Add Ernie-M Model to huggingface (#21349)

* config and tokenization(fast too) changed and ErnieEncoder added

* Slow Tokenization Added

* Tokenizer(slow) is now working and Fast Tokenizer removed

* Added Config code

* Added Base Model and utils

* ErnieMModel is now working

* All added except tests

* All tests passed except ErnieUIEM

* All tests passed

* all fixes done

* all fixes done

* fixed MAP

* fixed check_code_quality

* fixed Build PR Documentation issue

* Added changes(comments) and also updated to the latest upstream/main

* Added fixup

* Added # Copied comments

* Added fixup

* Added more comments and some nits

* Added fixup

* Fixed README_hd.md

* Added more fixes

* ErnieMTokenizer (being sentencepiece) protected and other docs edited

* Added code_quality fix

* Fixed for

* Added more fix

* modified AZ

* ernie-m tokenization test added!

* attention mask part fixed(with 0->self.config.pad_token_id)

* applied make fixup

src/transformers/models/ernie_m/configuration_ernie_m.py

+# coding=utf-8
+# Copyright 2023 Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" ErnieM model configuration"""
+# Adapted from original paddlenlp repository.(https://github.com/PaddlePaddle/PaddleNLP/blob/develop/paddlenlp/transformers/ernie_m/configuration.py)
+
+from __future__ import annotations
+
+from typing import Dict
+
+from ...configuration_utils import PretrainedConfig
+
+
+ERNIE_M_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "susnato/ernie-m-base_pytorch": "https://huggingface.co/susnato/ernie-m-base_pytorch/blob/main/config.json",
+    "susnato/ernie-m-large_pytorch": "https://huggingface.co/susnato/ernie-m-large_pytorch/blob/main/config.json",
+}
+
+
+class ErnieMConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ErnieMModel`]. It is used to instantiate a
+    Ernie-M 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 `Ernie-M`
+    [susnato/ernie-m-base_pytorch](https://huggingface.co/susnato/ernie-m-base_pytorch) architecture.
+
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 250002):
+            Vocabulary size of `inputs_ids` in [`ErnieMModel`]. Also is the vocab size of token embedding matrix.
+            Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling
+            [`ErnieMModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the embedding layer, encoder layers and 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 feed-forward (ff) layer in the encoder. Input tensors to feed-forward layers are
+            firstly projected from hidden_size to intermediate_size, and then projected back to hidden_size. Typically
+            intermediate_size is larger than hidden_size.
+        hidden_act (`str`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function in the feed-forward layer. `"gelu"`, `"relu"` and any other torch
+            supported activation functions are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings and encoder.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability used in `MultiHeadAttention` in all encoder layers to drop some attention target.
+        act_dropout (`float`, *optional*, defaults to 0.0):
+            This dropout probability is used in `ErnieMEncoderLayer` after activation.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum value of the dimensionality of position encoding, which dictates the maximum supported length
+            of an input sequence.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the layer normalization layers.
+        classifier_dropout (`float`, *optional*):
+            The dropout ratio for the classification head.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the normal initializer for initializing all weight matrices.
+        pad_token_id(`int`, *optional*, defaults to 1):
+            The index of padding token in the token vocabulary.
+
+    A normal_initializer initializes weight matrices as normal distributions. See
+    `ErnieMPretrainedModel._init_weights()` for how weights are initialized in `ErnieMModel`.
+    """
+    model_type = "ernie_m"
+    attribute_map: Dict[str, str] = {"dropout": "classifier_dropout", "num_classes": "num_labels"}
+
+    def __init__(
+        self,
+        vocab_size: int = 250002,
+        hidden_size: int = 768,
+        num_hidden_layers: int = 12,
+        num_attention_heads: int = 12,
+        intermediate_size: int = 3072,
+        hidden_act: str = "gelu",
+        hidden_dropout_prob: float = 0.1,
+        attention_probs_dropout_prob: float = 0.1,
+        max_position_embeddings: int = 514,
+        initializer_range: float = 0.02,
+        pad_token_id: int = 1,
+        layer_norm_eps: float = 1e-05,
+        classifier_dropout=None,
+        is_decoder=False,
+        act_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.classifier_dropout = classifier_dropout
+        self.is_decoder = is_decoder
+        self.act_dropout = act_dropout

src/transformers/models/ernie_m/modeling_ernie_m.py

+# coding=utf-8
+# Copyright 2023 Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch ErnieM model."""
+
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn, tensor
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from .configuration_ernie_m import ErnieMConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "susnato/ernie-m-base_pytorch"
+_CONFIG_FOR_DOC = "ErnieMConfig"
+_TOKENIZER_FOR_DOC = "ErnieMTokenizer"
+
+ERNIE_M_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "susnato/ernie-m-base_pytorch",
+    "susnato/ernie-m-large_pytorch",
+    # See all ErnieM models at https://huggingface.co/models?filter=ernie_m
+]
+
+
+# Adapted from paddlenlp.transformers.ernie_m.modeling.ErnieEmbeddings
+class ErnieMEmbeddings(nn.Module):
+    """Construct the embeddings from word and position embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size, padding_idx=config.pad_token_id
+        )
+        self.layer_norm = nn.LayerNorm(normalized_shape=config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(p=config.hidden_dropout_prob)
+        self.padding_idx = config.pad_token_id
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.LongTensor] = None,
+        past_key_values_length: int = 0,
+    ) -> torch.Tensor:
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        if position_ids is None:
+            input_shape = inputs_embeds.size()[:-1]
+            ones = torch.ones(input_shape, dtype=torch.int64)
+            seq_length = torch.cumsum(ones, dim=1)
+            position_ids = seq_length - ones
+
+            if past_key_values_length > 0:
+                position_ids = position_ids + past_key_values_length
+        # to mimic paddlenlp implementation
+        position_ids += 2
+        position_ids = position_ids.to(next(self.position_embeddings.parameters()).device)
+        position_embeddings = self.position_embeddings(position_ids)
+        embeddings = inputs_embeds + position_embeddings
+        embeddings = self.layer_norm(embeddings)
+        embeddings = self.dropout(embeddings)
+
+        return embeddings
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->ErnieM,self.value->self.v_proj,self.key->self.k_proj,self.query->self.q_proj
+class ErnieMSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.q_proj = nn.Linear(config.hidden_size, self.all_head_size)
+        self.k_proj = nn.Linear(config.hidden_size, self.all_head_size)
+        self.v_proj = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.q_proj(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.k_proj(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.v_proj(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.k_proj(hidden_states))
+            value_layer = self.transpose_for_scores(self.v_proj(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.k_proj(hidden_states))
+            value_layer = self.transpose_for_scores(self.v_proj(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        use_cache = past_key_value is not None
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            query_length, key_length = query_layer.shape[2], key_layer.shape[2]
+            if use_cache:
+                position_ids_l = torch.tensor(key_length - 1, dtype=torch.long, device=hidden_states.device).view(
+                    -1, 1
+                )
+            else:
+                position_ids_l = torch.arange(query_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(key_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in ErnieMModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        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] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+class ErnieMAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self_attn = ErnieMSelfAttention(config, position_embedding_type=position_embedding_type)
+        self.out_proj = nn.Linear(config.hidden_size, config.hidden_size)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self_attn.num_attention_heads, self.self_attn.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self_attn.q_proj = prune_linear_layer(self.self_attn.q_proj, index)
+        self.self_attn.k_proj = prune_linear_layer(self.self_attn.k_proj, index)
+        self.self_attn.v_proj = prune_linear_layer(self.self_attn.v_proj, index)
+        self.out_proj = prune_linear_layer(self.out_proj, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self_attn.num_attention_heads = self.self_attn.num_attention_heads - len(heads)
+        self.self_attn.all_head_size = self.self_attn.attention_head_size * self.self_attn.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self_attn(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.out_proj(self_outputs[0])
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class ErnieMEncoderLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        # to mimic paddlenlp implementation
+        dropout = 0.1 if config.hidden_dropout_prob is None else config.hidden_dropout_prob
+        act_dropout = config.hidden_dropout_prob if config.act_dropout is None else config.act_dropout
+
+        self.self_attn = ErnieMAttention(config)
+        self.linear1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.dropout = nn.Dropout(act_dropout)
+        self.linear2 = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.norm1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.norm2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        if isinstance(config.hidden_act, str):
+            self.activation = ACT2FN[config.hidden_act]
+        else:
+            self.activation = config.hidden_act
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = True,
+    ):
+        residual = hidden_states
+        if output_attentions:
+            hidden_states, attention_opt_weights = self.self_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                head_mask=head_mask,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+            )
+
+        else:
+            hidden_states = self.self_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                head_mask=head_mask,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+            )
+        hidden_states = residual + self.dropout1(hidden_states)
+        hidden_states = self.norm1(hidden_states)
+        residual = hidden_states
+
+        hidden_states = self.linear1(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.linear2(hidden_states)
+        hidden_states = residual + self.dropout2(hidden_states)
+        hidden_states = self.norm2(hidden_states)
+
+        if output_attentions:
+            return hidden_states, attention_opt_weights
+        else:
+            return hidden_states
+
+
+class ErnieMEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([ErnieMEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(
+        self,
+        input_embeds: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        hidden_states = () if output_hidden_states else None
+        attentions = () if output_attentions else None
+
+        output = input_embeds
+        if output_hidden_states:
+            hidden_states = hidden_states + (output,)
+        for i, layer in enumerate(self.layers):
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            output, opt_attn_weights = layer(
+                hidden_states=output,
+                attention_mask=attention_mask,
+                head_mask=layer_head_mask,
+                past_key_value=past_key_value,
+            )
+
+            if output_hidden_states:
+                hidden_states = hidden_states + (output,)
+            if output_attentions:
+                attentions = attentions + (opt_attn_weights,)
+
+        last_hidden_state = output
+        if not return_dict:
+            return tuple(v for v in [last_hidden_state, hidden_states, attentions] if v is not None)
+
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=last_hidden_state, hidden_states=hidden_states, attentions=attentions
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler with Bert->ErnieM
+class ErnieMPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class ErnieMPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = ErnieMConfig
+    base_model_prefix = "ernie_m"
+    supports_gradient_checkpointing = True
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, ErnieMEncoder):
+            module.gradient_checkpointing = value
+
+
+ERNIE_M_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
+    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`ErnieMConfig`]): 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.
+"""
+
+ERNIE_M_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`ErnieMTokenizer`]. 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)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_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 ErnieM Model transformer outputting raw hidden-states without any specific head on top.",
+    ERNIE_M_START_DOCSTRING,
+)
+class ErnieMModel(ErnieMPreTrainedModel):
+    def __init__(self, config, add_pooling_layer=True):
+        super(ErnieMModel, self).__init__(config)
+        self.initializer_range = config.initializer_range
+        self.embeddings = ErnieMEmbeddings(config)
+        self.encoder = ErnieMEncoder(config)
+        self.pooler = ErnieMPooler(config) if add_pooling_layer else None
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layers[layer].self_attn.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(ERNIE_M_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[tensor] = None,
+        position_ids: Optional[tensor] = None,
+        attention_mask: Optional[tensor] = None,
+        head_mask: Optional[tensor] = None,
+        inputs_embeds: Optional[tensor] = None,
+        past_key_values: Optional[Tuple[Tuple[tensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        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.")
+
+        # init the default bool value
+        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.return_dict
+
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        past_key_values_length = 0
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+
+        # Adapted from paddlenlp.transformers.ernie_m.ErnieMModel
+        if attention_mask is None:
+            attention_mask = (input_ids == self.config.pad_token_id).to(torch.float32)
+            attention_mask *= torch.finfo(attention_mask.dtype).min
+            if past_key_values is not None:
+                batch_size = past_key_values[0][0].shape[0]
+                past_mask = torch.zeros([batch_size, 1, 1, past_key_values_length], dtype=attention_mask.dtype)
+                attention_mask = torch.concat([past_mask, attention_mask], dim=-1)
+        # For 2D attention_mask from tokenizer
+        elif attention_mask.ndim == 2:
+            attention_mask = attention_mask.to(torch.float32)
+            attention_mask = 1.0 - attention_mask
+            attention_mask *= torch.finfo(attention_mask.dtype).min
+
+        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(1)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            past_key_values=past_key_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            sequence_output = encoder_outputs[0]
+            pooler_output = self.pooler(sequence_output) if self.pooler is not None else None
+            return (sequence_output, pooler_output) + encoder_outputs[1:]
+
+        sequence_output = encoder_outputs["last_hidden_state"]
+        pooler_output = self.pooler(sequence_output) if self.pooler is not None else None
+        hidden_states = None if not output_hidden_states else encoder_outputs["hidden_states"]
+        attentions = None if not output_attentions else encoder_outputs["attentions"]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooler_output,
+            hidden_states=hidden_states,
+            attentions=attentions,
+        )
+
+
+@add_start_docstrings(
+    """ErnieM Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks.""",
+    ERNIE_M_START_DOCSTRING,
+)
+class ErnieMForSequenceClassification(ErnieMPreTrainedModel):
+    # Copied from transformers.models.bert.modeling_bert.BertForSequenceClassification.__init__ with Bert->ErnieM,bert->ernie_m
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.ernie_m = ErnieMModel(config)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        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(ERNIE_M_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        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,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.Tensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = True,
+        labels: Optional[torch.Tensor] = None,
+    ):
+        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.ernie_m(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            output_hidden_states=output_hidden_states,
+            output_attentions=output_attentions,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        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:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            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(
+    """ErnieM Model with a multiple choice classification head on top (a linear layer on top of
+    the pooled output and a softmax) e.g. for RocStories/SWAG tasks.""",
+    ERNIE_M_START_DOCSTRING,
+)
+class ErnieMForMultipleChoice(ErnieMPreTrainedModel):
+    # Copied from transformers.models.bert.modeling_bert.BertForMultipleChoice.__init__ with Bert->ErnieM,bert->ernie_m
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.ernie_m = ErnieMModel(config)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ERNIE_M_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: 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] = True,
+    ):
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
+            num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
+            `input_ids` above)
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        outputs = self.ernie_m(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """ErnieM 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.""",
+    ERNIE_M_START_DOCSTRING,
+)
+class ErnieMForTokenClassification(ErnieMPreTrainedModel):
+    # Copied from transformers.models.bert.modeling_bert.BertForTokenClassification.__init__ with Bert->ErnieM,bert->ernie_m
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.ernie_m = ErnieMModel(config, add_pooling_layer=False)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        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(ERNIE_M_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        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,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.Tensor]] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = True,
+        labels: Optional[torch.Tensor] = None,
+    ):
+        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.ernie_m(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            past_key_values=past_key_values,
+            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[2:]
+            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(
+    """ErnieM 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`).""",
+    ERNIE_M_START_DOCSTRING,
+)
+class ErnieMForQuestionAnswering(ErnieMPreTrainedModel):
+    # Copied from transformers.models.bert.modeling_bert.BertForQuestionAnswering.__init__ with Bert->ErnieM,bert->ernie_m
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.ernie_m = ErnieMModel(config, add_pooling_layer=False)
+        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(ERNIE_M_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = True,
+    ):
+        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.ernie_m(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            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[2:]
+            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,
+        )
+
+
+@add_start_docstrings(
+    """ErnieMForInformationExtraction is a Ernie-M Model with two linear layer on top of the hidden-states output to
+    compute `start_prob` and `end_prob`, designed for Universal Information Extraction.""",
+    ERNIE_M_START_DOCSTRING,
+)
+# Copied from paddlenlp.transformers.ernie_m.modeling.UIEM
+class ErnieMForInformationExtraction(ErnieMPreTrainedModel):
+    def __init__(self, config):
+        super(ErnieMForInformationExtraction, self).__init__(config)
+        self.ernie_m = ErnieMModel(config)
+        self.linear_start = nn.Linear(config.hidden_size, 1)
+        self.linear_end = nn.Linear(config.hidden_size, 1)
+        self.sigmoid = nn.Sigmoid()
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ERNIE_M_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = True,
+    ):
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for position (index) for computing the start_positions loss. 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) for computing the end_positions loss. Position outside of the sequence are not
+            taken into account for computing the loss.
+        """
+
+        result = self.ernie_m(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        if return_dict:
+            sequence_output = result.last_hidden_state
+        elif not return_dict:
+            sequence_output = result[0]
+
+        start_logits = self.linear_start(sequence_output)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = self.linear_end(sequence_output)
+        end_logits = end_logits.squeeze(-1)
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = BCEWithLogitsLoss()
+            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:
+            return tuple(
+                i
+                for i in [total_loss, start_logits, end_logits, result.hidden_states, result.attentions]
+                if i is not None
+            )
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=result.hidden_states,
+            attentions=result.attentions,
+        )

src/transformers/models/ernie_m/tokenization_ernie_m.py

+# coding=utf-8
+# Copyright 2023 Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes for Ernie-M."""
+
+import io
+import os
+import unicodedata
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+
+from ...tokenization_utils import PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+SPIECE_UNDERLINE = "▁"
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt", "sentencepiece_model_ckpt": "sentencepiece.bpe.model"}
+
+RESOURCE_FILES_NAMES = {
+    "sentencepiece_model_file": "sentencepiece.bpe.model",
+    "vocab_file": "vocab.txt",
+}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "ernie-m-base": "https://huggingface.co/susnato/ernie-m-base_pytorch/blob/main/vocab.txt",
+        "ernie-m-large": "https://huggingface.co/susnato/ernie-m-base_pytorch/blob/main/vocab.txt",
+    },
+    "sentencepiece_model_file": {
+        "ernie-m-base": "https://huggingface.co/susnato/ernie-m-base_pytorch/blob/main/sentencepiece.bpe.model",
+        "ernie-m-large": "https://huggingface.co/susnato/ernie-m-base_pytorch/blob/main/sentencepiece.bpe.model",
+    },
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "ernie-m-base": 514,
+    "ernie-m-large": 514,
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    "ernie-m-base": {"do_lower_case": False},
+    "ernie-m-large": {"do_lower_case": False},
+}
+
+
+# Adapted from paddlenlp.transformers.ernie_m.tokenizer.ErnieMTokenizer
+class ErnieMTokenizer(PreTrainedTokenizer):
+    r"""
+    Constructs a Ernie-M tokenizer. It uses the `sentencepiece` tools to cut the words to sub-words.
+
+    Args:
+        sentencepiece_model_file (`str`):
+            The file path of sentencepiece model.
+        vocab_file (`str`, *optional*):
+            The file path of the vocabulary.
+        do_lower_case (`str`, *optional*, defaults to `True`):
+            Whether or not to lowercase the input when tokenizing.
+        unk_token (`str`, *optional*, defaults to `"[UNK]"`):
+            A special token representing the `unknown (out-of-vocabulary)` token. An unknown token is set to be
+            `unk_token` inorder to be converted to an ID.
+        sep_token (`str`, *optional*, defaults to `"[SEP]"`):
+            A special token separating two different sentences in the same input.
+        pad_token (`str`, *optional*, defaults to `"[PAD]"`):
+            A special token used to make arrays of tokens the same size for batching purposes.
+        cls_token (`str`, *optional*, defaults to `"[CLS]"`):
+            A special token used for sequence classification. It is the last token of the sequence when built with
+            special tokens.
+        mask_token (`str`, *optional*, defaults to `"[MASK]"`):
+            A special token representing a masked token. This is the token used in the masked language modeling task
+            which the model tries to predict the original unmasked ones.
+    """
+
+    # Ernie-M model doesn't have token_type embedding.
+    model_input_names: List[str] = ["input_ids"]
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    resource_files_names = RESOURCE_FILES_NAMES
+
+    def __init__(
+        self,
+        sentencepiece_model_ckpt,
+        vocab_file=None,
+        do_lower_case=False,
+        encoding="utf8",
+        unk_token="[UNK]",
+        sep_token="[SEP]",
+        pad_token="[PAD]",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        **kwargs,
+    ) -> None:
+        # Mask token behave like a normal word, i.e. include the space before it and
+        # is included in the raw text, there should be a match in a non-normalized sentence.
+
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        super().__init__(
+            do_lower_case=do_lower_case,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            vocab_file=vocab_file,
+            encoding=encoding,
+            sp_model_kwargs=self.sp_model_kwargs,
+            **kwargs,
+        )
+        self.do_lower_case = do_lower_case
+        self.sentencepiece_model_ckpt = sentencepiece_model_ckpt
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(sentencepiece_model_ckpt)
+
+        # to mimic paddlenlp.transformers.ernie_m.tokenizer.ErnieMTokenizer functioning
+        if vocab_file is not None:
+            self.vocab = self.load_vocab(filepath=vocab_file)
+        else:
+            self.vocab = dict((self.sp_model.id_to_piece(id), id) for id in range(self.sp_model.get_piece_size()))
+        self.reverse_vocab = dict((v, k) for k, v in self.vocab.items())
+
+    def get_offset_mapping(self, text):
+        if text is None:
+            return None
+
+        split_tokens = self.tokenize(text)
+        normalized_text, char_mapping = "", []
+
+        for i, ch in enumerate(text):
+            if ch in self.SP_CHAR_MAPPING:
+                ch = self.SP_CHAR_MAPPING.get(ch)
+            else:
+                ch = unicodedata.normalize("NFKC", ch)
+            if self.is_whitespace(ch):
+                continue
+            normalized_text += ch
+            char_mapping.extend([i] * len(ch))
+
+        text, token_mapping, offset = normalized_text, [], 0
+
+        if self.do_lower_case:
+            text = text.lower()
+
+        for token in split_tokens:
+            if token[:1] == "▁":
+                token = token[1:]
+            start = text[offset:].index(token) + offset
+            end = start + len(token)
+
+            token_mapping.append((char_mapping[start], char_mapping[end - 1] + 1))
+            offset = end
+        return token_mapping
+
+    @property
+    def vocab_size(self):
+        return len(self.vocab)
+
+    def get_vocab(self):
+        return dict(self.vocab, **self.added_tokens_encoder)
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.sentencepiece_model_ckpt)
+
+    def clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        return "".join((self.SP_CHAR_MAPPING.get(c, c) for c in text))
+
+    def _tokenize(self, text, enable_sampling=False, nbest_size=64, alpha=0.1):
+        """Tokenize a string."""
+
+        if self.sp_model_kwargs.get("enable_sampling") is True:
+            enable_sampling = True
+        if self.sp_model_kwargs.get("alpha") is not None:
+            alpha = self.sp_model_kwargs.get("alpha")
+        if self.sp_model_kwargs.get("nbest_size") is not None:
+            nbest_size = self.sp_model_kwargs.get("nbest_size")
+
+        if not enable_sampling:
+            pieces = self.sp_model.EncodeAsPieces(text)
+        else:
+            pieces = self.sp_model.SampleEncodeAsPieces(text, nbest_size, alpha)
+        new_pieces = []
+        for pi, piece in enumerate(pieces):
+            if piece == SPIECE_UNDERLINE:
+                if not pieces[pi + 1].startswith(SPIECE_UNDERLINE) and pi != 0:
+                    new_pieces.append(SPIECE_UNDERLINE)
+                    continue
+                else:
+                    continue
+            lst_i = 0
+            for i, chunk in enumerate(piece):
+                if chunk == SPIECE_UNDERLINE:
+                    continue
+                if self.is_ch_char(chunk) or self.is_punct(chunk):
+                    if i > lst_i and piece[lst_i:i] != SPIECE_UNDERLINE:
+                        new_pieces.append(piece[lst_i:i])
+                    new_pieces.append(chunk)
+                    lst_i = i + 1
+                elif chunk.isdigit() and i > 0 and not piece[i - 1].isdigit():
+                    if i > lst_i and piece[lst_i:i] != SPIECE_UNDERLINE:
+                        new_pieces.append(piece[lst_i:i])
+                    lst_i = i
+                elif not chunk.isdigit() and i > 0 and piece[i - 1].isdigit():
+                    if i > lst_i and piece[lst_i:i] != SPIECE_UNDERLINE:
+                        new_pieces.append(piece[lst_i:i])
+                    lst_i = i
+            if len(piece) > lst_i:
+                new_pieces.append(piece[lst_i:])
+        return new_pieces
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (strings for sub-words) in a single string."""
+        out_string = "".join(tokens).replace(SPIECE_UNDERLINE, " ").strip()
+        return out_string
+
+    def convert_ids_to_string(self, ids):
+        """
+        Converts a sequence of tokens (strings for sub-words) in a single string.
+        """
+        tokens = self.convert_ids_to_tokens(ids)
+        out_string = "".join(tokens).replace(SPIECE_UNDERLINE, " ").strip()
+        return out_string
+
+    # to mimic paddlenlp.transformers.ernie_m.tokenizer.ErnieMTokenizer functioning
+    def _convert_token_to_id(self, token):
+        return self.vocab.get(token, self.vocab.get(self.unk_token))
+
+    # to mimic paddlenlp.transformers.ernie_m.tokenizer.ErnieMTokenizer functioning
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.reverse_vocab.get(index, self.unk_token)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        r"""
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. An ErnieM sequence has the following format:
+
+        - single sequence: `[CLS] X [SEP]`
+        - pair of sequences: `[CLS] A [SEP] [SEP] B [SEP]`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+        Returns:
+            `List[int]`: List of input_id with the appropriate special tokens.
+        """
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        _cls = [self.cls_token_id]
+        _sep = [self.sep_token_id]
+        return _cls + token_ids_0 + _sep + _sep + token_ids_1 + _sep
+
+    def build_offset_mapping_with_special_tokens(self, offset_mapping_0, offset_mapping_1=None):
+        r"""
+        Build offset map from a pair of offset map by concatenating and adding offsets of special tokens. An Ernie-M
+        offset_mapping has the following format:
+
+        - single sequence: `(0,0) X (0,0)`
+        - pair of sequences: `(0,0) A (0,0) (0,0) B (0,0)`
+
+        Args:
+            offset_mapping_ids_0 (`List[tuple]`):
+                List of char offsets to which the special tokens will be added.
+            offset_mapping_ids_1 (`List[tuple]`, *optional*):
+                Optional second list of wordpiece offsets for offset mapping pairs.
+        Returns:
+            `List[tuple]`: List of wordpiece offsets with the appropriate offsets of special tokens.
+        """
+        if offset_mapping_1 is None:
+            return [(0, 0)] + offset_mapping_0 + [(0, 0)]
+
+        return [(0, 0)] + offset_mapping_0 + [(0, 0), (0, 0)] + offset_mapping_1 + [(0, 0)]
+
+    def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
+        r"""
+        Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `encode` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of ids of the first sequence.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`str`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+        Returns:
+            `List[int]`:
+                The list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            if token_ids_1 is not None:
+                raise ValueError(
+                    "You should not supply a second sequence if the provided sequence of "
+                    "ids is already formatted with special tokens for the model."
+                )
+            return list(map(lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0, token_ids_0))
+
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create the token type IDs corresponding to the sequences passed. [What are token type
+        IDs?](../glossary#token-type-ids) Should be overridden in a subclass if the model has a special way of
+        building: those.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                The first tokenized sequence.
+            token_ids_1 (`List[int]`, *optional*):
+                The second tokenized sequence.
+        Returns:
+            `List[int]`: The token type ids.
+        """
+        # called when `add_special_tokens` is True, so align with `build_inputs_with_special_tokens` method
+        if token_ids_1 is None:
+            # [CLS] X [SEP]
+            return (len(token_ids_0) + 2) * [0]
+
+        # [CLS] A [SEP] [SEP] B [SEP]
+        return [0] * (len(token_ids_0) + 1) + [1] * (len(token_ids_1) + 3)
+
+    def is_ch_char(self, char):
+        """
+        is_ch_char
+        """
+        if "\u4e00" <= char <= "\u9fff":
+            return True
+        return False
+
+    def is_alpha(self, char):
+        """
+        is_alpha
+        """
+        if ("a" <= char <= "z") or ("A" <= char <= "Z"):
+            return True
+        return False
+
+    def is_punct(self, char):
+        """
+        is_punct
+        """
+        if char in ",;:.?!~，；：。？！《》【】":
+            return True
+        return False
+
+    def is_whitespace(self, char):
+        """
+        is whitespace
+        """
+        if char == " " or char == "\t" or char == "\n" or char == "\r":
+            return True
+        if len(char) == 1:
+            cat = unicodedata.category(char)
+            if cat == "Zs":
+                return True
+        return False
+
+    def load_vocab(self, filepath):
+        token_to_idx = {}
+        with io.open(filepath, "r", encoding="utf-8") as f:
+            for index, line in enumerate(f):
+                token = line.rstrip("\n")
+                token_to_idx[token] = int(index)
+
+        return token_to_idx
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        index = 0
+        if os.path.isdir(save_directory):
+            vocab_file = os.path.join(
+                save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+            )
+        else:
+            vocab_file = (filename_prefix + "-" if filename_prefix else "") + save_directory
+        with open(vocab_file, "w", encoding="utf-8") as writer:
+            for token, token_index in sorted(self.vocab.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        f"Saving vocabulary to {vocab_file}: vocabulary indices are not consecutive."
+                        " Please check that the vocabulary is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(token + "\n")
+                index += 1
+
+        tokenizer_model_file = os.path.join(save_directory, "sentencepiece.bpe.model")
+        with open(tokenizer_model_file, "wb") as fi:
+            content_spiece_model = self.sp_model.serialized_model_proto()
+            fi.write(content_spiece_model)
+
+        return (vocab_file,)

src/transformers/utils/dummy_pt_objects.py

@@ -2415,6 +2415,58 @@ class ErniePreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+ERNIE_M_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class ErnieMForInformationExtraction(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ErnieMForMultipleChoice(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ErnieMForQuestionAnswering(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ErnieMForSequenceClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ErnieMForTokenClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ErnieMModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ErnieMPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 ESM_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_sentencepiece_objects.py

@@ -58,6 +58,13 @@ class DebertaV2Tokenizer(metaclass=DummyObject):
         requires_backends(self, ["sentencepiece"])
 
 
+class ErnieMTokenizer(metaclass=DummyObject):
+    _backends = ["sentencepiece"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["sentencepiece"])
+
+
 class FNetTokenizer(metaclass=DummyObject):
     _backends = ["sentencepiece"]
 

tests/models/ernie_m/test_modeling_ernie_m.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. and Baidu team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch ErnieM model. """
+
+
+import unittest
+
+from transformers import ErnieMConfig, is_torch_available
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        ErnieMForInformationExtraction,
+        ErnieMForMultipleChoice,
+        ErnieMForQuestionAnswering,
+        ErnieMForSequenceClassification,
+        ErnieMForTokenClassification,
+        ErnieMModel,
+    )
+    from transformers.models.ernie_m.modeling_ernie_m import ERNIE_M_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+class ErnieMModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def prepare_config_and_inputs_for_uiem(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return ErnieMConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
+        model = ErnieMModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, return_dict=True)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_for_question_answering(
+        self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = ErnieMForQuestionAnswering(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            start_positions=sequence_labels,
+            end_positions=sequence_labels,
+        )
+        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
+
+    def create_and_check_for_information_extraction(
+        self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = ErnieMForInformationExtraction(config=config)
+        model.to(torch_device)
+        model.eval()
+        sequence_labels = torch.ones_like(input_ids, dtype=torch.float32)
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            start_positions=sequence_labels,
+            end_positions=sequence_labels,
+        )
+        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
+
+    def create_and_check_for_sequence_classification(
+        self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_labels = self.num_labels
+        model = ErnieMForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, labels=sequence_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def create_and_check_for_token_classification(
+        self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_labels = self.num_labels
+        model = ErnieMForTokenClassification(config=config)
+        model.to(torch_device)
+        model.eval()
+        input_ids.to(torch_device)
+        input_mask.to(torch_device)
+        token_labels.to(torch_device)
+
+        result = model(input_ids, attention_mask=input_mask, labels=token_labels)
+
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
+
+    def create_and_check_for_multiple_choice(
+        self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_choices = self.num_choices
+        model = ErnieMForMultipleChoice(config=config)
+        model.to(torch_device)
+        model.eval()
+        multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        result = model(
+            multiple_choice_inputs_ids,
+            attention_mask=multiple_choice_input_mask,
+            labels=choice_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class ErnieMModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            ErnieMModel,
+            ErnieMForMultipleChoice,
+            ErnieMForQuestionAnswering,
+            ErnieMForSequenceClassification,
+            ErnieMForTokenClassification,
+        )
+        if is_torch_available()
+        else ()
+    )
+    all_generative_model_classes = ()
+    test_torchscript = False
+
+    def setUp(self):
+        self.model_tester = ErnieMModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=ErnieMConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_various_embeddings(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        for type in ["absolute", "relative_key", "relative_key_query"]:
+            config_and_inputs[0].position_embedding_type = type
+            self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_for_multiple_choice(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
+
+    def test_for_information_extraction(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_information_extraction(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in ERNIE_M_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = ErnieMModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+@require_torch
+class ErnieMModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference_model(self):
+        model = ErnieMModel.from_pretrained("susnato/ernie-m-base_pytorch")
+        model.eval()
+        input_ids = torch.tensor([[0, 1, 2, 3, 4, 5]])
+        output = model(input_ids)[0]
+
+        # TODO Replace vocab size
+        hidden_size = 768
+
+        expected_shape = torch.Size((1, 6, hidden_size))
+        self.assertEqual(output.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[[-0.0012, 0.1245, -0.0214], [-0.0742, 0.0244, -0.0771], [-0.0333, 0.1164, -0.1554]]]
+        )
+
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-3))

tests/models/ernie_m/test_tokenization_ernie_m.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. and Baidu team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch ErnieM model. """
+
+import unittest
+
+from transformers import ErnieMTokenizer
+from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_tokenizers, slow
+
+from ...test_tokenization_common import TokenizerTesterMixin
+
+
+SAMPLE_VOCAB = get_tests_dir("fixtures/spiece.model")
+
+
+@require_sentencepiece
+@require_tokenizers
+class ErnieMTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+    tokenizer_class = ErnieMTokenizer
+    test_seq2seq = False
+    test_sentencepiece = True
+    test_rust_tokenizer = False
+    test_sentencepiece_ignore_case = False
+
+    def setUp(self):
+        super().setUp()
+
+        # We have a SentencePiece fixture for testing
+        tokenizer = ErnieMTokenizer(SAMPLE_VOCAB, unk_token="<unk>", pad_token="<pad>")
+        tokenizer.save_pretrained(self.tmpdirname)
+
+    def get_input_output_texts(self, tokenizer):
+        input_text = "this is a test"
+        output_text = "this is a test"
+        return input_text, output_text
+
+    def test_convert_token_and_id(self):
+        """Test ``_convert_token_to_id`` and ``_convert_id_to_token``."""
+        token = "<pad>"
+        token_id = 0
+
+        self.assertEqual(self.get_tokenizer()._convert_token_to_id(token), token_id)
+        self.assertEqual(self.get_tokenizer()._convert_id_to_token(token_id), token)
+
+    def test_get_vocab(self):
+        vocab_keys = list(self.get_tokenizer().get_vocab().keys())
+
+        self.assertEqual(vocab_keys[0], "<pad>")
+        self.assertEqual(vocab_keys[1], "<unk>")
+        self.assertEqual(vocab_keys[-1], "▁eloquent")
+        self.assertEqual(len(vocab_keys), 30_000)
+
+    def test_vocab_size(self):
+        self.assertEqual(self.get_tokenizer().vocab_size, 30_000)
+
+    def test_rust_and_python_full_tokenizers(self):
+        if not self.test_rust_tokenizer:
+            return
+
+        tokenizer = self.get_tokenizer()
+        rust_tokenizer = self.get_rust_tokenizer()
+
+        sequence = "I was born in 92000, and this is falsé."
+
+        tokens = tokenizer.tokenize(sequence)
+        rust_tokens = rust_tokenizer.tokenize(sequence)
+        self.assertListEqual(tokens, rust_tokens)
+
+        ids = tokenizer.encode(sequence, add_special_tokens=False)
+        rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
+        self.assertListEqual(ids, rust_ids)
+
+        rust_tokenizer = self.get_rust_tokenizer()
+        ids = tokenizer.encode(sequence)
+        rust_ids = rust_tokenizer.encode(sequence)
+        self.assertListEqual(ids, rust_ids)
+
+    def test_full_tokenizer(self):
+        tokenizer = ErnieMTokenizer(SAMPLE_VOCAB, do_lower_case=True, unk_token="<unk>", pad_token="<pad>")
+
+        tokens = tokenizer.tokenize("This is a test")
+        self.assertListEqual(tokens, ["▁this", "▁is", "▁a", "▁test"])
+
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [48, 25, 21, 1289])
+
+        tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
+        # ErnieMTokenizer(paddlenlp implementation) outputs '9' instead of '_9' so to mimic that '_9' is changed to '9'
+        self.assertListEqual(
+            tokens, ["▁i", "▁was", "▁born", "▁in", "9", "2000", ",", "▁and", "▁this", "▁is", "▁fal", "s", "é", "."]
+        )
+        ids = tokenizer.convert_tokens_to_ids(tokens)
+        self.assertListEqual(ids, [31, 23, 386, 19, 518, 3050, 15, 17, 48, 25, 8256, 18, 1, 9])
+
+        back_tokens = tokenizer.convert_ids_to_tokens(ids)
+        self.assertListEqual(
+            back_tokens,
+            ["▁i", "▁was", "▁born", "▁in", "9", "2000", ",", "▁and", "▁this", "▁is", "▁fal", "s", "<unk>", "."],
+        )
+
+    def test_sequence_builders(self):
+        tokenizer = ErnieMTokenizer(SAMPLE_VOCAB, unk_token="<unk>", pad_token="<pad>")
+
+        text = tokenizer.encode("sequence builders")
+        text_2 = tokenizer.encode("multi-sequence build")
+
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
+
+        assert encoded_sentence == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id]
+        assert encoded_pair == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id] + [
+            tokenizer.sep_token_id
+        ] + text_2 + [tokenizer.sep_token_id]
+
+    @slow
+    def test_tokenizer_integration(self):
+        # fmt: off
+        expected_encoding = {'input_ids': [[0, 11062, 82772, 7, 15, 82772, 538, 51529, 237, 17198, 1290, 206, 9, 215175, 1314, 136, 17198, 1290, 206, 9, 56359, 42, 122009, 9, 16466, 16, 87344, 4537, 9, 4717, 78381, 6, 159958, 7, 15, 24480, 618, 4, 527, 22693, 9, 304, 4, 2777, 24480, 9874, 4, 43523, 594, 4, 803, 18392, 33189, 18, 4, 43523, 24447, 5, 5, 5, 16, 100, 24955, 83658, 9626, 144057, 15, 839, 22335, 16, 136, 24955, 83658, 83479, 15, 39102, 724, 16, 678, 645, 6460, 1328, 4589, 42, 122009, 115774, 23, 3559, 1328, 46876, 7, 136, 53894, 1940, 42227, 41159, 17721, 823, 425, 4, 27512, 98722, 206, 136, 5531, 4970, 919, 17336, 5, 2], [0, 20080, 618, 83, 82775, 47, 479, 9, 1517, 73, 53894, 333, 80581, 110117, 18811, 5256, 1295, 51, 152526, 297, 7986, 390, 124416, 538, 35431, 214, 98, 15044, 25737, 136, 7108, 43701, 23, 756, 135355, 7, 5, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 581, 63773, 119455, 6, 147797, 88203, 7, 645, 70, 21, 3285, 10269, 5, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]}
+        # fmt: on
+
+        self.tokenizer_integration_test_util(
+            expected_encoding=expected_encoding,
+            model_name="susnato/ernie-m-base_pytorch",
+            sequences=[
+                "Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides "
+                "general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet...) for Natural "
+                "Language Understanding (NLU) and Natural Language Generation (NLG) with over32+ pretrained "
+                "models in100+ languages and deep interoperability between Jax, PyTorch and TensorFlow.",
+                "BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly "
+                "conditioning on both left and right context in all layers.",
+                "The quick brown fox jumps over the lazy dog.",
+            ],
+        )

utils/check_repo.py

@@ -56,6 +56,7 @@ IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
     "Blip2QFormerModel",  # Building part of bigger (tested) model.
     "DetaEncoder",  # Building part of bigger (tested) model.
     "DetaDecoder",  # Building part of bigger (tested) model.
+    "ErnieMForInformationExtraction",
     "GraphormerEncoder",  # Building part of bigger (tested) model.
     "GraphormerDecoderHead",  # Building part of bigger (tested) model.
     "CLIPSegDecoder",  # Building part of bigger (tested) model.
@@ -175,6 +176,7 @@ IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     "Blip2ForConditionalGeneration",
     "Blip2QFormerModel",
     "Blip2VisionModel",
+    "ErnieMForInformationExtraction",
     "GitVisionModel",
     "GraphormerModel",
     "GraphormerForGraphClassification",

utils/documentation_tests.txt

@@ -80,6 +80,8 @@ src/transformers/models/electra/configuration_electra.py
 src/transformers/models/electra/modeling_electra.py
 src/transformers/models/electra/modeling_tf_electra.py
 src/transformers/models/ernie/configuration_ernie.py
+src/transformers/models/ernie_m/configuration_ernie_m.py
+src/transformers/models/ernie_m/modeling_ernie_m.py
 src/transformers/models/flava/configuration_flava.py
 src/transformers/models/fnet/configuration_fnet.py
 src/transformers/models/fsmt/configuration_fsmt.py