Traceary

Catalog / Transformers

5.0.0

Transformers v5

7 months agobreakingaddedfixedOriginal notes

Transformers v5 release notes

  • Highlights
  • Significant API changes: dynamic weight loading, tokenization
  • Backwards Incompatible Changes
  • Bugfixes and improvements

We have a migration guide that will be continuously updated available on the main branch, please check it out in case you're facing issues: migration guide.

Highlights

We are excited to announce the initial release of Transformers v5. This is the first major release in five years, and the release is significant: 1200 commits have been pushed to main since the latest minor release. This release removes a lot of long-due deprecations, introduces several refactors that significantly simplify our APIs and internals, and comes with a large number of bug fixes.

We give an overview of our focus for this release in the following blogpost. In these release notes, we'll focus directly on the refactors and new APIs coming with v5.

This release is the full V5 release. It sets in motion something bigger: going forward, starting with v5, we'll now release minor releases every week, rather than every 5 weeks. Expect v5.1 to follow next week, then v5.2 the week that follows, etc.

We're moving forward with this change to ensure you have access to models as soon as they're supported in the library, rather than a few weeks after.

In order to install this release, please do so with the following:

pip install transformers

For us to deliver the best package possible, it is imperative that we have feedback on how the toolkit is currently working for you. Please try it out, and open an issue in case you're facing something inconsistent/a bug.

Transformers version 5 is a community endeavor, and we couldn't have shipped such a massive release without the help of the entire community.

Significant API changes

Dynamic weight loading

We introduce a new weight loading API in transformers, which significantly improves on the previous API. This weight loading API is designed to apply operations to the checkpoints loaded by transformers.

Instead of loading the checkpoint exactly as it is serialized within the model, these operations can reshape, merge, and split the layers according to how they're defined in this new API. These operations are often a necessity when working with quantization or parallelism algorithms.

This new API is centered around the new WeightConverter class:

class WeightConverter(WeightTransform):
    operations: list[ConversionOps]
    source_keys: Union[str, list[str]]
    target_keys: Union[str, list[str]]

The weight converter is designed to apply a list of operations on the source keys, resulting in target keys. A common operation done on the attention layers is to fuse the query, key, values layers. Doing so with this API would amount to defining the following conversion:

conversion = WeightConverter(
    ["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"],  # The input layers
    "self_attn.qkv_proj",  # The single layer as output
    operations=[Concatenate(dim=0)],
)

In this situation, we apply the Concatenate operation, which accepts a list of layers as input and returns a single layer.

This allows us to define a mapping from architecture to a list of weight conversions. Applying those weight conversions can apply arbitrary transformations to the layers themselves. This significantly simplified the from_pretrained method and helped us remove a lot of technical debt that we accumulated over the past few years.

This results in several improvements:

  • Much cleaner definition of transformations applied to the checkpoint
  • Reversible transformations, so loading and saving a checkpoint should result in the same checkpoint
  • Faster model loading thanks to scheduling of tensor materialization
  • Enables complex mix of transformations that wouldn't otherwise be possible (such as quantization + MoEs, or TP + MoEs)

Linked PR: https://github.com/huggingface/transformers/pull/41580

Tokenization

Just as we moved towards a single backend library for model definition, we want our tokenizers, and the Tokenizer object to be a lot more intuitive. With v5, tokenizer definition is much simpler; one can now initialize an empty LlamaTokenizer and train it directly on your corpus.

Defining a new tokenizer object should be as simple as this:

from transformers import TokenizersBackend, generate_merges
from tokenizers import pre_tokenizers, Tokenizer
from tokenizers.model import BPE

class Llama5Tokenizer(TokenizersBackend):
    def __init__(self, unk_token="<unk>",bos_token="<s>", eos_token="</s>", vocab=None, merges=None ):
        if vocab is None:
            self._vocab = {
                str(unk_token): 0,
                str(bos_token): 1,
                str(eos_token): 2,
            }

        else:
            self._vocab = vocab

            self._merges = merges

        self._tokenizer = Tokenizer(
            BPE(vocab=self._vocab, merges=self._merges, fuse_unk=True)
        )
        self._tokenizer.pre_tokenizer = pre_tokenizers.Metaspace(
            replacement="▁", prepend_scheme=_get_prepend_scheme(self.add_prefix_space, self), split=False
        )
        super().__init__(
            tokenizer_object=self._tokenizer,
            unk_token=unk_token,
            bos_token=bos_token,
            eos_token=eos_token,
        )

Once the tokenizer is defined as above, you can load it with the following: Llama5Tokenizer(). Doing this returns you an empty, trainable tokenizer that follows the definition of the authors of Llama5 (it does not exist yet :wink:).

The above is the main motivation towards refactoring tokenization: we want tokenizers to behave similarly to models: trained or empty, and with exactly what is defined in their class definition.

Backend Architecture Changes: moving away from the slow/fast tokenizer separation

Up to now, transformers maintained two parallel implementations for many tokenizers:

  • "Slow" tokenizers (tokenization_<model>.py) - Python-based implementations, often using SentencePiece as the backend.
  • "Fast" tokenizers (tokenization_<model>_fast.py) - Rust-based implementations using the 🤗 tokenizers library.

In v5, we consolidate to a single tokenizer file per model: tokenization_<model>.py. This file will use the most appropriate backend available:

  1. TokenizersBackend (preferred): Rust-based tokenizers from the 🤗 tokenizers library. In general it provides optimal performance, but it also offers a lot more features that are commonly adopted across the ecosystem:
  • handling additional tokens
  • a full python API for setting and updating
  • automatic parallelization,
  • automatic offsets
  • customization
  • training
  1. SentencePieceBackend: for tokenizers requiring the sentencepiece library. It inherits from PythonBackend.
  2. PythonBackend: a Python implementations of the features provided by tokenizers. Basically allows adding tokens.
  3. MistralCommonBackend: relies on MistralCommon's tokenization library. (Previously known as the MistralCommonTokenizer)

The AutoTokenizer automatically selects the appropriate backend based on available files and dependencies. This is transparent, you continue to use AutoTokenizer.from_pretrained() as before. This allows transformers to be future-proof and modular to easily support future backends.

Defining a tokenizers outside of the existing backends

We enable users and tokenizer builders to define their own tokenizers from top to bottom. Tokenizers are usually defined using a backend such as tokenizers, sentencepiece or mistral-common, but we offer the possibility to design the tokenizer at a higher-level, without relying on those backends.

To do so, you can import the PythonBackend (which was previously known as PreTrainedTokenizer). This class encapsulates all the logic related to added tokens, encoding, and decoding.

If you want something even higher up the stack, then PreTrainedTokenizerBase is what PythonBackend inherits from. It contains the very basic tokenizer API features:

  • encode
  • decode
  • vocab_size
  • get_vocab
  • convert_tokens_to_ids
  • convert_ids_to_tokens
  • from_pretrained
  • save_pretrained
  • among a few others

API Changes

1. Direct tokenizer initialization with vocab and merges

Starting with v5, we now enable initializing blank, untrained tokenizers-backed tokenizers:

from transformers import LlamaTokenizer

tokenizer = LlamaTokenizer()

This tokenizer will therefore follow the definition of the LlamaTokenizer as defined in its class definition. It can then be trained on a corpus as can be seen in the tokenizers documentation.

These tokenizers can also be initialized from vocab and merges (if necessary), like the previous "slow" tokenizers:

from transformers import LlamaTokenizer

vocab = {"<unk>": 0, "<s>": 1, "</s>": 2, "hello": 3, "world": 4}
merges = [("h", "e"), ("l", "l"), ("o", " ")]

tokenizer = LlamaTokenizer(vocab=vocab, merges=merges)

This tokenizer will behave as a Llama-like tokenizer, with an updated vocabulary. This allows comparing different tokenizer classes with the same vocab; therefore enabling the comparison of different pre-tokenizers, normalizers, etc.

⚠️ The vocab_file (as in, a path towards a file containing the vocabulary) cannot be used to initialize the LlamaTokenizer as loading from files is reserved to the from_pretrained method.

2. Simplified decoding API

The batch_decode and decode methods have been unified to reflect behavior of the encode method. Both single and batch decoding now use the same decode method. See an example of the new behavior below:

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("t5-small") 
inputs = ["hey how are you?", "fine"]
tokenizer.decode(tokenizer.encode(inputs))

Gives:

- 'hey how are you?</s> fine</s>'
+ ['hey how are you?</s>', 'fine</s>']

We expect encode and decode to behave, as two sides of the same coin: encode, process, decode, should work.

[!NOTE] A common use-case would be: encode, model.generate, decode. However, using generate would return list[list[int]], which would then be incompatible with decode.

3. Unified encoding API

The encode_plus method is deprecated in favor of the single __call__ method.

4. apply_chat_template returns BatchEncoding

Previously, apply_chat_template returned input_ids for backward compatibility. Starting with v5, it now consistently returns a BatchEncoding dict like other tokenizer methods.

# v5
messages = [
    {"role": "user", "content": "Hello!"},
    {"role": "assistant", "content": "Hi there!"}
]

# Now returns BatchEncoding with input_ids, attention_mask, etc.
outputs = tokenizer.apply_chat_template(messages, return_tensors="pt")
print(outputs.keys())  # dict_keys(['input_ids', 'attention_mask'])

5. Removed legacy configuration file saving:

We simplify the serialization of tokenization attributes:

  • special_tokens_map.json - special tokens are now stored in tokenizer_config.json.
  • added_tokens.json - added tokens are now stored in tokenizer.json.
  • added_tokens_decoder is only stored when there is no tokenizer.json.

When loading older tokenizers, these files are still read for backward compatibility, but new saves use the consolidated format. We're gradually moving towards consolidating attributes to fewer files so that other libraries and implementations may depend on them more reliably.

6. Model-Specific Changes

Several models that had identical tokenizers now import from their base implementation:

  • LayoutLM → uses BertTokenizer
  • LED → uses BartTokenizer
  • Longformer → uses RobertaTokenizer
  • LXMert → uses BertTokenizer
  • MT5 → uses T5Tokenizer
  • MVP → uses BartTokenizer

These modules will eventually be removed altogether.

Removed T5-specific workarounds

The internal _eventually_correct_t5_max_length method has been removed. T5 tokenizers now handle max length consistently with other models.

Testing Changes

A few testing changes specific to tokenizers have been applied:

  • Model-specific tokenization test files now focus on integration tests.
  • Common tokenization API tests (e.g., add_tokens, encode, decode) are now centralized and automatically applied across all tokenizers. This reduces test duplication and ensures consistent behavior

For legacy implementations, the original BERT Python tokenizer code (including WhitespaceTokenizer, BasicTokenizer, etc.) is preserved in bert_legacy.py for reference purposes.

7. Deprecated / Modified Features

Special Tokens Structure:

  • SpecialTokensMixin: Merged into PreTrainedTokenizerBase to simplify the tokenizer architecture.
  • special_tokens_map: Now only stores named special token attributes (e.g., bos_token, eos_token). Use extra_special_tokens for additional special tokens (formerly additional_special_tokens). all_special_tokens includes both named and extra tokens.
# v4
tokenizer.special_tokens_map  # Included 'additional_special_tokens'

# v5
tokenizer.special_tokens_map  # Only named tokens
tokenizer.extra_special_tokens  # Additional tokens
  • special_tokens_map_extended and all_special_tokens_extended: Removed. Access AddedToken objects directly from _special_tokens_map or _extra_special_tokens if needed.
  • additional_special_tokens: Still accepted for backward compatibility but is automatically converted to extra_special_tokens.

Deprecated Methods:

  • sanitize_special_tokens(): Already deprecated in v4, removed in v5.
  • prepare_seq2seq_batch(): Deprecated; use __call__() with text_target parameter instead.
# v4
model_inputs = tokenizer.prepare_seq2seq_batch(src_texts, tgt_texts, max_length=128)

# v5
model_inputs = tokenizer(src_texts, text_target=tgt_texts, max_length=128, return_tensors="pt")
model_inputs["labels"] = model_inputs.pop("input_ids_target")
  • BatchEncoding.words(): Deprecated; use word_ids() instead.

Removed Methods:

  • create_token_type_ids_from_sequences(): Removed from base class. Subclasses that need custom token type ID creation should implement this method directly.
  • prepare_for_model(), build_inputs_with_special_tokens(), truncate_sequences(): Moved from tokenization_utils_base.py to tokenization_python.py for PythonBackend tokenizers. TokenizersBackend provides model-ready input via tokenize() and encode(), so these methods are no longer needed in the base class.
  • _switch_to_input_mode(), _switch_to_target_mode(), as_target_tokenizer(): Removed from base class. Use __call__() with text_target parameter instead.
# v4
with tokenizer.as_target_tokenizer():
    labels = tokenizer(tgt_texts, ...)

# v5
labels = tokenizer(text_target=tgt_texts, ...)
  • parse_response(): Removed from base class.

Performance

MoE Performance

The v5 release significantly improves the performance of the MoE models, as can be seen in the graphs below. We improve and optimize MoE performance through batched and grouped experts implementations, and we optimize them for decoding using batched_mm.

Core performance

We focus on improving the performance of loading weights on device (which gives speedups up to 6x in tensor parallel situations); this is preliminary work that we'll continue to work on in the coming weeks. Some notable improvements:

Library-wide changes with lesser impact

Default dtype update

We have updated the default dtype for all models loaded with from_pretrained to be auto. This will lead to model instantiations respecting the dtype in which the model was saved, rather than forcing it to load in float 32.

You can, of course, still specify the dtype in which you want to load your model by specifying it as an argument to the from_pretrained method.

Shard size

The Hugging Face Hub infrastructure has gradually moved to a XET backend. This will significantly simplify uploads and downloads, with higher download and upload speeds, partial uploads, and, most notably, a higher threshold for accepted file sizes on the Hugging Face Hub.

To reflect this, we're increasing the default shard size of models serialized on the Hub to 50GB (up from 5GB).

use_auth_token

The use_auth_token argument/parameter is deprecated in favor of token everywhere. You should be able to search and replace use_auth_token with token and get the same logic.

Linked PR: https://github.com/huggingface/transformers/pull/41666

Attention-related features

We decided to remove some features for the upcoming v5 as they are currently only supported in a few old models and no longer integrated in current model additions. It's recommended to stick to v4.x in case you need them. Following features are affected:

  • No more head masking, see #41076. This feature allowed to turn off certain heads during the attention calculation and only worked for eager.
  • No more relative positional biases in Bert-like models, see #41170. This feature was introduced to allow relative position scores within attention calculations (similar to T5). However, this feature is barely used in official models and a lot of complexity instead. It also only worked with eager.
  • No more head pruning, see #41417 by @gante. As the name suggests, it allowed to prune heads within your attention layers.

Updates to supported torch APIs

We dropped support for two torch APIs:

Those APIs were deprecated by the PyTorch team, and we're instead focusing on the supported APIs dynamo and export.

Quantization changes

We clean up the quantization API in transformers, and significantly refactor the weight loading as highlighted above.

We drop support for two quantization arguments that have been deprecated for some time:

  • load_in_4bit
  • load_in_8bit

We remove them in favor of the quantization_config argument which is much more complete. As an example, here is how you would load a 4-bit bitsandbytes model using this argument:

from transformers import AutoModelForCausalLM, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(load_in_4bit=True)

model_4bit = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.2-3B",
    device_map="auto",
    quantization_config=quantization_config
)

Configuration

  • Methods to init a nested config such as from_xxx_config are deleted. Configs can be init from the __init__ method in the same way. See #41314.
  • It is no longer possible to load a config class from a URL file. Configs must be loaded from either a local path or a repo on the Hub. See #42383.
  • All parameters for configuring model's rotary embedding are now stored under mode.rope_parameters, including the rope_theta and rope_type. Model's config.rope_parameters is a simple dictionaty in most cases, and can also be a nested dict in special cases (i.e. Gemma3 and ModernBert) with different rope parameterization for each layer type. Trying to get config.rope_theta will throw an attribute error from now on. See #39847 and #42255
  • Qwen-VL family configuration is in a nested format and trying to access keys directly will throw an error (e.g. config.vocab_size). Users are expected to access keys from their respective sub-configs (config.text_config.vocab_size).
  • Configurations of non-generative models (any model that doesn't call model.generate()) will no longer have a generation_config and model.config.generation_config will throw an attribute error.

Processing

Tokenization

  • Slow tokenizer files (aka: tokenization_<model>.py ) will be removed in favor of using fast tokenizer files tokenization_<model>_fast.py --> will be renamed to tokenization_<model>.py. As fast tokenizers are :hugs:tokenizers - backend, they include a wider range of features that are maintainable and reliable.
  • Other backends (sentence piece, tokenizers, etc.) will be supported with a light layer if loading a fast tokenizer fails
  • Remove legacy files like special_tokens_map.json and added_tokens.json
  • Remove _eventually_correct_t5_max_length
  • encode_plus --> __call__
  • batch_decode --> decode

apply_chat_template by default returns naked input_ids rather than a BatchEncoding dict. This was inconvenient - it should return a BatchEncoding dict like tokenizer.__call__(), but we were stuck with it for backward compatibility. The method now returns a BatchEncoding.

Linked PRs:

Processing classes

Modeling

  • Some RotaryEmbeddings layers will start returning a dict of tuples, in case the model uses several RoPE configurations (Gemma2, ModernBert). Each value will be a tuple of "cos, sin" per RoPE type.
  • Config attribute for RotaryEmbeddings layer will be unified and accessed via config.rope_parameters. Config attr for rope_theta might not be accessible anymore for some models, and instead will be in config.rope_parameters['rope_theta']. BC will be supported for a while as much as possible, and in the near future we'll gradually move to the new RoPE format (https://github.com/huggingface/transformers/pull/39847)
  • Vision Language models will not have a shortcut access to its language and vision component from the generative model via model.language_model. It is recommended to either access the module with model.model.language_model or model.get_decoder(). See #42156
  • All models now accept kwargs in their forward methods

Generate

  • Old, deprecated output type aliases were removed (e.g. GreedySearchEncoderDecoderOutput). We now only have 4 output classes built from the following matrix: decoder-only vs encoder-decoder, uses beams vs doesn't use beams (https://github.com/huggingface/transformers/pull/40998)
  • Removed deprecated classes regarding decoding methods that were moved to the Hub due to low usage (constraints and beam scores) (https://github.com/huggingface/transformers/pull/41223)
  • If generate doesn't receive any KV Cache argument, the default cache class used is now defined by the model (as opposed to always being DynamicCache) (https://github.com/huggingface/transformers/pull/41505)
  • Generation parameters are no longer accessible via model's config. If generation paramaters are serialized in config.json for any old model, it will be loaded back into model's generation config. Users are expected to access or modify generation parameters only with model.generation_config.do_sample = True.

Trainer

New Features

  • ALST/Ulysses Sequence Parallelism Integration
    • Added sequence parallelism support via HF Accelerate for training with longer sequences. Enables splitting sequences across devices using ALST (All-to-All Long Sequence Training) and Ulysses algorithms with DeepSpeed.
  • Improved compute_loss_func Handling
    • compute_loss_func now always takes priority over the model's built-in loss computation, giving users consistent control over custom loss functions.
  • num_items_in_batch in Prediction Step
    • The num_items_in_batch argument is now passed to compute_loss during prediction_step, enabling proper loss scaling during evaluation.

Breaking Changes

  • report_to now defaults to "none"
    • Logging integrations are no longer auto-detected by default; users must explicitly specify which reporting backends to use.

Removing arguments without deprecation cycle in TrainingArguments due to low usage

  • mp_parameters -> legacy param that was later on added to the Sagemaker trainer
  • _n_gpu -> not intended for users to set, we will initialize it correctly instead of putting it in the TrainingArguments
  • overwrite_output_dir - > replaced by resume_from_checkpoint, and it was only used in the examples script, no impact on Trainer.
  • logging_dir -> only used for tensorboard, set TENSORBOARD_LOGGING_DIR env var instead
  • jit_mode_eval -> use use_torch_compile instead, as torchscript is not recommended anymore
  • tpu_num_cores-> It is actually better to remove it, as it is not recommended to set the number of cores. By default, all TPU cores are used . Set TPU_NUM_CORES env var instead
  • past_index -> it was only used for a very small number of models that have special architecture like transformersxl + it was not documented at all how to train those models
  • ray_scope -> only for a minor arg for ray integration. Set RAY_SCOPE var env instead
  • warmup_ratio -> use warmup_step instead. We combined both args together by allowing passing float values in warmup_step.

Removing deprecated arguments in TrainingArguments

  • fsdp_min_num_params and fsdp_transformer_layer_cls_to_wrap -> use fsdp_config
  • tpu_metrics_debug -> debug
  • push_to_hub_token -> hub_token
  • push_to_hub_model_id and push_to_hub_organization -> hub_model_id
  • include_inputs_for_metrics -> include_for_metrics
  • per_gpu_train_batch_size -> per_device_train_batch_size
  • per_gpu_eval_batch_size -> per_device_eval_batch_size
  • use_mps_device -> mps will be used by default if detected
  • fp16_backend and half_precision_backend -> we will only rely on torch.amp as everything has been upstreamed to torch
  • no_cuda -> use_cpu
  • include_tokens_per_second -> include_num_input_tokens_seen
  • use_legacy_prediction_loop -> we only use evaluation_loop function from now on

Removing deprecated arguments in Trainer

  • tokenizer in initialization -> processing_class
  • model_path in train() -> resume_from_checkpoint

Removed features for Trainer

  • sigpot integration for hp search was removed as the library was archived + the api stopped working
  • drop support for sagemaker API <1.10
  • bump accelerate minimum version to 1.1.0
  • bump peft minimum version to 0.18.0
  • bump bitsandbytes minimum version to 0.46.1

New defaults for Trainer

  • use_cache in the model config will be set to False. You can still change the cache value through TrainingArguments usel_cache argument if needed.

Pipeline

  • Image text to text pipelines will no longer accept images as a separate argument along with conversation chats. Image data has to be embedded in the chat's "content" field. See #42359

PushToHubMixin

  • removed deprecated organization and repo_url from PushToHubMixin. You must pass a repo_id instead.
  • removed ignore_metadata_errors from PushToMixin. In practice if we ignore errors while loading the model card, we won't be able to push the card back to the Hub so it's better to fail early and not provide the option to fail later.
  • push_to_hub do not accept **kwargs anymore. All accepted parameters are explicitly documented.
  • arguments of push_to_hub are now keyword-only to avoid confusion. Only repo_id can be positional since it's the main arg.
  • removed use_temp_dir argument from push_to_hub. We now use a tmp dir in all cases.

Linked PR: https://github.com/huggingface/transformers/pull/42391.

CLI

The deprecated transformers-cli ... command was deprecated, transformers ... is now the only CLI entry point.

transformers CLI has been migrated to Typer, making it easier to maintain + adding some nice features out of the box (improved --help section, autocompletion).

Biggest breaking change is in transformers chat. This command starts a terminal UI to interact with a chat model. It used to also be able to start a Chat Completion server powered by transformers and chat with it. In this revamped version, this feature has been removed in favor of transformers serve. The goal of splitting transformers chat and transformers serve is to define clear boundaries between client and server code. It helps with maintenance but also makes the commands less bloated. The new signature of transformers chat is:

Usage: transformers chat [OPTIONS] BASE_URL MODEL_ID [GENERATE_FLAGS]...

Chat with a model from the command line.

It works hand in hand with transformers serve, which means that if transformers serve is running on its default endpoint, transformers chat can be launched as follows:

transformers chat HuggingFaceTB/SmolLM3-3B

It can however use any OpenAI API compatible HTTP endpoint:

transformers chat HuggingFaceTB/SmolLM3-3B https://router.huggingface.co/v1

Linked PRs:

Removal of the run method

The transformers run (previously transformers-cli run) is an artefact of the past, was not documented nor tested, and isn't part of any public documentation. We're removing it for now and ask you to please let us know in case this is a method you are using; in which case we should bring it back with better support.

Linked PR:

These notes run past the length kept in the archive. The rest is on the publisher’s page.