transformers/docs/source/en/model_doc/modernbert.md

This model was published in HF papers on 2024-12-18 and contributed to Hugging Face Transformers on 2024-12-19.

ModernBERT

ModernBERT is a modernized version of [BERT] trained on 2T tokens. It brings many improvements to the original architecture such as rotary positional embeddings to support sequences of up to 8192 tokens, unpadding to avoid wasting compute on padding tokens, GeGLU layers, and alternating attention.

You can find all the original ModernBERT checkpoints under the ModernBERT collection.

Tip

Click on the ModernBERT models in the right sidebar for more examples of how to apply ModernBERT to different language tasks.

The example below demonstrates how to predict the [MASK] token with [Pipeline], [AutoModel], and from the command line.

from transformers import pipeline


pipeline = pipeline(
    task="fill-mask",
    model="answerdotai/ModernBERT-base",
    device=0
)
pipeline("Plants create [MASK] through a process known as photosynthesis.")

import torch

from transformers import AutoModelForMaskedLM, AutoTokenizer


tokenizer = AutoTokenizer.from_pretrained(
    "answerdotai/ModernBERT-base",
)
model = AutoModelForMaskedLM.from_pretrained(
    "answerdotai/ModernBERT-base",
    device_map="auto",
    attn_implementation="sdpa"
)
inputs = tokenizer("Plants create [MASK] through a process known as photosynthesis.", return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = outputs.logits

masked_index = torch.where(inputs['input_ids'] == tokenizer.mask_token_id)[1]
predicted_token_id = predictions[0, masked_index].argmax(dim=-1)
predicted_token = tokenizer.decode(predicted_token_id)

print(f"The predicted token is: {predicted_token}")

Padding-free inference and training

ModernBERT supports padding-free inference and training. For example, you can leverage the [DataCollatorWithFlattening] to prepare your inputs:

Tip

Padding-free inference and training requires flash_attention_2 as the attention implementation. Since ModernBERT no longer defaults to FlashAttention2, you must explicitly set attn_implementation="flash_attention_2" when loading the model for padding-free usage.

import torch

from transformers import AutoModelForMaskedLM, AutoTokenizer, DataCollatorWithFlattening


model_id = "answerdotai/ModernBERT-base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
collator = DataCollatorWithFlattening(return_flash_attn_kwargs=True)


def prepare_text_for_padding_free(texts):
    # base tokenization with padding and subsequent flattening
    inputs_dict = tokenizer(texts, return_tensors="pt", padding=True).to(model.device)
    flattened_features = collator(
        [
            {"input_ids": i[a.bool()].tolist()}
            for i, a in zip(inputs_dict["input_ids"], inputs_dict["attention_mask"])
        ]
    )

    for k, v in flattened_features.items():
        if isinstance(v, torch.Tensor):
            flattened_features[k] = v.to(model.device)

    return flattened_features


inputs = prepare_text_for_padding_free(
    ["The capital of France is [MASK].", "ModernBERT is a [MASK] model."]
)
model = AutoModelForMaskedLM.from_pretrained(
    model_id, attn_implementation="flash_attention_2", device_map="cuda"
)

# Optional: use torch.compile for faster inference
# model.forward = torch.compile(model.forward, fullgraph=True)

out = model(**inputs)

ModernBertConfig

autodoc ModernBertConfig

ModernBertModel

autodoc ModernBertModel - forward

ModernBertForMaskedLM

autodoc ModernBertForMaskedLM - forward

ModernBertForSequenceClassification

autodoc ModernBertForSequenceClassification - forward

ModernBertForTokenClassification

autodoc ModernBertForTokenClassification - forward

ModernBertForMultipleChoice

autodoc ModernBertForMultipleChoice - forward

ModernBertForQuestionAnswering

autodoc ModernBertForQuestionAnswering - forward

Usage tips

The ModernBert model can be fine-tuned using the HuggingFace Transformers library with its official script for question-answering tasks.