Add Multi Resolution Analysis (MRA) (New PR) (#24513)

* Add all files

* Update masked_language_modeling.md

* fix mlm models

* fix conflicts

* fix conflicts

* fix copies

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Reduce seq_len and hidden_size in ModelTester

* remove output_attentions

* fix conflicts

* remove copied from statements

* Apply suggestions from code review
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

---------
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

src/transformers/kernels/mra/torch_extension.cpp

+#include <torch/extension.h>
+#include <ATen/ATen.h>
+#include "cuda_launch.h"
+#include <vector>
+
+std::vector<at::Tensor> index_max(
+  at::Tensor index_vals,
+  at::Tensor indices,
+  int A_num_block,
+  int B_num_block
+) {
+  return index_max_kernel(
+    index_vals,
+    indices,
+    A_num_block,
+    B_num_block
+  );
+}
+
+at::Tensor mm_to_sparse(
+  at::Tensor dense_A,
+  at::Tensor dense_B,
+  at::Tensor indices
+) {
+  return mm_to_sparse_kernel(
+    dense_A,
+    dense_B,
+    indices
+  );
+}
+
+at::Tensor sparse_dense_mm(
+  at::Tensor sparse_A,
+  at::Tensor indices,
+  at::Tensor dense_B,
+  int A_num_block
+) {
+  return sparse_dense_mm_kernel(
+    sparse_A,
+    indices,
+    dense_B,
+    A_num_block
+  );
+}
+
+at::Tensor reduce_sum(
+  at::Tensor sparse_A,
+  at::Tensor indices,
+  int A_num_block,
+  int B_num_block
+) {
+  return reduce_sum_kernel(
+    sparse_A,
+    indices,
+    A_num_block,
+    B_num_block
+  );
+}
+
+at::Tensor scatter(
+  at::Tensor dense_A,
+  at::Tensor indices,
+  int B_num_block
+) {
+  return scatter_kernel(
+    dense_A,
+    indices,
+    B_num_block
+  );
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("index_max", &index_max, "index_max (CUDA)");
+  m.def("mm_to_sparse", &mm_to_sparse, "mm_to_sparse (CUDA)");
+  m.def("sparse_dense_mm", &sparse_dense_mm, "sparse_dense_mm (CUDA)");
+  m.def("reduce_sum", &reduce_sum, "reduce_sum (CUDA)");
+  m.def("scatter", &scatter, "scatter (CUDA)");
+}

src/transformers/models/__init__.py

@@ -134,6 +134,7 @@ from . import (
     mobilevit,
     mobilevitv2,
     mpnet,
+    mra,
     mt5,
     musicgen,
     mvp,

src/transformers/models/auto/configuration_auto.py

@@ -137,6 +137,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("mobilevit", "MobileViTConfig"),
         ("mobilevitv2", "MobileViTV2Config"),
         ("mpnet", "MPNetConfig"),
+        ("mra", "MraConfig"),
         ("mt5", "MT5Config"),
         ("musicgen", "MusicgenConfig"),
         ("mvp", "MvpConfig"),
@@ -329,6 +330,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("mobilevit", "MOBILEVIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("mobilevitv2", "MOBILEVITV2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("mpnet", "MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mra", "MRA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("musicgen", "MUSICGEN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("mvp", "MVP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("nat", "NAT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -534,6 +536,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("mobilevit", "MobileViT"),
         ("mobilevitv2", "MobileViTV2"),
         ("mpnet", "MPNet"),
+        ("mra", "MRA"),
         ("mt5", "MT5"),
         ("musicgen", "MusicGen"),
         ("mvp", "MVP"),

src/transformers/models/auto/modeling_auto.py

@@ -134,6 +134,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("mobilevit", "MobileViTModel"),
         ("mobilevitv2", "MobileViTV2Model"),
         ("mpnet", "MPNetModel"),
+        ("mra", "MraModel"),
         ("mt5", "MT5Model"),
         ("mvp", "MvpModel"),
         ("nat", "NatModel"),
@@ -247,6 +248,7 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
         ("megatron-bert", "MegatronBertForPreTraining"),
         ("mobilebert", "MobileBertForPreTraining"),
         ("mpnet", "MPNetForMaskedLM"),
+        ("mra", "MraForMaskedLM"),
         ("mvp", "MvpForConditionalGeneration"),
         ("nezha", "NezhaForPreTraining"),
         ("nllb-moe", "NllbMoeForConditionalGeneration"),
@@ -326,6 +328,7 @@ MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
         ("megatron-bert", "MegatronBertForCausalLM"),
         ("mobilebert", "MobileBertForMaskedLM"),
         ("mpnet", "MPNetForMaskedLM"),
+        ("mra", "MraForMaskedLM"),
         ("mvp", "MvpForConditionalGeneration"),
         ("nezha", "NezhaForMaskedLM"),
         ("nllb-moe", "NllbMoeForConditionalGeneration"),
@@ -572,6 +575,7 @@ MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
         ("megatron-bert", "MegatronBertForMaskedLM"),
         ("mobilebert", "MobileBertForMaskedLM"),
         ("mpnet", "MPNetForMaskedLM"),
+        ("mra", "MraForMaskedLM"),
         ("mvp", "MvpForConditionalGeneration"),
         ("nezha", "NezhaForMaskedLM"),
         ("nystromformer", "NystromformerForMaskedLM"),
@@ -704,6 +708,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("megatron-bert", "MegatronBertForSequenceClassification"),
         ("mobilebert", "MobileBertForSequenceClassification"),
         ("mpnet", "MPNetForSequenceClassification"),
+        ("mra", "MraForSequenceClassification"),
         ("mvp", "MvpForSequenceClassification"),
         ("nezha", "NezhaForSequenceClassification"),
         ("nystromformer", "NystromformerForSequenceClassification"),
@@ -771,6 +776,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
         ("megatron-bert", "MegatronBertForQuestionAnswering"),
         ("mobilebert", "MobileBertForQuestionAnswering"),
         ("mpnet", "MPNetForQuestionAnswering"),
+        ("mra", "MraForQuestionAnswering"),
         ("mt5", "MT5ForQuestionAnswering"),
         ("mvp", "MvpForQuestionAnswering"),
         ("nezha", "NezhaForQuestionAnswering"),
@@ -856,6 +862,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("megatron-bert", "MegatronBertForTokenClassification"),
         ("mobilebert", "MobileBertForTokenClassification"),
         ("mpnet", "MPNetForTokenClassification"),
+        ("mra", "MraForTokenClassification"),
         ("nezha", "NezhaForTokenClassification"),
         ("nystromformer", "NystromformerForTokenClassification"),
         ("qdqbert", "QDQBertForTokenClassification"),
@@ -899,6 +906,7 @@ MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES = OrderedDict(
         ("megatron-bert", "MegatronBertForMultipleChoice"),
         ("mobilebert", "MobileBertForMultipleChoice"),
         ("mpnet", "MPNetForMultipleChoice"),
+        ("mra", "MraForMultipleChoice"),
         ("nezha", "NezhaForMultipleChoice"),
         ("nystromformer", "NystromformerForMultipleChoice"),
         ("qdqbert", "QDQBertForMultipleChoice"),

src/transformers/models/auto/tokenization_auto.py

@@ -214,6 +214,7 @@ else:
             ("mluke", ("MLukeTokenizer" if is_sentencepiece_available() else None, None)),
             ("mobilebert", ("MobileBertTokenizer", "MobileBertTokenizerFast" if is_tokenizers_available() else None)),
             ("mpnet", ("MPNetTokenizer", "MPNetTokenizerFast" if is_tokenizers_available() else None)),
+            ("mra", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
             (
                 "mt5",
                 (

src/transformers/models/mra/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2023 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.
+from typing import TYPE_CHECKING
+
+# rely on isort to merge the imports
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {"configuration_mra": ["MRA_PRETRAINED_CONFIG_ARCHIVE_MAP", "MraConfig"]}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_mra"] = [
+        "MRA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "MraForMaskedLM",
+        "MraForMultipleChoice",
+        "MraForQuestionAnswering",
+        "MraForSequenceClassification",
+        "MraForTokenClassification",
+        "MraLayer",
+        "MraModel",
+        "MraPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_mra import MRA_PRETRAINED_CONFIG_ARCHIVE_MAP, MraConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_mra import (
+            MRA_PRETRAINED_MODEL_ARCHIVE_LIST,
+            MraForMaskedLM,
+            MraForMultipleChoice,
+            MraForQuestionAnswering,
+            MraForSequenceClassification,
+            MraForTokenClassification,
+            MraLayer,
+            MraModel,
+            MraPreTrainedModel,
+        )
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)

src/transformers/models/mra/configuration_mra.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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.
+""" MRA model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+MRA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "uw-madison/mra-base-512-4": "https://huggingface.co/uw-madison/mra-base-512-4/resolve/main/config.json",
+}
+
+
+class MraConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`MraModel`]. It is used to instantiate an MRA
+    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 Mra
+    [uw-madison/mra-base-512-4](https://huggingface.co/uw-madison/mra-base-512-4) architecture.
+
+    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 50265):
+            Vocabulary size of the Mra model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`MraModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimension of the encoder layers and the pooler layer.
+        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.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 1):
+            The vocabulary size of the `token_type_ids` passed when calling [`MraModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`.
+        block_per_row (`int`, *optional*, defaults to 4):
+            Used to set the budget for the high resolution scale.
+        approx_mode (`str`, *optional*, defaults to `"full"`):
+            Controls whether both low and high resolution approximations are used. Set to `"full"` for both low and
+            high resolution and `"sparse"` for only low resolution.
+        initial_prior_first_n_blocks (`int`, *optional*, defaults to 0):
+            The initial number of blocks for which high resolution is used.
+        initial_prior_diagonal_n_blocks (`int`, *optional*, defaults to 0):
+            The number of diagonal blocks for which high resolution is used.
+
+    Example:
+
+    ```python
+    >>> from transformers import MraConfig, MraModel
+
+    >>> # Initializing a Mra uw-madison/mra-base-512-4 style configuration
+    >>> configuration = MraConfig()
+
+    >>> # Initializing a model (with random weights) from the uw-madison/mra-base-512-4 style configuration
+    >>> model = MraModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "mra"
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=1,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        position_embedding_type="absolute",
+        block_per_row=4,
+        approx_mode="full",
+        initial_prior_first_n_blocks=0,
+        initial_prior_diagonal_n_blocks=0,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        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.initializer_range = initializer_range
+        self.type_vocab_size = type_vocab_size
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.block_per_row = block_per_row
+        self.approx_mode = approx_mode
+        self.initial_prior_first_n_blocks = initial_prior_first_n_blocks
+        self.initial_prior_diagonal_n_blocks = initial_prior_diagonal_n_blocks

src/transformers/models/mra/convert_mra_pytorch_to_pytorch.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# 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.
+"""Convert MRA checkpoints from the original repository. URL: https://github.com/mlpen/mra-attention"""
+
+import argparse
+
+import torch
+
+from transformers import MraConfig, MraForMaskedLM
+
+
+def rename_key(orig_key):
+    if "model" in orig_key:
+        orig_key = orig_key.replace("model.", "")
+    if "norm1" in orig_key:
+        orig_key = orig_key.replace("norm1", "attention.output.LayerNorm")
+    if "norm2" in orig_key:
+        orig_key = orig_key.replace("norm2", "output.LayerNorm")
+    if "norm" in orig_key:
+        orig_key = orig_key.replace("norm", "LayerNorm")
+    if "transformer" in orig_key:
+        layer_num = orig_key.split(".")[0].split("_")[-1]
+        orig_key = orig_key.replace(f"transformer_{layer_num}", f"encoder.layer.{layer_num}")
+    if "mha.attn" in orig_key:
+        orig_key = orig_key.replace("mha.attn", "attention.self")
+    if "mha" in orig_key:
+        orig_key = orig_key.replace("mha", "attention")
+    if "W_q" in orig_key:
+        orig_key = orig_key.replace("W_q", "self.query")
+    if "W_k" in orig_key:
+        orig_key = orig_key.replace("W_k", "self.key")
+    if "W_v" in orig_key:
+        orig_key = orig_key.replace("W_v", "self.value")
+    if "ff.0" in orig_key:
+        orig_key = orig_key.replace("ff.0", "intermediate.dense")
+    if "ff.2" in orig_key:
+        orig_key = orig_key.replace("ff.2", "output.dense")
+    if "ff" in orig_key:
+        orig_key = orig_key.replace("ff", "output.dense")
+    if "mlm_class" in orig_key:
+        orig_key = orig_key.replace("mlm.mlm_class", "cls.predictions.decoder")
+    if "mlm" in orig_key:
+        orig_key = orig_key.replace("mlm", "cls.predictions.transform")
+    if "backbone.backbone.encoders" in orig_key:
+        orig_key = orig_key.replace("backbone.backbone.encoders", "encoder.layer")
+    if "cls" not in orig_key:
+        orig_key = "mra." + orig_key
+
+    return orig_key
+
+
+def convert_checkpoint_helper(max_position_embeddings, orig_state_dict):
+    for key in orig_state_dict.copy().keys():
+        val = orig_state_dict.pop(key)
+
+        if ("pooler" in key) or ("sen_class" in key):
+            continue
+        else:
+            orig_state_dict[rename_key(key)] = val
+
+    orig_state_dict["cls.predictions.bias"] = orig_state_dict["cls.predictions.decoder.bias"]
+    orig_state_dict["mra.embeddings.position_ids"] = torch.arange(max_position_embeddings).expand((1, -1)) + 2
+
+    return orig_state_dict
+
+
+def convert_mra_checkpoint(checkpoint_path, mra_config_file, pytorch_dump_path):
+    orig_state_dict = torch.load(checkpoint_path, map_location="cpu")["model_state_dict"]
+    config = MraConfig.from_json_file(mra_config_file)
+    model = MraForMaskedLM(config)
+
+    new_state_dict = convert_checkpoint_helper(config.max_position_embeddings, orig_state_dict)
+
+    print(model.load_state_dict(new_state_dict))
+    model.eval()
+    model.save_pretrained(pytorch_dump_path)
+
+    print(f"Checkpoint successfuly converted. Model saved at {pytorch_dump_path}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--pytorch_model_path", default=None, type=str, required=True, help="Path to Mra pytorch checkpoint."
+    )
+    parser.add_argument(
+        "--config_file",
+        default=None,
+        type=str,
+        required=True,
+        help="The json file for Mra model config.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    args = parser.parse_args()
+    convert_mra_checkpoint(args.pytorch_model_path, args.config_file, args.pytorch_dump_path)

src/transformers/models/mra/modeling_mra.py

+# coding=utf-8
+# Copyright 2023 University of Wisconsin-Madison and The HuggingFace Inc. 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.
+""" PyTorch MRA model."""
+
+
+import math
+from pathlib import Path
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from torch.utils.cpp_extension import load
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_ninja_available,
+    is_torch_cuda_available,
+    logging,
+)
+from .configuration_mra import MraConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "uw-madison/mra-base-512-4"
+_CONFIG_FOR_DOC = "MraConfig"
+_TOKENIZER_FOR_DOC = "AutoTokenizer"
+
+MRA_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "uw-madison/mra-base-512-4",
+    # See all Mra models at https://huggingface.co/models?filter=mra
+]
+
+
+def load_cuda_kernels():
+    global cuda_kernel
+    src_folder = Path(__file__).resolve().parent.parent.parent / "kernels" / "mra"
+
+    def append_root(files):
+        return [src_folder / file for file in files]
+
+    src_files = append_root(["cuda_kernel.cu", "cuda_launch.cu", "torch_extension.cpp"])
+
+    cuda_kernel = load("cuda_kernel", src_files, verbose=True)
+
+    import cuda_kernel
+
+
+cuda_kernel = None
+
+
+if is_torch_cuda_available() and is_ninja_available():
+    logger.info("Loading custom CUDA kernels...")
+
+    try:
+        load_cuda_kernels()
+    except Exception as e:
+        logger.warning(
+            "Failed to load CUDA kernels. Mra requires custom CUDA kernels. Please verify that compatible versions of"
+            f" PyTorch and CUDA Toolkit are installed: {e}"
+        )
+else:
+    pass
+
+
+def sparse_max(sparse_qk_prod, indices, query_num_block, key_num_block):
+    """
+    Computes maximum values for softmax stability.
+    """
+    if len(sparse_qk_prod.size()) != 4:
+        raise ValueError("sparse_qk_prod must be a 4-dimensional tensor.")
+
+    if len(indices.size()) != 2:
+        raise ValueError("indices must be a 2-dimensional tensor.")
+
+    if sparse_qk_prod.size(2) != 32:
+        raise ValueError("The size of the second dimension of sparse_qk_prod must be 32.")
+
+    if sparse_qk_prod.size(3) != 32:
+        raise ValueError("The size of the third dimension of sparse_qk_prod must be 32.")
+
+    index_vals = sparse_qk_prod.max(dim=-2).values.transpose(-1, -2)
+    index_vals = index_vals.contiguous()
+
+    indices = indices.int()
+    indices = indices.contiguous()
+
+    max_vals, max_vals_scatter = cuda_kernel.index_max(index_vals, indices, query_num_block, key_num_block)
+    max_vals_scatter = max_vals_scatter.transpose(-1, -2)[:, :, None, :]
+
+    return max_vals, max_vals_scatter
+
+
+def sparse_mask(mask, indices, block_size=32):
+    """
+    Converts attention mask to a sparse mask for high resolution logits.
+    """
+    if len(mask.size()) != 2:
+        raise ValueError("mask must be a 2-dimensional tensor.")
+
+    if len(indices.size()) != 2:
+        raise ValueError("indices must be a 2-dimensional tensor.")
+
+    if mask.shape[0] != indices.shape[0]:
+        raise ValueError("mask and indices must have the same size in the zero-th dimension.")
+
+    batch_size, seq_len = mask.shape
+    num_block = seq_len // block_size
+
+    batch_idx = torch.arange(indices.size(0), dtype=torch.long, device=indices.device)
+    mask = mask.reshape(batch_size, num_block, block_size)
+    mask = mask[batch_idx[:, None], (indices % num_block).long(), :]
+
+    return mask
+
+
+def mm_to_sparse(dense_query, dense_key, indices, block_size=32):
+    """
+    Performs Sampled Dense Matrix Multiplication.
+    """
+    batch_size, query_size, dim = dense_query.size()
+    _, key_size, dim = dense_key.size()
+
+    if query_size % block_size != 0:
+        raise ValueError("query_size (size of first dimension of dense_query) must be divisible by block_size.")
+
+    if key_size % block_size != 0:
+        raise ValueError("key_size (size of first dimension of dense_key) must be divisible by block_size.")
+
+    dense_query = dense_query.reshape(batch_size, query_size // block_size, block_size, dim).transpose(-1, -2)
+    dense_key = dense_key.reshape(batch_size, key_size // block_size, block_size, dim).transpose(-1, -2)
+
+    if len(dense_query.size()) != 4:
+        raise ValueError("dense_query must be a 4-dimensional tensor.")
+
+    if len(dense_key.size()) != 4:
+        raise ValueError("dense_key must be a 4-dimensional tensor.")
+
+    if len(indices.size()) != 2:
+        raise ValueError("indices must be a 2-dimensional tensor.")
+
+    if dense_query.size(3) != 32:
+        raise ValueError("The third dimension of dense_query must be 32.")
+
+    if dense_key.size(3) != 32:
+        raise ValueError("The third dimension of dense_key must be 32.")
+
+    dense_query = dense_query.contiguous()
+    dense_key = dense_key.contiguous()
+
+    indices = indices.int()
+    indices = indices.contiguous()
+
+    return cuda_kernel.mm_to_sparse(dense_query, dense_key, indices.int())
+
+
+def sparse_dense_mm(sparse_query, indices, dense_key, query_num_block, block_size=32):
+    """
+    Performs matrix multiplication of a sparse matrix with a dense matrix.
+    """
+    batch_size, key_size, dim = dense_key.size()
+
+    if key_size % block_size != 0:
+        raise ValueError("key_size (size of first dimension of dense_key) must be divisible by block_size.")
+
+    if sparse_query.size(2) != block_size:
+        raise ValueError("The size of the second dimension of sparse_query must be equal to the block_size.")
+
+    if sparse_query.size(3) != block_size:
+        raise ValueError("The size of the third dimension of sparse_query must be equal to the block_size.")
+
+    dense_key = dense_key.reshape(batch_size, key_size // block_size, block_size, dim).transpose(-1, -2)
+
+    if len(sparse_query.size()) != 4:
+        raise ValueError("sparse_query must be a 4-dimensional tensor.")
+
+    if len(dense_key.size()) != 4:
+        raise ValueError("dense_key must be a 4-dimensional tensor.")
+
+    if len(indices.size()) != 2:
+        raise ValueError("indices must be a 2-dimensional tensor.")
+
+    if dense_key.size(3) != 32:
+        raise ValueError("The size of the third dimension of dense_key must be 32.")
+
+    sparse_query = sparse_query.contiguous()
+
+    indices = indices.int()
+    indices = indices.contiguous()
+    dense_key = dense_key.contiguous()
+
+    dense_qk_prod = cuda_kernel.sparse_dense_mm(sparse_query, indices, dense_key, query_num_block)
+    dense_qk_prod = dense_qk_prod.transpose(-1, -2).reshape(batch_size, query_num_block * block_size, dim)
+    return dense_qk_prod
+
+
+def transpose_indices(indices, dim_1_block, dim_2_block):
+    return ((indices % dim_2_block) * dim_1_block + torch.div(indices, dim_2_block, rounding_mode="floor")).long()
+
+
+class MraSampledDenseMatMul(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, dense_query, dense_key, indices, block_size):
+        sparse_qk_prod = mm_to_sparse(dense_query, dense_key, indices, block_size)
+        ctx.save_for_backward(dense_query, dense_key, indices)
+        ctx.block_size = block_size
+        return sparse_qk_prod
+
+    @staticmethod
+    def backward(ctx, grad):
+        dense_query, dense_key, indices = ctx.saved_tensors
+        block_size = ctx.block_size
+        query_num_block = dense_query.size(1) // block_size
+        key_num_block = dense_key.size(1) // block_size
+        indices_T = transpose_indices(indices, query_num_block, key_num_block)
+        grad_key = sparse_dense_mm(grad.transpose(-1, -2), indices_T, dense_query, key_num_block)
+        grad_query = sparse_dense_mm(grad, indices, dense_key, query_num_block)
+        return grad_query, grad_key, None, None
+
+    @staticmethod
+    def operator_call(dense_query, dense_key, indices, block_size=32):
+        return MraSampledDenseMatMul.apply(dense_query, dense_key, indices, block_size)
+
+
+class MraSparseDenseMatMul(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, sparse_query, indices, dense_key, query_num_block):
+        sparse_qk_prod = sparse_dense_mm(sparse_query, indices, dense_key, query_num_block)
+        ctx.save_for_backward(sparse_query, indices, dense_key)
+        ctx.query_num_block = query_num_block
+        return sparse_qk_prod
+
+    @staticmethod
+    def backward(ctx, grad):
+        sparse_query, indices, dense_key = ctx.saved_tensors
+        query_num_block = ctx.query_num_block
+        key_num_block = dense_key.size(1) // sparse_query.size(-1)
+        indices_T = transpose_indices(indices, query_num_block, key_num_block)
+        grad_key = sparse_dense_mm(sparse_query.transpose(-1, -2), indices_T, grad, key_num_block)
+        grad_query = mm_to_sparse(grad, dense_key, indices)
+        return grad_query, None, grad_key, None
+
+    @staticmethod
+    def operator_call(sparse_query, indices, dense_key, query_num_block):
+        return MraSparseDenseMatMul.apply(sparse_query, indices, dense_key, query_num_block)
+
+
+class MraReduceSum:
+    @staticmethod
+    def operator_call(sparse_query, indices, query_num_block, key_num_block):
+        batch_size, num_block, block_size, _ = sparse_query.size()
+
+        if len(sparse_query.size()) != 4:
+            raise ValueError("sparse_query must be a 4-dimensional tensor.")
+
+        if len(indices.size()) != 2:
+            raise ValueError("indices must be a 2-dimensional tensor.")
+
+        _, _, block_size, _ = sparse_query.size()
+        batch_size, num_block = indices.size()
+
+        sparse_query = sparse_query.sum(dim=2).reshape(batch_size * num_block, block_size)
+
+        batch_idx = torch.arange(indices.size(0), dtype=torch.long, device=indices.device)
+        global_idxes = (
+            torch.div(indices, key_num_block, rounding_mode="floor").long() + batch_idx[:, None] * query_num_block
+        ).reshape(batch_size * num_block)
+        temp = torch.zeros(
+            (batch_size * query_num_block, block_size), dtype=sparse_query.dtype, device=sparse_query.device
+        )
+        output = temp.index_add(0, global_idxes, sparse_query).reshape(batch_size, query_num_block, block_size)
+
+        output = output.reshape(batch_size, query_num_block * block_size)
+        return output
+
+
+def get_low_resolution_logit(query, key, block_size, mask=None, value=None):
+    """
+    Compute low resolution approximation.
+    """
+    batch_size, seq_len, head_dim = query.size()
+
+    num_block_per_row = seq_len // block_size
+
+    value_hat = None
+    if mask is not None:
+        token_count = mask.reshape(batch_size, num_block_per_row, block_size).sum(dim=-1)
+        query_hat = query.reshape(batch_size, num_block_per_row, block_size, head_dim).sum(dim=-2) / (
+            token_count[:, :, None] + 1e-6
+        )
+        key_hat = key.reshape(batch_size, num_block_per_row, block_size, head_dim).sum(dim=-2) / (
+            token_count[:, :, None] + 1e-6
+        )
+        if value is not None:
+            value_hat = value.reshape(batch_size, num_block_per_row, block_size, head_dim).sum(dim=-2) / (
+                token_count[:, :, None] + 1e-6
+            )
+    else:
+        token_count = block_size * torch.ones(batch_size, num_block_per_row, dtype=torch.float, device=query.device)
+        query_hat = query.reshape(batch_size, num_block_per_row, block_size, head_dim).mean(dim=-2)
+        key_hat = key.reshape(batch_size, num_block_per_row, block_size, head_dim).mean(dim=-2)
+        if value is not None:
+            value_hat = value.reshape(batch_size, num_block_per_row, block_size, head_dim).mean(dim=-2)
+
+    low_resolution_logit = torch.matmul(query_hat, key_hat.transpose(-1, -2)) / math.sqrt(head_dim)
+
+    low_resolution_logit_row_max = low_resolution_logit.max(dim=-1, keepdims=True).values
+
+    if mask is not None:
+        low_resolution_logit = (
+            low_resolution_logit - 1e4 * ((token_count[:, None, :] * token_count[:, :, None]) < 0.5).float()
+        )
+
+    return low_resolution_logit, token_count, low_resolution_logit_row_max, value_hat
+
+
+def get_block_idxes(
+    low_resolution_logit, num_blocks, approx_mode, initial_prior_first_n_blocks, initial_prior_diagonal_n_blocks
+):
+    """
+    Compute the indices of the subset of components to be used in the approximation.
+    """
+    batch_size, total_blocks_per_row, _ = low_resolution_logit.shape
+
+    if initial_prior_diagonal_n_blocks > 0:
+        offset = initial_prior_diagonal_n_blocks // 2
+        temp_mask = torch.ones(total_blocks_per_row, total_blocks_per_row, device=low_resolution_logit.device)
+        diagonal_mask = torch.tril(torch.triu(temp_mask, diagonal=-offset), diagonal=offset)
+        low_resolution_logit = low_resolution_logit + diagonal_mask[None, :, :] * 5e3
+
+    if initial_prior_first_n_blocks > 0:
+        low_resolution_logit[:, :initial_prior_first_n_blocks, :] = (
+            low_resolution_logit[:, :initial_prior_first_n_blocks, :] + 5e3
+        )
+        low_resolution_logit[:, :, :initial_prior_first_n_blocks] = (
+            low_resolution_logit[:, :, :initial_prior_first_n_blocks] + 5e3
+        )
+
+    top_k_vals = torch.topk(
+        low_resolution_logit.reshape(batch_size, -1), num_blocks, dim=-1, largest=True, sorted=False
+    )
+    indices = top_k_vals.indices
+
+    if approx_mode == "full":
+        threshold = top_k_vals.values.min(dim=-1).values
+        high_resolution_mask = (low_resolution_logit >= threshold[:, None, None]).float()
+    elif approx_mode == "sparse":
+        high_resolution_mask = None
+    else:
+        raise ValueError(f"{approx_mode} is not a valid approx_model value.")
+
+    return indices, high_resolution_mask
+
+
+def mra2_attention(
+    query,
+    key,
+    value,
+    mask,
+    num_blocks,
+    approx_mode,
+    block_size=32,
+    initial_prior_first_n_blocks=0,
+    initial_prior_diagonal_n_blocks=0,
+):
+    """
+    Use Mra to approximate self-attention.
+    """
+    if cuda_kernel is None:
+        return torch.zeros_like(query).requires_grad_()
+
+    batch_size, num_head, seq_len, head_dim = query.size()
+    meta_batch = batch_size * num_head
+
+    if seq_len % block_size != 0:
+        raise ValueError("sequence length must be divisible by the block_size.")
+
+    num_block_per_row = seq_len // block_size
+
+    query = query.reshape(meta_batch, seq_len, head_dim)
+    key = key.reshape(meta_batch, seq_len, head_dim)
+    value = value.reshape(meta_batch, seq_len, head_dim)
+
+    if mask is not None:
+        query = query * mask[:, :, None]
+        key = key * mask[:, :, None]
+        value = value * mask[:, :, None]
+
+    if approx_mode == "full":
+        low_resolution_logit, token_count, low_resolution_logit_row_max, value_hat = get_low_resolution_logit(
+            query, key, block_size, mask, value
+        )
+    elif approx_mode == "sparse":
+        with torch.no_grad():
+            low_resolution_logit, token_count, low_resolution_logit_row_max, _ = get_low_resolution_logit(
+                query, key, block_size, mask
+            )
+    else:
+        raise Exception('approx_mode must be "full" or "sparse"')
+
+    with torch.no_grad():
+        low_resolution_logit_normalized = low_resolution_logit - low_resolution_logit_row_max
+        indices, high_resolution_mask = get_block_idxes(
+            low_resolution_logit_normalized,
+            num_blocks,
+            approx_mode,
+            initial_prior_first_n_blocks,
+            initial_prior_diagonal_n_blocks,
+        )
+
+    high_resolution_logit = MraSampledDenseMatMul.operator_call(
+        query, key, indices, block_size=block_size
+    ) / math.sqrt(head_dim)
+    max_vals, max_vals_scatter = sparse_max(high_resolution_logit, indices, num_block_per_row, num_block_per_row)
+    high_resolution_logit = high_resolution_logit - max_vals_scatter
+    if mask is not None:
+        high_resolution_logit = high_resolution_logit - 1e4 * (1 - sparse_mask(mask, indices)[:, :, :, None])
+    high_resolution_attn = torch.exp(high_resolution_logit)
+    high_resolution_attn_out = MraSparseDenseMatMul.operator_call(
+        high_resolution_attn, indices, value, num_block_per_row
+    )
+    high_resolution_normalizer = MraReduceSum.operator_call(
+        high_resolution_attn, indices, num_block_per_row, num_block_per_row
+    )
+
+    if approx_mode == "full":
+        low_resolution_attn = (
+            torch.exp(low_resolution_logit - low_resolution_logit_row_max - 1e4 * high_resolution_mask)
+            * token_count[:, None, :]
+        )
+
+        low_resolution_attn_out = (
+            torch.matmul(low_resolution_attn, value_hat)[:, :, None, :]
+            .repeat(1, 1, block_size, 1)
+            .reshape(meta_batch, seq_len, head_dim)
+        )
+        low_resolution_normalizer = (
+            low_resolution_attn.sum(dim=-1)[:, :, None].repeat(1, 1, block_size).reshape(meta_batch, seq_len)
+        )
+
+        log_correction = low_resolution_logit_row_max.repeat(1, 1, block_size).reshape(meta_batch, seq_len) - max_vals
+        if mask is not None:
+            log_correction = log_correction * mask
+
+        low_resolution_corr = torch.exp(log_correction * (log_correction <= 0).float())
+        low_resolution_attn_out = low_resolution_attn_out * low_resolution_corr[:, :, None]
+        low_resolution_normalizer = low_resolution_normalizer * low_resolution_corr
+
+        high_resolution_corr = torch.exp(-log_correction * (log_correction > 0).float())
+        high_resolution_attn_out = high_resolution_attn_out * high_resolution_corr[:, :, None]
+        high_resolution_normalizer = high_resolution_normalizer * high_resolution_corr
+
+        context_layer = (high_resolution_attn_out + low_resolution_attn_out) / (
+            high_resolution_normalizer[:, :, None] + low_resolution_normalizer[:, :, None] + 1e-6
+        )
+
+    elif approx_mode == "sparse":
+        context_layer = high_resolution_attn_out / (high_resolution_normalizer[:, :, None] + 1e-6)
+    else:
+        raise Exception('config.approx_mode must be "full" or "sparse"')
+
+    if mask is not None:
+        context_layer = context_layer * mask[:, :, None]
+
+    context_layer = context_layer.reshape(batch_size, num_head, seq_len, head_dim)
+
+    return context_layer
+
+
+class MraEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings + 2, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)) + 2)
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        self.register_buffer(
+            "token_type_ids",
+            torch.zeros(self.position_ids.size(), dtype=torch.long, device=self.position_ids.device),
+            persistent=False,
+        )
+
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class MraSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = (
+            position_embedding_type if position_embedding_type is not None else config.position_embedding_type
+        )
+
+        self.num_block = (config.max_position_embeddings // 32) * config.block_per_row
+        self.num_block = min(self.num_block, int((config.max_position_embeddings // 32) ** 2))
+
+        self.approx_mode = config.approx_mode
+        self.initial_prior_first_n_blocks = config.initial_prior_first_n_blocks
+        self.initial_prior_diagonal_n_blocks = config.initial_prior_diagonal_n_blocks
+
+    def transpose_for_scores(self, layer):
+        new_layer_shape = layer.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        layer = layer.view(*new_layer_shape)
+        return layer.permute(0, 2, 1, 3)
+
+    def forward(self, hidden_states, attention_mask=None):
+        mixed_query_layer = self.query(hidden_states)
+
+        key_layer = self.transpose_for_scores(self.key(hidden_states))
+        value_layer = self.transpose_for_scores(self.value(hidden_states))
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        batch_size, num_heads, seq_len, head_dim = query_layer.size()
+
+        # revert changes made by get_extended_attention_mask
+        attention_mask = 1.0 + attention_mask / 10000.0
+        attention_mask = (
+            attention_mask.squeeze().repeat(1, num_heads, 1).reshape(batch_size * num_heads, seq_len).int()
+        )
+
+        # The CUDA kernels are most efficient with inputs whose size is a multiple of a GPU's warp size (32). Inputs
+        # smaller than this are padded with zeros.
+        gpu_warp_size = 32
+
+        if head_dim < gpu_warp_size:
+            pad_size = batch_size, num_heads, seq_len, gpu_warp_size - head_dim
+
+            query_layer = torch.cat([query_layer, torch.zeros(pad_size, device=query_layer.device)], dim=-1)
+            key_layer = torch.cat([key_layer, torch.zeros(pad_size, device=key_layer.device)], dim=-1)
+            value_layer = torch.cat([value_layer, torch.zeros(pad_size, device=value_layer.device)], dim=-1)
+
+        context_layer = mra2_attention(
+            query_layer.float(),
+            key_layer.float(),
+            value_layer.float(),
+            attention_mask.float(),
+            self.num_block,
+            approx_mode=self.approx_mode,
+            initial_prior_first_n_blocks=self.initial_prior_first_n_blocks,
+            initial_prior_diagonal_n_blocks=self.initial_prior_diagonal_n_blocks,
+        )
+
+        if head_dim < gpu_warp_size:
+            context_layer = context_layer[:, :, :, :head_dim]
+
+        context_layer = context_layer.reshape(batch_size, num_heads, seq_len, head_dim)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer,)
+
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
+class MraSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class MraAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = MraSelfAttention(config, position_embedding_type=position_embedding_type)
+        self.output = MraSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(self, hidden_states, attention_mask=None):
+        self_outputs = self.self(hidden_states, attention_mask)
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate
+class MraIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput
+class MraOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class MraLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = MraAttention(config)
+        self.add_cross_attention = config.add_cross_attention
+        self.intermediate = MraIntermediate(config)
+        self.output = MraOutput(config)
+
+    def forward(self, hidden_states, attention_mask=None):
+        self_attention_outputs = self.attention(hidden_states, attention_mask)
+        attention_output = self_attention_outputs[0]
+
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
+        )
+        outputs = (layer_output,) + outputs
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+class MraEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([MraLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        output_hidden_states=False,
+        return_dict=True,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(hidden_states, attention_mask)
+
+            hidden_states = layer_outputs[0]
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states] if v is not None)
+        return BaseModelOutputWithCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPredictionHeadTransform
+class MraPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertLMPredictionHead with Bert->Mra
+class MraLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = MraPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOnlyMLMHead with Bert->Mra
+class MraOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = MraLMPredictionHead(config)
+
+    def forward(self, sequence_output: torch.Tensor) -> torch.Tensor:
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+# Copied from transformers.models.yoso.modeling_yoso.YosoPreTrainedModel with Yoso->Mra,yoso->mra
+class MraPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = MraConfig
+    base_model_prefix = "mra"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, MraEncoder):
+            module.gradient_checkpointing = value
+
+
+MRA_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
+    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`MraConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+MRA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare MRA Model transformer outputting raw hidden-states without any specific head on top.",
+    MRA_START_DOCSTRING,
+)
+class MraModel(MraPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = MraEmbeddings(config)
+        self.encoder = MraEncoder(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(MRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithCrossAttentions]:
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length)), device=device)
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+
+        if not return_dict:
+            return (sequence_output,) + encoder_outputs[1:]
+
+        return BaseModelOutputWithCrossAttentions(
+            last_hidden_state=sequence_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings("""MRA Model with a `language modeling` head on top.""", MRA_START_DOCSTRING)
+class MraForMaskedLM(MraPreTrainedModel):
+    _tied_weights_keys = ["cls.predictions.decoder.weight", "cls.predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.mra = MraModel(config)
+        self.cls = MraOnlyMLMHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(MRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MaskedLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (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]`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.mra(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[1:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+# Copied from transformers.models.yoso.modeling_yoso.YosoClassificationHead with Yoso->Mra
+class MraClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.config = config
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = ACT2FN[self.config.hidden_act](x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+@add_start_docstrings(
+    """MRA Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks.""",
+    MRA_START_DOCSTRING,
+)
+class MraForSequenceClassification(MraPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.mra = MraModel(config)
+        self.classifier = MraClassificationHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(MRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutput]:
+        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).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.mra(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            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)
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """MRA Model with a multiple choice classification head on top (a linear layer on top of
+    the pooled output and a softmax) e.g. for RocStories/SWAG tasks.""",
+    MRA_START_DOCSTRING,
+)
+class MraForMultipleChoice(MraPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.mra = MraModel(config)
+        self.pre_classifier = nn.Linear(config.hidden_size, config.hidden_size)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(MRA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MultipleChoiceModelOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
+            num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
+            `input_ids` above)
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        outputs = self.mra(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_state = outputs[0]  # (bs * num_choices, seq_len, dim)
+        pooled_output = hidden_state[:, 0]  # (bs * num_choices, dim)
+        pooled_output = self.pre_classifier(pooled_output)  # (bs * num_choices, dim)
+        pooled_output = nn.ReLU()(pooled_output)  # (bs * num_choices, dim)
+        logits = self.classifier(pooled_output)
+
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """MRA Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks.""",
+    MRA_START_DOCSTRING,
+)
+class MraForTokenClassification(MraPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.mra = MraModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(MRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.mra(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """MRA Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).""",
+    MRA_START_DOCSTRING,
+)
+class MraForQuestionAnswering(MraPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        config.num_labels = 2
+        self.num_labels = config.num_labels
+
+        self.mra = MraModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(MRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.mra(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[1:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

src/transformers/utils/dummy_pt_objects.py

@@ -4901,6 +4901,58 @@ class MPNetPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+MRA_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class MraForMaskedLM(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MraForMultipleChoice(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MraForQuestionAnswering(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MraForSequenceClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MraForTokenClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MraModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MraPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class MT5EncoderModel(metaclass=DummyObject):
     _backends = ["torch"]
 

tests/models/mra/test_modeling_mra.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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 MRA model. """
+
+
+import unittest
+
+from transformers import MraConfig, is_torch_available
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        MraForMaskedLM,
+        MraForMultipleChoice,
+        MraForQuestionAnswering,
+        MraForSequenceClassification,
+        MraForTokenClassification,
+        MraModel,
+    )
+    from transformers.models.mra.modeling_mra import MRA_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+class MraModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=2,
+        seq_length=8,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=16,
+        num_hidden_layers=5,
+        num_attention_heads=2,
+        intermediate_size=36,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        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.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])
+
+        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 MraConfig(
+            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,
+        )
+
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 300
+        return config
+
+    def prepare_config_and_inputs_for_decoder(self):
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = self.prepare_config_and_inputs()
+
+        config.is_decoder = True
+        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
+        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        return (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+    def create_and_check_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = MraModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        result = model(input_ids, token_type_ids=token_type_ids)
+        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_model_as_decoder(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.add_cross_attention = True
+        model = MraModel(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+        )
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_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, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = MraForMaskedLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_for_question_answering(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = MraForQuestionAnswering(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            start_positions=sequence_labels,
+            end_positions=sequence_labels,
+        )
+        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
+
+    def create_and_check_for_sequence_classification(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_labels = self.num_labels
+        model = MraForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def create_and_check_for_token_classification(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_labels = self.num_labels
+        model = MraForTokenClassification(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
+
+    def create_and_check_for_multiple_choice(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_choices = self.num_choices
+        model = MraForMultipleChoice(config=config)
+        model.to(torch_device)
+        model.eval()
+        multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        result = model(
+            multiple_choice_inputs_ids,
+            attention_mask=multiple_choice_input_mask,
+            token_type_ids=multiple_choice_token_type_ids,
+            labels=choice_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
+
+    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, "token_type_ids": token_type_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class MraModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            MraModel,
+            MraForMaskedLM,
+            MraForMultipleChoice,
+            MraForQuestionAnswering,
+            MraForSequenceClassification,
+            MraForTokenClassification,
+        )
+        if is_torch_available()
+        else ()
+    )
+    test_pruning = False
+    test_headmasking = False
+    test_torchscript = False
+    has_attentions = False
+
+    all_generative_model_classes = ()
+
+    def setUp(self):
+        self.model_tester = MraModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=MraConfig, 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_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
+
+    def test_for_multiple_choice(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in MRA_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = MraModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+    @unittest.skip(reason="MRA does not output attentions")
+    def test_attention_outputs(self):
+        return
+
+
+@require_torch
+class MraModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference_no_head(self):
+        model = MraModel.from_pretrained("uw-madison/mra-base-512-4")
+        input_ids = torch.arange(256).unsqueeze(0)
+
+        with torch.no_grad():
+            output = model(input_ids)[0]
+
+        expected_shape = torch.Size((1, 256, 768))
+        self.assertEqual(output.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[[-0.0140, 0.0830, -0.0381], [0.1546, 0.1402, 0.0220], [0.1162, 0.0851, 0.0165]]]
+        )
+
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_inference_masked_lm(self):
+        model = MraForMaskedLM.from_pretrained("uw-madison/mra-base-512-4")
+        input_ids = torch.arange(256).unsqueeze(0)
+
+        with torch.no_grad():
+            output = model(input_ids)[0]
+
+        vocab_size = 50265
+
+        expected_shape = torch.Size((1, 256, vocab_size))
+        self.assertEqual(output.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[[9.2595, -3.6038, 11.8819], [9.3869, -3.2693, 11.0956], [11.8524, -3.4938, 13.1210]]]
+        )
+
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_inference_masked_lm_long_input(self):
+        model = MraForMaskedLM.from_pretrained("uw-madison/mra-base-4096-8-d3")
+        input_ids = torch.arange(4096).unsqueeze(0)
+
+        with torch.no_grad():
+            output = model(input_ids)[0]
+
+        vocab_size = 50265
+
+        expected_shape = torch.Size((1, 4096, vocab_size))
+        self.assertEqual(output.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[[5.4789, -2.3564, 7.5064], [7.9067, -1.3369, 9.9668], [9.0712, -1.8106, 7.0380]]]
+        )
+
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))