more fixes

docs/source/en/_toctree.yml

@@ -445,6 +445,8 @@
         title: ErnieM
       - local: model_doc/esm
         title: ESM
+      - local: model_doc/eurobert
+        title: EuroBERT
       - local: model_doc/falcon
         title: Falcon
       - local: model_doc/falcon3

docs/source/en/index.md

@@ -19,7 +19,6 @@ rendered properly in your Markdown viewer.
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_as_a_model_definition.png"/>
 </h3>
 
-
 Transformers acts as the model-definition framework for state-of-the-art machine learning models in text, computer
 vision, audio, video, and multimodal model, for both inference and training.
 
@@ -61,4 +60,5 @@ Transformers is designed for developers and machine learning engineers and resea
 
 ## Learn
 
-If you're new to Transformers or want to learn more about transformer models, we recommend starting with the [LLM course](https://huggingface.co/learn/llm-course/chapter1/1?fw=pt). This comprehensive course covers everything from the fundamentals of how transformer models work to practical applications across various tasks. You'll learn the complete workflow, from curating high-quality datasets to fine-tuning large language models and implementing reasoning capabilities. The course contains both theoretical and hands-on exercises to build a solid foundational knowledge of transformer models as you learn.
+If you're new to Transformers or want to learn more about transformer models, we recommend starting with the [LLM course](https://huggingface.co/learn/llm-course/chapter1/1?fw=pt). This comprehensive course covers everything from the fundamentals of how transformer models work to practical applications across various tasks. You'll learn the complete workflow, from curating high-quality datasets to fine-tuning large language models and implementing reasoning capabilities. The course contains both theoretical and hands-on exercises to build a solid foundational knowledge of transformer models as you learn.
+

docs/source/en/model_doc/eurobert.md

@@ -0,0 +1,53 @@
+<!--Copyright 2024 The HuggingFace Team. 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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# EuroBERT
+
+<div class="flex flex-wrap space-x-1">
+<a href="https://huggingface.co/models?filter=eurobert">
+<img alt="Models" src="https://img.shields.io/badge/All_model_pages-eurobert-blueviolet">
+</a>
+
+## Overview
+
+EuroBERT is a multilingual encoder model based on a refreshed transformer architecture. It supports a mixture of European and widely spoken languages.
+
+Check out all the available models [here](https://huggingface.co/EuroBERT).
+
+## EuroBertConfig
+
+[[autodoc]] EuroBertConfig
+
+<frameworkcontent>
+<pt>
+
+## EuroBertModel
+
+[[autodoc]] EuroBertModel
+    - forward
+
+## EuroBertForMaskedLM
+
+[[autodoc]] EuroBertForMaskedLM
+    - forward
+
+## EuroBertForSequenceClassification
+
+[[autodoc]] EuroBertForSequenceClassification
+    - forward
+
+</pt>
+</frameworkcontent>

src/transformers/models/auto/configuration_auto.py

@@ -130,6 +130,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("ernie", "ErnieConfig"),
         ("ernie_m", "ErnieMConfig"),
         ("esm", "EsmConfig"),
+        ("eurobert", "EuroBertConfig"),
         ("falcon", "FalconConfig"),
         ("falcon_h1", "FalconH1Config"),
         ("falcon_mamba", "FalconMambaConfig"),
@@ -519,6 +520,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("ernie", "ERNIE"),
         ("ernie_m", "ErnieM"),
         ("esm", "ESM"),
+        ("eurobert", "EuroBERT"),
         ("falcon", "Falcon"),
         ("falcon3", "Falcon3"),
         ("falcon_h1", "FalconH1"),

src/transformers/models/auto/modeling_auto.py

@@ -121,6 +121,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("ernie", "ErnieModel"),
         ("ernie_m", "ErnieMModel"),
         ("esm", "EsmModel"),
+        ("eurobert", "EuroBertModel"),
         ("falcon", "FalconModel"),
         ("falcon_h1", "FalconH1Model"),
         ("falcon_mamba", "FalconMambaModel"),
@@ -975,6 +976,7 @@ MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
         ("electra", "ElectraForMaskedLM"),
         ("ernie", "ErnieForMaskedLM"),
         ("esm", "EsmForMaskedLM"),
+        ("eurobert", "EuroBertForMaskedLM"),
         ("flaubert", "FlaubertWithLMHeadModel"),
         ("fnet", "FNetForMaskedLM"),
         ("funnel", "FunnelForMaskedLM"),
@@ -1125,6 +1127,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("ernie", "ErnieForSequenceClassification"),
         ("ernie_m", "ErnieMForSequenceClassification"),
         ("esm", "EsmForSequenceClassification"),
+        ("eurobert", "EuroBertForSequenceClassification"),
         ("falcon", "FalconForSequenceClassification"),
         ("flaubert", "FlaubertForSequenceClassification"),
         ("fnet", "FNetForSequenceClassification"),

src/transformers/models/eurobert/__init__.py

@@ -0,0 +1,27 @@
+# Copyright 2025 Nicolas Boizard, Duarte M. Alves, Hippolyte Gisserot-Boukhlef and the EuroBert team. 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.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_eurobert import *
+    from .modeling_eurobert import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/eurobert/configuration_eurobert.py

@@ -0,0 +1,206 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/eurobert/modular_eurobert.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_eurobert.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 Nicolas Boizard, Duarte M. Alves, Hippolyte Gisserot-Boukhlef and the EuroBert team. 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.
+
+from ..llama import LlamaConfig
+
+
+class EuroBertConfig(LlamaConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`EuroBertModel`]. It is used to instantiate an EuroBert
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the EuroBERT-210m.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 128256):
+            Vocabulary size of the EuroBert model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`EuroBertModel`]
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the encoder and pooler.
+        max_position_embeddings (`int`, *optional*, defaults to 8192):
+            The maximum sequence length that this model might ever be used with. EuroBert supports up to 8192 tokens,
+            EuroBert-pretrained up to 2048.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the rms normalization layers.
+        bos_token_id (`int`, *optional*, defaults to 128000):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 128001):
+            End of stream token id.
+        pad_token_id (`int`, *optional*, defaults to 128001):
+            Padding token id.
+        mask_token_id (`int`, *optional*, defaults to 128002):
+            Mask token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to
+            understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
+            results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 250000.0):
+            The base period of the RoPE embeddings. EuroBert used base period of 250000.0,
+            EuroBert-pretrained 10000.0.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'eurobert3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'eurobert3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'eurobert3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'eurobert3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
+        head_dim (`int`, *optional*):
+            The attention head dimension. If None, it will default to hidden_size // num_attention_heads
+        classifier_pooling (`str`, *optional*, defaults to `"late"`):
+            The pooling strategy to use for the classifier. Can be one of ['bos', 'mean', 'late'].
+
+    ```python
+    >>> from transformers import EuroBertModel, EuroBertConfig
+
+    >>> # Initializing a EuroBert eurobert-base style configuration
+    >>> configuration = EuroBertConfig()
+
+    >>> # Initializing a model from the eurobert-base style configuration
+    >>> model = EuroBertModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "eurobert"
+
+    def __init__(
+        self,
+        vocab_size=128256,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=8192,
+        initializer_range=0.02,
+        rms_norm_eps=1e-05,
+        bos_token_id=128000,
+        eos_token_id=128001,
+        pad_token_id=128001,
+        mask_token_id=128002,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=250000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        head_dim=None,
+        classifier_pooling="late",
+        use_cache=False,
+        **kwargs,
+    ):
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        super().__init__(
+            vocab_size=vocab_size,
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_hidden_layers=num_hidden_layers,
+            num_attention_heads=num_attention_heads,
+            num_key_value_heads=num_key_value_heads,
+            hidden_act=hidden_act,
+            max_position_embeddings=max_position_embeddings,
+            initializer_range=initializer_range,
+            rms_norm_eps=rms_norm_eps,
+            use_cache=False,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pad_token_id=pad_token_id,
+            pretraining_tp=pretraining_tp,
+            tie_word_embeddings=tie_word_embeddings,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            attention_bias=attention_bias,
+            attention_dropout=attention_dropout,
+            mlp_bias=mlp_bias,
+            head_dim=head_dim,
+            **kwargs,
+        )
+        self.mask_token_id = mask_token_id
+        self.clf_pooling = classifier_pooling
+
+
+__all__ = ["EuroBertConfig"]

src/transformers/models/eurobert/modeling_eurobert.py

@@ -0,0 +1,644 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/eurobert/modular_eurobert.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_eurobert.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 Nicolas Boizard, Duarte M. Alves, Hippolyte Gisserot-Boukhlef and the EuroBert team. 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.
+
+from typing import Callable, Optional, Union
+
+import torch
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...cache_utils import Cache
+from ...generation import GenerationMixin
+from ...integrations import use_kernel_forward_from_hub
+from ...modeling_attn_mask_utils import AttentionMaskConverter
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_layers import GradientCheckpointingLayer
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
+from ...utils import auto_docstring, can_return_tuple
+from ...utils.generic import TransformersKwargs, check_model_inputs
+from .configuration_eurobert import EuroBertConfig
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class EuroBertRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-5):
+        """
+        EuroBertRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class EuroBertAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: EuroBertConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = False
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            is_causal=False,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class EuroBertMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class EuroBertDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: EuroBertConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = EuroBertAttention(config=config, layer_idx=layer_idx)
+
+        self.mlp = EuroBertMLP(config)
+        self.input_layernorm = EuroBertRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = EuroBertRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor]:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+@auto_docstring
+class EuroBertPreTrainedModel(PreTrainedModel):
+    config: EuroBertConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["EuroBertDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    _supports_static_cache = True
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": EuroBertDecoderLayer,
+        "attentions": EuroBertAttention,
+    }
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, EuroBertRMSNorm):
+            module.weight.data.fill_(1.0)
+
+
+class EuroBertRotaryEmbedding(nn.Module):
+    def __init__(self, config: EuroBertConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+@auto_docstring
+class EuroBertModel(EuroBertPreTrainedModel):
+    def __init__(self, config: EuroBertConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [EuroBertDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = EuroBertRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = EuroBertRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+        self.mask_converter = AttentionMaskConverter(is_causal=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[tuple, BaseModelOutput]:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if attention_mask is not None and self.config._attn_implementation != "flash_attention_2":
+            mask = self.mask_converter.to_4d(attention_mask, attention_mask.shape[1], inputs_embeds.dtype)
+        else:
+            mask = attention_mask
+
+        if position_ids is None:
+            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device).unsqueeze(0)
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        for encoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = encoder_layer(
+                hidden_states,
+                attention_mask=mask,
+                position_ids=position_ids,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+        hidden_states = self.norm(hidden_states)
+
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+        )
+
+
+@auto_docstring
+class EuroBertForMaskedLM(EuroBertPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config: EuroBertConfig):
+        super().__init__(config)
+        self.model = EuroBertModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, config.mlp_bias)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, EuroBertForMaskedLM
+
+        >>> model = EuroBertForMaskedLM.from_pretrained("meta-eurobert/EuroBert-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-eurobert/EuroBert-2-7b-hf")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class EuroBertForSequenceClassification(EuroBertPreTrainedModel):
+    def __init__(self, config: EuroBertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.clf_pooling = config.clf_pooling
+
+        self.model = EuroBertModel(config)
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.GELU()
+        self.classifier = nn.Linear(config.hidden_size, self.num_labels)
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+    ) -> Union[tuple[torch.Tensor], SequenceClassifierOutput]:
+        encoder_output = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+        )
+        last_hidden_state = encoder_output[0]
+
+        if self.clf_pooling in ["bos", "mean"]:
+            if self.clf_pooling == "bos":
+                pooled_output = last_hidden_state[:, 0]
+
+            elif self.clf_pooling == "mean":
+                if attention_mask is None:
+                    pooled_output = last_hidden_state.mean(dim=1)
+                else:
+                    pooled_output = (last_hidden_state * attention_mask.unsqueeze(-1)).sum(dim=1)
+                    pooled_output /= attention_mask.sum(dim=1, keepdim=True)
+
+            pooled_output = self.dense(pooled_output)
+            pooled_output = self.activation(pooled_output)
+            logits = self.classifier(pooled_output)
+
+        elif self.clf_pooling == "late":
+            x = self.dense(last_hidden_state)
+            x = self.activation(x)
+            logits = self.classifier(x)
+            if attention_mask is None:
+                logits = logits.mean(dim=1)
+            else:
+                logits = (logits * attention_mask.unsqueeze(-1)).sum(dim=1)
+                logits /= attention_mask.sum(dim=1, keepdim=True)
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=encoder_output.hidden_states,
+            attentions=encoder_output.attentions,
+        )
+
+
+class EuroBertForTokenClassification(EuroBertPreTrainedModel):
+    def __init__(self, config: EuroBertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = EuroBertModel(config)
+
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+    ) -> Union[tuple, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "EuroBertPreTrainedModel",
+    "EuroBertModel",
+    "EuroBertForMaskedLM",
+    "EuroBertForSequenceClassification",
+    "EuroBertForTokenClassification",
+]

src/transformers/models/eurobert/modular_eurobert.py

@@ -0,0 +1,480 @@
+# coding=utf-8
+# Copyright 2025 Nicolas Boizard, Duarte M. Alves, Hippolyte Gisserot-Boukhlef and the EuroBert team. 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.
+
+from typing import Callable, Optional, Union
+
+import torch
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...modeling_attn_mask_utils import AttentionMaskConverter
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_outputs import BaseModelOutput, SequenceClassifierOutput, TokenClassifierOutput
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
+from ...processing_utils import Unpack
+from ...utils import (
+    logging,
+)
+from ...utils.generic import TransformersKwargs
+from ..llama import LlamaConfig
+from ..llama.modeling_llama import (
+    LlamaAttention,
+    LlamaForCausalLM,
+    LlamaModel,
+    LlamaPreTrainedModel,
+    LlamaRMSNorm,
+    apply_rotary_pos_emb,
+    eager_attention_forward,
+)
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "EuroBERT/EuroBERT-210m"
+_CONFIG_FOR_DOC = "EuroBertConfig"
+
+
+class EuroBertConfig(LlamaConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`EuroBertModel`]. It is used to instantiate an EuroBert
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the EuroBERT-210m.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 128256):
+            Vocabulary size of the EuroBert model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`EuroBertModel`]
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the encoder and pooler.
+        max_position_embeddings (`int`, *optional*, defaults to 8192):
+            The maximum sequence length that this model might ever be used with. EuroBert supports up to 8192 tokens,
+            EuroBert-pretrained up to 2048.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the rms normalization layers.
+        bos_token_id (`int`, *optional*, defaults to 128000):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 128001):
+            End of stream token id.
+        pad_token_id (`int`, *optional*, defaults to 128001):
+            Padding token id.
+        mask_token_id (`int`, *optional*, defaults to 128002):
+            Mask token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to
+            understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
+            results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 250000.0):
+            The base period of the RoPE embeddings. EuroBert used base period of 250000.0,
+            EuroBert-pretrained 10000.0.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'eurobert3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'eurobert3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'eurobert3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'eurobert3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
+        head_dim (`int`, *optional*):
+            The attention head dimension. If None, it will default to hidden_size // num_attention_heads
+        classifier_pooling (`str`, *optional*, defaults to `"late"`):
+            The pooling strategy to use for the classifier. Can be one of ['bos', 'mean', 'late'].
+
+    ```python
+    >>> from transformers import EuroBertModel, EuroBertConfig
+
+    >>> # Initializing a EuroBert eurobert-base style configuration
+    >>> configuration = EuroBertConfig()
+
+    >>> # Initializing a model from the eurobert-base style configuration
+    >>> model = EuroBertModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "eurobert"
+
+    def __init__(
+        self,
+        vocab_size=128256,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=8192,
+        initializer_range=0.02,
+        rms_norm_eps=1e-05,
+        bos_token_id=128000,
+        eos_token_id=128001,
+        pad_token_id=128001,
+        mask_token_id=128002,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=250000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        head_dim=None,
+        classifier_pooling="late",
+        use_cache=False,
+        **kwargs,
+    ):
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        super().__init__(
+            vocab_size=vocab_size,
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_hidden_layers=num_hidden_layers,
+            num_attention_heads=num_attention_heads,
+            num_key_value_heads=num_key_value_heads,
+            hidden_act=hidden_act,
+            max_position_embeddings=max_position_embeddings,
+            initializer_range=initializer_range,
+            rms_norm_eps=rms_norm_eps,
+            use_cache=False,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pad_token_id=pad_token_id,
+            pretraining_tp=pretraining_tp,
+            tie_word_embeddings=tie_word_embeddings,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            attention_bias=attention_bias,
+            attention_dropout=attention_dropout,
+            mlp_bias=mlp_bias,
+            head_dim=head_dim,
+            **kwargs,
+        )
+        self.mask_token_id = mask_token_id
+        self.clf_pooling = classifier_pooling
+
+
+class EuroBertRMSNorm(LlamaRMSNorm):
+    def __init__(self, hidden_size, eps=1e-5):
+        super().__init__(hidden_size, eps)
+
+
+# llama without cache
+class EuroBertAttention(LlamaAttention):
+    def __init__(self, config: EuroBertConfig, layer_idx: int):
+        super().__init__(config, layer_idx)
+        self.is_causal = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            is_causal=False,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class EuroBertPreTrainedModel(LlamaPreTrainedModel):
+    pass
+
+
+class EuroBertModel(LlamaModel):
+    def __init__(self, config: EuroBertConfig):
+        super().__init__(config)
+        self.mask_converter = AttentionMaskConverter(is_causal=False)
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[tuple, BaseModelOutput]:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if attention_mask is not None and self.config._attn_implementation != "flash_attention_2":
+            mask = self.mask_converter.to_4d(attention_mask, attention_mask.shape[1], inputs_embeds.dtype)
+        else:
+            mask = attention_mask
+
+        if position_ids is None:
+            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device).unsqueeze(0)
+
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        for encoder_layer in self.layers[: self.config.num_hidden_layers]:
+            hidden_states = encoder_layer(
+                hidden_states,
+                attention_mask=mask,
+                position_ids=position_ids,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+        hidden_states = self.norm(hidden_states)
+
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+        )
+
+
+class EuroBertForMaskedLM(LlamaForCausalLM):
+    def __init__(self, config: EuroBertConfig):
+        super().__init__(config)
+        self.model = EuroBertModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, config.mlp_bias)
+        self.post_init()
+
+
+class EuroBertForSequenceClassification(EuroBertPreTrainedModel):
+    def __init__(self, config: EuroBertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.clf_pooling = config.clf_pooling
+
+        self.model = EuroBertModel(config)
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.GELU()
+        self.classifier = nn.Linear(config.hidden_size, self.num_labels)
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+    ) -> Union[tuple[torch.Tensor], SequenceClassifierOutput]:
+        encoder_output = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+        )
+        last_hidden_state = encoder_output[0]
+
+        if self.clf_pooling in ["bos", "mean"]:
+            if self.clf_pooling == "bos":
+                pooled_output = last_hidden_state[:, 0]
+
+            elif self.clf_pooling == "mean":
+                if attention_mask is None:
+                    pooled_output = last_hidden_state.mean(dim=1)
+                else:
+                    pooled_output = (last_hidden_state * attention_mask.unsqueeze(-1)).sum(dim=1)
+                    pooled_output /= attention_mask.sum(dim=1, keepdim=True)
+
+            pooled_output = self.dense(pooled_output)
+            pooled_output = self.activation(pooled_output)
+            logits = self.classifier(pooled_output)
+
+        elif self.clf_pooling == "late":
+            x = self.dense(last_hidden_state)
+            x = self.activation(x)
+            logits = self.classifier(x)
+            if attention_mask is None:
+                logits = logits.mean(dim=1)
+            else:
+                logits = (logits * attention_mask.unsqueeze(-1)).sum(dim=1)
+                logits /= attention_mask.sum(dim=1, keepdim=True)
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=encoder_output.hidden_states,
+            attentions=encoder_output.attentions,
+        )
+
+
+class EuroBertForTokenClassification(EuroBertPreTrainedModel):
+    def __init__(self, config: EuroBertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = EuroBertModel(config)
+
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+    ) -> Union[tuple, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "EuroBertConfig",
+    "EuroBertPreTrainedModel",
+    "EuroBertModel",
+    "EuroBertForMaskedLM",
+    "EuroBertForSequenceClassification",
+    "EuroBertForTokenClassification",
+]

tests/models/eurobert/test_modeling_eurobert.py

@@ -0,0 +1,444 @@
+# coding=utf-8
+# Copyright 2025 Nicolas Boizard, Duarte M. Alves, Hippolyte Gisserot-Boukhlef and the EuroBert team. 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.
+
+"""Testing suite for the PyTorch EuroBERT model."""
+
+import unittest
+
+from parameterized import parameterized
+
+from transformers import EuroBertConfig, is_torch_available, set_seed
+from transformers.testing_utils import (
+    require_torch,
+    require_torch_accelerator,
+    torch_device,
+)
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        EuroBertForMaskedLM,
+        EuroBertForSequenceClassification,
+        EuroBertModel,
+    )
+    from transformers.models.eurobert.modeling_eurobert import EuroBertRotaryEmbedding
+
+
+class EuroBertModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=False,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        pad_token_id=0,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.pad_token_id = pad_token_id
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length]).to(torch_device)
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = self.get_config()
+
+        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def get_config(self):
+        return EuroBertConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+            pad_token_id=self.pad_token_id,
+        )
+
+    def create_and_check_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = EuroBertModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask)
+        result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_for_masked_lm(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        model = EuroBertForMaskedLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_for_sequence_classification(
+        self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_labels = self.num_labels
+        model = EuroBertForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, labels=sequence_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class EuroBertModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            EuroBertModel,
+            EuroBertForMaskedLM,
+            EuroBertForSequenceClassification,
+        )
+        if is_torch_available()
+        else ()
+    )
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": EuroBertModel,
+            "text-classification": EuroBertForSequenceClassification,
+            "fill-mask": EuroBertForMaskedLM,
+            "zero-shot": EuroBertForSequenceClassification,
+        }
+        if is_torch_available()
+        else {}
+    )
+    test_headmasking = False
+    test_pruning = False
+    fx_compatible = False  # Broken by attention refactor cc @Cyrilvallez
+
+    # Need to use `0.8` instead of `0.9` for `test_cpu_offload`
+    # This is because we are hitting edge cases with the causal_mask buffer
+    model_split_percents = [0.5, 0.7, 0.8]
+
+    # used in `test_torch_compile_for_training`
+    _torch_compile_train_cls = EuroBertForMaskedLM if is_torch_available() else None
+
+    def setUp(self):
+        self.model_tester = EuroBertModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=EuroBertConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_various_embeddings(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        for type in ["absolute", "relative_key", "relative_key_query"]:
+            config_and_inputs[0].position_embedding_type = type
+            self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_eurobert_sequence_classification_model(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.num_labels = 3
+        input_ids = input_dict["input_ids"]
+        attention_mask = input_ids.ne(1).to(torch_device)
+        sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
+        model = EuroBertForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
+        self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
+
+    def test_eurobert_sequence_classification_model_for_single_label(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.num_labels = 3
+        config.problem_type = "single_label_classification"
+        input_ids = input_dict["input_ids"]
+        attention_mask = input_ids.ne(1).to(torch_device)
+        sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
+        model = EuroBertForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
+        self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
+
+    def test_eurobert_sequence_classification_model_for_multi_label(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.num_labels = 3
+        config.problem_type = "multi_label_classification"
+        input_ids = input_dict["input_ids"]
+        attention_mask = input_ids.ne(1).to(torch_device)
+        sequence_labels = ids_tensor(
+            [self.model_tester.batch_size, config.num_labels], self.model_tester.type_sequence_label_size
+        ).to(torch.float)
+        model = EuroBertForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
+        self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
+
+    @unittest.skip(reason="EuroBert buffers include complex numbers, which breaks this test")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @parameterized.expand([("linear",), ("dynamic",), ("yarn",)])
+    def test_model_rope_scaling_from_config(self, scaling_type):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+        short_input = ids_tensor([1, 10], config.vocab_size)
+        long_input = ids_tensor([1, int(config.max_position_embeddings * 1.5)], config.vocab_size)
+
+        set_seed(42)  # Fixed seed at init time so the two models get the same random weights
+        original_model = EuroBertModel(config)
+        original_model.to(torch_device)
+        original_model.eval()
+        original_short_output = original_model(short_input).last_hidden_state
+        original_long_output = original_model(long_input).last_hidden_state
+
+        set_seed(42)  # Fixed seed at init time so the two models get the same random weights
+        config.rope_scaling = {"type": scaling_type, "factor": 10.0}
+        scaled_model = EuroBertModel(config)
+        scaled_model.to(torch_device)
+        scaled_model.eval()
+        scaled_short_output = scaled_model(short_input).last_hidden_state
+        scaled_long_output = scaled_model(long_input).last_hidden_state
+
+        # Dynamic scaling does not change the RoPE embeddings until it receives an input longer than the original
+        # maximum sequence length, so the outputs for the short input should match.
+        if scaling_type == "dynamic":
+            torch.testing.assert_close(original_short_output, scaled_short_output, rtol=1e-5, atol=1e-5)
+        else:
+            self.assertFalse(torch.allclose(original_short_output, scaled_short_output, atol=1e-5))
+
+        # The output should be different for long inputs
+        self.assertFalse(torch.allclose(original_long_output, scaled_long_output, atol=1e-5))
+
+    def test_model_rope_scaling(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+        scaling_factor = 10
+        short_input_length = 10
+        long_input_length = int(config.max_position_embeddings * 1.5)
+
+        # Inputs
+        x = torch.randn(1, dtype=torch.float32, device=torch_device)  # used exlusively to get the dtype and the device
+        position_ids_short = torch.arange(short_input_length, dtype=torch.long, device=torch_device)
+        position_ids_short = position_ids_short.unsqueeze(0)
+        position_ids_long = torch.arange(long_input_length, dtype=torch.long, device=torch_device)
+        position_ids_long = position_ids_long.unsqueeze(0)
+
+        # Sanity check original RoPE
+        original_rope = EuroBertRotaryEmbedding(config=config).to(torch_device)
+        original_cos_short, original_sin_short = original_rope(x, position_ids_short)
+        original_cos_long, original_sin_long = original_rope(x, position_ids_long)
+        torch.testing.assert_close(original_cos_short, original_cos_long[:, :short_input_length, :])
+        torch.testing.assert_close(original_sin_short, original_sin_long[:, :short_input_length, :])
+
+        # Sanity check linear RoPE scaling
+        # New position "x" should match original position with index "x/scaling_factor"
+        config.rope_scaling = {"type": "linear", "factor": scaling_factor}
+        linear_scaling_rope = EuroBertRotaryEmbedding(config=config).to(torch_device)
+        linear_cos_short, linear_sin_short = linear_scaling_rope(x, position_ids_short)
+        linear_cos_long, linear_sin_long = linear_scaling_rope(x, position_ids_long)
+        torch.testing.assert_close(linear_cos_short, linear_cos_long[:, :short_input_length, :])
+        torch.testing.assert_close(linear_sin_short, linear_sin_long[:, :short_input_length, :])
+        for new_position in range(0, long_input_length, scaling_factor):
+            original_position = int(new_position // scaling_factor)
+            torch.testing.assert_close(linear_cos_long[:, new_position, :], original_cos_long[:, original_position, :])
+            torch.testing.assert_close(linear_sin_long[:, new_position, :], original_sin_long[:, original_position, :])
+
+        # Sanity check Dynamic NTK RoPE scaling
+        # Scaling should only be observed after a long input is fed. We can observe that the frequencies increase
+        # with scaling_factor (or that `inv_freq` decreases)
+        config.rope_scaling = {"type": "dynamic", "factor": scaling_factor}
+        ntk_scaling_rope = EuroBertRotaryEmbedding(config=config).to(torch_device)
+        ntk_cos_short, ntk_sin_short = ntk_scaling_rope(x, position_ids_short)
+        ntk_cos_long, ntk_sin_long = ntk_scaling_rope(x, position_ids_long)
+        torch.testing.assert_close(ntk_cos_short, original_cos_short)
+        torch.testing.assert_close(ntk_sin_short, original_sin_short)
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(ntk_cos_long, original_cos_long)
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(ntk_sin_long, original_sin_long)
+        self.assertTrue((ntk_scaling_rope.inv_freq <= original_rope.inv_freq).all())
+
+        # Sanity check Yarn RoPE scaling
+        # Scaling should be over the entire input
+        config.rope_scaling = {"type": "yarn", "factor": scaling_factor}
+        yarn_scaling_rope = EuroBertRotaryEmbedding(config=config).to(torch_device)
+        yarn_cos_short, yarn_sin_short = yarn_scaling_rope(x, position_ids_short)
+        yarn_cos_long, yarn_sin_long = yarn_scaling_rope(x, position_ids_long)
+        torch.testing.assert_close(yarn_cos_short, yarn_cos_long[:, :short_input_length, :])
+        torch.testing.assert_close(yarn_sin_short, yarn_sin_long[:, :short_input_length, :])
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(yarn_cos_short, original_cos_short)
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(yarn_sin_short, original_sin_short)
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(yarn_cos_long, original_cos_long)
+        with self.assertRaises(AssertionError):
+            torch.testing.assert_close(yarn_sin_long, original_sin_long)
+
+    def test_model_loading_old_rope_configs(self):
+        def _reinitialize_config(base_config, new_kwargs):
+            # Reinitialize the config with the new kwargs, forcing the config to go through its __init__ validation
+            # steps.
+            base_config_dict = base_config.to_dict()
+            new_config = EuroBertConfig.from_dict(config_dict={**base_config_dict, **new_kwargs})
+            return new_config
+
+        # from untouched config -> ✅
+        base_config, model_inputs = self.model_tester.prepare_config_and_inputs_for_common()
+        original_model = EuroBertForMaskedLM(base_config).to(torch_device)
+        original_model(**model_inputs)
+
+        # from a config with the expected rope configuration -> ✅
+        config = _reinitialize_config(base_config, {"rope_scaling": {"rope_type": "linear", "factor": 10.0}})
+        original_model = EuroBertForMaskedLM(config).to(torch_device)
+        original_model(**model_inputs)
+
+        # from a config with the old rope configuration ('type' instead of 'rope_type')  -> ✅ we gracefully handle BC
+        config = _reinitialize_config(base_config, {"rope_scaling": {"type": "linear", "factor": 10.0}})
+        original_model = EuroBertForMaskedLM(config).to(torch_device)
+        original_model(**model_inputs)
+
+        # from a config with both 'type' and 'rope_type'  -> ✅ they can coexist (and both are present in the config)
+        config = _reinitialize_config(
+            base_config, {"rope_scaling": {"type": "linear", "rope_type": "linear", "factor": 10.0}}
+        )
+        self.assertTrue(config.rope_scaling["type"] == "linear")
+        self.assertTrue(config.rope_scaling["rope_type"] == "linear")
+        original_model = EuroBertForMaskedLM(config).to(torch_device)
+        original_model(**model_inputs)
+
+        # from a config with parameters in a bad range ('factor' should be >= 1.0) -> ⚠️ throws a warning
+        with self.assertLogs("transformers.modeling_rope_utils", level="WARNING") as logs:
+            config = _reinitialize_config(base_config, {"rope_scaling": {"rope_type": "linear", "factor": -999.0}})
+            original_model = EuroBertForMaskedLM(config).to(torch_device)
+            original_model(**model_inputs)
+            self.assertEqual(len(logs.output), 1)
+            self.assertIn("factor field", logs.output[0])
+
+        # from a config with unknown parameters ('foo' isn't a rope option) -> ⚠️ throws a warning
+        with self.assertLogs("transformers.modeling_rope_utils", level="WARNING") as logs:
+            config = _reinitialize_config(
+                base_config, {"rope_scaling": {"rope_type": "linear", "factor": 10.0, "foo": "bar"}}
+            )
+            original_model = EuroBertForMaskedLM(config).to(torch_device)
+            original_model(**model_inputs)
+            self.assertEqual(len(logs.output), 1)
+            self.assertIn("Unrecognized keys", logs.output[0])
+
+        # from a config with specific rope type but missing one of its mandatory parameters -> ❌ throws exception
+        with self.assertRaises(KeyError):
+            config = _reinitialize_config(base_config, {"rope_scaling": {"rope_type": "linear"}})  # missing "factor"
+
+
+@require_torch_accelerator
+class EuroBertIntegrationTest(unittest.TestCase):
+    # This variable is used to determine which CUDA device are we using for our runners (A10 or T4)
+    # Depending on the hardware we get different logits / generations
+    cuda_compute_capability_major_version = None
+
+    @classmethod
+    def setUpClass(cls):
+        if is_torch_available() and torch.cuda.is_available():
+            # 8 is for A100 / A10 and 7 for T4
+            cls.cuda_compute_capability_major_version = torch.cuda.get_device_capability()[0]
+
+    # TODO: Implement integration tests