_token_id` expected to have length {batch_size} but got {decoder_start_token_id.shape[0]}" ) decoder_start_token_id = decoder_start_token_id.view(-1, 1) else: decoder_start_token_id = ( torch.ones((batch_size, 1), dtype=torch.long, device=device) * decoder_start_token_id ) # 3. Encoder-decoder models expect the `decoder_input_ids` to start with a special token. Let's ensure that. # no user input -> use decoder_start_token_id as decoder_input_ids if decoder_input_ids is None: decoder_input_ids = decoder_start_token_id # exception: Donut checkpoints have task-specific decoder starts and don't expect a BOS token. Note that the # original checkpoints can't be detected through `self.__class__.__name__.lower()`, needing custom logic. # See: https://github.com/huggingface/transformers/pull/31470 elif "donut" in self.__class__.__name__.lower() or ( self.config.model_type == "vision-encoder-decoder" and "donut" in self.config.encoder.model_type.lower() ): pass elif self.config.model_type == "whisper": 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[:, 0]).all().item(): decoder_input_ids = torch.cat([decoder_start_token_id, 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 @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, ...]""" # Do not call torch.repeat_interleave if expand_size is 1 because it clones # the input tensor and thus requires more memory although no change is applied if expand_size == 1: return input_ids, model_kwargs def _expand_dict_for_generation(dict_to_expand): for key in dict_to_expand: if ( key != "cache_position" and 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 _update_model_kwargs_for_generation( self, outputs: ModelOutput, model_kwargs: dict[str, Any], is_encoder_decoder: bool = False, num_new_tokens: int = 1, ) -> dict[str, Any]: # update past_key_values keeping its naming used in model code for possible_cache_name in ALL_CACHE_NAMES: if possible_cache_name in outputs: # TODO (joao): remove output/input mismatch when these old models (xlnet, reformer) are deprecated if possible_cache_name in ("past_buckets_states", "mems"): cache_name = "past_key_values" else: cache_name = possible_cache_name model_kwargs[cache_name] = getattr(outputs, possible_cache_name) break # 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, ) if model_kwargs.get("use_cache", True): model_kwargs["cache_position"] = model_kwargs["cache_position"][-1:] + num_new_tokens else: past_positions = model_kwargs.pop("cache_position") new_positions = torch.arange( past_positions[-1] + 1, past_positions[-1] + num_new_tokens + 1, dtype=past_positions.dtype ).to(past_positions.device) model_kwargs["cache_position"] = torch.cat((past_positions, new_positions)) return model_kwargs def _get_candidate_generator( self, generation_config: GenerationConfig, input_ids: torch.LongTensor, inputs_tensor: torch.Tensor, logits_processor: LogitsProcessorList, model_kwargs: dict[str, Any], assistant_model: Optional["PreTrainedModel"] = None, target_tokenizer: Optional["PreTrainedTokenizerBase"] = None, assistant_tokenizer: Optional["PreTrainedTokenizerBase"] = None, ) -> CandidateGenerator: """ Returns the candidate generator to be used in `assisted_generation` """ different_tokenizers = all(v is not None for v in (assistant_model, target_tokenizer, assistant_tokenizer)) if generation_config.assistant_early_exit is not None: candidate_generator = EarlyExitCandidateGenerator( input_ids=input_ids, assistant_model=self, generation_config=generation_config, model_kwargs=model_kwargs, inputs_tensor=inputs_tensor, logits_processor=logits_processor, ) elif generation_config.prompt_lookup_num_tokens is not None: candidate_generator = PromptLookupCandidateGenerator( eos_token_id=generation_config._eos_token_tensor, num_output_tokens=generation_config.prompt_lookup_num_tokens, max_matching_ngram_size=generation_config.max_matching_ngram_size or 2, max_length=generation_config.max_length, logits_processor=logits_processor, vocab_size=self.config.get_text_config().vocab_size, ) elif different_tokenizers: if generation_config.do_sample is True: atm_translator = AssistantVocabTranslatorCache.get_translator( target_tokenizer, assistant_tokenizer, self.config.get_text_config().vocab_size, assistant_model=assistant_model, assistant_prune_lm_head=True, # prune LM head of assistant model ) # Since we prune the LM head, we cannot use the repetition penalty on the assistant model due to mismatches between token ids and logits index assistant_model.generation_config.repetition_penalty = None candidate_generator = UniversalSpeculativeDecodingGenerator( input_ids=input_ids, assistant_model=assistant_model, generation_config=generation_config, model_kwargs=model_kwargs, inputs_tensor=inputs_tensor, logits_processor=logits_processor, target_tokenizer=target_tokenizer, assistant_tokenizer=assistant_tokenizer, atm_translator=atm_translator, ) elif generation_config.do_sample is False: candidate_generator = AssistedCandidateGeneratorDifferentTokenizers( input_ids=input_ids, assistant_model=assistant_model, generation_config=generation_config, model_kwargs=model_kwargs, inputs_tensor=inputs_tensor, logits_processor=logits_processor, target_tokenizer=target_tokenizer, assistant_tokenizer=assistant_tokenizer, ) else: raise ValueError( f"Invalid value for `do_sample`: expected a boolean, got {type(generation_config.do_sample).__name__}" ) else: candidate_generator = AssistedCandidateGenerator( input_ids=input_ids, assistant_model=assistant_model, generation_config=generation_config, model_kwargs=model_kwargs, inputs_tensor=inputs_tensor, logits_processor=logits_processor, ) return candidate_generator def _get_logits_processor( self, generation_config: GenerationConfig, input_ids_seq_length: Optional[int] = None, encoder_input_ids: Optional[torch.LongTensor] = None, prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], list[int]]] = None, logits_processor: Optional[LogitsProcessorList] = None, device: Optional[str] = None, model_kwargs: Optional[dict[str, Any]] = None, negative_prompt_ids: Optional[torch.Tensor] = None, negative_prompt_attention_mask: Optional[torch.Tensor] = None, ) -> 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 logits_processor is None: logits_processor = [] if generation_config.guidance_scale is not None and generation_config.guidance_scale != 1: processors.append( UnbatchedClassifierFreeGuidanceLogitsProcessor( generation_config.guidance_scale, self, unconditional_ids=negative_prompt_ids, unconditional_attention_mask=negative_prompt_attention_mask, use_cache=generation_config.use_cache, ) ) if generation_config.sequence_bias is not None: processors.append(SequenceBiasLogitsProcessor(sequence_bias=generation_config.sequence_bias)) if ( generation_config.encoder_repetition_penalty is not None and generation_config.encoder_repetition_penalty != 1.0 ): if len(encoder_input_ids.shape) == 2: processors.append( EncoderRepetitionPenaltyLogitsProcessor( penalty=generation_config.encoder_repetition_penalty, encoder_input_ids=encoder_input_ids, ) ) else: warnings.warn( "Passing `encoder_repetition_penalty` requires some form of `input_ids` to be passed to " "`generate`, ignoring the argument.", UserWarning, ) 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 len(encoder_input_ids.shape) == 2: processors.append( EncoderNoRepeatNGramLogitsProcessor( generation_config.encoder_no_repeat_ngram_size, encoder_input_ids, ) ) else: warnings.warn( "Passing `encoder_no_repeat_ngram_size` requires some form of `input_ids` to be passed to " "`generate`, ignoring the argument.", UserWarning, ) if generation_config.bad_words_ids is not None: processors.append( NoBadWordsLogitsProcessor( generation_config.bad_words_ids, generation_config._eos_token_tensor, ) ) if ( generation_config.min_length is not None and getattr(generation_config, "_eos_token_tensor", None) is not None and generation_config.min_length > 0 ): processors.append( MinLengthLogitsProcessor( generation_config.min_length, generation_config._eos_token_tensor, device=device, ) ) if ( generation_config.min_new_tokens is not None and getattr(generation_config, "_eos_token_tensor", None) 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_tensor, device=device, ) ) if prefix_allowed_tokens_fn is not None: processors.append( PrefixConstrainedLogitsProcessor( prefix_allowed_tokens_fn, generation_config.num_beams, ) ) 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, device=device, ) ) 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_tensor, input_ids_seq_length, ) ) if generation_config.suppress_tokens is not None: processors.append( SuppressTokensLogitsProcessor( generation_config.suppress_tokens, device=device, ) ) 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 ) processors.append( SuppressTokensAtBeginLogitsProcessor( generation_config.begin_suppress_tokens, begin_index, device=device, ) ) # TODO (joao): find a strategy to specify the order of the processors processors = self._merge_criteria_processor_list(processors, logits_processor) # Processors previously known as `LogitsWarpers`, only applied with sampling strategies if generation_config.do_sample: # In beam methods, we need to keep at least one non-eos token to explore continuations that might have a # better score (i.e. keep len(list(generation_config._eos_token_tensor)) + 1) if generation_config.num_beams > 1: if isinstance(generation_config._eos_token_tensor, list): min_tokens_to_keep = len(generation_config._eos_token_tensor) + 1 elif isinstance(generation_config._eos_token_tensor, torch.Tensor): min_tokens_to_keep = generation_config._eos_token_tensor.shape[0] + 1 else: min_tokens_to_keep = 2 else: min_tokens_to_keep = 1 # 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: processors.append(TemperatureLogitsWarper(generation_config.temperature)) if generation_config.top_k is not None and generation_config.top_k != 0: processors.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: processors.append( TopPLogitsWarper(top_p=generation_config.top_p, min_tokens_to_keep=min_tokens_to_keep) ) if generation_config.min_p is not None: # Applied after temperature scaling (see https://github.com/ggerganov/llama.cpp/pull/3841#issuecomment-2073826084) processors.append( MinPLogitsWarper(min_p=generation_config.min_p, min_tokens_to_keep=min_tokens_to_keep) ) if generation_config.typical_p is not None and generation_config.typical_p < 1.0: processors.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: processors.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: processors.append( EtaLogitsWarper( epsilon=generation_config.eta_cutoff, min_tokens_to_keep=min_tokens_to_keep, device=device ) ) # Watermarking should be after all logits processing is finished (see #34630) if generation_config.watermarking_config is not None: processors.append( generation_config.watermarking_config.construct_processor( self.config.get_text_config().vocab_size, device ) ) # `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], tokenizer: Optional["PreTrainedTokenizerBase"] = None, ) -> 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)) if generation_config.stop_strings is not None: if tokenizer is None: raise ValueError( "There are one or more stop strings, either in the arguments to `generate` or in the " "model's generation config, but we could not locate a tokenizer. When generating with " "stop strings, you must pass the model's tokenizer to the `tokenizer` argument of `generate`." ) criteria.append(StopStringCriteria(stop_strings=generation_config.stop_strings, tokenizer=tokenizer)) if generation_config._eos_token_tensor is not None: criteria.append(EosTokenCriteria(eos_token_id=generation_config._eos_token_tensor)) if ( generation_config.is_assistant and generation_config.assistant_confidence_threshold is not None and generation_config.assistant_confidence_threshold > 0 ): criteria.append( ConfidenceCriteria(assistant_confidence_threshold=generation_config.assistant_confidence_threshold) ) 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]: """ Merge user-defined processors/criteria with the ones instantiated inside `generate`. In case the same processor/criteria is present on both lists, use the user-defined one. (Note: up to v4.49.0, this function threw an exception is the same logit processor was found twice.) """ if len(custom_list) == 0: return default_list final_list = type(default_list)() for default in default_list: using_custom = False for custom in custom_list: if type(custom) is type(default): object_type = "stopping criteria" if isinstance(custom, StoppingCriteria) else "logits processor" logger.warning_once( f"A custom {object_type} of type {type(custom)} has been passed to `.generate()`, but it " f"was also created in `.generate()`, given its parameterization. The custom {type(custom)} " f"will take precedence. Please check the docstring of {type(custom)} to see related " "`.generate()` flags." ) final_list.append(custom) using_custom = True break if not using_custom: final_list.append(default) for custom in custom_list: if custom not in final_list: final_list.append(custom) return final_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 quickly 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 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*): 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("openai-community/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 | log probability | 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 1: 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`. >>> # Tip 2: the output length does NOT include the input length >>> output_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.get_text_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.get_text_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_generation_mode(self, generation_mode, generation_config, generation_mode_kwargs): if generation_mode == GenerationMode.BEAM_SEARCH and "streamer" in generation_mode_kwargs: raise ValueError( "`streamer` cannot be used with beam search (yet!). Make sure that `num_beams` is set to 1." ) if generation_mode == GenerationMode.ASSISTED_GENERATION: 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 self._is_stateful: # In assisted generation we need the ability to confirm whether the model would pick certain tokens, # which is not possible with stateful models (they can't reset to a previous subset of generated text) raise ValueError( f"assisted generation is not supported with stateful models, such as {self.__class__.__name__}" ) if (assistant_model := generation_mode_kwargs.get("assistant_model")) is not None: if self.config.is_encoder_decoder and not assistant_model.config.is_encoder_decoder: attributes_to_check = ["encoder_attention_heads", "encoder_ffn_dim", "encoder_layers"] attributes_to_check = [attr for attr in dir(assistant_model.config) if attr in attributes_to_check] are_equal = all( getattr(self.config, attr) == getattr(assistant_model.config, attr) for attr in attributes_to_check ) if not are_equal: raise ValueError( "The main model and the assistant don't have compatible encoder-dependent input shapes. " "Ensure you load the assistant with the correct encoder-decoder class, e.g. `AutoModelForSpeechSeq2Seq` for Whisper." ) doc_reference = ( "(see https://huggingface.co/docs/transformers/en/generation_strategies#universal-assisted-decoding)" ) if self.config.get_text_config().vocab_size == assistant_model.config.get_text_config().vocab_size: if "assistant_tokenizer" in generation_mode_kwargs: raise ValueError( f"`assistant_tokenizer` is not required when the main and assistant models use the same tokenizer. Please omit `assistant_tokenizer` from `generate()` {doc_reference}." ) else: if "tokenizer" not in generation_mode_kwargs or "assistant_tokenizer" not in generation_mode_kwargs: raise ValueError( f"The main and assistant models have different tokenizers. Please provide `tokenizer` and `assistant_tokenizer` to `generate()` {doc_reference}." ) 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} # TransformersKwargs are model-agnostic attention and generation arguments such as 'output_attentions' for key, value in model_kwargs.items(): if value is not None and key not in model_args and key not in TransformersKwargs.__optional_keys__: 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)" ) def _validate_generated_length(self, generation_config, input_ids_length, has_default_max_length): """Performs validation related to the resulting generated length""" # 1. Max length warnings related to poor parameterization 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. We recommend setting `max_new_tokens` to control the maximum length of the " "generation.", UserWarning, ) if input_ids_length >= generation_config.max_length: input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids" raise ValueError( f"Input length of {input_ids_string} is {input_ids_length}, but `max_length` is set to" f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider" " increasing `max_length` or, better yet, setting `max_new_tokens`." ) # 2. Min length warnings due to unfeasible parameter combinations min_length_error_suffix = ( " Generation will stop at the defined maximum length. You should decrease the minimum length and/or " "increase the maximum length." ) if has_default_max_length: min_length_error_suffix += ( f" Note that `max_length` is set to {generation_config.max_length}, its default value." ) if generation_config.min_length is not None and generation_config.min_length > generation_config.max_length: warnings.warn( f"Unfeasible length constraints: `min_length` ({generation_config.min_length}) is larger than" f" the maximum possible length ({generation_config.max_length})." + min_length_error_suffix, UserWarning, ) if generation_config.min_new_tokens is not None: min_length = generation_config.min_new_tokens + input_ids_length if min_length > generation_config.max_length: warnings.warn( f"Unfeasible length constraints: `min_new_tokens` ({generation_config.min_new_tokens}), when " f"added to the prompt length ({input_ids_length}), is larger than" f" the maximum possible length ({generation_config.max_length})." + min_length_error_suffix, UserWarning, ) def _prepare_generated_length( self, generation_config, has_default_max_length, has_default_min_length, model_input_name, input_ids_length, inputs_tensor, ): """Prepared max and min length in generation configs to avoid clashes between similar attributes""" if generation_config.max_new_tokens is not None: if not has_default_max_length and generation_config.max_length is not None: 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_length # if both `inputs_embeds` and `input_ids` are passed, we do not correct the length # otherwise we need total length [inputs-embeds-len + new-tokens-len] to not go beyond indicated `max_length`` elif ( model_input_name == "inputs_embeds" and input_ids_length != inputs_tensor.shape[1] and not self.config.is_encoder_decoder ): generation_config.max_length -= inputs_tensor.shape[1] elif has_default_max_length: # by default let's always generate 20 new tokens if generation_config.max_length == GenerationConfig().max_length: generation_config.max_length = generation_config.max_length + input_ids_length max_position_embeddings = getattr(self.config, "max_position_embeddings", None) if max_position_embeddings is not None: generation_config.max_length = min(generation_config.max_length, max_position_embeddings) # same for min length if generation_config.min_new_tokens is not None: if not has_default_min_length: logger.warning( f"Both `min_new_tokens` (={generation_config.min_new_tokens}) and `min_length`(=" f"{generation_config.min_length}) seem to have been set. `min_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.min_length = generation_config.min_new_tokens + input_ids_length elif ( model_input_name == "inputs_embeds" and input_ids_length != inputs_tensor.shape[1] and not self.config.is_encoder_decoder ): generation_config.min_length = max(generation_config.min_length - inputs_tensor.shape[1], 0) return generation_config def _prepare_generation_config( self, generation_config: Optional[GenerationConfig], use_model_defaults: Optional[bool] = None, **kwargs: Any, ) -> tuple[GenerationConfig, dict]: """ Prepares the base generation config, then applies any generation configuration options from kwargs. This function handles retrocompatibility with respect to configuration files. """ # parameterization priority: # kwargs > non-global default values in `generation_config` > `model.generation_config` > GenerationConfig() # TODO (joao): per-model generation config classes. using_model_generation_config = False if generation_config is None: # legacy: users may modify the model configuration to control generation. To trigger this legacy behavior, # the following conditions must be met # 1) the generation config must have been created from the model config (`_from_model_config` field); # 2) the generation config must have seen no modification since its creation (the hash is the same); # 3) there are non-default generation parameters in the model config. # 4) the user must have set new generation parameters in the model config. if ( self.generation_config._from_model_config # 1) and self.generation_config._original_object_hash == hash(self.generation_config) # 2) and len(self.config._get_non_default_generation_parameters()) > 0 # 3) ): new_generation_config = GenerationConfig.from_model_config(self.config) if new_generation_config != self.generation_config: # 4) warnings.warn( "You have modified the pretrained model configuration to control generation. This is a" " deprecated strategy to control generation and will be removed in v5." " Please use and modify the model generation configuration (see" " https://huggingface.co/docs/transformers/generation_strategies#default-text-generation-configuration )", UserWarning, ) self.generation_config = new_generation_config generation_config = self.generation_config using_model_generation_config = True # Related to #40039: prior to this PR, models with sliding window attention were forced to have # `cache_implementation="hybrid"` (the static sliding window cache). For these models, we now want to use # the dynamic sliding window cache by default, so we UNSET `cache_implementation` if it is a default value. # (if we're inside this branch, then it is because we're using default values from the Hub) if generation_config.cache_implementation == "hybrid": generation_config.cache_implementation = None # `torch.export.export` usually raises an exception if it is called # with ``strict=True``. deepcopy can only be processed if ``strict=False``. generation_config = copy.deepcopy(generation_config) if not using_model_generation_config: # If `generation_config` is provided: # - `use_model_defaults`: let's fallback ALL default values to the model's generation config # - otherwise: legacy behavior, let's just make sure we have the tokens defined model_base_version = version.parse(version.parse(self.generation_config.transformers_version).base_version) if use_model_defaults is True or ( use_model_defaults is None and model_base_version >= version.parse("4.50.0") ): modified_values = {} global_default_generation_config = GenerationConfig() model_generation_config = self.generation_config # we iterate over the model's generation config: it may hold custom keys, which we'll want to copy for key, model_gen_config_value in model_generation_config.__dict__.items(): if key.startswith("_") or key == "transformers_version": # metadata continue # Don't set `cache_implementation = 'hybrid'` from the model defaults, see #40135 if key == "cache_implementation" and model_generation_config.cache_implementation == "hybrid": continue global_default_value = getattr(global_default_generation_config, key, None) custom_gen_config_value = getattr(generation_config, key, None) if ( custom_gen_config_value == global_default_value and model_gen_config_value != global_default_value ): modified_values[key] = model_gen_config_value setattr(generation_config, key, model_gen_config_value) # edge case: we may set `temperature=0.0` and `do_sample=False`, but the model defaults to # `do_sample=True` if generation_config.temperature == 0.0: generation_config.do_sample = False if use_model_defaults is None and len(modified_values) > 0: logger.warning_once( f"`generation_config` default values have been modified to match model-specific defaults: " f"{modified_values}. If this is not desired, please set these values explicitly." ) else: if generation_config.bos_token_id is None: generation_config.bos_token_id = self.generation_config.bos_token_id if generation_config.eos_token_id is None: generation_config.eos_token_id = self.generation_config.eos_token_id if generation_config.pad_token_id is None: generation_config.pad_token_id = self.generation_config.pad_token_id if generation_config.decoder_start_token_id is None: generation_config.decoder_start_token_id = self.generation_config.decoder_start_token_id # Finally, apply any passed kwargs model_kwargs = generation_config.update(**kwargs) # And keep in model_kwargs variable output controls output_attentions = generation_config.output_attentions output_hidden_states = generation_config.output_hidden_states model_kwargs.update({"output_attentions": output_attentions} if output_attentions else {}) model_kwargs.update({"output_hidden_states": output_hidden_states} if output_hidden_states else {}) return generation_config, model_kwargs def _get_initial_cache_position(self, seq_length, device, model_kwargs): """Calculates `cache_position` for the pre-fill stage based on `input_ids` and optionally past length""" # `torch.compile`-friendly `torch.arange` from a shape -- the lines below are equivalent to `torch.arange` if "cache_position" in model_kwargs and model_kwargs["cache_position"] is not None: return model_kwargs if "inputs_embeds" in model_kwargs and not self.config.is_encoder_decoder: cache_position = torch.ones_like(model_kwargs["inputs_embeds"][0, :, 0], dtype=torch.int64).cumsum(0) - 1 elif "decoder_inputs_embeds" in model_kwargs and self.config.is_encoder_decoder: cache_position = ( torch.ones_like(model_kwargs["decoder_inputs_embeds"][0, :, 0], dtype=torch.int64).cumsum(0) - 1 ) else: cache_position = torch.ones(seq_length, dtype=torch.int64, device=device).cumsum(0) - 1 past_length = 0 if model_kwargs.get("past_key_values") is not None: cache = model_kwargs["past_key_values"] past_length = 0 # Support for BC tuple cache format if isinstance(cache, tuple): past_length = cache[0][0].shape[2] elif hasattr(cache, "get_seq_length"): past_length = cache.get_seq_length() cache_position = cache_position[past_length:] model_kwargs["cache_position"] = cache_position return model_kwargs def _get_cache(self, cache_implementation: str, batch_size: int, max_cache_len: int, model_kwargs) -> Cache: """ Sets a cache for `generate`, that will persist across calls. A new cache will only be initialized a new `generate` call requires a larger cache or uses a different batch size. Returns the resulting cache object. """ requires_cross_attention_cache = ( self.config.is_encoder_decoder or model_kwargs.get("encoder_outputs") is not None ) offload_cache = "offloaded" in cache_implementation if hasattr(self, "_cache"): cache_to_check = self._cache.self_attention_cache if requires_cross_attention_cache else self._cache need_new_cache = ( not hasattr(self, "_cache") or cache_to_check.offloading != offload_cache or cache_to_check.max_batch_size != batch_size or cache_to_check.max_cache_len < max_cache_len ) if requires_cross_attention_cache and hasattr(self, "_cache"): need_new_cache = ( need_new_cache or self._cache.cross_attention_cache.max_cache_len != model_kwargs["encoder_outputs"][0].shape[1] ) if need_new_cache: self_attention_cache_kwargs = { "config": self.config.get_text_config(decoder=True), "max_cache_len": max_cache_len, "offloading": offload_cache, } self._cache = StaticCache(**self_attention_cache_kwargs) if requires_cross_attention_cache: cross_attention_cache_kwargs = { "config": self.config.get_text_config(decoder=True), "max_cache_len": model_kwargs["encoder_outputs"][0].shape[1], "offloading": offload_cache, } self._cache = EncoderDecoderCache(self._cache, StaticCache(**cross_attention_cache_kwargs)) else: self._cache.reset() return self._cache @classmethod def _supports_default_dynamic_cache(cls) -> bool: """ Return `True` if current model can use a `DynamicCache` instance when initializing the `past_key_values`. This adds exception for some models like `Mamba` models which use their own caches and do not need to initialize the Cache in advance in order to save memory (because no back and forth `to_legacy_cache` and `from_legacy_cache` will be performed for mamba-based models). """ # NOTE: remove xlnet/reformer when the models are deprecated, non-standard model architecture/cache name return not cls._is_stateful and all( special_model_name not in cls.__name__.lower() for special_model_name in [ "reformer", "minimax", "xlnet", "lfm2", "lfm2-vl", ] ) def _prepare_cache_for_generation( self, generation_config: GenerationConfig, model_kwargs: dict, generation_mode: GenerationMode, batch_size: int, max_cache_length: int, ) -> bool: """ Prepares the cache for generation (if applicable), given `generate`'s parameterization. If a cache is instantiated, writes it to `model_kwargs`, under the name expected by the model. """ is_hybrid_cache = any(class_name in self.__class__.__name__.lower() for class_name in ["mamba", "falconh1"]) cache_name = "past_key_values" if not is_hybrid_cache else "cache_params" requires_cross_attention_cache = ( self.config.is_encoder_decoder or model_kwargs.get("encoder_outputs") is not None ) # Quick escape route 1: if the user specifies a cache, we only need to: # a) check for conflicting `generate` arguments # b) convert to the new cache format (if the user passes a legacy cache and model supports it) user_defined_cache = model_kwargs.get(cache_name) if user_defined_cache is not None: if generation_config.cache_implementation is not None: raise ValueError( f"Passing both `cache_implementation` (used to initialize certain caches) and `{cache_name}` (a " "Cache object) is unsupported. Please use only one of the two." ) if isinstance(user_defined_cache, tuple) and self._supports_default_dynamic_cache(): logger.warning_once( "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.58.0. " "You should pass an instance of `Cache` instead." ) model_kwargs[cache_name] = ( DynamicCache.from_legacy_cache(user_defined_cache) if not requires_cross_attention_cache else EncoderDecoderCache.from_legacy_cache(user_defined_cache) ) return # Quick escape route 2: if the user specifies no cache is to be used. (conflicting arguments are handled in # `generation_config.validate()`) if generation_config.use_cache is False: return # Quick escape route 3: model that only supports legacy caches or models that supply it in # `prepare_inputs_for_generation` (mamba, zamba, ...) if not self._supports_default_dynamic_cache(): if generation_config.cache_implementation is not None: logger.warning_once( "This model does not support `Cache` instances. `cache_implementation` (set to " f"{generation_config.cache_implementation}) will be ignored.", ) return # Otherwise we NEED to prepare a cache, based on `generation_config.cache_implementation` # TODO(joao): support static caches in assisted generation. assisted generation needs to roll back caches, # which is only supported in dynamic caches atm if ( generation_mode == GenerationMode.ASSISTED_GENERATION and generation_config.cache_implementation is not None ): logger.warning_once( "An assistant model is provided, using a dynamic cache instead of a cache of type=" f"'{generation_config.cache_implementation}'." ) generation_config.cache_implementation = None # Assisted decoding and contrastive search require cache rollback, which is incompatible with sliding layers. # To handle this, we skip passing the model config to DynamicCache (forcing a full-layer cache). # The "dynamic_full" option is a shortcut for generate() users to avoid sliding layers on their own. if ( generation_mode in (GenerationMode.ASSISTED_GENERATION, GenerationMode.CONTRASTIVE_SEARCH) or generation_config.cache_implementation == "dynamic_full" ): dynamic_cache_kwargs = {} else: dynamic_cache_kwargs = {"config": self.config.get_text_config(decoder=True)} if generation_config.cache_implementation is not None: if generation_config.cache_implementation in ALL_STATIC_CACHE_IMPLEMENTATIONS: if generation_config.cache_implementation in DEPRECATED_STATIC_CACHE_IMPLEMENTATIONS: logger.warning_once( f"Using `cache_implementation='{generation_config.cache_implementation}' is deprecated. " f"Please only use one of {STATIC_CACHE_IMPLEMENTATIONS}, and the layer structure will be " "inferred automatically." ) model_kwargs[cache_name] = self._get_cache( cache_implementation=generation_config.cache_implementation, batch_size=max(generation_config.num_beams, generation_config.num_return_sequences) * batch_size, max_cache_len=max_cache_length, model_kwargs=model_kwargs, ) elif generation_config.cache_implementation == "quantized": if self.config.is_encoder_decoder or not self._supports_default_dynamic_cache(): raise ValueError( "This model does not support the quantized cache. If you want your model to support quantized " "cache, please open an issue and tag @zucchini-nlp." ) cache_config = generation_config.cache_config if generation_config.cache_config is not None else {} # Add the config if it was not provided, as it's a required argument if "config" not in cache_config: cache_config["config"] = self.config.get_text_config() # Pop the backend from the config (defaults to quanto if not defined) backend = cache_config.pop("backend", "quanto") if backend == "quanto" and not is_optimum_quanto_available(): raise ImportError( "You need to install optimum-quanto in order to use KV cache quantization with optimum-quanto " "backend. Please install it via with `pip install optimum-quanto`" ) elif backend == "HQQ" and not is_hqq_available(): raise ImportError( "You need to install `HQQ` in order to use KV cache quantization with HQQ backend. " "Please install it via with `pip install hqq`" ) model_kwargs[cache_name] = QuantizedCache(backend=backend, **cache_config) elif generation_config.cache_implementation == "offloaded": model_kwargs[cache_name] = DynamicCache(**dynamic_cache_kwargs, offloading=True) elif "dynamic" in generation_config.cache_implementation: model_kwargs[cache_name] = DynamicCache(**dynamic_cache_kwargs) # Use DynamicCache instance by default. This will avoid back and forth from legacy format that # keeps copying the cache thus using much more memory # TODO (joao): remove this `else` when we remove the last traces of the legacy cache format (v4.58.0, search # for `instance(past_key_values, Cache)` as well). In general, if `cache_implementation` is unset, cache # initialization should happen inside the model at prefill time. else: model_kwargs[cache_name] = DynamicCache(**dynamic_cache_kwargs) # TODO (joao): this logic is incomplete, e.g. `offloaded` should apply to both caches. Refactor this function # to correctly pass parameterization to both caches. if requires_cross_attention_cache and not isinstance(model_kwargs[cache_name], EncoderDecoderCache): model_kwargs[cache_name] = EncoderDecoderCache( model_kwargs[cache_name], # self-attention cache DynamicCache(**dynamic_cache_kwargs), # cross-attention cache ) def _supports_logits_to_keep(self) -> bool: """ Return True if the current model supports the keyword argument `logits_to_keep` in forward() to save memory. Checking it in this way allows to avoid using a new model attribute. """ return "logits_to_keep" in set(inspect.signature(self.forward).parameters.keys()) def _prepare_special_tokens( self, generation_config: GenerationConfig, kwargs_has_attention_mask: Optional[bool] = None, device: Optional[Union[torch.device, str]] = None, ): """ Prepares the special tokens for generation, overwriting the generation config with their processed versions converted to tensor. Note that `generation_config` is changed in place and stops being serializable after this method is called. That is no problem if called within `generate` (`generation_config` is a local copy that doesn't leave the function). However, if called outside `generate`, consider creating a copy of `generation_config` first. """ # Convert special tokens to tensors def _tensor_or_none(token, device=None): if token is None: return token device = device if device is not None else self.device if isinstance(token, torch.Tensor): return token.to(device) return torch.tensor(token, device=device, dtype=torch.long) bos_token_tensor = _tensor_or_none(generation_config.bos_token_id, device=device) eos_token_tensor = _tensor_or_none(generation_config.eos_token_id, device=device) pad_token_tensor = _tensor_or_none(generation_config.pad_token_id, device=device) decoder_start_token_tensor = _tensor_or_none(generation_config.decoder_start_token_id, device=device) # for BC we also try to get `decoder_start_token_id` or `bos_token_id` (#30892) if self.config.is_encoder_decoder: decoder_start_token_tensor = ( decoder_start_token_tensor if decoder_start_token_tensor is not None else bos_token_tensor ) # We can have more than one eos token. Always treat it as a 1D tensor (when it exists). if eos_token_tensor is not None and eos_token_tensor.ndim == 0: eos_token_tensor = eos_token_tensor.unsqueeze(0) # Set pad token if unset (and there are conditions to do so) if pad_token_tensor is None and eos_token_tensor is not None: if kwargs_has_attention_mask is not None and not kwargs_has_attention_mask: 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." ) pad_token_tensor = eos_token_tensor[0] logger.warning(f"Setting `pad_token_id` to `eos_token_id`:{pad_token_tensor} for open-end generation.") # Sanity checks/warnings if self.config.is_encoder_decoder and decoder_start_token_tensor is None: raise ValueError( "`decoder_start_token_id` or `bos_token_id` has to be defined for encoder-decoder generation." ) if ( eos_token_tensor is not None and isin_mps_friendly(elements=eos_token_tensor, test_elements=pad_token_tensor).any() ): if kwargs_has_attention_mask is not None and not kwargs_has_attention_mask: logger.warning_once( "The attention mask is not set and cannot be inferred from input because pad token is same as " "eos token. As a consequence, you may observe unexpected behavior. Please pass your input's " "`attention_mask` to obtain reliable results." ) if eos_token_tensor is not None and ( torch.is_floating_point(eos_token_tensor) or (eos_token_tensor < 0).any() ): logger.warning( f"`eos_token_id` should consist of positive integers, but is {eos_token_tensor}. Your generation " "will not stop until the maximum length is reached. Depending on other flags, it may even crash." ) # Update generation config with the updated special tokens tensors # NOTE: this must be written into a different attribute name than the one holding the original special tokens # (in their non-tensor form), in order to enable end-to-end compilation. See # https://pytorch.org/docs/stable/torch.compiler_cudagraph_trees.html#limitations generation_config._bos_token_tensor = bos_token_tensor generation_config._eos_token_tensor = eos_token_tensor generation_config._pad_token_tensor = pad_token_tensor generation_config._decoder_start_token_tensor = decoder_start_token_tensor def _valid_auto_compile_criteria(self, model_kwargs: dict[str, Any], generation_config: GenerationConfig) -> bool: """ Determines whether to trigger auto-compilation of the model's forward pass at generation time. """ # Override: honor `disable_compile` flag if generation_config.disable_compile: return False # Base logic valid_hardware = self.device.type == "cuda" or bool( generation_config.compile_config is not None and generation_config.compile_config._compile_all_devices ) using_compilable_cache = ( isinstance(model_kwargs.get("past_key_values"), Cache) and model_kwargs["past_key_values"].is_compileable ) can_compile = valid_hardware and using_compilable_cache # Exception 1: Some quantization methods do not support compilation if getattr(self, "hf_quantizer", None) is not None: can_compile &= self.hf_quantizer.is_compileable if hasattr(self, "hf_device_map"): all_model_devices = set(self.hf_device_map.values()) # Exception 2: Don't compile if the model is using CPU offload (as of April 2025, this results in a crash) has_cpu_offload = "cpu" in all_model_devices and len(all_model_devices) > 1 can_compile &= not has_cpu_offload # Exception 3: Disk offload is not supported for compilation has_disk_offload = "disk" in all_model_devices can_compile &= not has_disk_offload # Finally: if the user has manually specified compilation options, but compilation is not possible, let's warn # them if generation_config.compile_config is not None and not can_compile: logger.warning_once( "You have set `compile_config`, but we are unable to meet the criteria for compilation. Compilation " "will be skipped." ) return can_compile def _get_deprecated_gen_repo( self, generation_mode: GenerationMode, trust_remote_code: bool, custom_generate: Optional[str] = None, ) -> Optional[str]: """ Returns the Hub repo for a deprecated generation mode, if any. """ if custom_generate is not None or "/" not in (repo := GENERATION_MODES_MAPPING[generation_mode]): return None logger.warning_once( f"{generation_mode.name.replace('_', ' ').title()} was moved to a `custom_generate` repo: https://hf.co/{repo}. " f"To prevent loss of backward compatibility, add `custom_generate='{repo}'` " "to your `generate` call before v4.62.0." ) if not trust_remote_code: raise ValueError( f"{generation_mode.name.replace('_', ' ').title()} requires `trust_remote_code=True` in your `generate` call, " f"since it loads https://hf.co/{repo}." ) return repo def _extract_generation_mode_kwargs( self, custom_generate, kwargs, synced_gpus, assistant_model, streamer, ) -> dict[str, Any]: """ Extracts and returns the generation mode related keyword arguments from the provided kwargs. """ generation_mode_kwargs = { "tokenizer": kwargs.pop("tokenizer", None), "assistant_tokenizer": kwargs.pop("assistant_tokenizer", None), "assistant_model": assistant_model, "streamer": streamer, } generation_mode_kwargs["synced_gpus"] = ( (is_deepspeed_zero3_enabled() or is_fsdp_managed_module(self)) and dist.get_world_size() > 1 if synced_gpus is None else synced_gpus ) generation_mode_kwargs = {k: v for k, v in generation_mode_kwargs.items() if v is not None} # Custom_generate callables can have their own set of arguments # To extract them, we compare the signature with the standard _sample method if isinstance(custom_generate, Callable): usual_mode_kwargs = inspect.signature(GenerationMixin._sample).parameters.keys() custom_generate_kwargs = inspect.signature(custom_generate).parameters.keys() new_custom_keys = custom_generate_kwargs - usual_mode_kwargs generation_mode_kwargs = {k: kwargs.pop(k) for k in new_custom_keys if k in kwargs} return generation_mode_kwargs @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, negative_prompt_ids: Optional[torch.Tensor] = None, negative_prompt_attention_mask: Optional[torch.Tensor] = None, use_model_defaults: Optional[bool] = None, custom_generate: Optional[Union[str, Callable]] = None, **kwargs, ) -> Union[GenerateOutput, torch.LongTensor]: r""" Generates sequences of token ids for models with a language modeling head. 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). 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 be 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 has 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 complements 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. If your stopping criteria depends on the `scores` input, make sure you pass `return_dict_in_generate=True, output_scores=True` to `generate`. 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://huggingface.co/papers/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` if using `FullyShardedDataParallel` or DeepSpeed ZeRO Stage 3 with multiple GPUs to avoid deadlocking if one GPU finishes generating before other GPUs. Otherwise, defaults 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 assistant 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. negative_prompt_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): The negative prompt needed for some processors such as CFG. The batch size must match the input batch size. This is an experimental feature, subject to breaking API changes in future versions. negative_prompt_attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Attention_mask for `negative_prompt_ids`. use_model_defaults (`bool`, *optional*): When it is `True`, unset parameters in `generation_config` will be set to the model-specific default generation configuration (`model.generation_config`), as opposed to the global defaults (`GenerationConfig()`). If unset, models saved starting from `v4.50` will consider this flag to be `True`. custom_generate (`str` or `Callable`, *optional*): One of the following: - `str` (Hugging Face Hub repository name): runs the custom `generate` function defined at `custom_generate/generate.py` in that repository instead of the standard `generate` method. The repository fully replaces the generation logic, and the return type may differ. - `str` (local repository path): same as above but from a local path, `trust_remote_code` not required. - `Callable`: `generate` will perform the usual input preparation steps, then call the provided callable to run the decoding loop. For more information, see [the docs](../../generation_strategies#custom-generation-methods). kwargs (`dict[str, Any]`, *optional*): Ad hoc parametrization of `generation_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.LongTensor`. If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible [`~utils.ModelOutput`] types are: - [`~generation.GenerateDecoderOnlyOutput`], - [`~generation.GenerateBeamDecoderOnlyOutput`] If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible [`~utils.ModelOutput`] types are: - [`~generation.GenerateEncoderDecoderOutput`], - [`~generation.GenerateBeamEncoderDecoderOutput`] """ # 0. If requested, load an arbitrary generation recipe from the Hub and run it instead trust_remote_code = kwargs.pop("trust_remote_code", None) if custom_generate is not None and isinstance(custom_generate, str): # Get all `generate` arguments in a single variable. Custom functions are responsible for handling them: # they receive the same inputs as `generate`, with `model` instead of `self` and excluding the arguments to # trigger the custom generation. They can access to methods from `GenerationMixin` through `model`. global_keys_to_exclude = { "self", "kwargs", "global_keys_to_exclude", "trust_remote_code", "custom_generate", } generate_arguments = {key: value for key, value in locals().items() if key not in global_keys_to_exclude} generate_arguments.update(kwargs) custom_generate_function = self.load_custom_generate( custom_generate, trust_remote_code=trust_remote_code, **kwargs ) return custom_generate_function(model=self, **generate_arguments) # 1. Handle kwargs, `generation_config`, validate them and obtain generation mode generation_mode_kwargs = self._extract_generation_mode_kwargs( custom_generate, kwargs, synced_gpus, assistant_model, streamer, ) generation_config, model_kwargs = self._prepare_generation_config( generation_config, use_model_defaults, **kwargs ) generation_mode = generation_config.get_generation_mode(assistant_model) if isinstance(custom_generate, Callable): decoding_method = custom_generate else: # type() required to access the unbound class-level method decoding_method = getattr(type(self), GENERATION_MODES_MAPPING[generation_mode]) self._validate_model_kwargs(model_kwargs.copy()) self._validate_generation_mode(generation_mode, generation_config, generation_mode_kwargs) # Deprecation-related step: set Hub repo for deprecated strategies. # NOTE: This must come after initializing generation_config, since we need it to determine if this is a deprecated mode. # It must also be before any preparation steps, since Hub repos expect to be loaded before preparation steps. # TODO joao, manuel: remove this in v4.62.0 if deprecated_mode_repo := self._get_deprecated_gen_repo(generation_mode, trust_remote_code, custom_generate): return GenerationMixin.generate( self, inputs=inputs, generation_config=generation_config, logits_processor=logits_processor, stopping_criteria=stopping_criteria, prefix_allowed_tokens_fn=prefix_allowed_tokens_fn, assistant_model=assistant_model, negative_prompt_ids=negative_prompt_ids, negative_prompt_attention_mask=negative_prompt_attention_mask, use_model_defaults=use_model_defaults, custom_generate=deprecated_mode_repo, trust_remote_code=trust_remote_code, **generation_mode_kwargs, **kwargs, ) # 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() accepts_attention_mask = "attention_mask" in set(inspect.signature(self.forward).parameters.keys()) requires_attention_mask = "encoder_outputs" not in model_kwargs kwargs_has_attention_mask = model_kwargs.get("attention_mask", None) is not None # 3. Define model inputs inputs_tensor, model_input_name, model_kwargs = self._prepare_model_inputs( inputs, generation_config.bos_token_id, model_kwargs ) # Some generation modes (e.g. assisted) need `inputs_tensor` to rerun encoder.forward() if "inputs_tensor" in inspect.signature(decoding_method).parameters.keys(): generation_mode_kwargs["inputs_tensor"] = inputs_tensor batch_size = inputs_tensor.shape[0] device = inputs_tensor.device self._prepare_special_tokens(generation_config, kwargs_has_attention_mask, device=device) # decoder-only models must use left-padding for batched 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_tensor is not None and batch_size > 1 and len(inputs_tensor.shape) == 2 and torch.sum(inputs_tensor[:, -1] == generation_config._pad_token_tensor) > 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." ) # 4. Define other model kwargs # 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": generation_config.use_cache = True if not kwargs_has_attention_mask and requires_attention_mask and accepts_attention_mask: model_kwargs["attention_mask"] = self._prepare_attention_mask_for_generation( inputs_tensor, generation_config, model_kwargs ) elif kwargs_has_attention_mask: # TODO (joao): generalize this check with other types of inputs if model_input_name == "input_ids" and len(model_kwargs["attention_mask"].shape) > 2: raise ValueError("`attention_mask` passed to `generate` must be 2D.") 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, generation_config ) # 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_tensor, device=inputs_tensor.device, ) else: input_ids = inputs_tensor if model_input_name == "input_ids" else model_kwargs.pop("input_ids") # Expand inputs depending on the generation mode input_ids, model_kwargs = self._expand_inputs_for_generation( input_ids=input_ids, expand_size=max(generation_config.num_beams, generation_config.num_return_sequences), is_encoder_decoder=self.config.is_encoder_decoder, **model_kwargs, ) if generation_config.token_healing: input_ids = self.heal_tokens(input_ids, generation_mode_kwargs.get("tokenizer")) if streamer is not None: streamer.put(input_ids.cpu()) # 6. Prepare `max_length` depending on other stopping criteria. input_ids_length = input_ids.shape[1] has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None has_default_min_length = kwargs.get("min_length") is None and generation_config.min_length is not None generation_config = self._prepare_generated_length( generation_config=generation_config, has_default_max_length=has_default_max_length, has_default_min_length=has_default_min_length, model_input_name=model_input_name, inputs_tensor=inputs_tensor, input_ids_length=input_ids_length, ) # If the model supports `logits_to_keep` in forward(), set it to 1 to avoid computing the whole # logit matrix. This can save a lot of memory during the first forward pass. Note that assisted decoding # dynamically overrides this value as it can need more than the last token logits if self._supports_logits_to_keep() and "logits_to_keep" not in model_kwargs: model_kwargs["logits_to_keep"] = 1 self._validate_generated_length(generation_config, input_ids_length, has_default_max_length) # 7. Prepare the cache. # - `model_kwargs` may be updated in place with a cache as defined by the parameters in `generation_config`. # - different models have a different cache name expected by the model (default = "past_key_values") # - `max_length`, prepared above, is used to determine the maximum cache length max_cache_length = generation_config.max_length - 1 if ( inputs_tensor.shape[1] != input_ids_length and model_input_name == "inputs_embeds" and not self.config.is_encoder_decoder ): max_cache_length += inputs_tensor.shape[1] self._prepare_cache_for_generation( generation_config, model_kwargs, generation_mode, batch_size, max_cache_length ) 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 logits processors and stopping criteria prepared_logits_processor = self._get_logits_processor( generation_config=generation_config, input_ids_seq_length=input_ids_length, encoder_input_ids=inputs_tensor, prefix_allowed_tokens_fn=prefix_allowed_tokens_fn, logits_processor=logits_processor, device=inputs_tensor.device, model_kwargs=model_kwargs, negative_prompt_ids=negative_prompt_ids, negative_prompt_attention_mask=negative_prompt_attention_mask, ) prepared_stopping_criteria = self._get_stopping_criteria( generation_config=generation_config, stopping_criteria=stopping_criteria, tokenizer=generation_mode_kwargs.get("tokenizer"), ) # Set model_kwargs `use_cache` so we can use it later in forward runs model_kwargs["use_cache"] = generation_config.use_cache # 9. Call generation mode result = decoding_method( self, input_ids, logits_processor=prepared_logits_processor, stopping_criteria=prepared_stopping_criteria, generation_config=generation_config, **generation_mode_kwargs, **model_kwargs, ) # Convert to legacy cache format if requested if ( generation_config.return_legacy_cache is True and hasattr(result, "past_key_values") and getattr(result.past_key_values, "to_legacy_cache") is not None ): result.past_key_values = result.past_key_values.to_legacy_cache() return result def _has_unfinished_sequences(self, this_peer_finished: bool, synced_gpus: bool, device: torch.device) -> bool: """ Returns whether there are still unfinished sequences in the device. The existence of unfinished sequences is fed through `this_peer_finished`. ZeRO stage 3-friendly. """ 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, device=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: return False elif this_peer_finished: return False return True def heal_tokens( self, input_ids: torch.LongTensor, tokenizer: Optional["PreTrainedTokenizerBase"] = None ) -> torch.LongTensor: r""" Generates sequences of token ids for models with a language modeling head. Parameters: input_ids (`torch.LongTensor`): The sequence used as a prompt for the generation. tokenizer (`PreTrainedTokenizerBase`, *optional*): The tokenizer used to decode the input ids. Return: `torch.LongTensor` where each sequence has its tail token replaced with its appropriate extension. """ if tokenizer is None: raise ValueError( " When generating with token healing, you must pass the model's tokenizer to the `tokenizer` " "argument of `generate`." ) bos_token_id, pad_token_id = tokenizer.bos_token_id, tokenizer.pad_token_id vocab_trie = ExtensionsTrie(tokenizer.get_vocab()) generation_config = GenerationConfig(max_new_tokens=1, pad_token_id=pad_token_id) # assumption: leading/trailing whitespace is not meaningful, so the prompts are # stripped before re-tokenizing to desensitize generation to whitespace artefacts prompts = [p.strip() for p in tokenizer.batch_decode(input_ids, skip_special_tokens=True)] input_ids = tokenizer( prompts, return_tensors="pt", padding=True, ).input_ids.to(input_ids.device) # replace bos with pad to not condition healing on it input_ids = torch.where(input_ids == bos_token_id, pad_token_id, input_ids) # the latter code assumes the input_ids is not empty, input_id has to be checked if contains elements if input_ids.numel() == 0: return input_ids tail_ids = input_ids[:, -1].tolist() # tail tokens are used for a prefix search, thus, whitespaces are replaced with # their tokenization (e.g. 'Ġ') to enable search for tokens prefixed with a whitespace if tokenizer.convert_tokens_to_ids(" ") is not None: space_tok = tokenizer.convert_ids_to_tokens(tokenizer.convert_tokens_to_ids(" "))[0] tail_toks = (tokenizer.decode(t).replace(" ", space_tok) for t in tail_ids) else: tail_toks = (tokenizer.decode(t) for t in tail_ids) for batch_idx, (tail_id, tail_tok) in enumerate(zip(tail_ids, tail_toks)): batch_ids = input_ids[batch_idx] if torch.all(batch_ids == pad_token_id).item(): continue # skip empty sequences (all pad ids) # apply bias for alternatives (extensions) to the tail token """ seq_bias key has to be tuple with int so have to use tokenizer function to convert str to int """ seq_bias = { (tokenizer.convert_tokens_to_ids(alt_tok),): 10.0 for alt_tok in vocab_trie.extensions(prefix=tail_tok) } if len(seq_bias) == 1: continue # skip if there are no token alternatives to heal with # slightly favor original token to limit aggressive healing e.g. 'http' -> 'https' seq_bias[(tail_id,)] += 1.0 generation_config.update(sequence_bias=seq_bias) trimmed_ids = batch_ids[:-1] """ the latter code assumes trimmed_ids is not empty so have to check the its element count """ if trimmed_ids.numel() == 0: continue # if the prompt is a single (non-pad) token, regenerate from bos if len(batch_ids[batch_ids != pad_token_id]) == 1: trimmed_ids[-1] = bos_token_id input_ids[batch_idx] = self.generate(trimmed_ids.unsqueeze(0), generation_config=generation_config) return input_ids def _sample( self, input_ids: torch.LongTensor, logits_processor: LogitsProcessorList, stopping_criteria: StoppingCriteriaList, generation_config: GenerationConfig, synced_gpus: bool = False, streamer: Optional["BaseStreamer"] = None, **model_kwargs, ) -> Union[GenerateNonBeamOutput, 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. Parameters: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): The sequence used as a prompt for the generation. logits_processor (`LogitsProcessorList`): 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`): An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`] used to tell if the generation loop should stop. generation_config ([`~generation.GenerationConfig`]): The generation configuration to be used as parametrization of the decoding method. synced_gpus (`bool`): Whether to continue running the while loop until max_length (needed to avoid deadlocking with `FullyShardedDataParallel` and DeepSpeed 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.GenerateDecoderOnlyOutput`], [`~generation.GenerateEncoderDecoderOutput`] or `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a [`~generation.GenerateDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and `return_dict_in_generate=True` or a [`~generation.GenerateEncoderDecoderOutput`] if `model.config.is_encoder_decoder=True`. """ # init values pad_token_id = generation_config._pad_token_tensor output_attentions = generation_config.output_attentions output_hidden_states = generation_config.output_hidden_states output_scores = generation_config.output_scores output_logits = generation_config.output_logits return_dict_in_generate = generation_config.return_dict_in_generate has_eos_stopping_criteria = any(hasattr(criteria, "eos_token_id") for criteria in stopping_criteria) do_sample = generation_config.do_sample # init attention / hidden states / scores tuples scores = () if (return_dict_in_generate and output_scores) else None raw_logits = () if (return_dict_in_generate and output_logits) 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 batch_size, cur_len = input_ids.shape[:2] this_peer_finished = False unfinished_sequences = torch.ones(batch_size, dtype=torch.long, device=input_ids.device) model_kwargs = self._get_initial_cache_position(cur_len, input_ids.device, model_kwargs) model_forward = self.__call__ compile_forward = self._valid_auto_compile_criteria(model_kwargs, generation_config) if compile_forward: os.environ["TOKENIZERS_PARALLELISM"] = "0" # If we use FA2 and a static cache, we cannot compile with fullgraph if self.config._attn_implementation == "flash_attention_2": # only raise warning if the user passed an explicit compile-config if generation_config.compile_config is not None and generation_config.compile_config.fullgraph: logger.warning_once( "When using Flash Attention 2 and a static cache, you cannot use the option `CompileConfig(fullgraph=True)` as " "FA2 introduces graph breaks. We overrode the option with `fullgraph=False`." ) generation_config.compile_config.fullgraph = False model_forward = self.get_compiled_call(generation_config.compile_config) if generation_config.prefill_chunk_size is not None: model_kwargs = self._prefill_chunking(input_ids, generation_config, **model_kwargs) is_prefill = False else: is_prefill = True while self._has_unfinished_sequences(this_peer_finished, synced_gpus, device=input_ids.device): # prepare model inputs model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs) if is_prefill: outputs = self(**model_inputs, return_dict=True) is_prefill = False else: outputs = model_forward(**model_inputs, return_dict=True) # synced_gpus: don't waste resources running the code we don't need; kwargs must be updated before skipping model_kwargs = self._update_model_kwargs_for_generation( outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder, ) if synced_gpus and this_peer_finished: continue # Copy is needed to avoid keeping a hanging ref to outputs.logits which may be very large for first iteration # (the clone itself is always small) next_token_logits = outputs.logits[:, -1, :].to(copy=True, dtype=torch.float32, device=input_ids.device) # pre-process distribution next_token_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_token_scores,) if output_logits: raw_logits += (next_token_logits,) 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,) ) # token selection if do_sample: probs = nn.functional.softmax(next_token_scores, dim=-1) # TODO (joao): this OP throws "skipping cudagraphs due to ['incompatible ops']", find solution next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1) else: next_tokens = torch.argmax(next_token_scores, dim=-1) # finished sentences should have their next token be a padding token if has_eos_stopping_criteria: 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()) unfinished_sequences = unfinished_sequences & ~stopping_criteria(input_ids, scores) this_peer_finished = unfinished_sequences.max() == 0 cur_len += 1 # This is needed to properly delete outputs.logits which may be very large for first iteration # Otherwise a reference to outputs is kept which keeps the logits alive in the next iteration del outputs if streamer is not None: streamer.end() if return_dict_in_generate: if self.config.is_encoder_decoder: return GenerateEncoderDecoderOutput( sequences=input_ids, scores=scores, logits=raw_logits, encoder_attentions=encoder_attentions, encoder_hidden_states=encoder_hidden_states, decoder_attentions=decoder_attentions, cross_attentions=cross_attentions, decoder_hidden_states=decoder_hidden_states, past_key_values=model_kwargs.get("past_key_values"), ) else: return GenerateDecoderOnlyOutput( sequences=input_ids, scores=scores, logits=raw_logits, attentions=decoder_attentions, hidden_states=decoder_hidden_states, past_key_values=model_kwargs.get("past_key_values"), ) else: return input_ids @staticmethod def _flatten_beam_dim(tensor: torch.Tensor) -> torch.Tensor: """[batch_size, num_beams, ...] -> [batch_size * num_beams, ...]""" shape = list(tensor.shape) return torch.reshape(tensor, [shape[0] * shape[1]] + shape[2:]) @staticmethod def _unflatten_beam_dim(tensor: torch.Tensor, batch_size: int, num_beams: int) -> torch.Tensor: """[batch_size * num_beams, ...] -> [batch_size, num_beams, ...]""" shape = list(tensor.shape) return torch.reshape(tensor, [batch_size, num_beams] + shape[1:]) @staticmethod def _gather_beams(tensor: torch.Tensor, beam_indices: torch.Tensor) -> torch.Tensor: """ Gathers the beam slices indexed by beam_indices into new beam array. Args: tensor (`torch.Tensor`): A tensor containing data to be gathered. The tensor is a 2D or a 3D tensor with the two first dimensions depicting the batch and the beam dimensions. beam_indices (`torch.Tensor` of shape `(batch_size, num_beams_to_select)`): The indices of the beams to select . Returns: A tensor with the selected beams """ # `take_along_dim` requires its indices arg to have the same number of dims as `input` while len(beam_indices.shape) < len(tensor.shape): beam_indices = beam_indices.unsqueeze(-1) gathered_tensor = torch.take_along_dim(input=tensor, indices=beam_indices, dim=1) return gathered_tensor @staticmethod def _check_early_stop_heuristic( is_early_stop_heuristic_unsatisfied: torch.Tensor, running_beam_scores: torch.Tensor, beam_scores: torch.Tensor, is_sent_finished: torch.Tensor, cur_len: int, max_length: int, decoder_prompt_len: int, early_stopping: Union[bool, str], length_penalty: float, ): """ Determine whether early stopping is possible by checking if the best possible score of running beams could still improve upon the finished ones. Mechanism: - Without a length penalty, beam scores typically decrease as more tokens are generated. So, if the *best possible* score from any running beam is already worse than the *worst* finished beam, we can safely stop early. - With a length penalty, scores may increase with longer sequences. In this case, we use heuristics to estimate the best possible score — though this estimate may not always be correct — and stop if no further improvement seems likely. We apply different heuristics depending on the value of `early_stopping`: 1. `early_stopping == False`: -> Use a heuristic that assumes the best score comes from the current length minus the decoder prompt length. -> See detailed discussion: https://github.com/huggingface/transformers/pull/20901#issuecomment-1369845565 2. `early_stopping == "never"`: -> Estimate the best score using either `max_length` or `cur_len`, depending on the sign of `length_penalty`. -> A positive length penalty favors longer sequences, so we use `max_length` in that case. NOTE: the canonical beam search implementation can be replicated with `early_stopping="never"` and `length_penalty=0.0`, which are NOT the default flags. The default behavior was empirically found to produce better sequences (prior to 2022), and changing it is BC breaking. """ if early_stopping == "never" and length_penalty > 0.0: best_hypothetical_length = max_length - decoder_prompt_len else: best_hypothetical_length = cur_len - decoder_prompt_len best_possible_running_score = running_beam_scores[:, :1] / (best_hypothetical_length**length_penalty) worst_finished_score = torch.where(is_sent_finished, torch.min(beam_scores, dim=1, keepdim=True)[0], -1.0e9) return is_early_stop_heuristic_unsatisfied & torch.any( best_possible_running_score > worst_finished_score, dim=-1, keepdim=True ) @staticmethod def _beam_search_has_unfinished_sequences( is_early_stop_heuristic_unsatisfied: torch.Tensor, is_sent_finished: torch.Tensor, next_token_hits_stopping_criteria: torch.Tensor, early_stopping: Union[bool, str], ): """ Beam Search stopping condition -- halts the generation loop if any of these conditions becomes False """ # a. Can the open beams improve the top completed scores? improvement_possible = torch.any(is_early_stop_heuristic_unsatisfied) # b. Is there still a beam without fully completed sequences? This is only relevant if early_stopping is # enabled, where we want to finish as soon as all beams have a completed sequence. exists_open_beam = ~(torch.all(is_sent_finished) & (early_stopping is True)) # c. Have we hit a stopping criteria with all running sequences and have no way to continue? e.g. we have # reached `max_length`` valid_continuations = ~torch.all(next_token_hits_stopping_criteria) return improvement_possible & exists_open_beam & valid_continuations def _get_top_k_continuations( self, accumulated_log_probs: torch.Tensor, running_sequences: torch.Tensor, running_beam_indices: torch.Tensor, cur_len: int, decoder_prompt_len: int, do_sample: bool, beams_to_keep: int, num_beams: int, vocab_size: int, batch_size: int, ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: """ Get top-K continuations given the accumulated log probs on the next token. A few notes to understand what's going on: 1. Each item in batch has `num_beams` * `vocab_size` candidate continuations. For each item, get the top K [K = (number of EOS tokens + 1) * `num_beams`] candidates with the highest accumulated log-probabilities, or sample them without replacement using the accumulated scores 2. We gather the top K (as opposed to `num_beams`, or any number lower than K) here so that we have at least `num_beams` sequences remaining to continue the live beam search. 3. Note that other stopping criteria might result in impossible to continue beams, i.e. all continuations selected in this step hit the stopping criteria. """ # TODO (joao): This function should take an optional beam scorer function, to manipulate the scores after # token selection. The function should be an argument exposed, so that custom scoring functions can be # defined. # Gather the top K scores from _all_ beams. if do_sample: topk_indices = torch.multinomial( nn.functional.softmax(accumulated_log_probs, dim=-1), num_samples=beams_to_keep ) topk_log_probs = torch.gather(input=accumulated_log_probs, dim=1, index=topk_indices) else: topk_log_probs, topk_indices = torch.topk(accumulated_log_probs, k=beams_to_keep) # Gather K top beams, recover the beam index by floor division and token id by modulo division topk_current_beam_indices = topk_indices // vocab_size topk_running_beam_indices = self._gather_beams(running_beam_indices, topk_current_beam_indices) topk_running_sequences = self._gather_beams(running_sequences, topk_current_beam_indices) topk_ids = topk_indices % vocab_size # Update sequences for the K top-k new sequences. topk_running_sequences[:, :, cur_len] = topk_ids # we want to store the beam indices with batch information -> real beam index = beam index % num beams batch_offset = torch.arange(batch_size, device=topk_ids.device).view(-1, 1) * num_beams batch_modified_indices = topk_current_beam_indices + batch_offset topk_running_beam_indices[:, :, cur_len - decoder_prompt_len] = batch_modified_indices return topk_log_probs, topk_running_sequences, topk_running_beam_indices def _get_running_beams_for_next_iteration( self, topk_log_probs: torch.Tensor, topk_running_sequences: torch.Tensor, topk_running_beam_indices: torch.Tensor, next_token_hits_stopping_criteria: torch.Tensor, num_beams: int, ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: """ Given the top-K continuations, their scores, and whether they hit a stopping criteria, select the best non-finished beams to continue beam search in the next iteration. """ # To prevent these just finished sequences from being used in subsequent iterations, set their log probs # to a very large negative value topk_running_log_probs = topk_log_probs + next_token_hits_stopping_criteria.to(torch.float32) * -1.0e9 next_topk_indices = torch.topk(topk_running_log_probs, k=num_beams)[1] running_sequences = self._gather_beams(topk_running_sequences, next_topk_indices) running_beam_scores = self._gather_beams(topk_running_log_probs, next_topk_indices) running_beam_indices = self._gather_beams(topk_running_beam_indices, next_topk_indices) return running_sequences, running_beam_scores, running_beam_indices def _update_finished_beams( self, sequences: torch.Tensor, topk_running_sequences: torch.Tensor, beam_scores: torch.Tensor, topk_log_probs: torch.Tensor, beam_indices: torch.Tensor, topk_running_beam_indices: torch.Tensor, is_early_stop_heuristic_unsatisfied: torch.Tensor, is_sent_finished: torch.Tensor, next_token_hits_stopping_criteria: torch.Tensor, top_num_beam_mask: torch.Tensor, num_beams: int, cur_len: int, decoder_prompt_len: int, length_penalty: float, early_stopping: Union[bool, str], ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: """ Updates the finished beams if (and only if) there are new completed sequences that have a higher score than the current finished sequences. """ # Only the top `num_beam` sequences can be considered for the final returned sequences. Remember: the # remaining sequences only exist as a backup to ensure that we have at least `num_beams` sequences to # continue. did_top_num_beams_just_finished = next_token_hits_stopping_criteria & top_num_beam_mask[None, :] # Further process topk logits for the finished beams # - add length penalty topk_log_probs = topk_log_probs / ((cur_len + 1 - decoder_prompt_len) ** length_penalty) # - make sure no scores can be added anymore if beam is full and early stopping is on beams_in_batch_are_full = torch.all(is_sent_finished, axis=-1, keepdims=True) & (early_stopping is True) topk_log_probs += beams_in_batch_are_full.to(torch.float32) * -1.0e9 # - make sure no scores can be added anymore if improvement is not possible topk_log_probs += (~is_early_stop_heuristic_unsatisfied).to(torch.float32) * -1.0e9 # - make sure still running sequences cannot be chosen as finalized beam topk_log_probs += (~did_top_num_beams_just_finished) * -1.0e9 # Get finalized `num_beam` sequences for the next generation step -- combine the previous finalized # data with the new finalized sequences (if any, non-finalized sequences have a very large negative score # in this step), and keep the best `num_beams` sequences. merged_sequences = torch.cat((sequences, topk_running_sequences), dim=1) merged_scores = torch.cat((beam_scores, topk_log_probs), dim=1) merged_beam_indices = torch.cat((beam_indices, topk_running_beam_indices), dim=1) merged_is_sent_finished = torch.cat((is_sent_finished, did_top_num_beams_just_finished), dim=1) topk_merged_indices = torch.topk(merged_scores, k=num_beams)[1] sequences = self._gather_beams(merged_sequences, topk_merged_indices) beam_scores = self._gather_beams(merged_scores, topk_merged_indices) beam_indices = self._gather_beams(merged_beam_indices, topk_merged_indices) is_sent_finished = self._gather_beams(merged_is_sent_finished, topk_merged_indices) return sequences, beam_scores, beam_indices, is_sent_finished # end of auxiliary functions for beam search def _beam_search( self, input_ids: torch.LongTensor, logits_processor: LogitsProcessorList, stopping_criteria: StoppingCriteriaList, generation_config: GenerationConfig, synced_gpus: bool = False, **model_kwargs, ) -> Union[GenerateBeamOutput, 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. If it's the first time you're diving into Beam Search, we recommend you read the following blog post: https://huggingface.co/blog/how-to-generate (especially the beam search section). You can recompute the sequence scores from the individual scores using the `compute_transition_scores` function (https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationMixin.compute_transition_scores) Parameters: input_ids (`torch.LongTensor` of shape `(batch_size*num_beams, sequence_length)`): The sequence used as a prompt for the generation. logits_processor (`LogitsProcessorList`): 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`: An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`] used to tell if the generation loop should stop. generation_config ([`~generation.GenerationConfig`]): The generation configuration to be used as parametrization of the decoding method. synced_gpus (`bool`): Whether to continue running the while loop until max_length (needed to avoid deadlocking with `FullyShardedDataParallel` and DeepSpeed 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.GenerateBeamDecoderOnlyOutput`], [`~generation.GenerateBeamEncoderDecoderOutput`] or `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a [`~generation.GenerateBeamDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and `return_dict_in_generate=True` or a [`~generation.GenerateBeamEncoderDecoderOutput`] if `model.config.is_encoder_decoder=True`. """ # 1. init beam_search values pad_token_id = generation_config._pad_token_tensor eos_token_id = generation_config._eos_token_tensor output_attentions = generation_config.output_attentions output_hidden_states = generation_config.output_hidden_states output_scores = generation_config.output_scores output_logits = generation_config.output_logits return_dict_in_generate = generation_config.return_dict_in_generate do_sample = generation_config.do_sample early_stopping = generation_config.early_stopping length_penalty = generation_config.length_penalty max_length = generation_config.max_length num_beams = generation_config.num_beams num_return_sequences = generation_config.num_return_sequences batch_size_unflattened, cur_len = input_ids.shape[:2] batch_size = batch_size_unflattened // num_beams # TODO (joao): standardize special cases if self.__class__.__name__ == "MoshiDepthDecoder": vocab_size = self.config.audio_vocab_size elif self.__class__.__name__ == "ImageGPTForCausalImageModeling": vocab_size = self.get_output_embeddings().out_features elif self.__class__.__name__ == "BarkSemanticModel": vocab_size = self.config.output_vocab_size else: vocab_size = self.config.get_text_config().vocab_size decoder_prompt_len = cur_len this_peer_finished = False # At each beam search step, we want to keep top K [K = (number of EOS tokens + 1) * `num_beams`] candidates # with the highest log-probabilities, or sample K continuations without replacement. We gather the top K # (as opposed to `num_beams`, or any number lower than K) so that we have at least `num_beams` sequences # non-finished to continue the live beam search, in case the top `num_beams` all select an EOS token. n_eos_tokens = eos_token_id.shape[0] if eos_token_id is not None else 0 beams_to_keep = max(2, 1 + n_eos_tokens) * num_beams top_num_beam_mask = torch.cat( (torch.ones((num_beams), dtype=torch.bool), torch.zeros((beams_to_keep - num_beams), dtype=torch.bool)), dim=0, ).to(input_ids.device) model_kwargs = self._get_initial_cache_position(cur_len, input_ids.device, model_kwargs) # (joao) feature lost in the refactor. Probably won't implement, hurts readability with minimal gains (there # are newer low-memory alternatives like the offloaded cache) sequential = generation_config.low_memory if sequential: raise ValueError( "`low_memory=True` is not supported after the beam search refactor. Please check the discussion in " "#35802 *after the PR got merged*, and add a comment there if your questions are not yet answered." ) # 2. init output tuples all_scores = () if (return_dict_in_generate and output_scores) else None raw_logits = () if (return_dict_in_generate and output_logits) else None beam_indices = () if (return_dict_in_generate and output_logits) 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 ) # 3. init running tensors and static-shaped placeholders # per batch, beam-item holding current token in loop and completed sequences output_fill_value = pad_token_id or eos_token_id[0] if eos_token_id is not None else -1 running_sequences = torch.full( (batch_size, num_beams, max_length), fill_value=output_fill_value, dtype=torch.int64, device=input_ids.device, ) running_sequences[:, :, :cur_len] = self._unflatten_beam_dim(input_ids, batch_size, num_beams) sequences = running_sequences.detach().clone() # per batch, beam-item score, logprobs # 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. running_beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device) running_beam_scores[:, 1:] = -1e9 beam_scores = torch.full((batch_size, num_beams), fill_value=-1e9, dtype=torch.float, device=input_ids.device) # per batch, beam-item state bit indicating if sentence has finished. is_sent_finished = torch.zeros((batch_size, num_beams), dtype=torch.bool, device=input_ids.device) # per batch state bit indicating if there is a possibility to improve the best finished sentence. is_early_stop_heuristic_unsatisfied = torch.ones((batch_size, 1), dtype=torch.bool, device=input_ids.device) # per batch, beam-item state bit indicating if there are valid continuations. next_token_hits_stopping_criteria = torch.zeros( (batch_size, num_beams), dtype=torch.bool, device=input_ids.device ) # per batch selected beam indices running_beam_indices = torch.full( (batch_size, num_beams, max_length - cur_len), fill_value=-1, dtype=torch.int32, device=input_ids.device ) beam_indices = running_beam_indices.detach().clone() # 4. run the generation loop while self._has_unfinished_sequences(this_peer_finished, synced_gpus, device=input_ids.device): # a. Forward current tokens, obtain the logits flat_running_sequences = self._flatten_beam_dim(running_sequences[:, :, :cur_len]) model_inputs = self.prepare_inputs_for_generation(flat_running_sequences, **model_kwargs) model_outputs = self(**model_inputs, return_dict=True) # synced_gpus: don't waste resources running the code we don't need; kwargs must be updated before skipping model_kwargs = self._update_model_kwargs_for_generation( model_outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder, ) if synced_gpus and this_peer_finished: continue # Copy is needed to avoid keeping a hanging ref logits = model_outputs.logits[:, -1, :].to(copy=True, dtype=torch.float32, device=input_ids.device) # b. Compute log probs -- get log probabilities from logits, process logits with processors (*e.g.* # `temperature`, ...), and add new logprobs to existing running logprobs scores. log_probs = nn.functional.log_softmax(logits, dim=-1) log_probs = logits_processor(flat_running_sequences, log_probs) # Store logits, attentions and hidden_states when required if return_dict_in_generate: if output_logits: raw_logits += (logits.clone(),) if return_dict_in_generate and output_scores: all_scores += (log_probs.clone(),) if output_attentions: decoder_attentions += ( (model_outputs.decoder_attentions,) if self.config.is_encoder_decoder else (model_outputs.attentions,) ) if self.config.is_encoder_decoder: cross_attentions += (model_outputs.cross_attentions,) if output_hidden_states: decoder_hidden_states += ( (model_outputs.decoder_hidden_states,) if self.config.is_encoder_decoder else (model_outputs.hidden_states,) ) # This is needed to properly delete logits which may be very large for first iteration # Otherwise a reference to outputs is kept which keeps the logits alive in the next iteration del model_outputs log_probs = self._unflatten_beam_dim(log_probs, batch_size, num_beams) log_probs = log_probs + running_beam_scores[:, :, None] log_probs = torch.reshape(log_probs, (batch_size, num_beams * vocab_size)) # c. Retrieve top-K continuations, i.e. select the next token (greedy or sampling) and then keep the best # continuations among all beams based on the accumulated scores. topk_log_probs, topk_running_sequences, topk_running_beam_indices = self._get_top_k_continuations( accumulated_log_probs=log_probs, running_sequences=running_sequences, running_beam_indices=running_beam_indices, cur_len=cur_len, decoder_prompt_len=decoder_prompt_len, do_sample=do_sample, beams_to_keep=beams_to_keep, num_beams=num_beams, vocab_size=vocab_size, batch_size=batch_size, ) # d. Check which running sequences have finished next_token_hits_stopping_criteria = stopping_criteria( self._flatten_beam_dim(topk_running_sequences[:, :, : cur_len + 1]), # remove unfilled token indexes all_scores, ) next_token_hits_stopping_criteria = self._unflatten_beam_dim( next_token_hits_stopping_criteria, batch_size, beams_to_keep ) # e. Get the non-finished running `num_beams` sequences for the next generation step running_sequences, running_beam_scores, running_beam_indices = self._get_running_beams_for_next_iteration( topk_log_probs=topk_log_probs, topk_running_sequences=topk_running_sequences, topk_running_beam_indices=topk_running_beam_indices, next_token_hits_stopping_criteria=next_token_hits_stopping_criteria, num_beams=num_beams, ) # f. Update the completed beams if a new high score in a finished sequence is found sequences, beam_scores, beam_indices, is_sent_finished = self._update_finished_beams( sequences=sequences, topk_running_sequences=topk_running_sequences, beam_scores=beam_scores, topk_log_probs=topk_log_probs, beam_indices=beam_indices, topk_running_beam_indices=topk_running_beam_indices, is_early_stop_heuristic_unsatisfied=is_early_stop_heuristic_unsatisfied, is_sent_finished=is_sent_finished, next_token_hits_stopping_criteria=next_token_hits_stopping_criteria, top_num_beam_mask=top_num_beam_mask, num_beams=num_beams, cur_len=cur_len, decoder_prompt_len=decoder_prompt_len, length_penalty=length_penalty, early_stopping=early_stopping, ) # g. Prepare remaining data for the next iteration, including computing the stopping condition for # beam search as a whole (as opposed to individual beams, i.e. `stopping_criteria`) # pluck the cache from the beam indices that will be used in the next iteration # NOTE: we need to check if `self._reorder_cache` exists for special models like RAG, RecurrentGemma etc. if model_kwargs.get("past_key_values", None) is not None: beam_idx = self._flatten_beam_dim(running_beam_indices[..., cur_len - decoder_prompt_len]) if hasattr(self, "_reorder_cache"): model_kwargs["past_key_values"] = self._reorder_cache(model_kwargs["past_key_values"], beam_idx) else: model_kwargs["past_key_values"].reorder_cache(beam_idx) cur_len = cur_len + 1 is_early_stop_heuristic_unsatisfied = self._check_early_stop_heuristic( is_early_stop_heuristic_unsatisfied=is_early_stop_heuristic_unsatisfied, running_beam_scores=running_beam_scores, beam_scores=beam_scores, is_sent_finished=is_sent_finished, cur_len=cur_len, max_length=max_length, decoder_prompt_len=decoder_prompt_len, early_stopping=early_stopping, length_penalty=length_penalty, ) this_peer_finished = not self._beam_search_has_unfinished_sequences( is_early_stop_heuristic_unsatisfied, is_sent_finished, next_token_hits_stopping_criteria, early_stopping, ) # 5. prepare outputs # Take best beams for each batch (the score is sorted in descending order) sequences = self._flatten_beam_dim(sequences[:, :num_return_sequences, :]) beam_scores = self._flatten_beam_dim(beam_scores[:, :num_return_sequences]) beam_indices = self._flatten_beam_dim(beam_indices[:, :num_return_sequences, :]) # Crop the static-shaped tensors to the actual size. # `beam_indices` is initialized with -1s, and is updated with the beam index of the generated token at each # step. We can use it to detect the generated length, which may be != `cur_len` (e.g. selected beam is from a # previous decoding iteration) max_generated_length = ((beam_indices + 1).bool()).sum(dim=1).max() output_length = decoder_prompt_len + max_generated_length sequences = sequences[:, :output_length] beam_indices = beam_indices[:, :max_generated_length] if return_dict_in_generate: if not output_scores: beam_scores = None if self.config.is_encoder_decoder: return GenerateBeamEncoderDecoderOutput( sequences=sequences, sequences_scores=beam_scores, scores=all_scores, logits=raw_logits, beam_indices=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, past_key_values=model_kwargs.get("past_key_values"), ) else: return GenerateBeamDecoderOnlyOutput( sequences=sequences, sequences_scores=beam_scores, scores=all_scores, logits=raw_logits, beam_indices=beam_indices, attentions=decoder_attentions, hidden_states=decoder_hidden_states, past_key_values=model_kwargs.get("past_key_values"), ) else: return sequences def _assisted_decoding( self, input_ids: torch.LongTensor, logits_processor: LogitsProcessorList, stopping_criteria: StoppingCriteriaList, generation_config: GenerationConfig, synced_gpus: bool = False, streamer: Optional["BaseStreamer"] = None, inputs_tensor: Optional[torch.FloatTensor] = None, assistant_model: Optional["PreTrainedModel"] = None, assistant_tokenizer: Optional["PreTrainedTokenizerBase"] = None, tokenizer: Optional["PreTrainedTokenizerBase"] = None, **model_kwargs, ) -> Union[GenerateNonBeamOutput, torch.LongTensor]: r""" Generates sequences of token ids for models with a language modeling head using **greedy decoding** or **sample** (depending on `do_sample`), assisted by candidate sequences. Assisted generation is an example of a candidate decoding strategy. Can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models. Parameters: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): The sequence used as a prompt for the generation. logits_processor (`LogitsProcessorList`): 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`): An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`] used to tell if the generation loop should stop. generation_config ([`~generation.GenerationConfig`]): The generation configuration to be used as parametrization of the decoding method. synced_gpus (`bool`): Whether to continue running the while loop until max_length (needed to avoid deadlocking with `FullyShardedDataParallel` and DeepSpeed 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. inputs_tensor (`torch.FloatTensor`, *optional*): The input tensor for generation. For decoder models, usually `input_ids`. For encoder-decoder models, the tensor that produced `model_kwargs["encoder_outputs"]`. assistant_model (`PreTrainedModel`, *optional*): The model used to assist the generation process. If not provided, the main model will be used. assistant_tokenizer (`PreTrainedTokenizerBase`, *optional*): The tokenizer used for the assistant model. If not provided, the token space is assumed to be the same. tokenizer (`PreTrainedTokenizerBase`, *optional*): The tokenizer used for the main model. If not provided, the token space is assumed to be the same. 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.GenerateDecoderOnlyOutput`], [`~generation.GenerateEncoderDecoderOutput`] or `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a [`~generation.GenerateDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and `return_dict_in_generate=True` or a [`~generation.GenerateEncoderDecoderOutput`] if `model.config.is_encoder_decoder=True`. """ # The cache must be dynamic for assisted generation, and the check must happen AFTER preparing cache if not model_kwargs["use_cache"]: raise ValueError("assisted generate requires `use_cache=True`") if generation_config.cache_implementation in ["static", "hybrid", "sliding_window"] or ( "past_key_values" in model_kwargs and hasattr(model_kwargs["past_key_values"], "layers") and any(getattr(l, "is_compileable", False) for l in model_kwargs["past_key_values"].layers) ): raise ValueError("assisted generate is not supported with Static cache classes`") # Get the candidate generator, given the parameterization candidate_generator = self._get_candidate_generator( generation_config=generation_config, input_ids=input_ids, inputs_tensor=inputs_tensor, assistant_model=assistant_model, logits_processor=logits_processor, target_tokenizer=tokenizer, assistant_tokenizer=assistant_tokenizer, model_kwargs=model_kwargs, ) # init values do_sample = generation_config.do_sample output_attentions = generation_config.output_attentions output_hidden_states = generation_config.output_hidden_states output_scores = generation_config.output_scores output_logits = generation_config.output_logits return_dict_in_generate = generation_config.return_dict_in_generate # init attention / hidden states / scores tuples scores = () if (return_dict_in_generate and output_scores) else None raw_logits = () if (return_dict_in_generate and output_logits) 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 batch_size, cur_len = input_ids.shape[:2] if batch_size > 1: raise ValueError("assisted generate is only supported for batch_size = 1") unfinished_sequences = torch.ones(batch_size, dtype=torch.long, device=input_ids.device) model_kwargs = self._get_initial_cache_position(cur_len, input_ids.device, model_kwargs) this_peer_finished = False is_first_iteration = True # to preserve the same API in the output as other generation methods while self._has_unfinished_sequences(this_peer_finished, synced_gpus, device=input_ids.device): cur_len = input_ids.shape[1] # 1. Fetch candidate sequences from a `CandidateGenerator` and move to the correct device candidate_input_ids, candidate_logits = candidate_generator.get_candidates(input_ids) candidate_input_ids = candidate_input_ids.to(self.device) if candidate_logits is not None: candidate_logits = candidate_logits.to(self.device) candidate_length = candidate_input_ids.shape[1] - input_ids.shape[1] is_done_candidate = stopping_criteria(candidate_input_ids, None) # 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. Prepare the model inputs candidate_kwargs = copy.copy(model_kwargs) candidate_kwargs = _prepare_attention_mask( candidate_kwargs, candidate_input_ids.shape[1], self.config.is_encoder_decoder ) candidate_kwargs = _prepare_token_type_ids(candidate_kwargs, candidate_input_ids.shape[1]) if "cache_position" in candidate_kwargs: candidate_kwargs["cache_position"] = torch.cat( ( candidate_kwargs["cache_position"], torch.arange(cur_len, cur_len + candidate_length, device=input_ids.device, dtype=torch.long), ), dim=0, ) model_inputs = self.prepare_inputs_for_generation(candidate_input_ids, **candidate_kwargs) if "logits_to_keep" in model_inputs: model_inputs["logits_to_keep"] = candidate_length + 1 # 2.2. Run a forward pass on the candidate sequence outputs = self(**model_inputs) # 2.3. Process the new logits # .float() is needed to retain precision for later logits manipulations new_logits = outputs.logits[:, -candidate_length - 1 :].to( dtype=torch.float32, device=input_ids.device ) # excludes the input prompt if present next_token_logits = new_logits.clone() if len(logits_processor) > 0: for i in range(candidate_length + 1): new_logits[:, i, :] = logits_processor(candidate_input_ids[:, : cur_len + i], new_logits[:, i, :]) # 3. Select the accepted tokens. There are two possible cases: # Case 1: `do_sample=True` and we have logits for the candidates (originally from speculative decoding) # 👉 Apply algorithm 1 from the speculative decoding paper (https://huggingface.co/papers/2211.17192). if do_sample and candidate_logits is not None: valid_tokens, n_matches = _speculative_sampling( candidate_input_ids, candidate_logits, candidate_length, new_logits, is_done_candidate, ) # Case 2: all other cases (originally from assisted generation) 👉 Compare the tokens selected from the # original model logits with the candidate tokens. We can keep the candidate tokens until the first # mismatch, or until the max length is reached. else: if do_sample: probs = new_logits.softmax(dim=-1) selected_tokens = torch.multinomial(probs[0, :, :], num_samples=1).squeeze(1)[None, :] else: selected_tokens = new_logits.argmax(dim=-1) candidate_new_tokens = candidate_input_ids[:, cur_len:] n_matches = ((~(candidate_new_tokens == selected_tokens[:, :-1])).cumsum(dim=-1) < 1).sum() # Ensure we don't generate beyond max_len or an EOS token if is_done_candidate and n_matches == candidate_length: n_matches -= 1 valid_tokens = selected_tokens[:, : n_matches + 1] # 4. 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. # 4.1. Get the valid continuation, after the matching tokens 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] # 4.2. Discard past key values relative to unused assistant tokens outputs.past_key_values.crop(new_cur_len - 1) # 5. Update the candidate generation strategy if needed candidate_generator.update_candidate_strategy(input_ids, new_logits, n_matches) # synced_gpus: don't waste resources running the code we don't need; kwargs must be updated before skipping model_kwargs = self._update_model_kwargs_for_generation( outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder, num_new_tokens=n_matches + 1, ) if synced_gpus and this_peer_finished: continue # 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: newly_added_length = n_matches + 1 if output_scores: scores += tuple(new_logits[:, i, :] for i in range(newly_added_length)) if output_logits: raw_logits += tuple(next_token_logits[:, i, :] for i in range(newly_added_length)) newly_added_length = new_cur_len if is_first_iteration else newly_added_length if output_attentions: if self.config.is_encoder_decoder: cross_attentions = _split_model_outputs( cross_attentions, outputs.cross_attentions, cur_len, newly_added_length ) decoder_attentions = _split_model_outputs( decoder_attentions, outputs.decoder_attentions, cur_len, newly_added_length, is_decoder_attention=True, ) # some (V)LLMs have hard requirement on SDPA and thus never return attn elif outputs.attentions[0] is not None: decoder_attentions = _split_model_outputs( decoder_attentions, outputs.attentions, cur_len, newly_added_length, 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, newly_added_length ) else: decoder_hidden_states = _split_model_outputs( decoder_hidden_states, outputs.hidden_states, cur_len, newly_added_length ) unfinished_sequences = unfinished_sequences & ~stopping_criteria(input_ids, scores) this_peer_finished = unfinished_sequences.max() == 0 is_first_iteration = False if streamer is not None: streamer.end() if ( hasattr(candidate_generator, "assistant_model") and candidate_generator.assistant_model.generation_config.num_assistant_tokens_schedule == "heuristic" ): candidate_generator.assistant_model.generation_config.num_assistant_tokens = ( candidate_generator.num_assistant_tokens ) if return_dict_in_generate: if self.config.is_encoder_decoder: return GenerateEncoderDecoderOutput( sequences=input_ids, scores=scores, logits=raw_logits, encoder_attentions=encoder_attentions, encoder_hidden_states=encoder_hidden_states, decoder_attentions=decoder_attentions, cross_attentions=cross_attentions, decoder_hidden_states=decoder_hidden_states, past_key_values=model_kwargs.get("past_key_values"), ) else: return GenerateDecoderOnlyOutput( sequences=input_ids, scores=scores, logits=raw_logits, attentions=decoder_attentions, hidden_states=decoder_hidden_states, past_key_values=model_kwargs.get("past_key_values"), ) else: return input_ids def _prefill_chunking(self, input_ids: torch.LongTensor, generation_config: GenerationConfig, **model_kwargs): # Even if we are not compiling the forward, flex is always compiled when used. With chunk prefill, we may # end up needing just a bit more graphs than the default (which is 8). Doing this avoids very cryptic warnings torch._dynamo.config.cache_size_limit = 64 chunk_size = generation_config.prefill_chunk_size # Only chunk up the token just before last, so that decoding is completely performed outside this function # (here we simply prefill the cache) input_chunks = torch.split(input_ids[:, :-1], chunk_size, dim=-1) if "past_key_values" not in model_kwargs: raise ValueError("Cannot use prefill chunking without a cache") model_forward = self.forward compile_forward = self._valid_auto_compile_criteria(model_kwargs, generation_config) if compile_forward: model_forward = self.get_compiled_call(generation_config.compile_config) attention_mask = model_kwargs.pop("attention_mask", None) past_length = 0 for input_chunk in input_chunks: current_length = past_length + input_chunk.shape[-1] # Prepare inputs if attention_mask is not None: model_kwargs["attention_mask"] = attention_mask[:, :current_length] model_kwargs["cache_position"] = torch.arange( past_length, current_length, dtype=torch.long, device=input_chunk.device ) model_kwargs["position_ids"] = model_kwargs["cache_position"].unsqueeze(0) model_inputs = self.prepare_inputs_for_generation(input_chunk, **model_kwargs) outputs = model_forward(**model_inputs, return_dict=True) model_kwargs["past_key_values"] = outputs.past_key_values past_length = current_length model_kwargs["attention_mask"] = attention_mask model_kwargs["cache_position"] = model_kwargs["cache_position"][-1:] + 1 _ = model_kwargs.pop("position_ids", None) return model_kwargs def _speculative_sampling( candidate_input_ids, candidate_logits, candidate_length, new_logits, is_done_candidate, ): """ Applies sampling as in the speculative decoding paper (https://huggingface.co/papers/2211.17192, algorithm 1). Returns the selected tokens, as well as the number of candidate matches. NOTE: Unless otherwise stated, the variable names match those in the paper. """ new_candidate_input_ids = candidate_input_ids[:, -candidate_length:] # Gets the probabilities from the logits. q_i and p_i denote the assistant and model probabilities of the tokens # selected by the assistant, respectively. q = candidate_logits.softmax(dim=-1) q_i = q[:, torch.arange(candidate_length), new_candidate_input_ids].squeeze(0, 1) p = new_logits.softmax(dim=-1) p_i = p[:, torch.arange(candidate_length), new_candidate_input_ids].squeeze(0, 1) probability_ratio = p_i / q_i # When probability_ratio > 1 (i.e. q_i(x) < p_i(x), or "assistant probability of the candidate token is smaller # than the model probability for the same token"), keep the token. Otherwise reject with p = 1 - probability_ratio # (= keep with p = probability_ratio). Keep all the tokens until the first rejection r_i = torch.rand_like(probability_ratio) is_accepted = r_i <= probability_ratio n_matches = ((~is_accepted).cumsum(dim=-1) < 1).sum() # this is `n` in algorithm 1 # Ensure we don't generate beyond max_len or an EOS token (not in algorithm 1, but needed for correct behavior) if is_done_candidate and n_matches == candidate_length: # Output length is assumed to be `n_matches + 1`. Since we won't generate another token with the target model # due to acceptance on EOS we fix `n_matches` n_matches -= 1 valid_tokens = new_candidate_input_ids[:, : n_matches + 1] else: # Next token selection: if there is a rejection, adjust the distribution from the main model before sampling. gamma = candidate_logits.shape[1] p_n_plus_1 = p[:, n_matches, :] if n_matches < gamma: q_n_plus_1 = q[:, n_matches, :] p_prime = torch.clamp((p_n_plus_1 - q_n_plus_1), min=0) p_prime.div_(p_prime.sum()) else: p_prime = p_n_plus_1 t = torch.multinomial(p_prime, num_samples=1).squeeze(1)[None, :] # The selected tokens include the matches (if any) plus the next sampled tokens if n_matches > 0: valid_tokens = torch.cat((new_candidate_input_ids[:, :n_matches], t), dim=-1) else: valid_tokens = t return valid_tokens, n_matches 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