Upload DebertaForMaskedLM

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,36 @@
+{
+  "_name_or_path": "babylm24_small_mlsm0.15_babylm_l12_1500_mlm0_not-reinit_10M_final/10000/",
+  "architectures": [
+    "DebertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "auto_map": {
+    "AutoModelForMaskedLM": "modeling_deberta.DebertaForMaskedLM"
+  },
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-07,
+  "max_position_embeddings": 512,
+  "max_relative_positions": -1,
+  "model_type": "deberta",
+  "num_attention_heads": 12,
+  "num_concepts": 1500,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "pooler_dropout": 0,
+  "pooler_hidden_act": "gelu",
+  "pooler_hidden_size": 768,
+  "pos_att_type": [
+    "c2p",
+    "p2c"
+  ],
+  "position_biased_input": false,
+  "relative_attention": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.41.2",
+  "type_vocab_size": 0,
+  "vocab_size": 26500
+}

configuration_deberta.py

@@ -0,0 +1,191 @@
+# coding=utf-8
+# Copyright 2020, Microsoft and 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.
+"""DeBERTa model configuration"""
+
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Any, Mapping, Optional, Union
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.onnx import OnnxConfig
+from transformers.utils import logging
+
+
+if TYPE_CHECKING:
+    from transformers import FeatureExtractionMixin, PreTrainedTokenizerBase, TensorType
+
+
+logger = logging.get_logger(__name__)
+
+
+class DebertaConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DebertaModel`] or a [`TFDebertaModel`]. It is
+    used to instantiate a DeBERTa 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 DeBERTa
+    [microsoft/deberta-base](https://huggingface.co/microsoft/deberta-base) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Arguments:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the DeBERTa model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`DebertaModel`] or [`TFDebertaModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality 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):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"`, `"gelu"`, `"tanh"`, `"gelu_fast"`, `"mish"`, `"linear"`, `"sigmoid"` and `"gelu_new"`
+            are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability 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 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`DebertaModel`] or [`TFDebertaModel`].
+        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-12):
+            The epsilon used by the layer normalization layers.
+        relative_attention (`bool`, *optional*, defaults to `False`):
+            Whether use relative position encoding.
+        max_relative_positions (`int`, *optional*, defaults to 1):
+            The range of relative positions `[-max_position_embeddings, max_position_embeddings]`. Use the same value
+            as `max_position_embeddings`.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            The value used to pad input_ids.
+        position_biased_input (`bool`, *optional*, defaults to `True`):
+            Whether add absolute position embedding to content embedding.
+        pos_att_type (`List[str]`, *optional*):
+            The type of relative position attention, it can be a combination of `["p2c", "c2p"]`, e.g. `["p2c"]`,
+            `["p2c", "c2p"]`.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+
+    Example:
+
+    ```python
+    >>> from transformers import DebertaConfig, DebertaModel
+
+    >>> # Initializing a DeBERTa microsoft/deberta-base style configuration
+    >>> configuration = DebertaConfig()
+
+    >>> # Initializing a model (with random weights) from the microsoft/deberta-base style configuration
+    >>> model = DebertaModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "deberta"
+
+    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=0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-7,
+        relative_attention=False,
+        max_relative_positions=-1,
+        pad_token_id=0,
+        position_biased_input=True,
+        pos_att_type=None,
+        pooler_dropout=0,
+        pooler_hidden_act="gelu",
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        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.initializer_range = initializer_range
+        self.relative_attention = relative_attention
+        self.max_relative_positions = max_relative_positions
+        self.pad_token_id = pad_token_id
+        self.position_biased_input = position_biased_input
+
+        # Backwards compatibility
+        if isinstance(pos_att_type, str):
+            pos_att_type = [x.strip() for x in pos_att_type.lower().split("|")]
+
+        self.pos_att_type = pos_att_type
+        self.vocab_size = vocab_size
+        self.layer_norm_eps = layer_norm_eps
+
+        self.pooler_hidden_size = kwargs.get("pooler_hidden_size", hidden_size)
+        self.pooler_dropout = pooler_dropout
+        self.pooler_hidden_act = pooler_hidden_act
+
+
+# Copied from transformers.models.deberta_v2.configuration_deberta_v2.DebertaV2OnnxConfig
+class DebertaOnnxConfig(OnnxConfig):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        if self.task == "multiple-choice":
+            dynamic_axis = {0: "batch", 1: "choice", 2: "sequence"}
+        else:
+            dynamic_axis = {0: "batch", 1: "sequence"}
+        if self._config.type_vocab_size > 0:
+            return OrderedDict(
+                [("input_ids", dynamic_axis), ("attention_mask", dynamic_axis), ("token_type_ids", dynamic_axis)]
+            )
+        else:
+            return OrderedDict([("input_ids", dynamic_axis), ("attention_mask", dynamic_axis)])
+
+    @property
+    def default_onnx_opset(self) -> int:
+        return 12
+
+    def generate_dummy_inputs(
+        self,
+        preprocessor: Union["PreTrainedTokenizerBase", "FeatureExtractionMixin"],
+        batch_size: int = -1,
+        seq_length: int = -1,
+        num_choices: int = -1,
+        is_pair: bool = False,
+        framework: Optional["TensorType"] = None,
+        num_channels: int = 3,
+        image_width: int = 40,
+        image_height: int = 40,
+        tokenizer: "PreTrainedTokenizerBase" = None,
+    ) -> Mapping[str, Any]:
+        dummy_inputs = super().generate_dummy_inputs(preprocessor=preprocessor, framework=framework)
+        if self._config.type_vocab_size == 0 and "token_type_ids" in dummy_inputs:
+            del dummy_inputs["token_type_ids"]
+        return dummy_inputs

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8a1b6e450889d22ab47122c4a7998498be289b152a98ea42b91733d4ac51019
+size 633999824

modeling_deberta.py

@@ -0,0 +1,1500 @@
+# coding=utf-8
+# Copyright 2020 Microsoft and the Hugging Face 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.
+"""PyTorch DeBERTa model."""
+
+from collections.abc import Sequence
+from typing import Optional, Tuple, Union
+
+import os, pickle
+import transformers
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutput,
+    MaskedLMOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import softmax_backward_data
+from transformers.utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from .configuration_deberta import DebertaConfig
+
+
+logger = logging.get_logger(__name__)
+_CONFIG_FOR_DOC = "DebertaConfig"
+_CHECKPOINT_FOR_DOC = "microsoft/deberta-base"
+
+# Masked LM docstring
+_CHECKPOINT_FOR_MASKED_LM = "lsanochkin/deberta-large-feedback"
+_MASKED_LM_EXPECTED_OUTPUT = "' Paris'"
+_MASKED_LM_EXPECTED_LOSS = "0.54"
+
+# QuestionAnswering docstring
+_CHECKPOINT_FOR_QA = "Palak/microsoft_deberta-large_squad"
+_QA_EXPECTED_OUTPUT = "' a nice puppet'"
+_QA_EXPECTED_LOSS = 0.14
+_QA_TARGET_START_INDEX = 12
+_QA_TARGET_END_INDEX = 14
+
+
+class ContextPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.pooler_hidden_size, config.pooler_hidden_size)
+        self.dropout = StableDropout(config.pooler_dropout)
+        self.config = config
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+
+        context_token = hidden_states[:, 0]
+        context_token = self.dropout(context_token)
+        pooled_output = self.dense(context_token)
+        pooled_output = ACT2FN[self.config.pooler_hidden_act](pooled_output)
+        return pooled_output
+
+    @property
+    def output_dim(self):
+        return self.config.hidden_size
+
+
+class XSoftmax(torch.autograd.Function):
+    """
+    Masked Softmax which is optimized for saving memory
+
+    Args:
+        input (`torch.tensor`): The input tensor that will apply softmax.
+        mask (`torch.IntTensor`):
+            The mask matrix where 0 indicate that element will be ignored in the softmax calculation.
+        dim (int): The dimension that will apply softmax
+
+    Example:
+
+    ```python
+    >>> import torch
+    >>> from transformers.models.deberta.modeling_deberta import XSoftmax
+
+    >>> # Make a tensor
+    >>> x = torch.randn([4, 20, 100])
+
+    >>> # Create a mask
+    >>> mask = (x > 0).int()
+
+    >>> # Specify the dimension to apply softmax
+    >>> dim = -1
+
+    >>> y = XSoftmax.apply(x, mask, dim)
+    ```"""
+
+    @staticmethod
+    def forward(ctx, input, mask, dim):
+        ctx.dim = dim
+        rmask = ~(mask.to(torch.bool))
+
+        output = input.masked_fill(rmask, torch.tensor(torch.finfo(input.dtype).min))
+        output = torch.softmax(output, ctx.dim)
+        output.masked_fill_(rmask, 0)
+        ctx.save_for_backward(output)
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        (output,) = ctx.saved_tensors
+        inputGrad = softmax_backward_data(ctx, grad_output, output, ctx.dim, output)
+        return inputGrad, None, None
+
+    @staticmethod
+    def symbolic(g, self, mask, dim):
+        import torch.onnx.symbolic_helper as sym_help
+        from torch.onnx.symbolic_opset9 import masked_fill, softmax
+
+        mask_cast_value = g.op("Cast", mask, to_i=sym_help.cast_pytorch_to_onnx["Long"])
+        r_mask = g.op(
+            "Cast",
+            g.op("Sub", g.op("Constant", value_t=torch.tensor(1, dtype=torch.int64)), mask_cast_value),
+            to_i=sym_help.cast_pytorch_to_onnx["Bool"],
+        )
+        output = masked_fill(
+            g, self, r_mask, g.op("Constant", value_t=torch.tensor(torch.finfo(self.type().dtype()).min))
+        )
+        output = softmax(g, output, dim)
+        return masked_fill(g, output, r_mask, g.op("Constant", value_t=torch.tensor(0, dtype=torch.bool)))
+
+
+class DropoutContext:
+    def __init__(self):
+        self.dropout = 0
+        self.mask = None
+        self.scale = 1
+        self.reuse_mask = True
+
+
+def get_mask(input, local_context):
+    if not isinstance(local_context, DropoutContext):
+        dropout = local_context
+        mask = None
+    else:
+        dropout = local_context.dropout
+        dropout *= local_context.scale
+        mask = local_context.mask if local_context.reuse_mask else None
+
+    if dropout > 0 and mask is None:
+        mask = (1 - torch.empty_like(input).bernoulli_(1 - dropout)).to(torch.bool)
+
+    if isinstance(local_context, DropoutContext):
+        if local_context.mask is None:
+            local_context.mask = mask
+
+    return mask, dropout
+
+
+class XDropout(torch.autograd.Function):
+    """Optimized dropout function to save computation and memory by using mask operation instead of multiplication."""
+
+    @staticmethod
+    def forward(ctx, input, local_ctx):
+        mask, dropout = get_mask(input, local_ctx)
+        ctx.scale = 1.0 / (1 - dropout)
+        if dropout > 0:
+            ctx.save_for_backward(mask)
+            return input.masked_fill(mask, 0) * ctx.scale
+        else:
+            return input
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        if ctx.scale > 1:
+            (mask,) = ctx.saved_tensors
+            return grad_output.masked_fill(mask, 0) * ctx.scale, None
+        else:
+            return grad_output, None
+
+    @staticmethod
+    def symbolic(g: torch._C.Graph, input: torch._C.Value, local_ctx: Union[float, DropoutContext]) -> torch._C.Value:
+        from torch.onnx import symbolic_opset12
+
+        dropout_p = local_ctx
+        if isinstance(local_ctx, DropoutContext):
+            dropout_p = local_ctx.dropout
+        # StableDropout only calls this function when training.
+        train = True
+        # TODO: We should check if the opset_version being used to export
+        # is > 12 here, but there's no good way to do that. As-is, if the
+        # opset_version < 12, export will fail with a CheckerError.
+        # Once https://github.com/pytorch/pytorch/issues/78391 is fixed, do something like:
+        # if opset_version < 12:
+        #   return torch.onnx.symbolic_opset9.dropout(g, input, dropout_p, train)
+        return symbolic_opset12.dropout(g, input, dropout_p, train)
+
+
+class StableDropout(nn.Module):
+    """
+    Optimized dropout module for stabilizing the training
+
+    Args:
+        drop_prob (float): the dropout probabilities
+    """
+
+    def __init__(self, drop_prob):
+        super().__init__()
+        self.drop_prob = drop_prob
+        self.count = 0
+        self.context_stack = None
+
+    def forward(self, x):
+        """
+        Call the module
+
+        Args:
+            x (`torch.tensor`): The input tensor to apply dropout
+        """
+        if self.training and self.drop_prob > 0:
+            return XDropout.apply(x, self.get_context())
+        return x
+
+    def clear_context(self):
+        self.count = 0
+        self.context_stack = None
+
+    def init_context(self, reuse_mask=True, scale=1):
+        if self.context_stack is None:
+            self.context_stack = []
+        self.count = 0
+        for c in self.context_stack:
+            c.reuse_mask = reuse_mask
+            c.scale = scale
+
+    def get_context(self):
+        if self.context_stack is not None:
+            if self.count >= len(self.context_stack):
+                self.context_stack.append(DropoutContext())
+            ctx = self.context_stack[self.count]
+            ctx.dropout = self.drop_prob
+            self.count += 1
+            return ctx
+        else:
+            return self.drop_prob
+
+
+class DebertaLayerNorm(nn.Module):
+    """LayerNorm module in the TF style (epsilon inside the square root)."""
+
+    def __init__(self, size, eps=1e-12):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(size))
+        self.bias = nn.Parameter(torch.zeros(size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_type = hidden_states.dtype
+        hidden_states = hidden_states.float()
+        mean = hidden_states.mean(-1, keepdim=True)
+        variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True)
+        hidden_states = (hidden_states - mean) / torch.sqrt(variance + self.variance_epsilon)
+        hidden_states = hidden_states.to(input_type)
+        y = self.weight * hidden_states + self.bias
+        return y
+
+
+class DebertaSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = DebertaLayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_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 DebertaAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = DisentangledSelfAttention(config)
+        self.output = DebertaSelfOutput(config)
+        self.config = config
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_attentions=False,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+    ):
+        self_output = self.self(
+            hidden_states,
+            attention_mask,
+            output_attentions,
+            query_states=query_states,
+            relative_pos=relative_pos,
+            rel_embeddings=rel_embeddings,
+        )
+        if output_attentions:
+            self_output, att_matrix = self_output
+        if query_states is None:
+            query_states = hidden_states
+        attention_output = self.output(self_output, query_states)
+
+        if output_attentions:
+            return (attention_output, att_matrix)
+        else:
+            return attention_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->Deberta
+class DebertaIntermediate(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
+
+
+class DebertaOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = DebertaLayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def forward(self, hidden_states, input_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 DebertaLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = DebertaAttention(config)
+        self.intermediate = DebertaIntermediate(config)
+        self.output = DebertaOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+        output_attentions=False,
+    ):
+        attention_output = self.attention(
+            hidden_states,
+            attention_mask,
+            output_attentions=output_attentions,
+            query_states=query_states,
+            relative_pos=relative_pos,
+            rel_embeddings=rel_embeddings,
+        )
+        if output_attentions:
+            attention_output, att_matrix = attention_output
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        if output_attentions:
+            return (layer_output, att_matrix)
+        else:
+            return layer_output
+
+
+class DebertaEncoder(nn.Module):
+    """Modified BertEncoder with relative position bias support"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.layer = nn.ModuleList([DebertaLayer(config) for _ in range(config.num_hidden_layers)])
+        self.relative_attention = getattr(config, "relative_attention", False)
+        if self.relative_attention:
+            self.max_relative_positions = getattr(config, "max_relative_positions", -1)
+            if self.max_relative_positions < 1:
+                self.max_relative_positions = config.max_position_embeddings
+            self.rel_embeddings = nn.Embedding(self.max_relative_positions * 2, config.hidden_size)
+        self.gradient_checkpointing = False
+
+    def get_rel_embedding(self):
+        rel_embeddings = self.rel_embeddings.weight if self.relative_attention else None
+        return rel_embeddings
+
+    def get_attention_mask(self, attention_mask):
+        if attention_mask.dim() <= 2:
+            extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+            attention_mask = extended_attention_mask * extended_attention_mask.squeeze(-2).unsqueeze(-1)
+        elif attention_mask.dim() == 3:
+            attention_mask = attention_mask.unsqueeze(1)
+
+        return attention_mask
+
+    def get_rel_pos(self, hidden_states, query_states=None, relative_pos=None):
+        if self.relative_attention and relative_pos is None:
+            q = query_states.size(-2) if query_states is not None else hidden_states.size(-2)
+            relative_pos = build_relative_position(q, hidden_states.size(-2), hidden_states.device)
+        return relative_pos
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_hidden_states=True,
+        output_attentions=False,
+        query_states=None,
+        relative_pos=None,
+        return_dict=True,
+    ):
+        attention_mask = self.get_attention_mask(attention_mask)
+        relative_pos = self.get_rel_pos(hidden_states, query_states, relative_pos)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        if isinstance(hidden_states, Sequence):
+            next_kv = hidden_states[0]
+        else:
+            next_kv = hidden_states
+        rel_embeddings = self.get_rel_embedding()
+        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:
+                hidden_states = self._gradient_checkpointing_func(
+                    layer_module.__call__,
+                    next_kv,
+                    attention_mask,
+                    query_states,
+                    relative_pos,
+                    rel_embeddings,
+                    output_attentions,
+                )
+            else:
+                hidden_states = layer_module(
+                    next_kv,
+                    attention_mask,
+                    query_states=query_states,
+                    relative_pos=relative_pos,
+                    rel_embeddings=rel_embeddings,
+                    output_attentions=output_attentions,
+                )
+
+            if output_attentions:
+                hidden_states, att_m = hidden_states
+
+            if query_states is not None:
+                query_states = hidden_states
+                if isinstance(hidden_states, Sequence):
+                    next_kv = hidden_states[i + 1] if i + 1 < len(self.layer) else None
+            else:
+                next_kv = hidden_states
+
+            if output_attentions:
+                all_attentions = all_attentions + (att_m,)
+
+        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, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+def build_relative_position(query_size, key_size, device):
+    """
+    Build relative position according to the query and key
+
+    We assume the absolute position of query \\(P_q\\) is range from (0, query_size) and the absolute position of key
+    \\(P_k\\) is range from (0, key_size), The relative positions from query to key is \\(R_{q \\rightarrow k} = P_q -
+    P_k\\)
+
+    Args:
+        query_size (int): the length of query
+        key_size (int): the length of key
+
+    Return:
+        `torch.LongTensor`: A tensor with shape [1, query_size, key_size]
+
+    """
+
+    q_ids = torch.arange(query_size, dtype=torch.long, device=device)
+    k_ids = torch.arange(key_size, dtype=torch.long, device=device)
+    rel_pos_ids = q_ids[:, None] - k_ids.view(1, -1).repeat(query_size, 1)
+    rel_pos_ids = rel_pos_ids[:query_size, :]
+    rel_pos_ids = rel_pos_ids.unsqueeze(0)
+    return rel_pos_ids
+
+
+@torch.jit.script
+def c2p_dynamic_expand(c2p_pos, query_layer, relative_pos):
+    return c2p_pos.expand([query_layer.size(0), query_layer.size(1), query_layer.size(2), relative_pos.size(-1)])
+
+
+@torch.jit.script
+def p2c_dynamic_expand(c2p_pos, query_layer, key_layer):
+    return c2p_pos.expand([query_layer.size(0), query_layer.size(1), key_layer.size(-2), key_layer.size(-2)])
+
+
+@torch.jit.script
+def pos_dynamic_expand(pos_index, p2c_att, key_layer):
+    return pos_index.expand(p2c_att.size()[:2] + (pos_index.size(-2), key_layer.size(-2)))
+
+
+class DisentangledSelfAttention(nn.Module):
+    """
+    Disentangled self-attention module
+
+    Parameters:
+        config (`str`):
+            A model config class instance with the configuration to build a new model. The schema is similar to
+            *BertConfig*, for more details, please refer [`DebertaConfig`]
+
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0:
+            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.in_proj = nn.Linear(config.hidden_size, self.all_head_size * 3, bias=False)
+        self.q_bias = nn.Parameter(torch.zeros((self.all_head_size), dtype=torch.float))
+        self.v_bias = nn.Parameter(torch.zeros((self.all_head_size), dtype=torch.float))
+        self.pos_att_type = config.pos_att_type if config.pos_att_type is not None else []
+
+        self.relative_attention = getattr(config, "relative_attention", False)
+        self.talking_head = getattr(config, "talking_head", False)
+
+        if self.talking_head:
+            self.head_logits_proj = nn.Linear(config.num_attention_heads, config.num_attention_heads, bias=False)
+            self.head_weights_proj = nn.Linear(config.num_attention_heads, config.num_attention_heads, bias=False)
+
+        if self.relative_attention:
+            self.max_relative_positions = getattr(config, "max_relative_positions", -1)
+            if self.max_relative_positions < 1:
+                self.max_relative_positions = config.max_position_embeddings
+            self.pos_dropout = StableDropout(config.hidden_dropout_prob)
+
+            if "c2p" in self.pos_att_type:
+                self.pos_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=False)
+            if "p2c" in self.pos_att_type:
+                self.pos_q_proj = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = StableDropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, -1)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_attentions=False,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+    ):
+        """
+        Call the module
+
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input states to the module usually the output from previous layer, it will be the Q,K and V in
+                *Attention(Q,K,V)*
+
+            attention_mask (`torch.BoolTensor`):
+                An attention mask matrix of shape [*B*, *N*, *N*] where *B* is the batch size, *N* is the maximum
+                sequence length in which element [i,j] = *1* means the *i* th token in the input can attend to the *j*
+                th token.
+
+            output_attentions (`bool`, *optional*):
+                Whether return the attention matrix.
+
+            query_states (`torch.FloatTensor`, *optional*):
+                The *Q* state in *Attention(Q,K,V)*.
+
+            relative_pos (`torch.LongTensor`):
+                The relative position encoding between the tokens in the sequence. It's of shape [*B*, *N*, *N*] with
+                values ranging in [*-max_relative_positions*, *max_relative_positions*].
+
+            rel_embeddings (`torch.FloatTensor`):
+                The embedding of relative distances. It's a tensor of shape [\\(2 \\times
+                \\text{max_relative_positions}\\), *hidden_size*].
+
+
+        """
+        if query_states is None:
+            qp = self.in_proj(hidden_states)  # .split(self.all_head_size, dim=-1)
+            query_layer, key_layer, value_layer = self.transpose_for_scores(qp).chunk(3, dim=-1)
+        else:
+
+            def linear(w, b, x):
+                if b is not None:
+                    return torch.matmul(x, w.t()) + b.t()
+                else:
+                    return torch.matmul(x, w.t())  # + b.t()
+
+            ws = self.in_proj.weight.chunk(self.num_attention_heads * 3, dim=0)
+            qkvw = [torch.cat([ws[i * 3 + k] for i in range(self.num_attention_heads)], dim=0) for k in range(3)]
+            qkvb = [None] * 3
+
+            q = linear(qkvw[0], qkvb[0], query_states.to(dtype=qkvw[0].dtype))
+            k, v = [linear(qkvw[i], qkvb[i], hidden_states.to(dtype=qkvw[i].dtype)) for i in range(1, 3)]
+            query_layer, key_layer, value_layer = [self.transpose_for_scores(x) for x in [q, k, v]]
+
+        query_layer = query_layer + self.transpose_for_scores(self.q_bias[None, None, :])
+        value_layer = value_layer + self.transpose_for_scores(self.v_bias[None, None, :])
+
+        rel_att = None
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        scale_factor = 1 + len(self.pos_att_type)
+        scale = torch.sqrt(torch.tensor(query_layer.size(-1), dtype=torch.float) * scale_factor)
+        query_layer = query_layer / scale.to(dtype=query_layer.dtype)
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        if self.relative_attention:
+            rel_embeddings = self.pos_dropout(rel_embeddings)
+            rel_att = self.disentangled_att_bias(query_layer, key_layer, relative_pos, rel_embeddings, scale_factor)
+
+        if rel_att is not None:
+            attention_scores = attention_scores + rel_att
+
+        # bxhxlxd
+        if self.talking_head:
+            attention_scores = self.head_logits_proj(attention_scores.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+
+        attention_probs = XSoftmax.apply(attention_scores, attention_mask, -1)
+        attention_probs = self.dropout(attention_probs)
+        if self.talking_head:
+            attention_probs = self.head_weights_proj(attention_probs.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (-1,)
+        context_layer = context_layer.view(new_context_layer_shape)
+        if output_attentions:
+            return (context_layer, attention_probs)
+        else:
+            return context_layer
+
+    def disentangled_att_bias(self, query_layer, key_layer, relative_pos, rel_embeddings, scale_factor):
+        if relative_pos is None:
+            q = query_layer.size(-2)
+            relative_pos = build_relative_position(q, key_layer.size(-2), query_layer.device)
+        if relative_pos.dim() == 2:
+            relative_pos = relative_pos.unsqueeze(0).unsqueeze(0)
+        elif relative_pos.dim() == 3:
+            relative_pos = relative_pos.unsqueeze(1)
+        # bxhxqxk
+        elif relative_pos.dim() != 4:
+            raise ValueError(f"Relative position ids must be of dim 2 or 3 or 4. {relative_pos.dim()}")
+
+        att_span = min(max(query_layer.size(-2), key_layer.size(-2)), self.max_relative_positions)
+        relative_pos = relative_pos.long().to(query_layer.device)
+        rel_embeddings = rel_embeddings[
+            self.max_relative_positions - att_span : self.max_relative_positions + att_span, :
+        ].unsqueeze(0)
+
+        score = 0
+
+        # content->position
+        if "c2p" in self.pos_att_type:
+            pos_key_layer = self.pos_proj(rel_embeddings)
+            pos_key_layer = self.transpose_for_scores(pos_key_layer)
+            c2p_att = torch.matmul(query_layer, pos_key_layer.transpose(-1, -2))
+            c2p_pos = torch.clamp(relative_pos + att_span, 0, att_span * 2 - 1)
+            c2p_att = torch.gather(c2p_att, dim=-1, index=c2p_dynamic_expand(c2p_pos, query_layer, relative_pos))
+            score += c2p_att
+
+        # position->content
+        if "p2c" in self.pos_att_type:
+            pos_query_layer = self.pos_q_proj(rel_embeddings)
+            pos_query_layer = self.transpose_for_scores(pos_query_layer)
+            pos_query_layer /= torch.sqrt(torch.tensor(pos_query_layer.size(-1), dtype=torch.float) * scale_factor)
+            if query_layer.size(-2) != key_layer.size(-2):
+                r_pos = build_relative_position(key_layer.size(-2), key_layer.size(-2), query_layer.device)
+            else:
+                r_pos = relative_pos
+            p2c_pos = torch.clamp(-r_pos + att_span, 0, att_span * 2 - 1)
+            p2c_att = torch.matmul(key_layer, pos_query_layer.transpose(-1, -2).to(dtype=key_layer.dtype))
+            p2c_att = torch.gather(
+                p2c_att, dim=-1, index=p2c_dynamic_expand(p2c_pos, query_layer, key_layer)
+            ).transpose(-1, -2)
+
+            if query_layer.size(-2) != key_layer.size(-2):
+                pos_index = relative_pos[:, :, :, 0].unsqueeze(-1)
+                p2c_att = torch.gather(p2c_att, dim=-2, index=pos_dynamic_expand(pos_index, p2c_att, key_layer))
+            score += p2c_att
+
+        return score
+
+
+class DebertaEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        pad_token_id = getattr(config, "pad_token_id", 0)
+        self.embedding_size = getattr(config, "embedding_size", config.hidden_size)
+        self.word_embeddings = nn.Embedding(config.vocab_size, self.embedding_size, padding_idx=pad_token_id)
+
+        self.position_biased_input = getattr(config, "position_biased_input", True)
+        if not self.position_biased_input:
+            self.position_embeddings = None
+        else:
+            self.position_embeddings = nn.Embedding(config.max_position_embeddings, self.embedding_size)
+
+        if config.type_vocab_size > 0:
+            self.token_type_embeddings = nn.Embedding(config.type_vocab_size, self.embedding_size)
+
+        if self.embedding_size != config.hidden_size:
+            self.embed_proj = nn.Linear(self.embedding_size, config.hidden_size, bias=False)
+        self.LayerNorm = DebertaLayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+        self.config = config
+
+        # 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)), persistent=False
+        )
+
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, mask=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]
+
+        if token_type_ids is None:
+            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)
+
+        if self.position_embeddings is not None:
+            position_embeddings = self.position_embeddings(position_ids.long())
+        else:
+            position_embeddings = torch.zeros_like(inputs_embeds)
+
+        embeddings = inputs_embeds
+        if self.position_biased_input:
+            embeddings += position_embeddings
+        if self.config.type_vocab_size > 0:
+            token_type_embeddings = self.token_type_embeddings(token_type_ids)
+            embeddings += token_type_embeddings
+
+        if self.embedding_size != self.config.hidden_size:
+            embeddings = self.embed_proj(embeddings)
+
+        embeddings = self.LayerNorm(embeddings)
+
+        if mask is not None:
+            if mask.dim() != embeddings.dim():
+                if mask.dim() == 4:
+                    mask = mask.squeeze(1).squeeze(1)
+                mask = mask.unsqueeze(2)
+            mask = mask.to(embeddings.dtype)
+
+            embeddings = embeddings * mask
+
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class DebertaPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = DebertaConfig
+    base_model_prefix = "deberta"
+    _keys_to_ignore_on_load_unexpected = ["position_embeddings"]
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            if module.weight.requires_grad==False: # a hack for skipping the nb params
+                return
+            # 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_()
+
+
+DEBERTA_START_DOCSTRING = r"""
+    The DeBERTa model was proposed in [DeBERTa: Decoding-enhanced BERT with Disentangled
+    Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen. It's build
+    on top of BERT/RoBERTa with two improvements, i.e. disentangled attention and enhanced mask decoder. With those two
+    improvements, it out perform BERT/RoBERTa on a majority of tasks with 80GB pretraining data.
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+
+    Parameters:
+        config ([`DebertaConfig`]): 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.
+"""
+
+DEBERTA_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)
+        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_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        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 DeBERTa Model transformer outputting raw hidden-states without any specific head on top.",
+    DEBERTA_START_DOCSTRING,
+)
+class DebertaModel(DebertaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = DebertaEmbeddings(config)
+        self.encoder = DebertaEncoder(config)
+        self.z_steps = 0
+        self.config = 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, new_embeddings):
+        self.embeddings.word_embeddings = new_embeddings
+
+    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
+        """
+        raise NotImplementedError("The prune function is not implemented in DeBERTa model.")
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutput,
+        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,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        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:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            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")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+        )
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask,
+            output_hidden_states=True,
+            output_attentions=output_attentions,
+            return_dict=return_dict,
+        )
+        encoded_layers = encoder_outputs[1]
+
+        if self.z_steps > 1:
+            hidden_states = encoded_layers[-2]
+            layers = [self.encoder.layer[-1] for _ in range(self.z_steps)]
+            query_states = encoded_layers[-1]
+            rel_embeddings = self.encoder.get_rel_embedding()
+            attention_mask = self.encoder.get_attention_mask(attention_mask)
+            rel_pos = self.encoder.get_rel_pos(embedding_output)
+            for layer in layers[1:]:
+                query_states = layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions=False,
+                    query_states=query_states,
+                    relative_pos=rel_pos,
+                    rel_embeddings=rel_embeddings,
+                )
+                encoded_layers.append(query_states)
+
+        sequence_output = encoded_layers[-1]
+
+        if not return_dict:
+            return (sequence_output,) + encoder_outputs[(1 if output_hidden_states else 2) :]
+
+        return BaseModelOutput(
+            last_hidden_state=sequence_output,
+            hidden_states=encoder_outputs.hidden_states if output_hidden_states else None,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings("""DeBERTa Model with a `language modeling` head on top.""", DEBERTA_START_DOCSTRING)
+class DebertaForMaskedLM(DebertaPreTrainedModel):
+    _tied_weights_keys = ["cls.predictions.decoder.weight", "cls.predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.deberta = DebertaModel(config)
+        self.cls = DebertaOnlyMLMHead(config)
+
+        self.post_cls = DebertaFinalMLMHead(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+        self.num_concepts = config.num_concepts
+        #self.nb = DebertaNB(config)
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_MASKED_LM,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+        mask="[MASK]",
+        expected_output=_MASKED_LM_EXPECTED_OUTPUT,
+        expected_loss=_MASKED_LM_EXPECTED_LOSS,
+    )
+    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,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = 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.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+        #prediction_scores = self.nb(prediction_scores)
+        prediction_scores = self.post_cls(prediction_scores)
+
+        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 - self.num_concepts), 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,
+        )
+
+class DebertaFinalMLMHead(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.num_concepts = config.num_concepts
+        self.head = torch.nn.Linear(self.num_concepts, config.vocab_size - self.num_concepts)
+
+    def forward(self, pre_logits):
+        concept_scores = pre_logits[:,:,-self.num_concepts:]
+        return self.head(concept_scores)
+
+class DebertaNB(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.top_k = config.top_k
+        self.prob_threshold = config.prob_threshold
+        print(self.top_k, self.prob_threshold)
+        nb = pickle.load(open(f'{os.path.dirname(os.path.abspath(__file__))}/nb_final_multinomial_{self.top_k}_{self.prob_threshold}.pickle', 'rb'))
+        #nb = pickle.load(open(f'nb_final_multinomial_{self.top_k}_1.0.pickle', 'rb'))
+        self.effective_vocab, self.num_concepts = nb.feature_count_.shape
+        #self.nb = torch.nn.Linear(self.num_concepts, self.effective_vocab)
+        with torch.no_grad(): 
+            class_log_prior = torch.from_numpy(nb.class_log_prior_).float()
+            #smallest_non_inf_prior = torch.min(class_log_prior[class_log_prior!=-torch.inf])
+            class_log_prior[class_log_prior==-torch.inf] = -1000 #5 * smallest_non_inf_prior
+            
+            self.nb_features_log_prob = torch.from_numpy(nb.feature_log_prob_.T).float()
+            self.nb_class_log_prior = class_log_prior
+
+            #self.nb.bias.copy_(class_log_prior)
+            #self.nb.weight.copy_(torch.from_numpy(nb.feature_log_prob_))
+            #for param in self.nb.parameters():
+            #    param.requires_grad = False
+    
+    def forward(self, prediction_scores):
+        #print(self.nb_class_log_prior.max(), self.nb_class_log_prior.min())
+        #print(self.nb.bias.max(), self.nb.bias.min())
+        #import sys
+        #sys.exit(2)
+        num_sequences, num_tokens, _ = prediction_scores.shape
+        concept_scores = prediction_scores[:,:,self.effective_vocab:]
+        batch_size, token_num, _ = concept_scores.shape
+        concept_probs = torch.nn.functional.softmax(concept_scores, dim=-1).view(-1, self.num_concepts)
+        probs, relevant_features = torch.topk(concept_probs, self.top_k, dim=-1)
+        
+        thresholds = torch.tensor([[self.prob_threshold] for _ in range(batch_size * token_num)])
+        limits = torch.searchsorted(torch.cumsum(probs, dim=-1),
+                                    torch.tensor([[self.prob_threshold] for _ in range(batch_size*token_num)], device=probs.device))
+
+        filtered_relevant_features = []
+        for feats, lims in zip(relevant_features, limits):
+            limit = min(self.top_k, lims[0].item())
+            filtered_relevant_features.append(torch.nn.functional.pad(feats[0:limit], pad=[0, self.top_k - limit], value=feats[0]))
+        relevant_features = torch.vstack(filtered_relevant_features).view(batch_size, token_num, -1)
+        device = concept_scores.device
+
+        features = torch.zeros((num_sequences, num_tokens, self.num_concepts), device=device, dtype=self.nb_features_log_prob.dtype)
+        features.scatter_(dim=2, index=relevant_features, src=torch.ones_like(relevant_features, device=features.device, dtype=features.dtype))
+        
+        #modified_prediction_scores = self.nb(features)
+        modified_prediction_scores = features @ self.nb_features_log_prob.to(features.device) + self.nb_class_log_prior.to(features.device)
+        #print(modified_prediction_scores.shape, modified_prediction_scores[0], "\n\n==============\n\n",  modified_prediction_scores2.shape, modified_prediction_scores2[0], "\n\n~~~~~~~~~~~~~~~~~~~~~~")
+        #import sys
+        #sys.exit(2)
+        return modified_prediction_scores
+
+class DebertaPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.embedding_size = getattr(config, "embedding_size", config.hidden_size)
+
+        self.dense = nn.Linear(config.hidden_size, self.embedding_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(self.embedding_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class DebertaLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = DebertaPredictionHeadTransform(config)
+
+        self.embedding_size = getattr(config, "embedding_size", config.hidden_size)
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(self.embedding_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 _tie_weights(self):
+        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.BertOnlyMLMHead with bert -> deberta
+class DebertaOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = DebertaLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+@add_start_docstrings(
+    """
+    DeBERTa Model transformer with a sequence classification/regression head on top (a linear layer on top of the
+    pooled output) e.g. for GLUE tasks.
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+class DebertaForSequenceClassification(DebertaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        num_labels = getattr(config, "num_labels", 2)
+        self.num_labels = num_labels
+
+        self.deberta = DebertaModel(config)
+        self.pooler = ContextPooler(config)
+        output_dim = self.pooler.output_dim
+
+        self.classifier = nn.Linear(output_dim, num_labels)
+        drop_out = getattr(config, "cls_dropout", None)
+        drop_out = self.config.hidden_dropout_prob if drop_out is None else drop_out
+        self.dropout = StableDropout(drop_out)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.deberta.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        self.deberta.set_input_embeddings(new_embeddings)
+
+    @add_start_docstrings_to_model_forward(DEBERTA_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,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = 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.deberta(
+            input_ids,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        encoder_layer = outputs[0]
+        pooled_output = self.pooler(encoder_layer)
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    # regression task
+                    loss_fn = nn.MSELoss()
+                    logits = logits.view(-1).to(labels.dtype)
+                    loss = loss_fn(logits, labels.view(-1))
+                elif labels.dim() == 1 or labels.size(-1) == 1:
+                    label_index = (labels >= 0).nonzero()
+                    labels = labels.long()
+                    if label_index.size(0) > 0:
+                        labeled_logits = torch.gather(
+                            logits, 0, label_index.expand(label_index.size(0), logits.size(1))
+                        )
+                        labels = torch.gather(labels, 0, label_index.view(-1))
+                        loss_fct = CrossEntropyLoss()
+                        loss = loss_fct(labeled_logits.view(-1, self.num_labels).float(), labels.view(-1))
+                    else:
+                        loss = torch.tensor(0).to(logits)
+                else:
+                    log_softmax = nn.LogSoftmax(-1)
+                    loss = -((log_softmax(logits) * labels).sum(-1)).mean()
+            elif 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(
+    """
+    DeBERTa 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.
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+class DebertaForTokenClassification(DebertaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.deberta = DebertaModel(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(DEBERTA_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,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = 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.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            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()
+            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(
+    """
+    DeBERTa 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`).
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+class DebertaForQuestionAnswering(DebertaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.deberta = DebertaModel(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(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_QA,
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output=_QA_EXPECTED_OUTPUT,
+        expected_loss=_QA_EXPECTED_LOSS,
+        qa_target_start_index=_QA_TARGET_START_INDEX,
+        qa_target_end_index=_QA_TARGET_END_INDEX,
+    )
+    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,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = 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.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            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).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        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,
+        )