Upload folder using huggingface_hub

.gitattributes

@@ -1,35 +1,16 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin.* filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
 *.h5 filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tar.gz filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
 *.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
 *.model filter=lfs diff=lfs merge=lfs -text
 *.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
 *.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,91 @@
 ---
+language: en
+tags: 
+- deberta
+- fill-mask
+thumbnail: https://huggingface.co/front/thumbnails/microsoft.png
 license: mit
 ---
+
+## DeBERTa: Decoding-enhanced BERT with Disentangled Attention
+
+[DeBERTa](https://arxiv.org/abs/2006.03654) improves the BERT and RoBERTa models using disentangled attention and enhanced mask decoder. It outperforms BERT and RoBERTa on  majority of NLU tasks with 80GB training data. 
+
+Please check the [official repository](https://github.com/microsoft/DeBERTa) for more details and updates.
+
+This is the DeBERTa V2 xxlarge model with 48 layers, 1536 hidden size. The total parameters are 1.5B and it is trained with 160GB raw data.
+
+
+### Fine-tuning on NLU tasks
+
+We present the dev results on SQuAD 1.1/2.0 and several GLUE benchmark tasks.
+
+| Model                     | SQuAD 1.1 | SQuAD 2.0 | MNLI-m/mm   | SST-2 | QNLI | CoLA | RTE    | MRPC  | QQP   |STS-B |
+|---------------------------|-----------|-----------|-------------|-------|------|------|--------|-------|-------|------|
+|                           | F1/EM     | F1/EM     | Acc         | Acc   | Acc  | MCC  | Acc    |Acc/F1 |Acc/F1 |P/S   |
+| BERT-Large                | 90.9/84.1 | 81.8/79.0 | 86.6/-      | 93.2  | 92.3 | 60.6 | 70.4   | 88.0/-       | 91.3/- |90.0/- |
+| RoBERTa-Large             | 94.6/88.9 | 89.4/86.5 | 90.2/-      | 96.4  | 93.9 | 68.0 | 86.6   | 90.9/-       | 92.2/- |92.4/- |
+| XLNet-Large               | 95.1/89.7 | 90.6/87.9 | 90.8/-      | 97.0  | 94.9 | 69.0 | 85.9   | 90.8/-       | 92.3/- |92.5/- |
+| [DeBERTa-Large](https://huggingface.co/microsoft/deberta-large)<sup>1</sup> | 95.5/90.1 | 90.7/88.0 | 91.3/91.1| 96.5|95.3| 69.5| 91.0| 92.6/94.6| 92.3/- |92.8/92.5 |
+| [DeBERTa-XLarge](https://huggingface.co/microsoft/deberta-xlarge)<sup>1</sup> | -/-  | -/-  | 91.5/91.2| 97.0 | - | -    | 93.1   | 92.1/94.3    | -    |92.9/92.7|
+| [DeBERTa-V2-XLarge](https://huggingface.co/microsoft/deberta-v2-xlarge)<sup>1</sup>|95.8/90.8| 91.4/88.9|91.7/91.6| **97.5**| 95.8|71.1|**93.9**|92.0/94.2|92.3/89.8|92.9/92.9|
+|**[DeBERTa-V2-XXLarge](https://huggingface.co/microsoft/deberta-v2-xxlarge)<sup>1,2</sup>**|**96.1/91.4**|**92.2/89.7**|**91.7/91.9**|97.2|**96.0**|**72.0**| 93.5| **93.1/94.9**|**92.7/90.3** |**93.2/93.1** |
+--------
+#### Notes.
+ - <sup>1</sup> Following RoBERTa, for RTE, MRPC, STS-B, we fine-tune the tasks based on [DeBERTa-Large-MNLI](https://huggingface.co/microsoft/deberta-large-mnli), [DeBERTa-XLarge-MNLI](https://huggingface.co/microsoft/deberta-xlarge-mnli), [DeBERTa-V2-XLarge-MNLI](https://huggingface.co/microsoft/deberta-v2-xlarge-mnli), [DeBERTa-V2-XXLarge-MNLI](https://huggingface.co/microsoft/deberta-v2-xxlarge-mnli). The results of SST-2/QQP/QNLI/SQuADv2 will also be slightly improved when start from MNLI fine-tuned models, however, we only report the numbers fine-tuned from pretrained base models for those 4 tasks.
+ - <sup>2</sup> To try the **XXLarge** model with **[HF transformers](https://huggingface.co/transformers/main_classes/trainer.html)**, we recommand using **deepspeed** as it's faster and saves memory.
+ 
+Run with `Deepspeed`,
+
+```bash
+pip install datasets
+pip install deepspeed
+
+# Download the deepspeed config file
+wget https://huggingface.co/microsoft/deberta-v2-xxlarge/resolve/main/ds_config.json -O ds_config.json
+
+export TASK_NAME=mnli
+output_dir="ds_results"
+num_gpus=8
+batch_size=8
+python -m torch.distributed.launch --nproc_per_node=${num_gpus} \\
+  run_glue.py \\
+  --model_name_or_path microsoft/deberta-v2-xxlarge \\
+  --task_name $TASK_NAME \\
+  --do_train \\
+  --do_eval \\
+  --max_seq_length 256 \\
+  --per_device_train_batch_size ${batch_size} \\
+  --learning_rate 3e-6 \\
+  --num_train_epochs 3 \\
+  --output_dir $output_dir \\
+  --overwrite_output_dir \\
+  --logging_steps 10 \\
+  --logging_dir $output_dir \\
+  --deepspeed ds_config.json
+```
+
+You can also run with `--sharded_ddp`
+```bash  
+cd transformers/examples/text-classification/
+export TASK_NAME=mnli
+python -m torch.distributed.launch --nproc_per_node=8 run_glue.py   --model_name_or_path microsoft/deberta-v2-xxlarge   \\
+--task_name $TASK_NAME   --do_train   --do_eval   --max_seq_length 256   --per_device_train_batch_size 8   \\
+--learning_rate 3e-6   --num_train_epochs 3   --output_dir /tmp/$TASK_NAME/ --overwrite_output_dir --sharded_ddp --fp16
+```
+
+
+### Citation
+
+If you find DeBERTa useful for your work, please cite the following paper:
+
+``` latex
+@inproceedings{
+he2021deberta,
+title={DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION},
+author={Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen},
+booktitle={International Conference on Learning Representations},
+year={2021},
+url={https://openreview.net/forum?id=XPZIaotutsD}
+}
+```

config.json

@@ -0,0 +1,35 @@
+{
+    "architectures": [
+        "DebertaV2ForCausalLM"
+    ],
+    "auto_map": {
+        "AutoConfig": "configuration_deberta.DebertaV2Config",
+        "AutoModel": "modeling_deberta.DebertaV2Model",
+        "AutoModelForMaskedLM": "modeling_deberta.DebertaV2ForMaskedLM",
+        "AutoModelForCausalLM": "modeling_deberta.DebertaV2ForCausalLM"
+    },
+    "sep_token_id": 2,
+    "mask_token_id": 128000,
+    "attention_probs_dropout_prob": 0.1,
+    "hidden_act": "gelu",
+    "hidden_dropout_prob": 0.1,
+    "hidden_size": 1536,
+    "initializer_range": 0.02,
+    "intermediate_size": 6144,
+    "max_position_embeddings": 512,
+    "relative_attention": true,
+    "position_buckets": 256,
+    "norm_rel_ebd": "layer_norm",
+    "share_att_key": true,
+    "pos_att_type": "p2c|c2p",
+    "layer_norm_eps": 1e-7,
+    "conv_kernel_size": 3,
+    "conv_act": "gelu",
+    "max_relative_positions": -1,
+    "position_biased_input": false,
+    "num_attention_heads": 24,
+    "attention_head_size": 64,
+    "num_hidden_layers": 48,
+    "type_vocab_size": 0,
+    "vocab_size": 128100
+}

configuration_deberta.py

@@ -0,0 +1,60 @@
+from transformers import PretrainedConfig
+
+class DebertaV2Config(PretrainedConfig):
+    model_type = "deberta-v2"
+
+    def __init__(
+        self,
+        vocab_size=128100,
+        hidden_size=1536,
+        sep_token_id=2,
+        mask_token_id=128000,
+        num_hidden_layers=24,
+        num_attention_heads=24,
+        intermediate_size=6144,
+        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.mask_token_id = mask_token_id
+        self.sep_token_id = sep_token_id
+        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

ds_config.json

@@ -0,0 +1,23 @@
+{
+  "fp16": {
+
+    "enabled": true,
+
+    "initial_scale_power": 12
+
+  },
+  "zero_optimization": {
+
+    "stage": 2,
+
+    "reduce_bucket_size": 5e7,
+
+    "allgather_bucket_size": 1.25e9,
+
+    "overlap_comm": true,
+
+    "contiguous_gradients": true
+
+  },
+  "zero_allow_untested_optimizer": true
+}

generation_config.json

@@ -0,0 +1,14 @@
+{
+    "_from_model_config": true,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "pad_token_id": 0,
+    "transformers_version": "4.36.0",
+    "top_k": 64,
+    "top_p": 0.9,
+    "repetition_penalty": 1.0,
+    "sample": true,
+    "temperature": 0.3,
+    "max_new_tokens": 512,
+    "use_cache": false
+}

modeling_deberta.py

@@ -0,0 +1,1391 @@
+# 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-v2 model."""
+
+from collections.abc import Sequence
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutput,
+    MaskedLMOutput,
+    CausalLMOutput,
+    MultipleChoiceModelOutput,
+    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 transformers.models.deberta_v2.modeling_deberta_v2 import DebertaV2Config
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "DebertaV2Config"
+_CHECKPOINT_FOR_DOC = "microsoft/deberta-v2-xlarge"
+_QA_TARGET_START_INDEX = 2
+_QA_TARGET_END_INDEX = 9
+
+DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "microsoft/deberta-v2-xlarge",
+    "microsoft/deberta-v2-xxlarge",
+    "microsoft/deberta-v2-xlarge-mnli",
+    "microsoft/deberta-v2-xxlarge-mnli",
+]
+
+
+# Copied from transformers.models.deberta.modeling_deberta.ContextPooler
+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
+
+
+# Copied from transformers.models.deberta.modeling_deberta.XSoftmax with deberta->deberta_v2
+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_v2.modeling_deberta_v2 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(self, input, mask, dim):
+        self.dim = dim
+        rmask = ~(mask.to(torch.bool))
+
+        output = input.masked_fill(rmask, torch.tensor(torch.finfo(input.dtype).min))
+        output = torch.softmax(output, self.dim)
+        output.masked_fill_(rmask, 0)
+        self.save_for_backward(output)
+        return output
+
+    @staticmethod
+    def backward(self, grad_output):
+        (output,) = self.saved_tensors
+        inputGrad = softmax_backward_data(self, grad_output, output, self.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)))
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DropoutContext
+class DropoutContext(object):
+    def __init__(self):
+        self.dropout = 0
+        self.mask = None
+        self.scale = 1
+        self.reuse_mask = True
+
+
+# Copied from transformers.models.deberta.modeling_deberta.get_mask
+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
+
+
+# Copied from transformers.models.deberta.modeling_deberta.XDropout
+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)
+
+
+# Copied from transformers.models.deberta.modeling_deberta.StableDropout
+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
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaSelfOutput with DebertaLayerNorm->LayerNorm
+class DebertaV2SelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(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
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaAttention with Deberta->DebertaV2
+class DebertaV2Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = DisentangledSelfAttention(config)
+        self.output = DebertaV2SelfOutput(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->DebertaV2
+class DebertaV2Intermediate(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.deberta.modeling_deberta.DebertaOutput with DebertaLayerNorm->LayerNorm
+class DebertaV2Output(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(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
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaLayer with Deberta->DebertaV2
+class DebertaV2Layer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = DebertaV2Attention(config)
+        self.intermediate = DebertaV2Intermediate(config)
+        self.output = DebertaV2Output(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 ConvLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        kernel_size = getattr(config, "conv_kernel_size", 3)
+        groups = getattr(config, "conv_groups", 1)
+        self.conv_act = getattr(config, "conv_act", "tanh")
+        self.conv = nn.Conv1d(
+            config.hidden_size, config.hidden_size, kernel_size, padding=(kernel_size - 1) // 2, groups=groups
+        )
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def forward(self, hidden_states, residual_states, input_mask):
+        out = self.conv(hidden_states.permute(0, 2, 1).contiguous()).permute(0, 2, 1).contiguous()
+        rmask = (1 - input_mask).bool()
+        out.masked_fill_(rmask.unsqueeze(-1).expand(out.size()), 0)
+        out = ACT2FN[self.conv_act](self.dropout(out))
+
+        layer_norm_input = residual_states + out
+        output = self.LayerNorm(layer_norm_input).to(layer_norm_input)
+
+        if input_mask is None:
+            output_states = output
+        else:
+            if input_mask.dim() != layer_norm_input.dim():
+                if input_mask.dim() == 4:
+                    input_mask = input_mask.squeeze(1).squeeze(1)
+                input_mask = input_mask.unsqueeze(2)
+
+            input_mask = input_mask.to(output.dtype)
+            output_states = output * input_mask
+
+        return output_states
+
+
+class DebertaV2Encoder(nn.Module):
+    """Modified BertEncoder with relative position bias support"""
+
+    def __init__(self, config):
+        super().__init__()
+
+        self.layer = nn.ModuleList([DebertaV2Layer(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.position_buckets = getattr(config, "position_buckets", -1)
+            pos_ebd_size = self.max_relative_positions * 2
+
+            if self.position_buckets > 0:
+                pos_ebd_size = self.position_buckets * 2
+
+            self.rel_embeddings = nn.Embedding(pos_ebd_size, config.hidden_size)
+
+        self.norm_rel_ebd = [x.strip() for x in getattr(config, "norm_rel_ebd", "none").lower().split("|")]
+
+        if "layer_norm" in self.norm_rel_ebd:
+            self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=True)
+
+        self.conv = ConvLayer(config) if getattr(config, "conv_kernel_size", 0) > 0 else None
+        self.gradient_checkpointing = False
+
+    def get_rel_embedding(self):
+        rel_embeddings = self.rel_embeddings.weight if self.relative_attention else None
+        if rel_embeddings is not None and ("layer_norm" in self.norm_rel_ebd):
+            rel_embeddings = self.LayerNorm(rel_embeddings)
+        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)
+            attention_mask = attention_mask.triu(diagonal=-510).tril(diagonal=510)
+            attention_mask[:, :, :, 0] = 1
+            attention_mask[:, :, :, -1] = 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),
+                bucket_size=self.position_buckets,
+                max_position=self.max_relative_positions,
+                device=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,
+    ):
+        if attention_mask.dim() <= 2:
+            input_mask = attention_mask
+        else:
+            input_mask = attention_mask.sum(-2) > 0
+        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()
+        output_states = next_kv
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (output_states,)
+
+            if self.gradient_checkpointing and self.training:
+                output_states = self._gradient_checkpointing_func(
+                    layer_module.__call__,
+                    next_kv,
+                    attention_mask,
+                    query_states,
+                    relative_pos,
+                    rel_embeddings,
+                    output_attentions,
+                )
+            else:
+                output_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:
+                output_states, att_m = output_states
+
+            if i == 0 and self.conv is not None:
+                output_states = self.conv(hidden_states, output_states, input_mask)
+
+            if query_states is not None:
+                query_states = output_states
+                if isinstance(hidden_states, Sequence):
+                    next_kv = hidden_states[i + 1] if i + 1 < len(self.layer) else None
+            else:
+                next_kv = output_states
+
+            if output_attentions:
+                all_attentions = all_attentions + (att_m,)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (output_states,)
+
+        if not return_dict:
+            return tuple(v for v in [output_states, all_hidden_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=output_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+def make_log_bucket_position(relative_pos, bucket_size, max_position):
+    sign = torch.sign(relative_pos)
+    mid = bucket_size // 2
+    abs_pos = torch.where(
+        (relative_pos < mid) & (relative_pos > -mid),
+        torch.tensor(mid - 1).type_as(relative_pos),
+        torch.abs(relative_pos),
+    )
+    log_pos = (
+        torch.ceil(torch.log(abs_pos / mid) / torch.log(torch.tensor((max_position - 1) / mid)) * (mid - 1)) + mid
+    )
+    bucket_pos = torch.where(abs_pos <= mid, relative_pos.type_as(log_pos), log_pos * sign)
+    bucket_pos = bucket_pos.clamp(min=-bucket_size+1, max=bucket_size-1)
+    return bucket_pos
+
+
+def build_relative_position(query_size, key_size, bucket_size=-1, max_position=-1, device=None):
+    """
+    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
+        bucket_size (int): the size of position bucket
+        max_position (int): the maximum allowed absolute position
+        device (`torch.device`): the device on which tensors will be created.
+
+    Return:
+        `torch.LongTensor`: A tensor with shape [1, query_size, key_size]
+    """
+
+    q_ids = torch.arange(0, query_size, device=device)
+    k_ids = torch.arange(0, key_size, device=device)
+    rel_pos_ids = q_ids[:, None] - k_ids[None, :]
+    if bucket_size > 0 and max_position > 0:
+        rel_pos_ids = make_log_bucket_position(rel_pos_ids, bucket_size, max_position)
+    rel_pos_ids = rel_pos_ids.to(torch.long)
+    rel_pos_ids = rel_pos_ids[:query_size, :]
+    rel_pos_ids = rel_pos_ids.unsqueeze(0)
+    return rel_pos_ids
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.c2p_dynamic_expand
+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
+# Copied from transformers.models.deberta.modeling_deberta.p2c_dynamic_expand
+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
+# Copied from transformers.models.deberta.modeling_deberta.pos_dynamic_expand
+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 (`DebertaV2Config`):
+            A model config class instance with the configuration to build a new model. The schema is similar to
+            *BertConfig*, for more details, please refer [`DebertaV2Config`]
+
+    """
+
+    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
+        _attention_head_size = config.hidden_size // config.num_attention_heads
+        self.attention_head_size = getattr(config, "attention_head_size", _attention_head_size)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+        self.query_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+        self.key_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+        self.value_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+
+        self.share_att_key = getattr(config, "share_att_key", False)
+        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)
+
+        if self.relative_attention:
+            self.position_buckets = getattr(config, "position_buckets", -1)
+            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_ebd_size = self.max_relative_positions
+            if self.position_buckets > 0:
+                self.pos_ebd_size = self.position_buckets
+
+            self.pos_dropout = StableDropout(config.hidden_dropout_prob)
+
+            if not self.share_att_key:
+                if "c2p" in self.pos_att_type:
+                    self.pos_key_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+                if "p2c" in self.pos_att_type:
+                    self.pos_query_proj = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = StableDropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x, attention_heads):
+        new_x_shape = x.size()[:-1] + (attention_heads, -1)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3).contiguous().view(-1, x.size(1), x.size(-1))
+
+    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:
+            query_states = hidden_states
+        query_layer = self.transpose_for_scores(self.query_proj(query_states), self.num_attention_heads)
+        key_layer = self.transpose_for_scores(self.key_proj(hidden_states), self.num_attention_heads)
+        value_layer = self.transpose_for_scores(self.value_proj(hidden_states), self.num_attention_heads)
+
+        rel_att = None
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        scale_factor = 1
+        if "c2p" in self.pos_att_type:
+            scale_factor += 1
+        if "p2c" in self.pos_att_type:
+            scale_factor += 1
+        scale = torch.sqrt(torch.tensor(query_layer.size(-1), dtype=torch.float) * scale_factor)
+        attention_scores = torch.bmm(query_layer, key_layer.transpose(-1, -2) / scale.to(dtype=query_layer.dtype))
+        if self.relative_attention:
+            rel_embeddings = self.pos_dropout(rel_embeddings)
+            rel_att = self.disentangled_attention_bias(
+                query_layer, key_layer, relative_pos, rel_embeddings, scale_factor
+            )
+
+        if rel_att is not None:
+            attention_scores = attention_scores + rel_att
+        attention_scores = attention_scores
+        attention_scores = attention_scores.view(
+            -1, self.num_attention_heads, attention_scores.size(-2), attention_scores.size(-1)
+        )
+
+        # bsz x height x length x dimension
+        attention_probs = XSoftmax.apply(attention_scores, attention_mask, -1)
+        attention_probs = self.dropout(attention_probs)
+        context_layer = torch.bmm(
+            attention_probs.view(-1, attention_probs.size(-2), attention_probs.size(-1)), value_layer
+        )
+        context_layer = (
+            context_layer.view(-1, self.num_attention_heads, context_layer.size(-2), context_layer.size(-1))
+            .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_attention_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),
+                bucket_size=self.position_buckets,
+                max_position=self.max_relative_positions,
+                device=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)
+        # bsz x height x query x key
+        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 = self.pos_ebd_size
+        relative_pos = relative_pos.long().to(query_layer.device)
+
+        rel_embeddings = rel_embeddings[0 : att_span * 2, :].unsqueeze(0)
+        if self.share_att_key:
+            pos_query_layer = self.transpose_for_scores(
+                self.query_proj(rel_embeddings), self.num_attention_heads
+            ).repeat(query_layer.size(0) // self.num_attention_heads, 1, 1)
+            pos_key_layer = self.transpose_for_scores(self.key_proj(rel_embeddings), self.num_attention_heads).repeat(
+                query_layer.size(0) // self.num_attention_heads, 1, 1
+            )
+        else:
+            if "c2p" in self.pos_att_type:
+                pos_key_layer = self.transpose_for_scores(
+                    self.pos_key_proj(rel_embeddings), self.num_attention_heads
+                ).repeat(query_layer.size(0) // self.num_attention_heads, 1, 1)  # .split(self.all_head_size, dim=-1)
+            if "p2c" in self.pos_att_type:
+                pos_query_layer = self.transpose_for_scores(
+                    self.pos_query_proj(rel_embeddings), self.num_attention_heads
+                ).repeat(query_layer.size(0) // self.num_attention_heads, 1, 1)  # .split(self.all_head_size, dim=-1)
+
+        score = 0
+        # content->position
+        if "c2p" in self.pos_att_type:
+            scale = torch.sqrt(torch.tensor(pos_key_layer.size(-1), dtype=torch.float) * scale_factor)
+            c2p_att = torch.bmm(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_pos.squeeze(0).expand([query_layer.size(0), query_layer.size(1), relative_pos.size(-1)]),
+            )
+            score += c2p_att / scale.to(dtype=c2p_att.dtype)
+
+        # position->content
+        if "p2c" in self.pos_att_type:
+            scale = torch.sqrt(torch.tensor(pos_query_layer.size(-1), dtype=torch.float) * scale_factor)
+            if key_layer.size(-2) != query_layer.size(-2):
+                r_pos = build_relative_position(
+                    key_layer.size(-2),
+                    key_layer.size(-2),
+                    bucket_size=self.position_buckets,
+                    max_position=self.max_relative_positions,
+                    device=query_layer.device,
+                )
+                r_pos = r_pos.unsqueeze(0)
+            else:
+                r_pos = relative_pos
+
+            p2c_pos = torch.clamp(-r_pos + att_span, 0, att_span * 2 - 1)
+            p2c_att = torch.bmm(key_layer, pos_query_layer.transpose(-1, -2))
+            p2c_att = torch.gather(
+                p2c_att,
+                dim=-1,
+                index=p2c_pos.squeeze(0).expand([query_layer.size(0), key_layer.size(-2), key_layer.size(-2)]),
+            ).transpose(-1, -2)
+            score += p2c_att / scale.to(dtype=p2c_att.dtype)
+
+        return score
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaEmbeddings with DebertaLayerNorm->LayerNorm
+class DebertaV2Embeddings(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)
+        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 = LayerNorm(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:
+            if seq_length > self.position_ids.size(1):
+                position_ids = torch.cat([
+                    torch.zeros(1, seq_length - self.position_ids.size(1), dtype=self.position_ids.dtype, device=self.position_ids.device),
+                    self.position_ids
+                ], dim=1)
+            else:   
+                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, position_embeddings
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaPreTrainedModel with Deberta->DebertaV2
+class DebertaV2PreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = DebertaV2Config
+    base_model_prefix = "deberta"
+    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_()
+
+
+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 ([`DebertaV2Config`]): 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,
+)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaModel with Deberta->DebertaV2
+class DebertaV2Model(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = DebertaV2Embeddings(config)
+        self.encoder = DebertaV2Encoder(config)
+        self.z_steps = 4
+        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, position_embeddings = 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 = list(encoder_outputs[1])
+
+#        print(self.z_steps)
+
+        if self.z_steps > 0:
+            hidden_states = encoded_layers[-2]
+            layers = [self.encoder.layer[-1] for _ in range(self.z_steps)]
+            query_states = position_embeddings + encoded_layers[-2]
+            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:
+                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 DebertaV2ForMaskedLM(DebertaV2PreTrainedModel):
+    _tied_weights_keys = ["cls.predictions.decoder.weight", "cls.predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.deberta = DebertaV2Model(config)
+        self.cls = DebertaV2OnlyMLMHead(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(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+        mask="[MASK]",
+    )
+    # Copied from transformers.models.deberta.modeling_deberta.DebertaForMaskedLM.forward with Deberta->DebertaV2
+    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)
+
+        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,
+        )
+
+@add_start_docstrings("""DeBERTa Model with a `language modeling` head on top.""", DEBERTA_START_DOCSTRING)
+class DebertaV2ForCausalLM(DebertaV2ForMaskedLM):
+    _tied_weights_keys = ["cls.predictions.decoder.weight", "cls.predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        config.is_decoder = True
+        self.mask_token_id = config.mask_token_id
+        self.sep_token_id = config.sep_token_id
+        self.n_masks = 3
+
+    def set_decoder(self, decoder):
+        self.deberta = decoder
+
+    def get_decoder(self):
+        return self.deberta
+
+    def can_generate(self):
+        return True
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        position_ids = kwargs.get("position_ids", None)
+
+        if input_ids[0, -1] == 2:
+            input_ids = input_ids[:, :-1]
+            if attention_mask is not None:
+                attention_mask = attention_mask[:, :-1]
+            if position_ids is not None:
+                position_ids = position_ids[:, :-1]
+
+        # Omit tokens covered by past_key_values
+        if past_key_values is not None:
+            pass  # should never happen
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=CausalLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+        mask="[MASK]",
+    )
+    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,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutput]:
+        
+        assert labels is None, "only inference is supported for now"
+        assert inputs_embeds is None, "inputs_embeds is not supported for now"
+        assert token_type_ids is None, "token_type_ids is not supported for now"
+        assert past_key_values is None, "past_key_values is not supported for now"
+        assert use_cache is None, "use_cache is not supported for now"
+
+        assert input_ids[0, -1] != self.sep_token_id, "remove the last token if it is a sep token"
+         
+        batch_size, seq_length = input_ids.shape
+        input_ids = torch.cat(
+            [
+                input_ids,
+                torch.full((batch_size, self.n_masks), self.mask_token_id, device=input_ids.device),
+                torch.full((batch_size, 1), self.sep_token_id, device=input_ids.device)
+            ],
+            dim=-1
+        )
+        attention_mask = torch.cat(
+            [
+                attention_mask,
+                torch.full((batch_size, self.n_masks + 1), attention_mask[0, -1], device=attention_mask.device),
+            ],
+            dim=-1
+        )
+        position_ids = torch.cat(
+            [
+                position_ids,
+                torch.arange(0, self.n_masks + 1, device=position_ids.device).unsqueeze(0) + position_ids[:, -1:],
+            ],
+            dim=-1
+        )
+
+        outputs = super().forward(
+            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,
+        )
+
+        # shift the outputs and skip excess masks
+        logits = outputs.logits[:, 1:-(self.n_masks), :].contiguous()
+
+        loss = None
+        if labels is not None:
+            pass
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaPredictionHeadTransform with Deberta->DebertaV2
+class DebertaV2PredictionHeadTransform(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
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaLMPredictionHead with Deberta->DebertaV2
+class DebertaV2LMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = DebertaV2PredictionHeadTransform(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=True)
+
+        #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.BertOnlyMLMHead with bert -> deberta
+class DebertaV2OnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = DebertaV2LMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1b1b0c5fcd18795a94c7bedd61d0ba84b3682444e36beedf88aa118df4ec99b
+size 3135918282

spm.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5598d5e96f339a8d980c15f9afd405a2e5e1be7db41de3ed13b0f03fac1e8c17
+size 2447305

tokenizer_config.json

@@ -0,0 +1,4 @@
+{
+  "do_lower_case": false,
+  "vocab_type": "spm"
+}