led

mindnlp.transformers.models.led.configuration_led

LED model configuration

mindnlp.transformers.models.led.configuration_led.LEDConfig

Bases: PretrainedConfig

This is the configuration class to store the configuration of a [LEDModel]. It is used to instantiate an LED 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 LED allenai/led-base-16384 architecture.

Configuration objects inherit from [PretrainedConfig] and can be used to control the model outputs. Read the documentation from [PretrainedConfig] for more information.

PARAMETER	DESCRIPTION
vocab_size	Vocabulary size of the LED model. Defines the number of different tokens that can be represented by the inputs_ids passed when calling [LEDModel] or [TFLEDModel]. TYPE: `int`, *optional*, defaults to 50265 DEFAULT: 50265
d_model	Dimensionality of the layers and the pooler layer. TYPE: `int`, *optional*, defaults to 1024 DEFAULT: 1024
encoder_layers	Number of encoder layers. TYPE: `int`, *optional*, defaults to 12 DEFAULT: 12
decoder_layers	Number of decoder layers. TYPE: `int`, *optional*, defaults to 12 DEFAULT: 12
encoder_attention_heads	Number of attention heads for each attention layer in the Transformer encoder. TYPE: `int`, *optional*, defaults to 16 DEFAULT: 16
decoder_attention_heads	Number of attention heads for each attention layer in the Transformer decoder. TYPE: `int`, *optional*, defaults to 16 DEFAULT: 16
decoder_ffn_dim	Dimensionality of the "intermediate" (often named feed-forward) layer in decoder. TYPE: `int`, *optional*, defaults to 4096 DEFAULT: 4096
encoder_ffn_dim	Dimensionality of the "intermediate" (often named feed-forward) layer in decoder. TYPE: `int`, *optional*, defaults to 4096 DEFAULT: 4096
activation_function	The non-linear activation function (function or string) in the encoder and pooler. If string, "gelu", "relu", "silu" and "gelu_new" are supported. TYPE: `str` or `function`, *optional*, defaults to `"gelu"` DEFAULT: 'gelu'
dropout	The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. TYPE: `float`, *optional*, defaults to 0.1 DEFAULT: 0.1
attention_dropout	The dropout ratio for the attention probabilities. TYPE: `float`, *optional*, defaults to 0.0 DEFAULT: 0.0
activation_dropout	The dropout ratio for activations inside the fully connected layer. TYPE: `float`, *optional*, defaults to 0.0 DEFAULT: 0.0
classifier_dropout	The dropout ratio for classifier. TYPE: `float`, *optional*, defaults to 0.0 DEFAULT: 0.0
max_encoder_position_embeddings	The maximum sequence length that the encoder might ever be used with. TYPE: `int`, *optional*, defaults to 16384 DEFAULT: 16384
max_decoder_position_embeddings	The maximum sequence length that the decoder might ever be used with. TYPE: `int`, *optional*, defaults to 16384 DEFAULT: 1024
init_std	The standard deviation of the truncated_normal_initializer for initializing all weight matrices. TYPE: `float`, *optional*, defaults to 0.02 DEFAULT: 0.02
encoder_layerdrop	The LayerDrop probability for the encoder. See the LayerDrop paper for more details. TYPE: `float`, *optional*, defaults to 0.0 DEFAULT: 0.0
decoder_layerdrop	The LayerDrop probability for the decoder. See the LayerDrop paper for more details. TYPE: `float`, *optional*, defaults to 0.0 DEFAULT: 0.0
use_cache	Whether or not the model should return the last key/values attentions (not used by all models) TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True

Example

>>> from transformers import LEDModel, LEDConfig
...
>>> # Initializing a LED allenai/led-base-16384 style configuration
>>> configuration = LEDConfig()
...
>>> # Initializing a model from the allenai/led-base-16384 style configuration
>>> model = LEDModel(configuration)
...
>>> # Accessing the model configuration
>>> configuration = model.config

Source code in mindnlp/transformers/models/led/configuration_led.py

class LEDConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`LEDModel`]. It is used to instantiate an LED
    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 LED
    [allenai/led-base-16384](https://huggingface.co/allenai/led-base-16384) architecture.

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.

    Args:
        vocab_size (`int`, *optional*, defaults to 50265):
            Vocabulary size of the LED model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`LEDModel`] or [`TFLEDModel`].
        d_model (`int`, *optional*, defaults to 1024):
            Dimensionality of the layers and the pooler layer.
        encoder_layers (`int`, *optional*, defaults to 12):
            Number of encoder layers.
        decoder_layers (`int`, *optional*, defaults to 12):
            Number of decoder layers.
        encoder_attention_heads (`int`, *optional*, defaults to 16):
            Number of attention heads for each attention layer in the Transformer encoder.
        decoder_attention_heads (`int`, *optional*, defaults to 16):
            Number of attention heads for each attention layer in the Transformer decoder.
        decoder_ffn_dim (`int`, *optional*, defaults to 4096):
            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
        encoder_ffn_dim (`int`, *optional*, defaults to 4096):
            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
        activation_function (`str` or `function`, *optional*, defaults to `"gelu"`):
            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
            `"relu"`, `"silu"` and `"gelu_new"` are supported.
        dropout (`float`, *optional*, defaults to 0.1):
            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        activation_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for activations inside the fully connected layer.
        classifier_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for classifier.
        max_encoder_position_embeddings (`int`, *optional*, defaults to 16384):
            The maximum sequence length that the encoder might ever be used with.
        max_decoder_position_embeddings (`int`, *optional*, defaults to 16384):
            The maximum sequence length that the decoder might ever be used with.
        init_std (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        encoder_layerdrop (`float`, *optional*, defaults to 0.0):
            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
            for more details.
        decoder_layerdrop (`float`, *optional*, defaults to 0.0):
            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
            for more details.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models)

    Example:
        ```python
        >>> from transformers import LEDModel, LEDConfig
        ...
        >>> # Initializing a LED allenai/led-base-16384 style configuration
        >>> configuration = LEDConfig()
        ...
        >>> # Initializing a model from the allenai/led-base-16384 style configuration
        >>> model = LEDModel(configuration)
        ...
        >>> # Accessing the model configuration
        >>> configuration = model.config
        ```
    """

    model_type = "led"
    attribute_map = {
        "num_attention_heads": "encoder_attention_heads",
        "hidden_size": "d_model",
        "attention_probs_dropout_prob": "attention_dropout",
        "initializer_range": "init_std",
    }

    def __init__(
        self,
        vocab_size=50265,
        max_encoder_position_embeddings=16384,
        max_decoder_position_embeddings=1024,
        encoder_layers=12,
        encoder_ffn_dim=4096,
        encoder_attention_heads=16,
        decoder_layers=12,
        decoder_ffn_dim=4096,
        decoder_attention_heads=16,
        encoder_layerdrop=0.0,
        decoder_layerdrop=0.0,
        use_cache=True,
        is_encoder_decoder=True,
        activation_function="gelu",
        d_model=1024,
        dropout=0.1,
        attention_dropout=0.0,
        activation_dropout=0.0,
        init_std=0.02,
        decoder_start_token_id=2,
        classifier_dropout=0.0,
        pad_token_id=1,
        bos_token_id=0,
        eos_token_id=2,
        attention_window: Union[List[int], int] = 512,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_encoder_position_embeddings = max_encoder_position_embeddings
        self.max_decoder_position_embeddings = max_decoder_position_embeddings
        self.d_model = d_model
        self.encoder_ffn_dim = encoder_ffn_dim
        self.encoder_layers = encoder_layers
        self.encoder_attention_heads = encoder_attention_heads
        self.decoder_ffn_dim = decoder_ffn_dim
        self.decoder_layers = decoder_layers
        self.decoder_attention_heads = decoder_attention_heads
        self.dropout = dropout
        self.attention_dropout = attention_dropout
        self.activation_dropout = activation_dropout
        self.activation_function = activation_function
        self.init_std = init_std
        self.encoder_layerdrop = encoder_layerdrop
        self.decoder_layerdrop = decoder_layerdrop
        self.classifier_dropout = classifier_dropout
        self.use_cache = use_cache
        self.num_hidden_layers = encoder_layers
        self.attention_window = attention_window

        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            is_encoder_decoder=is_encoder_decoder,
            decoder_start_token_id=decoder_start_token_id,
            **kwargs,
        )

mindnlp.transformers.models.led.modeling_led

MindSpore LED model.

mindnlp.transformers.models.led.modeling_led.LEDClassificationHead

Bases: Module

Head for sentence-level classification tasks.

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDClassificationHead(nn.Module):
    """Head for sentence-level classification tasks."""

    def __init__(
        self,
        input_dim: int,
        inner_dim: int,
        num_classes: int,
        pooler_dropout: float,
    ):
        super().__init__()
        self.dense = nn.Linear(input_dim, inner_dim)
        self.dropout = nn.Dropout(p=pooler_dropout)
        self.out_proj = nn.Linear(inner_dim, num_classes)

    def forward(self, hidden_states: mindspore.Tensor):
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.dense(hidden_states)
        hidden_states = ops.tanh(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.out_proj(hidden_states)
        return hidden_states

mindnlp.transformers.models.led.modeling_led.LEDDecoder

Bases: LEDPreTrainedModel

Transformer decoder consisting of config.decoder_layers layers. Each layer is a [LEDDecoderLayer]

PARAMETER	DESCRIPTION
config	LEDConfig TYPE: LEDConfig
embed_tokens	output embedding TYPE: Embedding DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDDecoder(LEDPreTrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`LEDDecoderLayer`]

    Args:
        config: LEDConfig
        embed_tokens (nn.Embedding): output embedding
    """

    def __init__(self, config: LEDConfig, embed_tokens: Optional[nn.Embedding] = None):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_decoder_position_embeddings

        if embed_tokens is not None:
            self.embed_tokens = embed_tokens
        else:
            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)

        self.embed_positions = LEDLearnedPositionalEmbedding(
            self.max_target_positions,
            config.d_model,
        )
        self.layers = nn.ModuleList([LEDDecoderLayer(config) for _ in range(config.decoder_layers)])
        self.layernorm_embedding = nn.LayerNorm(config.d_model)

        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        global_attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        head_mask=None,
        cross_attn_head_mask=None,
        past_key_values=None,
        inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        Args:
            input_ids (`mindspore.Tensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *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)
            global_attention_mask (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to decide the attention given on each token, local attention or global attention. Tokens with
                global attention attends to all other tokens, and all other tokens attend to them. This is important
                for task-specific finetuning because it makes the model more flexible at representing the task. For
                example, for classification, the <s> token should be given global attention. For QA, all question
                tokens should also have global attention. Please refer to the [Longformer
                paper](https://arxiv.org/abs/2004.05150) for more details. Mask values selected in `[0, 1]`:

                - 0 for local attention (a sliding window attention),
                - 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
            encoder_hidden_states (`mindspore.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                of the decoder.
            encoder_attention_mask (`mindspore.Tensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. 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)
            head_mask (`mindspore.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            cross_attn_head_mask (`mindspore.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            past_key_values (`tuple(tuple(mindspore.Tensor))`, *optional*, returned when `use_cache=True`
                is passed or when `config.use_cache=True`):
                Tuple of `tuple(mindspore.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of
                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
            inputs_embeds (`mindspore.Tensor` of shape `(batch_size, sequence_length, 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.
        """
        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
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.shape
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.shape[:-1]
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        # past_key_values_length
        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        # create causal mask
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        combined_attention_mask = None
        if input_shape[-1] > 1:
            combined_attention_mask = _create_4d_causal_attention_mask(
                input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
            )

        if attention_mask is not None and combined_attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            combined_attention_mask = combined_attention_mask + _prepare_4d_attention_mask_inverted(
                attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
            )

        # expand encoder attention mask
        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            encoder_attention_mask = _prepare_4d_attention_mask_inverted(
                encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
            )

        # embed positions
        positions = self.embed_positions(input_shape, past_key_values_length)

        hidden_states = inputs_embeds + positions
        hidden_states = self.layernorm_embedding(hidden_states)

        hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_cross_attentions = () if output_attentions else None
        next_decoder_cache = () if use_cache else None

        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
        for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
            if attn_mask is not None:
                if attn_mask.shape[0] != len(self.layers):
                    raise ValueError(
                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
                        f" {head_mask.shape[0]}."
                    )
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            if self.training:
                dropout_probability = ops.rand([])
                if dropout_probability < self.layerdrop:
                    continue

            past_key_value = past_key_values[idx] if past_key_values is not None else None

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states,
                    combined_attention_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
                    head_mask[idx] if head_mask is not None else None,
                    cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
                    None,
                    output_attentions,
                    use_cache,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=combined_attention_mask,
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=encoder_attention_mask,
                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                    cross_attn_layer_head_mask=(
                        cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
                    ),
                    past_key_value=past_key_value,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                )

            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)

            if output_attentions:
                all_self_attns += (layer_outputs[1],)
                all_cross_attentions += (layer_outputs[2],)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None
        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )

mindnlp.transformers.models.led.modeling_led.LEDDecoder.forward(input_ids=None, attention_mask=None, global_attention_mask=None, encoder_hidden_states=None, encoder_attention_mask=None, head_mask=None, cross_attn_head_mask=None, past_key_values=None, inputs_embeds=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None)

PARAMETER	DESCRIPTION
input_ids	Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [AutoTokenizer]. See [PreTrainedTokenizer.encode] and [PreTrainedTokenizer.__call__] for details. What are input IDs? TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length)` DEFAULT: None
attention_mask	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? TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional* DEFAULT: None
global_attention_mask	Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]: 0 for local attention (a sliding window attention), 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them). TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional* DEFAULT: None
encoder_hidden_states	Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. TYPE: `mindspore.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional* DEFAULT: None
encoder_attention_mask	Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values selected in [0, 1]: 1 for tokens that are not masked, 0 for tokens that are masked. What are attention masks? TYPE: `mindspore.Tensor` of shape `(batch_size, encoder_sequence_length)`, *optional* DEFAULT: None
head_mask	Mask to nullify selected heads of the attention modules. Mask values selected in [0, 1]: 1 indicates the head is not masked, 0 indicates the head is masked. TYPE: `mindspore.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional* DEFAULT: None
cross_attn_head_mask	Mask to nullify selected heads of the cross-attention modules. Mask values selected in [0, 1]: 1 indicates the head is not masked, 0 indicates the head is masked. TYPE: `mindspore.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional* DEFAULT: None
inputs_embeds	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. TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional* DEFAULT: None
output_attentions	Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail. TYPE: `bool`, *optional* DEFAULT: None
output_hidden_states	Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail. TYPE: `bool`, *optional* DEFAULT: None
return_dict	Whether or not to return a [~utils.ModelOutput] instead of a plain tuple. TYPE: `bool`, *optional* DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    input_ids=None,
    attention_mask=None,
    global_attention_mask=None,
    encoder_hidden_states=None,
    encoder_attention_mask=None,
    head_mask=None,
    cross_attn_head_mask=None,
    past_key_values=None,
    inputs_embeds=None,
    use_cache=None,
    output_attentions=None,
    output_hidden_states=None,
    return_dict=None,
):
    r"""
    Args:
        input_ids (`mindspore.Tensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
            provide it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *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)
        global_attention_mask (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to decide the attention given on each token, local attention or global attention. Tokens with
            global attention attends to all other tokens, and all other tokens attend to them. This is important
            for task-specific finetuning because it makes the model more flexible at representing the task. For
            example, for classification, the <s> token should be given global attention. For QA, all question
            tokens should also have global attention. Please refer to the [Longformer
            paper](https://arxiv.org/abs/2004.05150) for more details. Mask values selected in `[0, 1]`:

            - 0 for local attention (a sliding window attention),
            - 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
        encoder_hidden_states (`mindspore.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
            of the decoder.
        encoder_attention_mask (`mindspore.Tensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
            Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. 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)
        head_mask (`mindspore.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        cross_attn_head_mask (`mindspore.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        past_key_values (`tuple(tuple(mindspore.Tensor))`, *optional*, returned when `use_cache=True`
            is passed or when `config.use_cache=True`):
            Tuple of `tuple(mindspore.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of
            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
            shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
            cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
            that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
            all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        inputs_embeds (`mindspore.Tensor` of shape `(batch_size, sequence_length, 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.
    """
    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
    )
    use_cache = use_cache if use_cache is not None else self.config.use_cache
    return_dict = return_dict if return_dict is not None else self.config.use_return_dict

    # retrieve input_ids and inputs_embeds
    if input_ids is not None and inputs_embeds is not None:
        raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
    elif input_ids is not None:
        input_shape = input_ids.shape
        input_ids = input_ids.view(-1, input_shape[-1])
    elif inputs_embeds is not None:
        input_shape = inputs_embeds.shape[:-1]
    else:
        raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

    # past_key_values_length
    past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0

    if inputs_embeds is None:
        inputs_embeds = self.embed_tokens(input_ids)

    # create causal mask
    # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
    combined_attention_mask = None
    if input_shape[-1] > 1:
        combined_attention_mask = _create_4d_causal_attention_mask(
            input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
        )

    if attention_mask is not None and combined_attention_mask is not None:
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        combined_attention_mask = combined_attention_mask + _prepare_4d_attention_mask_inverted(
            attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
        )

    # expand encoder attention mask
    if encoder_hidden_states is not None and encoder_attention_mask is not None:
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        encoder_attention_mask = _prepare_4d_attention_mask_inverted(
            encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
        )

    # embed positions
    positions = self.embed_positions(input_shape, past_key_values_length)

    hidden_states = inputs_embeds + positions
    hidden_states = self.layernorm_embedding(hidden_states)

    hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)

    if self.gradient_checkpointing and self.training:
        if use_cache:
            logger.warning_once(
                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
            )
            use_cache = False

    # decoder layers
    all_hidden_states = () if output_hidden_states else None
    all_self_attns = () if output_attentions else None
    all_cross_attentions = () if output_attentions else None
    next_decoder_cache = () if use_cache else None

    # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
    for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
        if attn_mask is not None:
            if attn_mask.shape[0] != len(self.layers):
                raise ValueError(
                    f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
                    f" {head_mask.shape[0]}."
                )
    for idx, decoder_layer in enumerate(self.layers):
        # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
        if output_hidden_states:
            all_hidden_states += (hidden_states,)
        if self.training:
            dropout_probability = ops.rand([])
            if dropout_probability < self.layerdrop:
                continue

        past_key_value = past_key_values[idx] if past_key_values is not None else None

        if self.gradient_checkpointing and self.training:
            layer_outputs = self._gradient_checkpointing_func(
                decoder_layer.__call__,
                hidden_states,
                combined_attention_mask,
                encoder_hidden_states,
                encoder_attention_mask,
                head_mask[idx] if head_mask is not None else None,
                cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
                None,
                output_attentions,
                use_cache,
            )
        else:
            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=combined_attention_mask,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=encoder_attention_mask,
                layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                cross_attn_layer_head_mask=(
                    cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
                ),
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
            )

        hidden_states = layer_outputs[0]

        if use_cache:
            next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)

        if output_attentions:
            all_self_attns += (layer_outputs[1],)
            all_cross_attentions += (layer_outputs[2],)

    # add hidden states from the last decoder layer
    if output_hidden_states:
        all_hidden_states += (hidden_states,)

    next_cache = next_decoder_cache if use_cache else None
    if not return_dict:
        return tuple(
            v
            for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
            if v is not None
        )
    return BaseModelOutputWithPastAndCrossAttentions(
        last_hidden_state=hidden_states,
        past_key_values=next_cache,
        hidden_states=all_hidden_states,
        attentions=all_self_attns,
        cross_attentions=all_cross_attentions,
    )

mindnlp.transformers.models.led.modeling_led.LEDDecoderAttention

Bases: Module

Multi-headed attention from 'Attention Is All You Need' paper

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDDecoderAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        if self.head_dim * num_heads != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
                f" {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

    def _shape(self, tensor: mindspore.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).swapaxes(1, 2)

    def forward(
        self,
        hidden_states: mindspore.Tensor,
        key_value_states: Optional[mindspore.Tensor] = None,
        past_key_value: Optional[Tuple[mindspore.Tensor]] = None,
        attention_mask: Optional[mindspore.Tensor] = None,
        layer_head_mask: Optional[mindspore.Tensor] = None,
        output_attentions: bool = False,
    ) -> Tuple[mindspore.Tensor, Optional[mindspore.Tensor], Optional[Tuple[mindspore.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None
        bsz, tgt_len, embed_dim = hidden_states.shape

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling
        # get key, value proj
        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = ops.cat([past_key_value[0], key_states], axis=2)
            value_states = ops.cat([past_key_value[1], value_states], axis=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(mindspore.Tensor, mindspore.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(mindspore.Tensor, mindspore.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_states, value_states)

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
        key_states = key_states.view(*proj_shape)
        value_states = value_states.view(*proj_shape)

        src_len = key_states.shape[1]
        attn_weights = ops.bmm(query_states, key_states.swapaxes(1, 2))

        if attn_weights.shape != (bsz * self.num_heads, tgt_len, src_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
                f" {attn_weights.shape}"
            )

        if attention_mask is not None:
            if attention_mask.shape != (bsz, 1, tgt_len, src_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.shape}"
                )
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        attn_weights = ops.softmax(attn_weights, axis=-1)
        if layer_head_mask is not None:
            if layer_head_mask.shape != (self.num_heads,):
                raise ValueError(
                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
                    f" {layer_head_mask.shape}"
                )
            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        if output_attentions:
            # this operation is a bit awkward, but it's required to
            # make sure that attn_weights keeps its gradient.
            # In order to do so, attn_weights have to be reshaped
            # twice and have to be reused in the following
            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
        else:
            attn_weights_reshaped = None

        attn_probs = ops.dropout(attn_weights, p=self.dropout, training=self.training)

        attn_output = ops.bmm(attn_probs, value_states)

        if attn_output.shape != (bsz * self.num_heads, tgt_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {attn_output.shape}"
            )

        attn_output = (
            attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
            .swapaxes(1, 2)
            .reshape(bsz, tgt_len, embed_dim)
        )

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped, past_key_value

mindnlp.transformers.models.led.modeling_led.LEDDecoderAttention.forward(hidden_states, key_value_states=None, past_key_value=None, attention_mask=None, layer_head_mask=None, output_attentions=False)

Input shape: Batch x Time x Channel

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    hidden_states: mindspore.Tensor,
    key_value_states: Optional[mindspore.Tensor] = None,
    past_key_value: Optional[Tuple[mindspore.Tensor]] = None,
    attention_mask: Optional[mindspore.Tensor] = None,
    layer_head_mask: Optional[mindspore.Tensor] = None,
    output_attentions: bool = False,
) -> Tuple[mindspore.Tensor, Optional[mindspore.Tensor], Optional[Tuple[mindspore.Tensor]]]:
    """Input shape: Batch x Time x Channel"""

    # if key_value_states are provided this layer is used as a cross-attention layer
    # for the decoder
    is_cross_attention = key_value_states is not None
    bsz, tgt_len, embed_dim = hidden_states.shape

    # get query proj
    query_states = self.q_proj(hidden_states) * self.scaling
    # get key, value proj
    if is_cross_attention and past_key_value is not None:
        # reuse k,v, cross_attentions
        key_states = past_key_value[0]
        value_states = past_key_value[1]
    elif is_cross_attention:
        # cross_attentions
        key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
        value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
    elif past_key_value is not None:
        # reuse k, v, self_attention
        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
        key_states = ops.cat([past_key_value[0], key_states], axis=2)
        value_states = ops.cat([past_key_value[1], value_states], axis=2)
    else:
        # self_attention
        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

    if self.is_decoder:
        # if cross_attention save Tuple(mindspore.Tensor, mindspore.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(mindspore.Tensor, mindspore.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_states, value_states)

    proj_shape = (bsz * self.num_heads, -1, self.head_dim)
    query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
    key_states = key_states.view(*proj_shape)
    value_states = value_states.view(*proj_shape)

    src_len = key_states.shape[1]
    attn_weights = ops.bmm(query_states, key_states.swapaxes(1, 2))

    if attn_weights.shape != (bsz * self.num_heads, tgt_len, src_len):
        raise ValueError(
            f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
            f" {attn_weights.shape}"
        )

    if attention_mask is not None:
        if attention_mask.shape != (bsz, 1, tgt_len, src_len):
            raise ValueError(
                f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.shape}"
            )
        attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
        attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

    attn_weights = ops.softmax(attn_weights, axis=-1)
    if layer_head_mask is not None:
        if layer_head_mask.shape != (self.num_heads,):
            raise ValueError(
                f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
                f" {layer_head_mask.shape}"
            )
        attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
        attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

    if output_attentions:
        # this operation is a bit awkward, but it's required to
        # make sure that attn_weights keeps its gradient.
        # In order to do so, attn_weights have to be reshaped
        # twice and have to be reused in the following
        attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
        attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
    else:
        attn_weights_reshaped = None

    attn_probs = ops.dropout(attn_weights, p=self.dropout, training=self.training)

    attn_output = ops.bmm(attn_probs, value_states)

    if attn_output.shape != (bsz * self.num_heads, tgt_len, self.head_dim):
        raise ValueError(
            f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
            f" {attn_output.shape}"
        )

    attn_output = (
        attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
        .swapaxes(1, 2)
        .reshape(bsz, tgt_len, embed_dim)
    )

    attn_output = self.out_proj(attn_output)

    return attn_output, attn_weights_reshaped, past_key_value

mindnlp.transformers.models.led.modeling_led.LEDDecoderLayer

Bases: Module

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDDecoderLayer(nn.Module):
    def __init__(self, config: LEDConfig):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = LEDDecoderAttention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
        )
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout

        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.encoder_attn = LEDDecoderAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
        )
        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: mindspore.Tensor,
        attention_mask: Optional[mindspore.Tensor] = None,
        encoder_hidden_states: Optional[mindspore.Tensor] = None,
        encoder_attention_mask: Optional[mindspore.Tensor] = None,
        layer_head_mask: Optional[mindspore.Tensor] = None,
        cross_attn_layer_head_mask: Optional[mindspore.Tensor] = None,
        past_key_value: Optional[Tuple[mindspore.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = True,
    ):
        """
        Args:
            hidden_states (`mindspore.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)*
            attention_mask (`mindspore.Tensor`): attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            encoder_hidden_states (`mindspore.Tensor`):
                cross attention input to the layer of shape *(batch, seq_len, embed_dim)*
            encoder_attention_mask (`mindspore.Tensor`): encoder attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            layer_head_mask (`mindspore.Tensor`): mask for attention heads in a given layer of size
                *(decoder_attention_heads,)*.
            cross_attn_layer_head_mask (`mindspore.Tensor`): mask for encoder attention heads in a given layer of
                size *(decoder_attention_heads,)*.
            past_key_value (`Tuple(mindspore.Tensor)`): cached past key and value projection states
            output_attentions (`bool`): Whether the base model outputs attentions.
                This requires the attentions tensor to be reshaped in this function.
        """
        residual = hidden_states

        # Self-Attention
        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
        # add present self-attn cache to positions 1,2 of present_key_value tuple
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=self_attn_past_key_value,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Cross-Attention Block
        cross_attn_present_key_value = None
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states

            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                layer_head_mask=cross_attn_layer_head_mask,
                past_key_value=cross_attn_past_key_value,
                output_attentions=output_attentions,
            )
            hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
            hidden_states = residual + hidden_states
            hidden_states = self.encoder_attn_layer_norm(hidden_states)

            # add cross-attn to positions 3,4 of present_key_value tuple
            present_key_value = present_key_value + cross_attn_present_key_value

        # Fully Connected
        residual = hidden_states
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = ops.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.final_layer_norm(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)

        if use_cache:
            outputs += (present_key_value,)

        return outputs

mindnlp.transformers.models.led.modeling_led.LEDDecoderLayer.forward(hidden_states, attention_mask=None, encoder_hidden_states=None, encoder_attention_mask=None, layer_head_mask=None, cross_attn_layer_head_mask=None, past_key_value=None, output_attentions=False, use_cache=True)

PARAMETER	DESCRIPTION
hidden_states	input to the layer of shape (batch, seq_len, embed_dim) TYPE: `mindspore.Tensor`
attention_mask	attention mask of size (batch, 1, tgt_len, src_len) where padding elements are indicated by very large negative values. TYPE: `mindspore.Tensor` DEFAULT: None
encoder_hidden_states	cross attention input to the layer of shape (batch, seq_len, embed_dim) TYPE: `mindspore.Tensor` DEFAULT: None
encoder_attention_mask	encoder attention mask of size (batch, 1, tgt_len, src_len) where padding elements are indicated by very large negative values. TYPE: `mindspore.Tensor` DEFAULT: None
layer_head_mask	mask for attention heads in a given layer of size (decoder_attention_heads,). TYPE: `mindspore.Tensor` DEFAULT: None
cross_attn_layer_head_mask	mask for encoder attention heads in a given layer of size (decoder_attention_heads,). TYPE: `mindspore.Tensor` DEFAULT: None
past_key_value	cached past key and value projection states TYPE: `Tuple(mindspore.Tensor)` DEFAULT: None
output_attentions	Whether the base model outputs attentions. This requires the attentions tensor to be reshaped in this function. TYPE: `bool` DEFAULT: False

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    hidden_states: mindspore.Tensor,
    attention_mask: Optional[mindspore.Tensor] = None,
    encoder_hidden_states: Optional[mindspore.Tensor] = None,
    encoder_attention_mask: Optional[mindspore.Tensor] = None,
    layer_head_mask: Optional[mindspore.Tensor] = None,
    cross_attn_layer_head_mask: Optional[mindspore.Tensor] = None,
    past_key_value: Optional[Tuple[mindspore.Tensor]] = None,
    output_attentions: Optional[bool] = False,
    use_cache: Optional[bool] = True,
):
    """
    Args:
        hidden_states (`mindspore.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)*
        attention_mask (`mindspore.Tensor`): attention mask of size
            *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
        encoder_hidden_states (`mindspore.Tensor`):
            cross attention input to the layer of shape *(batch, seq_len, embed_dim)*
        encoder_attention_mask (`mindspore.Tensor`): encoder attention mask of size
            *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
        layer_head_mask (`mindspore.Tensor`): mask for attention heads in a given layer of size
            *(decoder_attention_heads,)*.
        cross_attn_layer_head_mask (`mindspore.Tensor`): mask for encoder attention heads in a given layer of
            size *(decoder_attention_heads,)*.
        past_key_value (`Tuple(mindspore.Tensor)`): cached past key and value projection states
        output_attentions (`bool`): Whether the base model outputs attentions.
            This requires the attentions tensor to be reshaped in this function.
    """
    residual = hidden_states

    # Self-Attention
    # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
    self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
    # add present self-attn cache to positions 1,2 of present_key_value tuple
    hidden_states, self_attn_weights, present_key_value = self.self_attn(
        hidden_states=hidden_states,
        past_key_value=self_attn_past_key_value,
        attention_mask=attention_mask,
        layer_head_mask=layer_head_mask,
        output_attentions=output_attentions,
    )
    hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
    hidden_states = residual + hidden_states
    hidden_states = self.self_attn_layer_norm(hidden_states)

    # Cross-Attention Block
    cross_attn_present_key_value = None
    cross_attn_weights = None
    if encoder_hidden_states is not None:
        residual = hidden_states

        # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
        cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
        hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
            hidden_states=hidden_states,
            key_value_states=encoder_hidden_states,
            attention_mask=encoder_attention_mask,
            layer_head_mask=cross_attn_layer_head_mask,
            past_key_value=cross_attn_past_key_value,
            output_attentions=output_attentions,
        )
        hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.encoder_attn_layer_norm(hidden_states)

        # add cross-attn to positions 3,4 of present_key_value tuple
        present_key_value = present_key_value + cross_attn_present_key_value

    # Fully Connected
    residual = hidden_states
    hidden_states = self.activation_fn(self.fc1(hidden_states))
    hidden_states = ops.dropout(hidden_states, p=self.activation_dropout, training=self.training)
    hidden_states = self.fc2(hidden_states)
    hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
    hidden_states = residual + hidden_states
    hidden_states = self.final_layer_norm(hidden_states)

    outputs = (hidden_states,)

    if output_attentions:
        outputs += (self_attn_weights, cross_attn_weights)

    if use_cache:
        outputs += (present_key_value,)

    return outputs

mindnlp.transformers.models.led.modeling_led.LEDEncoder

Bases: LEDPreTrainedModel

Transformer encoder consisting of config.encoder_layers self-attention layers. Each layer is a [LEDEncoderLayer].

PARAMETER	DESCRIPTION
config	LEDConfig TYPE: LEDConfig
embed_tokens	output embedding TYPE: Embedding DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDEncoder(LEDPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self-attention layers. Each layer is a
    [`LEDEncoderLayer`].

    Args:
        config: LEDConfig
        embed_tokens (nn.Embedding): output embedding
    """

    def __init__(self, config: LEDConfig, embed_tokens: Optional[nn.Embedding] = None):
        super().__init__(config)

        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop

        embed_dim = config.d_model
        self.padding_idx = config.pad_token_id
        self.max_source_positions = config.max_encoder_position_embeddings

        if isinstance(config.attention_window, int):
            if config.attention_window % 2 != 0:
                raise ValueError("`config.attention_window` has to be an even value")
            if config.attention_window <= 0:
                raise ValueError("`config.attention_window` has to be positive")
            config.attention_window = [config.attention_window] * config.num_hidden_layers  # one value per layer
        else:
            if len(config.attention_window) != config.num_hidden_layers:
                raise ValueError(
                    "`len(config.attention_window)` should equal `config.num_hidden_layers`. "
                    f"Expected {config.num_hidden_layers}, given {len(config.attention_window)}"
                )

        if embed_tokens is not None:
            self.embed_tokens = embed_tokens
        else:
            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)

        self.embed_positions = LEDLearnedPositionalEmbedding(
            self.max_source_positions,
            embed_dim,
        )
        self.layers = nn.ModuleList([LEDEncoderLayer(config, i) for i in range(config.encoder_layers)])
        self.layernorm_embedding = nn.LayerNorm(embed_dim)

        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    def _merge_to_attention_mask(self, attention_mask: mindspore.Tensor, global_attention_mask: mindspore.Tensor):
        # longformer self-attention expects attention mask to have 0 (no attn), 1 (local attn), 2 (global attn)
        # (global_attention_mask + 1) => 1 for local attention, 2 for global attention
        # => final attention_mask => 0 for no attention, 1 for local attention 2 for global attention
        if attention_mask is not None:
            attention_mask = attention_mask * (global_attention_mask + 1)
        else:
            # simply use `global_attention_mask` as `attention_mask`
            # if no `attention_mask` is given
            attention_mask = global_attention_mask + 1
        return attention_mask

    def _pad_to_window_size(
        self,
        input_ids: mindspore.Tensor,
        attention_mask: mindspore.Tensor,
        inputs_embeds: mindspore.Tensor,
        pad_token_id: int,
    ):
        """A helper function to pad tokens and mask to work with implementation of Longformer self-attention."""
        # padding
        attention_window = (
            self.config.attention_window
            if isinstance(self.config.attention_window, int)
            else max(self.config.attention_window)
        )

        if attention_window % 2 != 0:
            raise ValueError(f"`attention_window` should be an even value. Given {attention_window}")
        input_shape = input_ids.shape if input_ids is not None else inputs_embeds.shape
        batch_size, seq_len = input_shape[:2]

        padding_len = (attention_window - seq_len % attention_window) % attention_window
        if padding_len > 0:
            logger.warning_once(
                f"Input ids are automatically padded from {seq_len} to {seq_len + padding_len} to be a multiple of "
                f"`config.attention_window`: {attention_window}"
            )
            if input_ids is not None:
                input_ids = ops.pad(input_ids, (0, padding_len), value=pad_token_id)
            if inputs_embeds is not None:
                input_ids_padding = ops.full((batch_size, padding_len), self.config.pad_token_id, dtype=mindspore.int64)
                inputs_embeds_padding = self.embed_tokens(input_ids_padding)
                inputs_embeds = ops.cat([inputs_embeds, inputs_embeds_padding], axis=-2)

            attention_mask = ops.pad(
                attention_mask, (0, padding_len), value=False
            )  # no attention on the padding tokens

        return padding_len, input_ids, attention_mask, inputs_embeds

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        global_attention_mask=None,
        head_mask=None,
        inputs_embeds=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        Args:
            input_ids (`mindspore.Tensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *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)
            global_attention_mask (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to decide the attention given on each token, local attention or global attention for the encoder.
                Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is
                important for task-specific finetuning because it makes the model more flexible at representing the
                task. For example, for classification, the `<s>` token should be given global attention. For QA, all
                question tokens should also have global attention. Please refer to the [Longformer
                paper](https://arxiv.org/abs/2004.05150) for more details. Mask values selected in `[0, 1]`:

                - 0 for local attention (a sliding window attention),
                - 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
            head_mask (`mindspore.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.
            inputs_embeds (`mindspore.Tensor` of shape `(batch_size, sequence_length, 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.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # check input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is None and inputs_embeds is None:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        # create default attention_mask
        if attention_mask is None:
            attention_mask = ops.ones(inputs_embeds.shape[:-1], dtype=mindspore.int64)

        # merge `global_attention_mask` and `attention_mask`
        if global_attention_mask is not None:
            attention_mask = self._merge_to_attention_mask(attention_mask, global_attention_mask)

        # pad input if necessary
        padding_len, input_ids, attention_mask, inputs_embeds = self._pad_to_window_size(
            input_ids=input_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            pad_token_id=self.config.pad_token_id,
        )

        # retrieve input_shape
        if input_ids is not None:
            input_shape = input_ids.shape
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.shape[:-1]

        # convert attention_mask to float
        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, seq_len]; 1 -> 0.0; 0 -> "-inf"
            attention_mask = _prepare_4d_attention_mask_inverted(attention_mask, inputs_embeds.dtype)[:, 0, 0, :]

        # get masking tensors
        is_index_masked = attention_mask < 0
        is_index_global_attn = attention_mask > 0
        is_global_attn = is_index_global_attn.flatten().any().item()

        embed_pos = self.embed_positions(input_shape)

        hidden_states = inputs_embeds + embed_pos
        hidden_states = self.layernorm_embedding(hidden_states)
        hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
        all_global_attentions = () if (output_attentions and is_global_attn) else None

        # check if head_mask has a correct number of layers specified if desired
        if head_mask is not None:
            if head_mask.shape[0] != len(self.layers):
                raise ValueError(
                    f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
                    f" {head_mask.shape[0]}."
                )
        for idx, encoder_layer in enumerate(self.layers):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            dropout_probability = ops.rand([])

            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
                layer_outputs = (None, None, None)
            else:
                if self.gradient_checkpointing and self.training:
                    layer_outputs = self._gradient_checkpointing_func(
                        encoder_layer.__call__,
                        hidden_states,
                        attention_mask,
                        head_mask[idx] if head_mask is not None else None,
                        is_index_masked,
                        is_index_global_attn,
                        is_global_attn,
                        output_attentions,
                    )
                else:
                    layer_outputs = encoder_layer(
                        hidden_states,
                        attention_mask=attention_mask,
                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                        is_index_masked=is_index_masked,
                        is_index_global_attn=is_index_global_attn,
                        is_global_attn=is_global_attn,
                        output_attentions=output_attentions,
                    )
                hidden_states = layer_outputs[0]

            if output_attentions:
                # bzs x seq_len x num_attn_heads x (num_global_attn + attention_window_len + 1) => bzs x num_attn_heads x seq_len x (num_global_attn + attention_window_len + 1)
                all_attentions = all_attentions + (layer_outputs[1].swapaxes(1, 2),)

                if is_global_attn:
                    # bzs x num_attn_heads x num_global_attn x seq_len => bzs x num_attn_heads x seq_len x num_global_attn
                    all_global_attentions = all_global_attentions + (layer_outputs[2].swapaxes(2, 3),)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        # undo padding
        if padding_len > 0:
            # unpad `hidden_states` because the calling function is expecting a length == input_ids.shape[1]
            hidden_states = hidden_states[:, :-padding_len]
            if output_hidden_states:
                encoder_states = tuple(state[:, :-padding_len] for state in encoder_states)

            if output_attentions:
                all_attentions = tuple(state[:, :, :-padding_len, :] for state in all_attentions)

        if not return_dict:
            return tuple(
                v for v in [hidden_states, encoder_states, all_attentions, all_global_attentions] if v is not None
            )
        return LEDEncoderBaseModelOutput(
            last_hidden_state=hidden_states,
            hidden_states=encoder_states,
            attentions=all_attentions,
            global_attentions=all_global_attentions,
        )

mindnlp.transformers.models.led.modeling_led.LEDEncoder.forward(input_ids=None, attention_mask=None, global_attention_mask=None, head_mask=None, inputs_embeds=None, output_attentions=None, output_hidden_states=None, return_dict=None)

PARAMETER	DESCRIPTION
input_ids	Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [AutoTokenizer]. See [PreTrainedTokenizer.encode] and [PreTrainedTokenizer.__call__] for details. What are input IDs? TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length)` DEFAULT: None
attention_mask	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? TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional* DEFAULT: None
global_attention_mask	Mask to decide the attention given on each token, local attention or global attention for the encoder. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the <s> token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]: 0 for local attention (a sliding window attention), 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them). TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional* DEFAULT: None
head_mask	Mask to nullify selected heads of the attention modules. Mask values selected in [0, 1]: 1 indicates the head is not masked, 0 indicates the head is masked. TYPE: `mindspore.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional* DEFAULT: None
inputs_embeds	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. TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional* DEFAULT: None
output_attentions	Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail. TYPE: `bool`, *optional* DEFAULT: None
output_hidden_states	Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail. TYPE: `bool`, *optional* DEFAULT: None
return_dict	Whether or not to return a [~utils.ModelOutput] instead of a plain tuple. TYPE: `bool`, *optional* DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    input_ids=None,
    attention_mask=None,
    global_attention_mask=None,
    head_mask=None,
    inputs_embeds=None,
    output_attentions=None,
    output_hidden_states=None,
    return_dict=None,
):
    r"""
    Args:
        input_ids (`mindspore.Tensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
            provide it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *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)
        global_attention_mask (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to decide the attention given on each token, local attention or global attention for the encoder.
            Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is
            important for task-specific finetuning because it makes the model more flexible at representing the
            task. For example, for classification, the `<s>` token should be given global attention. For QA, all
            question tokens should also have global attention. Please refer to the [Longformer
            paper](https://arxiv.org/abs/2004.05150) for more details. Mask values selected in `[0, 1]`:

            - 0 for local attention (a sliding window attention),
            - 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
        head_mask (`mindspore.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        inputs_embeds (`mindspore.Tensor` of shape `(batch_size, sequence_length, 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.
    """
    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
    output_hidden_states = (
        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
    )
    return_dict = return_dict if return_dict is not None else self.config.use_return_dict

    # check input_ids and inputs_embeds
    if input_ids is not None and inputs_embeds is not None:
        raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
    elif input_ids is None and inputs_embeds is None:
        raise ValueError("You have to specify either input_ids or inputs_embeds")

    if inputs_embeds is None:
        inputs_embeds = self.embed_tokens(input_ids)

    # create default attention_mask
    if attention_mask is None:
        attention_mask = ops.ones(inputs_embeds.shape[:-1], dtype=mindspore.int64)

    # merge `global_attention_mask` and `attention_mask`
    if global_attention_mask is not None:
        attention_mask = self._merge_to_attention_mask(attention_mask, global_attention_mask)

    # pad input if necessary
    padding_len, input_ids, attention_mask, inputs_embeds = self._pad_to_window_size(
        input_ids=input_ids,
        attention_mask=attention_mask,
        inputs_embeds=inputs_embeds,
        pad_token_id=self.config.pad_token_id,
    )

    # retrieve input_shape
    if input_ids is not None:
        input_shape = input_ids.shape
        input_ids = input_ids.view(-1, input_shape[-1])
    elif inputs_embeds is not None:
        input_shape = inputs_embeds.shape[:-1]

    # convert attention_mask to float
    if attention_mask is not None:
        # [bsz, seq_len] -> [bsz, seq_len]; 1 -> 0.0; 0 -> "-inf"
        attention_mask = _prepare_4d_attention_mask_inverted(attention_mask, inputs_embeds.dtype)[:, 0, 0, :]

    # get masking tensors
    is_index_masked = attention_mask < 0
    is_index_global_attn = attention_mask > 0
    is_global_attn = is_index_global_attn.flatten().any().item()

    embed_pos = self.embed_positions(input_shape)

    hidden_states = inputs_embeds + embed_pos
    hidden_states = self.layernorm_embedding(hidden_states)
    hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)

    encoder_states = () if output_hidden_states else None
    all_attentions = () if output_attentions else None
    all_global_attentions = () if (output_attentions and is_global_attn) else None

    # check if head_mask has a correct number of layers specified if desired
    if head_mask is not None:
        if head_mask.shape[0] != len(self.layers):
            raise ValueError(
                f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
                f" {head_mask.shape[0]}."
            )
    for idx, encoder_layer in enumerate(self.layers):
        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)
        # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
        dropout_probability = ops.rand([])

        if self.training and (dropout_probability < self.layerdrop):  # skip the layer
            layer_outputs = (None, None, None)
        else:
            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    encoder_layer.__call__,
                    hidden_states,
                    attention_mask,
                    head_mask[idx] if head_mask is not None else None,
                    is_index_masked,
                    is_index_global_attn,
                    is_global_attn,
                    output_attentions,
                )
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                    is_index_masked=is_index_masked,
                    is_index_global_attn=is_index_global_attn,
                    is_global_attn=is_global_attn,
                    output_attentions=output_attentions,
                )
            hidden_states = layer_outputs[0]

        if output_attentions:
            # bzs x seq_len x num_attn_heads x (num_global_attn + attention_window_len + 1) => bzs x num_attn_heads x seq_len x (num_global_attn + attention_window_len + 1)
            all_attentions = all_attentions + (layer_outputs[1].swapaxes(1, 2),)

            if is_global_attn:
                # bzs x num_attn_heads x num_global_attn x seq_len => bzs x num_attn_heads x seq_len x num_global_attn
                all_global_attentions = all_global_attentions + (layer_outputs[2].swapaxes(2, 3),)

    if output_hidden_states:
        encoder_states = encoder_states + (hidden_states,)

    # undo padding
    if padding_len > 0:
        # unpad `hidden_states` because the calling function is expecting a length == input_ids.shape[1]
        hidden_states = hidden_states[:, :-padding_len]
        if output_hidden_states:
            encoder_states = tuple(state[:, :-padding_len] for state in encoder_states)

        if output_attentions:
            all_attentions = tuple(state[:, :, :-padding_len, :] for state in all_attentions)

    if not return_dict:
        return tuple(
            v for v in [hidden_states, encoder_states, all_attentions, all_global_attentions] if v is not None
        )
    return LEDEncoderBaseModelOutput(
        last_hidden_state=hidden_states,
        hidden_states=encoder_states,
        attentions=all_attentions,
        global_attentions=all_global_attentions,
    )

mindnlp.transformers.models.led.modeling_led.LEDEncoderAttention

Bases: Module

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDEncoderAttention(nn.Module):
    def __init__(self, config, layer_id):
        super().__init__()
        self.longformer_self_attn = LEDEncoderSelfAttention(config, layer_id=layer_id)
        self.output = nn.Linear(config.d_model, config.d_model)

    def forward(
        self,
        hidden_states: mindspore.Tensor,
        attention_mask: Optional[mindspore.Tensor] = None,
        layer_head_mask: Optional[mindspore.Tensor] = None,
        is_index_masked: Optional[mindspore.Tensor] = None,
        is_index_global_attn: Optional[mindspore.Tensor] = None,
        is_global_attn: Optional[bool] = None,
        output_attentions: bool = False,
    ) -> Tuple[mindspore.Tensor, Optional[mindspore.Tensor], Optional[Tuple[mindspore.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        self_outputs = self.longformer_self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            is_index_masked=is_index_masked,
            is_index_global_attn=is_index_global_attn,
            is_global_attn=is_global_attn,
            output_attentions=output_attentions,
        )

        attn_output = self.output(self_outputs[0])
        outputs = (attn_output,) + self_outputs[1:]

        return outputs

mindnlp.transformers.models.led.modeling_led.LEDEncoderAttention.forward(hidden_states, attention_mask=None, layer_head_mask=None, is_index_masked=None, is_index_global_attn=None, is_global_attn=None, output_attentions=False)

Input shape: Batch x Time x Channel

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    hidden_states: mindspore.Tensor,
    attention_mask: Optional[mindspore.Tensor] = None,
    layer_head_mask: Optional[mindspore.Tensor] = None,
    is_index_masked: Optional[mindspore.Tensor] = None,
    is_index_global_attn: Optional[mindspore.Tensor] = None,
    is_global_attn: Optional[bool] = None,
    output_attentions: bool = False,
) -> Tuple[mindspore.Tensor, Optional[mindspore.Tensor], Optional[Tuple[mindspore.Tensor]]]:
    """Input shape: Batch x Time x Channel"""

    self_outputs = self.longformer_self_attn(
        hidden_states=hidden_states,
        attention_mask=attention_mask,
        layer_head_mask=layer_head_mask,
        is_index_masked=is_index_masked,
        is_index_global_attn=is_index_global_attn,
        is_global_attn=is_global_attn,
        output_attentions=output_attentions,
    )

    attn_output = self.output(self_outputs[0])
    outputs = (attn_output,) + self_outputs[1:]

    return outputs

mindnlp.transformers.models.led.modeling_led.LEDEncoderBaseModelOutput dataclass

Bases: ModelOutput

Base class for LEDEncoder's outputs, with potential hidden states, local and global attentions.

PARAMETER	DESCRIPTION
last_hidden_state	Sequence of hidden-states at the output of the last layer of the model. TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`

Source code in mindnlp/transformers/models/led/modeling_led.py

@dataclass
# Copied from transformers.models.longformer.modeling_longformer.LongformerBaseModelOutput with Longformer->LEDEncoder
class LEDEncoderBaseModelOutput(ModelOutput):
    """
    Base class for LEDEncoder's outputs, with potential hidden states, local and global attentions.

    Args:
        last_hidden_state (`mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.
        hidden_states (`tuple(mindspore.Tensor)`, *optional*, returned when `output_hidden_states=True`
            is passed or when `config.output_hidden_states=True`):
            Tuple of `mindspore.Tensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True` is passed
            or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x +
            attention_window + 1)`, where `x` is the number of tokens with global attention mask.

            Local attentions weights after the attention softmax, used to compute the weighted average in the
            self-attention heads. Those are the attention weights from every token in the sequence to every token with
            global attention (first `x` values) and to every token in the attention window (remaining `attention_window
            + 1` values). Note that the first `x` values refer to tokens with fixed positions in the text, but the
            remaining `attention_window + 1` values refer to tokens with relative positions: the attention weight of a
            token to itself is located at index `x + attention_window / 2` and the `attention_window / 2` preceding
            (succeeding) values are the attention weights to the `attention_window / 2` preceding (succeeding) tokens.
            If the attention window contains a token with global attention, the attention weight at the corresponding
            index is set to 0; the value should be accessed from the first `x` attention weights. If a token has global
            attention, the attention weights to all other tokens in `attentions` is set to 0, the values should be
            accessed from `global_attentions`.
        global_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True` is passed
            or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x)`,
            where `x` is the number of tokens with global attention mask.

            Global attentions weights after the attention softmax, used to compute the weighted average in the
            self-attention heads. Those are the attention weights from every token with global attention to every token
            in the sequence.
    """

    last_hidden_state: mindspore.Tensor
    hidden_states: Optional[Tuple[mindspore.Tensor, ...]] = None
    attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    global_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None

mindnlp.transformers.models.led.modeling_led.LEDEncoderLayer

Bases: Module

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDEncoderLayer(nn.Module):
    def __init__(self, config: LEDConfig, layer_id: int):
        super().__init__()
        self.embed_dim = config.d_model
        self.self_attn = LEDEncoderAttention(config, layer_id)
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: mindspore.Tensor,
        attention_mask: mindspore.Tensor,
        layer_head_mask: mindspore.Tensor,
        is_index_masked=None,
        is_index_global_attn=None,
        is_global_attn=None,
        output_attentions=False,
    ):
        """
        Args:
            hidden_states (`mindspore.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)*
            attention_mask (`mindspore.Tensor`): attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            layer_head_mask (`mindspore.Tensor`): mask for attention heads in a given layer of size
                *(encoder_attention_heads,)*.
        """
        residual = hidden_states
        attn_outputs = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            is_index_masked=is_index_masked,
            is_index_global_attn=is_index_global_attn,
            is_global_attn=is_global_attn,
            output_attentions=output_attentions,
        )
        hidden_states = attn_outputs[0]
        hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        residual = hidden_states
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = ops.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.final_layer_norm(hidden_states)

        if hidden_states.dtype == mindspore.float16 and (
            ops.isinf(hidden_states).any() or ops.isnan(hidden_states).any()
        ):
            clamp_value = finfo(hidden_states.dtype, 'max') - 1000
            hidden_states = ops.clamp(hidden_states, min=-clamp_value, max=clamp_value)
        return (hidden_states,) + attn_outputs[1:]

mindnlp.transformers.models.led.modeling_led.LEDEncoderLayer.forward(hidden_states, attention_mask, layer_head_mask, is_index_masked=None, is_index_global_attn=None, is_global_attn=None, output_attentions=False)

PARAMETER	DESCRIPTION
hidden_states	input to the layer of shape (batch, seq_len, embed_dim) TYPE: `mindspore.Tensor`
attention_mask	attention mask of size (batch, 1, tgt_len, src_len) where padding elements are indicated by very large negative values. TYPE: `mindspore.Tensor`
layer_head_mask	mask for attention heads in a given layer of size (encoder_attention_heads,). TYPE: `mindspore.Tensor`

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    hidden_states: mindspore.Tensor,
    attention_mask: mindspore.Tensor,
    layer_head_mask: mindspore.Tensor,
    is_index_masked=None,
    is_index_global_attn=None,
    is_global_attn=None,
    output_attentions=False,
):
    """
    Args:
        hidden_states (`mindspore.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)*
        attention_mask (`mindspore.Tensor`): attention mask of size
            *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
        layer_head_mask (`mindspore.Tensor`): mask for attention heads in a given layer of size
            *(encoder_attention_heads,)*.
    """
    residual = hidden_states
    attn_outputs = self.self_attn(
        hidden_states=hidden_states,
        attention_mask=attention_mask,
        layer_head_mask=layer_head_mask,
        is_index_masked=is_index_masked,
        is_index_global_attn=is_index_global_attn,
        is_global_attn=is_global_attn,
        output_attentions=output_attentions,
    )
    hidden_states = attn_outputs[0]
    hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
    hidden_states = residual + hidden_states
    hidden_states = self.self_attn_layer_norm(hidden_states)

    residual = hidden_states
    hidden_states = self.activation_fn(self.fc1(hidden_states))
    hidden_states = ops.dropout(hidden_states, p=self.activation_dropout, training=self.training)
    hidden_states = self.fc2(hidden_states)
    hidden_states = ops.dropout(hidden_states, p=self.dropout, training=self.training)
    hidden_states = residual + hidden_states
    hidden_states = self.final_layer_norm(hidden_states)

    if hidden_states.dtype == mindspore.float16 and (
        ops.isinf(hidden_states).any() or ops.isnan(hidden_states).any()
    ):
        clamp_value = finfo(hidden_states.dtype, 'max') - 1000
        hidden_states = ops.clamp(hidden_states, min=-clamp_value, max=clamp_value)
    return (hidden_states,) + attn_outputs[1:]

mindnlp.transformers.models.led.modeling_led.LEDEncoderSelfAttention

Bases: Module

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDEncoderSelfAttention(nn.Module):
    def __init__(self, config, layer_id):
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0:
            raise ValueError(
                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
                f"heads ({config.num_attention_heads})"
            )
        self.num_heads = config.num_attention_heads
        self.head_dim = int(config.hidden_size / config.num_attention_heads)
        self.embed_dim = config.hidden_size

        self.query = nn.Linear(config.hidden_size, self.embed_dim)
        self.key = nn.Linear(config.hidden_size, self.embed_dim)
        self.value = nn.Linear(config.hidden_size, self.embed_dim)

        # separate projection layers for tokens with global attention
        self.query_global = nn.Linear(config.hidden_size, self.embed_dim)
        self.key_global = nn.Linear(config.hidden_size, self.embed_dim)
        self.value_global = nn.Linear(config.hidden_size, self.embed_dim)

        self.dropout = config.attention_probs_dropout_prob

        self.layer_id = layer_id
        attention_window = config.attention_window[self.layer_id]
        assert (
            attention_window % 2 == 0
        ), f"`attention_window` for layer {self.layer_id} has to be an even value. Given {attention_window}"
        assert (
            attention_window > 0
        ), f"`attention_window` for layer {self.layer_id} has to be positive. Given {attention_window}"

        self.one_sided_attn_window_size = attention_window // 2

        self.config = config

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        layer_head_mask=None,
        is_index_masked=None,
        is_index_global_attn=None,
        is_global_attn=None,
        output_attentions=False,
    ):
        """
        [`LEDEncoderSelfAttention`] expects *len(hidden_states)* to be multiple of *attention_window*. Padding to
        *attention_window* happens in [`LEDEncoderModel.forward`] to avoid redoing the padding on each layer.

        The *attention_mask* is changed in [`LEDEncoderModel.forward`] from 0, 1, 2 to:

        - -10000: no attention
        - 0: local attention
        - +10000: global attention
        """
        hidden_states = hidden_states.swapaxes(0, 1)

        # project hidden states
        query_vectors = self.query(hidden_states)
        key_vectors = self.key(hidden_states)
        value_vectors = self.value(hidden_states)

        seq_len, batch_size, embed_dim = hidden_states.shape
        assert (
            embed_dim == self.embed_dim
        ), f"hidden_states should have embed_dim = {self.embed_dim}, but has {embed_dim}"

        # normalize query
        query_vectors /= math.sqrt(self.head_dim)

        query_vectors = query_vectors.view(seq_len, batch_size, self.num_heads, self.head_dim).swapaxes(0, 1)
        key_vectors = key_vectors.view(seq_len, batch_size, self.num_heads, self.head_dim).swapaxes(0, 1)

        attn_scores = self._sliding_chunks_query_key_matmul(
            query_vectors, key_vectors, self.one_sided_attn_window_size
        )

        # values to pad for attention probs
        remove_from_windowed_attention_mask = (attention_mask != 0)[:, :, None, None]

        # cast to fp32/fp16 then replace 1's with -inf
        float_mask = remove_from_windowed_attention_mask.type_as(query_vectors).masked_fill(
            remove_from_windowed_attention_mask, finfo(query_vectors.dtype, 'min')
        )
        # diagonal mask with zeros everywhere and -inf inplace of padding
        diagonal_mask = self._sliding_chunks_query_key_matmul(
            float_mask.new_ones(size=float_mask.shape), float_mask, self.one_sided_attn_window_size
        )

        # pad local attention probs
        attn_scores += diagonal_mask

        assert list(attn_scores.shape) == [
            batch_size,
            seq_len,
            self.num_heads,
            self.one_sided_attn_window_size * 2 + 1,
        ], (
            f"local_attn_probs should be of size ({batch_size}, {seq_len}, {self.num_heads},"
            f" {self.one_sided_attn_window_size * 2 + 1}), but is of size {attn_scores.shape}"
        )

        # compute local attention probs from global attention keys and contact over window dim
        if is_global_attn:
            # compute global attn indices required through out forward fn
            (
                max_num_global_attn_indices,
                is_index_global_attn_nonzero,
                is_local_index_global_attn_nonzero,
                is_local_index_no_global_attn_nonzero,
            ) = self._get_global_attn_indices(is_index_global_attn)
            # calculate global attn probs from global key

            global_key_attn_scores = self._concat_with_global_key_attn_probs(
                query_vectors=query_vectors,
                key_vectors=key_vectors,
                max_num_global_attn_indices=max_num_global_attn_indices,
                is_index_global_attn_nonzero=is_index_global_attn_nonzero,
                is_local_index_global_attn_nonzero=is_local_index_global_attn_nonzero,
                is_local_index_no_global_attn_nonzero=is_local_index_no_global_attn_nonzero,
            )
            # concat to local_attn_probs
            # (batch_size, seq_len, num_heads, extra attention count + 2*window+1)
            attn_scores = ops.cat((global_key_attn_scores, attn_scores), axis=-1)

            # free memory
            del global_key_attn_scores

        attn_probs = ops.softmax(
            attn_scores, axis=-1, dtype=mindspore.float32
        )  # use fp32 for numerical stability

        if layer_head_mask is not None:
            assert layer_head_mask.shape == (
                self.num_heads,
            ), f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.shape}"
            attn_probs = layer_head_mask.view(1, 1, -1, 1) * attn_probs

        # softmax sometimes inserts NaN if all positions are masked, replace them with 0
        attn_probs = ops.masked_fill(attn_probs, is_index_masked[:, :, None, None], 0.0)
        attn_probs = attn_probs.type_as(attn_scores)

        # free memory
        del attn_scores

        # apply dropout
        attn_probs = ops.dropout(attn_probs, p=self.dropout, training=self.training)

        value_vectors = value_vectors.view(seq_len, batch_size, self.num_heads, self.head_dim).swapaxes(0, 1)

        # compute local attention output with global attention value and add
        if is_global_attn:
            # compute sum of global and local attn
            attn_output = self._compute_attn_output_with_global_indices(
                value_vectors=value_vectors,
                attn_probs=attn_probs,
                max_num_global_attn_indices=max_num_global_attn_indices,
                is_index_global_attn_nonzero=is_index_global_attn_nonzero,
                is_local_index_global_attn_nonzero=is_local_index_global_attn_nonzero,
            )
        else:
            # compute local attn only
            attn_output = self._sliding_chunks_matmul_attn_probs_value(
                attn_probs, value_vectors, self.one_sided_attn_window_size
            )

        assert attn_output.shape == (batch_size, seq_len, self.num_heads, self.head_dim), "Unexpected size"
        attn_output = attn_output.swapaxes(0, 1).reshape(seq_len, batch_size, embed_dim)

        # compute value for global attention and overwrite to attention output
        # TODO: remove the redundant computation
        if is_global_attn:
            global_attn_output, global_attn_probs = self._compute_global_attn_output_from_hidden(
                hidden_states=hidden_states,
                max_num_global_attn_indices=max_num_global_attn_indices,
                layer_head_mask=layer_head_mask,
                is_local_index_global_attn_nonzero=is_local_index_global_attn_nonzero,
                is_index_global_attn_nonzero=is_index_global_attn_nonzero,
                is_local_index_no_global_attn_nonzero=is_local_index_no_global_attn_nonzero,
                is_index_masked=is_index_masked,
            )

            # get only non zero global attn output
            nonzero_global_attn_output = global_attn_output[
                is_local_index_global_attn_nonzero[0], :, is_local_index_global_attn_nonzero[1]
            ]

            # overwrite values with global attention
            attn_output[is_index_global_attn_nonzero[::-1]] = nonzero_global_attn_output.view(
                len(is_local_index_global_attn_nonzero[0]), -1
            )
            # The attention weights for tokens with global attention are
            # just filler values, they were never used to compute the output.
            # Fill with 0 now, the correct values are in 'global_attn_probs'.
            attn_probs[is_index_global_attn_nonzero] = 0

        outputs = (attn_output.swapaxes(0, 1),)

        if output_attentions:
            outputs += (attn_probs,)

        return outputs + (global_attn_probs,) if (is_global_attn and output_attentions) else outputs

    @staticmethod
    def _pad_and_transpose_last_two_dims(hidden_states_padded, padding):
        """pads rows and then flips rows and columns"""
        hidden_states_padded = ops.pad(
            hidden_states_padded, padding
        )  # padding value is not important because it will be overwritten
        hidden_states_padded = hidden_states_padded.view(
            *hidden_states_padded.shape[:-2], hidden_states_padded.shape[-1], hidden_states_padded.shape[-2]
        )
        return hidden_states_padded

    @staticmethod
    def _pad_and_diagonalize(chunked_hidden_states):
        """
        shift every row 1 step right, converting columns into diagonals.

        Example:
            ```python
            >>> chunked_hidden_states: [
            ...     0.4983,
            ...     2.6918,
            ...     -0.0071,
            ...     1.0492,
            ...     -1.8348,
            ...     0.7672,
            ...     0.2986,
            ...     0.0285,
            ...     -0.7584,
            ...     0.4206,
            ...     -0.0405,
            ...     0.1599,
            ...     2.0514,
            ...     -1.1600,
            ...     0.5372,
            ...     0.2629,
            ... ]
            >>> window_overlap = num_rows = 4
            ```
             (pad & diagonalize) => [ 0.4983, 2.6918, -0.0071, 1.0492, 0.0000, 0.0000, 0.0000
               0.0000, -1.8348, 0.7672, 0.2986, 0.0285, 0.0000, 0.0000 0.0000, 0.0000, -0.7584, 0.4206,
               -0.0405, 0.1599, 0.0000 0.0000, 0.0000, 0.0000, 2.0514, -1.1600, 0.5372, 0.2629 ]
        """
        total_num_heads, num_chunks, window_overlap, hidden_dim = chunked_hidden_states.shape
        chunked_hidden_states = ops.pad(
            chunked_hidden_states, (0, window_overlap + 1)
        )  # total_num_heads x num_chunks x window_overlap x (hidden_dim+window_overlap+1). Padding value is not important because it'll be overwritten
        chunked_hidden_states = chunked_hidden_states.view(
            total_num_heads, num_chunks, -1
        )  # total_num_heads x num_chunks x window_overlap*window_overlap+window_overlap
        chunked_hidden_states = chunked_hidden_states[
            :, :, :-window_overlap
        ]  # total_num_heads x num_chunks x window_overlap*window_overlap
        chunked_hidden_states = chunked_hidden_states.view(
            total_num_heads, num_chunks, window_overlap, window_overlap + hidden_dim
        )
        chunked_hidden_states = chunked_hidden_states[:, :, :, :-1]
        return chunked_hidden_states

    @staticmethod
    def _chunk(hidden_states, window_overlap, onnx_export: bool = False):
        """convert into overlapping chunks. Chunk size = 2w, overlap size = w"""
        if not onnx_export:
            # non-overlapping chunks of size = 2w
            hidden_states = hidden_states.view(
                hidden_states.shape[0],
                scalar_div(hidden_states.shape[1], (window_overlap * 2), rounding_mode="trunc"),
                window_overlap * 2,
                hidden_states.shape[2],
            )
            # use `as_strided` to make the chunks overlap with an overlap size = window_overlap
            chunk_size = list(hidden_states.shape)
            chunk_size[1] = chunk_size[1] * 2 - 1

            chunk_stride = list(hidden_states.stride())
            chunk_stride[1] = chunk_stride[1] // 2
            return hidden_states.as_strided(size=chunk_size, stride=chunk_stride)

        # When exporting to ONNX, use this separate logic
        # have to use slow implementation since as_strided, unfold and 2d-tensor indexing aren't supported (yet) in ONNX export

        # TODO replace this with
        # > return hidden_states.unfold(dimension=1, size=window_overlap * 2, step=window_overlap).swapaxes(2, 3)
        # once `unfold` is supported
        # the case hidden_states.shape[1] == window_overlap * 2 can also simply return hidden_states.unsqueeze(1), but that's control flow

        chunk_size = [
            hidden_states.shape[0],
            scalar_div(hidden_states.shape[1], window_overlap, rounding_mode="trunc") - 1,
            window_overlap * 2,
            hidden_states.shape[2],
        ]

        overlapping_chunks = ops.zeros(chunk_size)
        for chunk in range(chunk_size[1]):
            overlapping_chunks[:, chunk, :, :] = hidden_states[
                :, chunk * window_overlap : chunk * window_overlap + 2 * window_overlap, :
            ]
        return overlapping_chunks

    @staticmethod
    def _mask_invalid_locations(input_tensor, affected_seq_len) -> mindspore.Tensor:
        beginning_mask_2d = input_tensor.new_ones((affected_seq_len, affected_seq_len + 1)).tril().flip(dims=[0])
        beginning_mask = beginning_mask_2d[None, :, None, :]
        ending_mask = beginning_mask.flip(dims=(1, 3))
        beginning_input = input_tensor[:, :affected_seq_len, :, : affected_seq_len + 1]
        beginning_mask = beginning_mask.expand(beginning_input.shape)
        input_tensor[:, :affected_seq_len, :, : affected_seq_len + 1] = ops.full_like(
            beginning_input, -float("inf")
        ).where(beginning_mask.bool(), beginning_input)
        ending_input = input_tensor[:, -affected_seq_len:, :, -(affected_seq_len + 1) :]
        ending_mask = ending_mask.expand(ending_input.shape)
        input_tensor[:, -affected_seq_len:, :, -(affected_seq_len + 1) :] = ops.full_like(
            ending_input, -float("inf")
        ).where(ending_mask.bool(), ending_input)

    def _sliding_chunks_query_key_matmul(self, query: mindspore.Tensor, key: mindspore.Tensor, window_overlap: int):
        """
        Matrix multiplication of query and key tensors using with a sliding window attention pattern. This
        implementation splits the input into overlapping chunks of size 2w (e.g. 512 for pretrained LEDEncoder) with an
        overlap of size window_overlap
        """
        batch_size, seq_len, num_heads, head_dim = query.shape
        assert (
            seq_len % (window_overlap * 2) == 0
        ), f"Sequence length should be multiple of {window_overlap * 2}. Given {seq_len}"
        assert query.shape == key.shape

        chunks_count = scalar_div(seq_len, window_overlap, rounding_mode="trunc") - 1

        # group batch_size and num_heads dimensions into one, then chunk seq_len into chunks of size window_overlap * 2
        query = query.swapaxes(1, 2).reshape(batch_size * num_heads, seq_len, head_dim)
        key = key.swapaxes(1, 2).reshape(batch_size * num_heads, seq_len, head_dim)

        query = self._chunk(query, window_overlap, getattr(self.config, "onnx_export", False))
        key = self._chunk(key, window_overlap, getattr(self.config, "onnx_export", False))

        # matrix multiplication
        # bcxd: batch_size * num_heads x chunks x 2window_overlap x head_dim
        # bcyd: batch_size * num_heads x chunks x 2window_overlap x head_dim
        # bcxy: batch_size * num_heads x chunks x 2window_overlap x 2window_overlap
        diagonal_chunked_attention_scores = ops.einsum("bcxd,bcyd->bcxy", (query, key))  # multiply

        # convert diagonals into columns
        diagonal_chunked_attention_scores = self._pad_and_transpose_last_two_dims(
            diagonal_chunked_attention_scores, padding=(0, 0, 0, 1)
        )

        # allocate space for the overall attention matrix where the chunks are combined. The last dimension
        # has (window_overlap * 2 + 1) columns. The first (window_overlap) columns are the window_overlap lower triangles (attention from a word to
        # window_overlap previous words). The following column is attention score from each word to itself, then
        # followed by window_overlap columns for the upper triangle.

        diagonal_attention_scores = diagonal_chunked_attention_scores.new_zeros(
            (batch_size * num_heads, chunks_count + 1, window_overlap, window_overlap * 2 + 1)
        )

        # copy parts from diagonal_chunked_attention_scores into the combined matrix of attentions
        # - copying the main diagonal and the upper triangle
        diagonal_attention_scores[:, :-1, :, window_overlap:] = diagonal_chunked_attention_scores[
            :, :, :window_overlap, : window_overlap + 1
        ]
        diagonal_attention_scores[:, -1, :, window_overlap:] = diagonal_chunked_attention_scores[
            :, -1, window_overlap:, : window_overlap + 1
        ]
        # - copying the lower triangle
        diagonal_attention_scores[:, 1:, :, :window_overlap] = diagonal_chunked_attention_scores[
            :, :, -(window_overlap + 1) : -1, window_overlap + 1 :
        ]

        diagonal_attention_scores[:, 0, 1:window_overlap, 1:window_overlap] = diagonal_chunked_attention_scores[
            :, 0, : window_overlap - 1, 1 - window_overlap :
        ]

        # separate batch_size and num_heads dimensions again
        diagonal_attention_scores = diagonal_attention_scores.view(
            batch_size, num_heads, seq_len, 2 * window_overlap + 1
        ).swapaxes(2, 1)

        self._mask_invalid_locations(diagonal_attention_scores, window_overlap)
        return diagonal_attention_scores

    def _sliding_chunks_matmul_attn_probs_value(
        self, attn_probs: mindspore.Tensor, value: mindspore.Tensor, window_overlap: int
    ):
        """
        Same as _sliding_chunks_query_key_matmul but for attn_probs and value tensors. Returned tensor will be of the
        same shape as `attn_probs`
        """
        batch_size, seq_len, num_heads, head_dim = value.shape

        assert seq_len % (window_overlap * 2) == 0
        assert attn_probs.shape[:3] == value.shape[:3]
        assert attn_probs.shape[3] == 2 * window_overlap + 1
        chunks_count = scalar_div(seq_len, window_overlap, rounding_mode="trunc") - 1
        # group batch_size and num_heads dimensions into one, then chunk seq_len into chunks of size 2 window overlap

        chunked_attn_probs = attn_probs.swapaxes(1, 2).reshape(
            batch_size * num_heads,
            scalar_div(seq_len, window_overlap, rounding_mode="trunc"),
            window_overlap,
            2 * window_overlap + 1,
        )

        # group batch_size and num_heads dimensions into one
        value = value.swapaxes(1, 2).reshape(batch_size * num_heads, seq_len, head_dim)

        # pad seq_len with w at the beginning of the sequence and another window overlap at the end
        padded_value = ops.pad(value, (0, 0, window_overlap, window_overlap), value=-1)

        # chunk padded_value into chunks of size 3 window overlap and an overlap of size window overlap
        chunked_value_size = (batch_size * num_heads, chunks_count + 1, 3 * window_overlap, head_dim)
        chunked_value_stride = padded_value.stride()
        chunked_value_stride = (
            chunked_value_stride[0],
            window_overlap * chunked_value_stride[1],
            chunked_value_stride[1],
            chunked_value_stride[2],
        )
        chunked_value = padded_value.as_strided(size=chunked_value_size, stride=chunked_value_stride)

        chunked_attn_probs = self._pad_and_diagonalize(chunked_attn_probs)

        context = ops.einsum("bcwd,bcdh->bcwh", (chunked_attn_probs, chunked_value))
        return context.view(batch_size, num_heads, seq_len, head_dim).swapaxes(1, 2)

    @staticmethod
    def _get_global_attn_indices(is_index_global_attn):
        """compute global attn indices required throughout forward pass"""
        # helper variable
        num_global_attn_indices = is_index_global_attn.long().sum(axis=1)

        # max number of global attn indices in batch
        max_num_global_attn_indices = num_global_attn_indices.max().item()

        # indices of global attn
        is_index_global_attn_nonzero = is_index_global_attn.nonzero(as_tuple=True)

        # helper variable
        is_local_index_global_attn = ops.arange(max_num_global_attn_indices) < num_global_attn_indices.unsqueeze(dim=-1)

        # location of the non-padding values within global attention indices
        is_local_index_global_attn_nonzero = is_local_index_global_attn.nonzero(as_tuple=True)

        # location of the padding values within global attention indices
        is_local_index_no_global_attn_nonzero = (is_local_index_global_attn == 0).nonzero(as_tuple=True)
        return (
            max_num_global_attn_indices,
            is_index_global_attn_nonzero,
            is_local_index_global_attn_nonzero,
            is_local_index_no_global_attn_nonzero,
        )

    def _concat_with_global_key_attn_probs(
        self,
        key_vectors,
        query_vectors,
        max_num_global_attn_indices,
        is_index_global_attn_nonzero,
        is_local_index_global_attn_nonzero,
        is_local_index_no_global_attn_nonzero,
    ):
        batch_size = key_vectors.shape[0]

        # create only global key vectors
        key_vectors_only_global = key_vectors.new_zeros(
            (batch_size, max_num_global_attn_indices, self.num_heads, self.head_dim)
        )

        key_vectors_only_global[is_local_index_global_attn_nonzero] = key_vectors[is_index_global_attn_nonzero]

        # (batch_size, seq_len, num_heads, max_num_global_attn_indices)
        attn_probs_from_global_key = ops.einsum("blhd,bshd->blhs", (query_vectors, key_vectors_only_global))

        # need to transpose since ONNX export only supports consecutive indexing: https://pytorch.org/docs/stable/onnx.html#writes-sets
        attn_probs_from_global_key = attn_probs_from_global_key.swapaxes(1, 3)
        attn_probs_from_global_key[
            is_local_index_no_global_attn_nonzero[0], is_local_index_no_global_attn_nonzero[1], :, :
        ] = finfo(attn_probs_from_global_key.dtype, 'min')
        attn_probs_from_global_key = attn_probs_from_global_key.swapaxes(1, 3)

        return attn_probs_from_global_key

    def _compute_attn_output_with_global_indices(
        self,
        value_vectors,
        attn_probs,
        max_num_global_attn_indices,
        is_index_global_attn_nonzero,
        is_local_index_global_attn_nonzero,
    ):
        batch_size = attn_probs.shape[0]

        # cut local attn probs to global only
        attn_probs_only_global = attn_probs.narrow(-1, 0, max_num_global_attn_indices)
        # get value vectors for global only
        value_vectors_only_global = value_vectors.new_zeros(
            (batch_size, max_num_global_attn_indices, self.num_heads, self.head_dim)
        )
        value_vectors_only_global[is_local_index_global_attn_nonzero] = value_vectors[is_index_global_attn_nonzero]

        # use `matmul` because `einsum` crashes sometimes with fp16
        # attn = torch.einsum('blhs,bshd->blhd', (selected_attn_probs, selected_v))
        # compute attn output only global
        attn_output_only_global = ops.matmul(
            attn_probs_only_global.swapaxes(1, 2).copy(), value_vectors_only_global.swapaxes(1, 2).copy()
        ).swapaxes(1, 2)

        # reshape attn probs
        attn_probs_without_global = attn_probs.narrow(
            -1, max_num_global_attn_indices, attn_probs.shape[-1] - max_num_global_attn_indices
        )

        # compute attn output with global
        attn_output_without_global = self._sliding_chunks_matmul_attn_probs_value(
            attn_probs_without_global, value_vectors, self.one_sided_attn_window_size
        )
        return attn_output_only_global + attn_output_without_global

    def _compute_global_attn_output_from_hidden(
        self,
        hidden_states,
        max_num_global_attn_indices,
        layer_head_mask,
        is_local_index_global_attn_nonzero,
        is_index_global_attn_nonzero,
        is_local_index_no_global_attn_nonzero,
        is_index_masked,
    ):
        seq_len, batch_size = hidden_states.shape[:2]

        # prepare global hidden states
        global_attn_hidden_states = hidden_states.new_zeros((max_num_global_attn_indices, batch_size, self.embed_dim))
        global_attn_hidden_states[is_local_index_global_attn_nonzero[::-1]] = hidden_states[
            is_index_global_attn_nonzero[::-1]
        ]

        # global key, query, value
        global_query_vectors_only_global = self.query_global(global_attn_hidden_states)
        global_key_vectors = self.key_global(hidden_states)
        global_value_vectors = self.value_global(hidden_states)

        # normalize
        global_query_vectors_only_global /= math.sqrt(self.head_dim)

        # reshape
        global_query_vectors_only_global = (
            global_query_vectors_only_global
            .view(max_num_global_attn_indices, batch_size * self.num_heads, self.head_dim)
            .swapaxes(0, 1)
        )  # (batch_size * self.num_heads, max_num_global_attn_indices, head_dim)
        global_key_vectors = (
            global_key_vectors.view(-1, batch_size * self.num_heads, self.head_dim).swapaxes(0, 1)
        )  # batch_size * self.num_heads, seq_len, head_dim)
        global_value_vectors = (
            global_value_vectors.view(-1, batch_size * self.num_heads, self.head_dim).swapaxes(0, 1)
        )  # batch_size * self.num_heads, seq_len, head_dim)

        # compute attn scores
        global_attn_scores = ops.bmm(global_query_vectors_only_global, global_key_vectors.swapaxes(1, 2))

        assert list(global_attn_scores.shape) == [
            batch_size * self.num_heads,
            max_num_global_attn_indices,
            seq_len,
        ], (
            "global_attn_scores have the wrong size. Size should be"
            f" {(batch_size * self.num_heads, max_num_global_attn_indices, seq_len)}, but is"
            f" {global_attn_scores.shape}."
        )

        global_attn_scores = global_attn_scores.view(batch_size, self.num_heads, max_num_global_attn_indices, seq_len)

        # need to transpose since ONNX export only supports consecutive indexing: https://pytorch.org/docs/stable/onnx.html#writes-sets
        global_attn_scores = global_attn_scores.swapaxes(1, 2)
        global_attn_scores[
            is_local_index_no_global_attn_nonzero[0], is_local_index_no_global_attn_nonzero[1], :, :
        ] = finfo(global_attn_scores.dtype, 'min')
        global_attn_scores = global_attn_scores.swapaxes(1, 2)

        global_attn_scores = global_attn_scores.masked_fill(
            is_index_masked[:, None, None, :],
            finfo(global_attn_scores.dtype, 'min'),
        )

        global_attn_scores = global_attn_scores.view(batch_size * self.num_heads, max_num_global_attn_indices, seq_len)

        # compute global attn probs
        global_attn_probs_float = ops.softmax(
            global_attn_scores, axis=-1, dtype=mindspore.float32
        )  # use fp32 for numerical stability

        # apply layer head masking
        if layer_head_mask is not None:
            assert layer_head_mask.shape == (
                self.num_heads,
            ), f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.shape}"
            global_attn_probs_float = layer_head_mask.view(1, -1, 1, 1) * global_attn_probs_float.view(
                batch_size, self.num_heads, max_num_global_attn_indices, seq_len
            )
            global_attn_probs_float = global_attn_probs_float.view(
                batch_size * self.num_heads, max_num_global_attn_indices, seq_len
            )

        global_attn_probs = ops.dropout(
            global_attn_probs_float.type_as(global_attn_scores), p=self.dropout, training=self.training
        )

        # global attn output
        global_attn_output = ops.bmm(global_attn_probs, global_value_vectors)

        assert list(global_attn_output.shape) == [
            batch_size * self.num_heads,
            max_num_global_attn_indices,
            self.head_dim,
        ], (
            "global_attn_output tensor has the wrong size. Size should be"
            f" {(batch_size * self.num_heads, max_num_global_attn_indices, self.head_dim)}, but is"
            f" {global_attn_output.shape}."
        )

        global_attn_probs = global_attn_probs.view(batch_size, self.num_heads, max_num_global_attn_indices, seq_len)
        global_attn_output = global_attn_output.view(
            batch_size, self.num_heads, max_num_global_attn_indices, self.head_dim
        )
        return global_attn_output, global_attn_probs

mindnlp.transformers.models.led.modeling_led.LEDEncoderSelfAttention.forward(hidden_states, attention_mask=None, layer_head_mask=None, is_index_masked=None, is_index_global_attn=None, is_global_attn=None, output_attentions=False)

[LEDEncoderSelfAttention] expects len(hidden_states) to be multiple of attention_window. Padding to attention_window happens in [LEDEncoderModel.forward] to avoid redoing the padding on each layer.

The attention_mask is changed in [LEDEncoderModel.forward] from 0, 1, 2 to:

• -10000: no attention
• 0: local attention
• +10000: global attention

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    hidden_states,
    attention_mask=None,
    layer_head_mask=None,
    is_index_masked=None,
    is_index_global_attn=None,
    is_global_attn=None,
    output_attentions=False,
):
    """
    [`LEDEncoderSelfAttention`] expects *len(hidden_states)* to be multiple of *attention_window*. Padding to
    *attention_window* happens in [`LEDEncoderModel.forward`] to avoid redoing the padding on each layer.

    The *attention_mask* is changed in [`LEDEncoderModel.forward`] from 0, 1, 2 to:

    - -10000: no attention
    - 0: local attention
    - +10000: global attention
    """
    hidden_states = hidden_states.swapaxes(0, 1)

    # project hidden states
    query_vectors = self.query(hidden_states)
    key_vectors = self.key(hidden_states)
    value_vectors = self.value(hidden_states)

    seq_len, batch_size, embed_dim = hidden_states.shape
    assert (
        embed_dim == self.embed_dim
    ), f"hidden_states should have embed_dim = {self.embed_dim}, but has {embed_dim}"

    # normalize query
    query_vectors /= math.sqrt(self.head_dim)

    query_vectors = query_vectors.view(seq_len, batch_size, self.num_heads, self.head_dim).swapaxes(0, 1)
    key_vectors = key_vectors.view(seq_len, batch_size, self.num_heads, self.head_dim).swapaxes(0, 1)

    attn_scores = self._sliding_chunks_query_key_matmul(
        query_vectors, key_vectors, self.one_sided_attn_window_size
    )

    # values to pad for attention probs
    remove_from_windowed_attention_mask = (attention_mask != 0)[:, :, None, None]

    # cast to fp32/fp16 then replace 1's with -inf
    float_mask = remove_from_windowed_attention_mask.type_as(query_vectors).masked_fill(
        remove_from_windowed_attention_mask, finfo(query_vectors.dtype, 'min')
    )
    # diagonal mask with zeros everywhere and -inf inplace of padding
    diagonal_mask = self._sliding_chunks_query_key_matmul(
        float_mask.new_ones(size=float_mask.shape), float_mask, self.one_sided_attn_window_size
    )

    # pad local attention probs
    attn_scores += diagonal_mask

    assert list(attn_scores.shape) == [
        batch_size,
        seq_len,
        self.num_heads,
        self.one_sided_attn_window_size * 2 + 1,
    ], (
        f"local_attn_probs should be of size ({batch_size}, {seq_len}, {self.num_heads},"
        f" {self.one_sided_attn_window_size * 2 + 1}), but is of size {attn_scores.shape}"
    )

    # compute local attention probs from global attention keys and contact over window dim
    if is_global_attn:
        # compute global attn indices required through out forward fn
        (
            max_num_global_attn_indices,
            is_index_global_attn_nonzero,
            is_local_index_global_attn_nonzero,
            is_local_index_no_global_attn_nonzero,
        ) = self._get_global_attn_indices(is_index_global_attn)
        # calculate global attn probs from global key

        global_key_attn_scores = self._concat_with_global_key_attn_probs(
            query_vectors=query_vectors,
            key_vectors=key_vectors,
            max_num_global_attn_indices=max_num_global_attn_indices,
            is_index_global_attn_nonzero=is_index_global_attn_nonzero,
            is_local_index_global_attn_nonzero=is_local_index_global_attn_nonzero,
            is_local_index_no_global_attn_nonzero=is_local_index_no_global_attn_nonzero,
        )
        # concat to local_attn_probs
        # (batch_size, seq_len, num_heads, extra attention count + 2*window+1)
        attn_scores = ops.cat((global_key_attn_scores, attn_scores), axis=-1)

        # free memory
        del global_key_attn_scores

    attn_probs = ops.softmax(
        attn_scores, axis=-1, dtype=mindspore.float32
    )  # use fp32 for numerical stability

    if layer_head_mask is not None:
        assert layer_head_mask.shape == (
            self.num_heads,
        ), f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.shape}"
        attn_probs = layer_head_mask.view(1, 1, -1, 1) * attn_probs

    # softmax sometimes inserts NaN if all positions are masked, replace them with 0
    attn_probs = ops.masked_fill(attn_probs, is_index_masked[:, :, None, None], 0.0)
    attn_probs = attn_probs.type_as(attn_scores)

    # free memory
    del attn_scores

    # apply dropout
    attn_probs = ops.dropout(attn_probs, p=self.dropout, training=self.training)

    value_vectors = value_vectors.view(seq_len, batch_size, self.num_heads, self.head_dim).swapaxes(0, 1)

    # compute local attention output with global attention value and add
    if is_global_attn:
        # compute sum of global and local attn
        attn_output = self._compute_attn_output_with_global_indices(
            value_vectors=value_vectors,
            attn_probs=attn_probs,
            max_num_global_attn_indices=max_num_global_attn_indices,
            is_index_global_attn_nonzero=is_index_global_attn_nonzero,
            is_local_index_global_attn_nonzero=is_local_index_global_attn_nonzero,
        )
    else:
        # compute local attn only
        attn_output = self._sliding_chunks_matmul_attn_probs_value(
            attn_probs, value_vectors, self.one_sided_attn_window_size
        )

    assert attn_output.shape == (batch_size, seq_len, self.num_heads, self.head_dim), "Unexpected size"
    attn_output = attn_output.swapaxes(0, 1).reshape(seq_len, batch_size, embed_dim)

    # compute value for global attention and overwrite to attention output
    # TODO: remove the redundant computation
    if is_global_attn:
        global_attn_output, global_attn_probs = self._compute_global_attn_output_from_hidden(
            hidden_states=hidden_states,
            max_num_global_attn_indices=max_num_global_attn_indices,
            layer_head_mask=layer_head_mask,
            is_local_index_global_attn_nonzero=is_local_index_global_attn_nonzero,
            is_index_global_attn_nonzero=is_index_global_attn_nonzero,
            is_local_index_no_global_attn_nonzero=is_local_index_no_global_attn_nonzero,
            is_index_masked=is_index_masked,
        )

        # get only non zero global attn output
        nonzero_global_attn_output = global_attn_output[
            is_local_index_global_attn_nonzero[0], :, is_local_index_global_attn_nonzero[1]
        ]

        # overwrite values with global attention
        attn_output[is_index_global_attn_nonzero[::-1]] = nonzero_global_attn_output.view(
            len(is_local_index_global_attn_nonzero[0]), -1
        )
        # The attention weights for tokens with global attention are
        # just filler values, they were never used to compute the output.
        # Fill with 0 now, the correct values are in 'global_attn_probs'.
        attn_probs[is_index_global_attn_nonzero] = 0

    outputs = (attn_output.swapaxes(0, 1),)

    if output_attentions:
        outputs += (attn_probs,)

    return outputs + (global_attn_probs,) if (is_global_attn and output_attentions) else outputs

mindnlp.transformers.models.led.modeling_led.LEDForConditionalGeneration

Bases: LEDPreTrainedModel

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDForConditionalGeneration(LEDPreTrainedModel):
    base_model_prefix = "led"
    # _keys_to_ignore_on_load_missing = ["final_logits_bias"]
    _tied_weights_keys = ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "lm_head.weight"]

    def __init__(self, config: LEDConfig):
        super().__init__(config)
        self.led = LEDModel(config)
        self.final_logits_bias = ops.zeros((1, self.led.shared.vocab_size))
        self.lm_head = nn.Linear(config.d_model, self.led.shared.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

    def get_encoder(self):
        return self.led.get_encoder()

    def get_decoder(self):
        return self.led.get_decoder()

    def resize_token_embeddings(self, new_num_tokens: int, pad_to_multiple_of: Optional[int] = None) -> nn.Embedding:
        new_embeddings = super().resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
        self._resize_final_logits_bias(new_embeddings.weight.shape[0])
        return new_embeddings

    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
        old_num_tokens = self.final_logits_bias.shape[-1]
        if new_num_tokens <= old_num_tokens:
            new_bias = self.final_logits_bias[:, :new_num_tokens]
        else:
            extra_bias = ops.zeros((1, new_num_tokens - old_num_tokens))
            new_bias = ops.cat([self.final_logits_bias, extra_bias], axis=1)
        self.final_logits_bias = new_bias

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def forward(
        self,
        input_ids: Optional[mindspore.Tensor] = None,
        attention_mask: Optional[mindspore.Tensor] = None,
        decoder_input_ids: Optional[mindspore.Tensor] = None,
        decoder_attention_mask: Optional[mindspore.Tensor] = None,
        head_mask: Optional[mindspore.Tensor] = None,
        decoder_head_mask: Optional[mindspore.Tensor] = None,
        cross_attn_head_mask: Optional[mindspore.Tensor] = None,
        encoder_outputs: Optional[Tuple[Tuple[mindspore.Tensor]]] = None,
        global_attention_mask: Optional[mindspore.Tensor] = None,
        past_key_values: Optional[Tuple[Tuple[mindspore.Tensor]]] = None,
        inputs_embeds: Optional[mindspore.Tensor] = None,
        decoder_inputs_embeds: Optional[mindspore.Tensor] = None,
        labels: Optional[mindspore.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[mindspore.Tensor], LEDSeq2SeqLMOutput]:
        r"""
        Args:
            labels (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Returns:
            `Union[Tuple[mindspore.Tensor], LEDSeq2SeqLMOutput]`

        Example:
            ```python
            >>> from transformers import AutoTokenizer, LEDForConditionalGeneration
            ...
            >>> tokenizer = AutoTokenizer.from_pretrained("allenai/led-base-16384")
            >>> TXT = "My friends are <mask> but they eat too many carbs."
            ...
            >>> model = LEDForConditionalGeneration.from_pretrained("allenai/led-base-16384")
            >>> input_ids = tokenizer([TXT], return_tensors="ms")["input_ids"]
            ...
            >>> prediction = model.generate(input_ids)[0]
            >>> print(tokenizer.decode(prediction, skip_special_tokens=True))
            ```
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if labels is not None:
            if use_cache:
                logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.")
            use_cache = False
            if decoder_input_ids is None and decoder_inputs_embeds is None:
                decoder_input_ids = shift_tokens_right(
                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
                )

        outputs = self.led(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            encoder_outputs=encoder_outputs,
            global_attention_mask=global_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias

        masked_lm_loss = None
        if labels is not None:
            masked_lm_loss = ops.cross_entropy(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))

        if not return_dict:
            output = (lm_logits,) + outputs[1:]
            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output

        return LEDSeq2SeqLMOutput(
            loss=masked_lm_loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
            encoder_global_attentions=outputs.encoder_global_attentions,
        )

    def prepare_inputs_for_generation(
        self,
        decoder_input_ids,
        past_key_values=None,
        attention_mask=None,
        global_attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        use_cache=None,
        encoder_outputs=None,
        **kwargs,
    ):
        # cut decoder_input_ids if past is used
        if past_key_values is not None:
            decoder_input_ids = decoder_input_ids[:, -1:]

        return {
            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
            "encoder_outputs": encoder_outputs,
            "past_key_values": past_key_values,
            "decoder_input_ids": decoder_input_ids,
            "attention_mask": attention_mask,
            "global_attention_mask": global_attention_mask,
            "head_mask": head_mask,
            "decoder_head_mask": decoder_head_mask,
            "cross_attn_head_mask": cross_attn_head_mask,
            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
        }

    def prepare_decoder_input_ids_from_labels(self, labels: mindspore.Tensor):
        return shift_tokens_right(labels, self.config.pad_token_id, self.config.decoder_start_token_id)

    @staticmethod
    def _reorder_cache(past_key_values, beam_idx):
        reordered_past = ()
        for layer_past in past_key_values:
            # cached cross_attention states don't have to be reordered -> they are always the same
            reordered_past += (
                tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2])
                + layer_past[2:],
            )
        return reordered_past

mindnlp.transformers.models.led.modeling_led.LEDForConditionalGeneration.forward(input_ids=None, attention_mask=None, decoder_input_ids=None, decoder_attention_mask=None, head_mask=None, decoder_head_mask=None, cross_attn_head_mask=None, encoder_outputs=None, global_attention_mask=None, past_key_values=None, inputs_embeds=None, decoder_inputs_embeds=None, labels=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None)

PARAMETER	DESCRIPTION
labels	Labels for computing the masked language modeling loss. Indices should either be in [0, ..., config.vocab_size] or -100 (see input_ids docstring). Tokens with indices set to -100 are ignored (masked), the loss is only computed for the tokens with labels in [0, ..., config.vocab_size]. TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional* DEFAULT: None

RETURNS	DESCRIPTION
Union[Tuple[Tensor], LEDSeq2SeqLMOutput]	Union[Tuple[mindspore.Tensor], LEDSeq2SeqLMOutput]

Example

>>> from transformers import AutoTokenizer, LEDForConditionalGeneration
...
>>> tokenizer = AutoTokenizer.from_pretrained("allenai/led-base-16384")
>>> TXT = "My friends are <mask> but they eat too many carbs."
...
>>> model = LEDForConditionalGeneration.from_pretrained("allenai/led-base-16384")
>>> input_ids = tokenizer([TXT], return_tensors="ms")["input_ids"]
...
>>> prediction = model.generate(input_ids)[0]
>>> print(tokenizer.decode(prediction, skip_special_tokens=True))

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    input_ids: Optional[mindspore.Tensor] = None,
    attention_mask: Optional[mindspore.Tensor] = None,
    decoder_input_ids: Optional[mindspore.Tensor] = None,
    decoder_attention_mask: Optional[mindspore.Tensor] = None,
    head_mask: Optional[mindspore.Tensor] = None,
    decoder_head_mask: Optional[mindspore.Tensor] = None,
    cross_attn_head_mask: Optional[mindspore.Tensor] = None,
    encoder_outputs: Optional[Tuple[Tuple[mindspore.Tensor]]] = None,
    global_attention_mask: Optional[mindspore.Tensor] = None,
    past_key_values: Optional[Tuple[Tuple[mindspore.Tensor]]] = None,
    inputs_embeds: Optional[mindspore.Tensor] = None,
    decoder_inputs_embeds: Optional[mindspore.Tensor] = None,
    labels: Optional[mindspore.Tensor] = None,
    use_cache: Optional[bool] = None,
    output_attentions: Optional[bool] = None,
    output_hidden_states: Optional[bool] = None,
    return_dict: Optional[bool] = None,
) -> Union[Tuple[mindspore.Tensor], LEDSeq2SeqLMOutput]:
    r"""
    Args:
        labels (`mindspore.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

    Returns:
        `Union[Tuple[mindspore.Tensor], LEDSeq2SeqLMOutput]`

    Example:
        ```python
        >>> from transformers import AutoTokenizer, LEDForConditionalGeneration
        ...
        >>> tokenizer = AutoTokenizer.from_pretrained("allenai/led-base-16384")
        >>> TXT = "My friends are <mask> but they eat too many carbs."
        ...
        >>> model = LEDForConditionalGeneration.from_pretrained("allenai/led-base-16384")
        >>> input_ids = tokenizer([TXT], return_tensors="ms")["input_ids"]
        ...
        >>> prediction = model.generate(input_ids)[0]
        >>> print(tokenizer.decode(prediction, skip_special_tokens=True))
        ```
    """
    return_dict = return_dict if return_dict is not None else self.config.use_return_dict

    if labels is not None:
        if use_cache:
            logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.")
        use_cache = False
        if decoder_input_ids is None and decoder_inputs_embeds is None:
            decoder_input_ids = shift_tokens_right(
                labels, self.config.pad_token_id, self.config.decoder_start_token_id
            )

    outputs = self.led(
        input_ids,
        attention_mask=attention_mask,
        decoder_input_ids=decoder_input_ids,
        decoder_attention_mask=decoder_attention_mask,
        encoder_outputs=encoder_outputs,
        global_attention_mask=global_attention_mask,
        head_mask=head_mask,
        decoder_head_mask=decoder_head_mask,
        cross_attn_head_mask=cross_attn_head_mask,
        past_key_values=past_key_values,
        inputs_embeds=inputs_embeds,
        decoder_inputs_embeds=decoder_inputs_embeds,
        use_cache=use_cache,
        output_attentions=output_attentions,
        output_hidden_states=output_hidden_states,
        return_dict=return_dict,
    )
    lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias

    masked_lm_loss = None
    if labels is not None:
        masked_lm_loss = ops.cross_entropy(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))

    if not return_dict:
        output = (lm_logits,) + outputs[1:]
        return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output

    return LEDSeq2SeqLMOutput(
        loss=masked_lm_loss,
        logits=lm_logits,
        past_key_values=outputs.past_key_values,
        decoder_hidden_states=outputs.decoder_hidden_states,
        decoder_attentions=outputs.decoder_attentions,
        cross_attentions=outputs.cross_attentions,
        encoder_last_hidden_state=outputs.encoder_last_hidden_state,
        encoder_hidden_states=outputs.encoder_hidden_states,
        encoder_attentions=outputs.encoder_attentions,
        encoder_global_attentions=outputs.encoder_global_attentions,
    )

mindnlp.transformers.models.led.modeling_led.LEDForQuestionAnswering

Bases: LEDPreTrainedModel

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDForQuestionAnswering(LEDPreTrainedModel):
    _tied_weights_keys = ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight"]

    def __init__(self, config):
        super().__init__(config)

        config.num_labels = 2
        self.num_labels = config.num_labels

        self.led = LEDModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Optional[mindspore.Tensor] = None,
        attention_mask: Optional[mindspore.Tensor] = None,
        decoder_input_ids: Optional[mindspore.Tensor] = None,
        decoder_attention_mask: Optional[mindspore.Tensor] = None,
        head_mask: Optional[mindspore.Tensor] = None,
        decoder_head_mask: Optional[mindspore.Tensor] = None,
        cross_attn_head_mask: Optional[mindspore.Tensor] = None,
        encoder_outputs: Optional[Tuple[Tuple[mindspore.Tensor]]] = None,
        global_attention_mask: Optional[mindspore.Tensor] = None,
        start_positions: Optional[mindspore.Tensor] = None,
        end_positions: Optional[mindspore.Tensor] = None,
        inputs_embeds: Optional[mindspore.Tensor] = None,
        decoder_inputs_embeds: Optional[mindspore.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[mindspore.Tensor], LEDSeq2SeqQuestionAnsweringModelOutput]:
        r"""
        Args:
            start_positions (`mindspore.Tensor` 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 (`mindspore.Tensor` 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
        if start_positions is not None and end_positions is not None:
            use_cache = False

        outputs = self.led(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            global_attention_mask=global_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            encoder_outputs=encoder_outputs,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            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, axis=-1)
        start_logits = start_logits.squeeze(-1)
        end_logits = end_logits.squeeze(-1)

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.shape) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.shape) > 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.shape[1]
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)

            start_loss = ops.cross_entropy(start_logits, start_positions, ignore_index=ignored_index)
            end_loss = ops.cross_entropy(end_logits, end_positions, ignore_index=ignored_index)
            total_loss = (start_loss + end_loss) / 2

        if not return_dict:
            output = (
                start_logits,
                end_logits,
            ) + outputs[1:]
            return ((total_loss,) + output) if total_loss is not None else output

        return LEDSeq2SeqQuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
            encoder_global_attentions=outputs.encoder_global_attentions,
        )

mindnlp.transformers.models.led.modeling_led.LEDForQuestionAnswering.forward(input_ids=None, attention_mask=None, decoder_input_ids=None, decoder_attention_mask=None, head_mask=None, decoder_head_mask=None, cross_attn_head_mask=None, encoder_outputs=None, global_attention_mask=None, start_positions=None, end_positions=None, inputs_embeds=None, decoder_inputs_embeds=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None)

PARAMETER	DESCRIPTION
start_positions	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. TYPE: `mindspore.Tensor` of shape `(batch_size,)`, *optional* DEFAULT: None
end_positions	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. TYPE: `mindspore.Tensor` of shape `(batch_size,)`, *optional* DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    input_ids: Optional[mindspore.Tensor] = None,
    attention_mask: Optional[mindspore.Tensor] = None,
    decoder_input_ids: Optional[mindspore.Tensor] = None,
    decoder_attention_mask: Optional[mindspore.Tensor] = None,
    head_mask: Optional[mindspore.Tensor] = None,
    decoder_head_mask: Optional[mindspore.Tensor] = None,
    cross_attn_head_mask: Optional[mindspore.Tensor] = None,
    encoder_outputs: Optional[Tuple[Tuple[mindspore.Tensor]]] = None,
    global_attention_mask: Optional[mindspore.Tensor] = None,
    start_positions: Optional[mindspore.Tensor] = None,
    end_positions: Optional[mindspore.Tensor] = None,
    inputs_embeds: Optional[mindspore.Tensor] = None,
    decoder_inputs_embeds: Optional[mindspore.Tensor] = None,
    use_cache: Optional[bool] = None,
    output_attentions: Optional[bool] = None,
    output_hidden_states: Optional[bool] = None,
    return_dict: Optional[bool] = None,
) -> Union[Tuple[mindspore.Tensor], LEDSeq2SeqQuestionAnsweringModelOutput]:
    r"""
    Args:
        start_positions (`mindspore.Tensor` 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 (`mindspore.Tensor` 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
    if start_positions is not None and end_positions is not None:
        use_cache = False

    outputs = self.led(
        input_ids,
        attention_mask=attention_mask,
        decoder_input_ids=decoder_input_ids,
        decoder_attention_mask=decoder_attention_mask,
        global_attention_mask=global_attention_mask,
        head_mask=head_mask,
        decoder_head_mask=decoder_head_mask,
        cross_attn_head_mask=cross_attn_head_mask,
        encoder_outputs=encoder_outputs,
        inputs_embeds=inputs_embeds,
        decoder_inputs_embeds=decoder_inputs_embeds,
        use_cache=use_cache,
        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, axis=-1)
    start_logits = start_logits.squeeze(-1)
    end_logits = end_logits.squeeze(-1)

    total_loss = None
    if start_positions is not None and end_positions is not None:
        # If we are on multi-GPU, split add a dimension
        if len(start_positions.shape) > 1:
            start_positions = start_positions.squeeze(-1)
        if len(end_positions.shape) > 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.shape[1]
        start_positions = start_positions.clamp(0, ignored_index)
        end_positions = end_positions.clamp(0, ignored_index)

        start_loss = ops.cross_entropy(start_logits, start_positions, ignore_index=ignored_index)
        end_loss = ops.cross_entropy(end_logits, end_positions, ignore_index=ignored_index)
        total_loss = (start_loss + end_loss) / 2

    if not return_dict:
        output = (
            start_logits,
            end_logits,
        ) + outputs[1:]
        return ((total_loss,) + output) if total_loss is not None else output

    return LEDSeq2SeqQuestionAnsweringModelOutput(
        loss=total_loss,
        start_logits=start_logits,
        end_logits=end_logits,
        past_key_values=outputs.past_key_values,
        decoder_hidden_states=outputs.decoder_hidden_states,
        decoder_attentions=outputs.decoder_attentions,
        cross_attentions=outputs.cross_attentions,
        encoder_last_hidden_state=outputs.encoder_last_hidden_state,
        encoder_hidden_states=outputs.encoder_hidden_states,
        encoder_attentions=outputs.encoder_attentions,
        encoder_global_attentions=outputs.encoder_global_attentions,
    )

mindnlp.transformers.models.led.modeling_led.LEDForSequenceClassification

Bases: LEDPreTrainedModel

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDForSequenceClassification(LEDPreTrainedModel):
    _tied_weights_keys = ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight"]

    def __init__(self, config: LEDConfig, **kwargs):
        warnings.warn(
            "The `transformers.LEDForSequenceClassification` class is deprecated and will be removed in version 5 of"
            " Transformers. No actual method were provided in the original paper on how to perfom"
            " sequence classification.",
            FutureWarning,
        )
        super().__init__(config, **kwargs)
        self.led = LEDModel(config)
        self.classification_head = LEDClassificationHead(
            config.d_model,
            config.d_model,
            config.num_labels,
            config.classifier_dropout,
        )

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Optional[mindspore.Tensor] = None,
        attention_mask: Optional[mindspore.Tensor] = None,
        decoder_input_ids: Optional[mindspore.Tensor] = None,
        decoder_attention_mask: Optional[mindspore.Tensor] = None,
        head_mask: Optional[mindspore.Tensor] = None,
        decoder_head_mask: Optional[mindspore.Tensor] = None,
        cross_attn_head_mask: Optional[mindspore.Tensor] = None,
        encoder_outputs: Optional[Tuple[Tuple[mindspore.Tensor]]] = None,
        global_attention_mask: Optional[mindspore.Tensor] = None,
        inputs_embeds: Optional[mindspore.Tensor] = None,
        decoder_inputs_embeds: Optional[mindspore.Tensor] = None,
        labels: Optional[mindspore.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[mindspore.Tensor], LEDSeq2SeqSequenceClassifierOutput]:
        r"""
        Args:
            labels (`mindspore.Tensor` 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 classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        if labels is not None:
            use_cache = False

        if input_ids is None and inputs_embeds is not None:
            raise NotImplementedError(
                f"Passing input embeddings is currently not supported for {self.__class__.__name__}"
            )

        outputs = self.led(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            global_attention_mask=global_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            encoder_outputs=encoder_outputs,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = outputs[0]  # last hidden state

        eos_mask = input_ids.eq(self.config.eos_token_id)
        if len(ops.unique_consecutive(eos_mask.sum(1))) > 1:
            raise ValueError("All examples must have the same number of <eos> tokens.")
        sentence_representation = hidden_states[eos_mask].view(hidden_states.shape[0], -1, hidden_states.shape[-1])[:, -1, :]
        logits = self.classification_head(sentence_representation)

        loss = None
        if labels is not None:
            if self.config.problem_type is None:
                if self.config.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.config.num_labels > 1 and labels.dtype in (mindspore.int64, mindspore.int32):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                if self.config.num_labels == 1:
                    loss = ops.mse_loss(logits.squeeze(), labels.squeeze())
                else:
                    loss = ops.mse_loss(logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss = ops.cross_entropy(logits.view(-1, self.config.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss = ops.binary_cross_entropy_with_logits(logits, labels)
        if not return_dict:
            output = (logits,) + outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return LEDSeq2SeqSequenceClassifierOutput(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
            encoder_global_attentions=outputs.encoder_global_attentions,
        )

mindnlp.transformers.models.led.modeling_led.LEDForSequenceClassification.forward(input_ids=None, attention_mask=None, decoder_input_ids=None, decoder_attention_mask=None, head_mask=None, decoder_head_mask=None, cross_attn_head_mask=None, encoder_outputs=None, global_attention_mask=None, inputs_embeds=None, decoder_inputs_embeds=None, labels=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None)

PARAMETER	DESCRIPTION
labels	Labels for computing the sequence classification/regression loss. Indices should be in [0, ..., config.num_labels - 1]. If config.num_labels > 1 a classification loss is computed (Cross-Entropy). TYPE: `mindspore.Tensor` of shape `(batch_size,)`, *optional* DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(
    self,
    input_ids: Optional[mindspore.Tensor] = None,
    attention_mask: Optional[mindspore.Tensor] = None,
    decoder_input_ids: Optional[mindspore.Tensor] = None,
    decoder_attention_mask: Optional[mindspore.Tensor] = None,
    head_mask: Optional[mindspore.Tensor] = None,
    decoder_head_mask: Optional[mindspore.Tensor] = None,
    cross_attn_head_mask: Optional[mindspore.Tensor] = None,
    encoder_outputs: Optional[Tuple[Tuple[mindspore.Tensor]]] = None,
    global_attention_mask: Optional[mindspore.Tensor] = None,
    inputs_embeds: Optional[mindspore.Tensor] = None,
    decoder_inputs_embeds: Optional[mindspore.Tensor] = None,
    labels: Optional[mindspore.Tensor] = None,
    use_cache: Optional[bool] = None,
    output_attentions: Optional[bool] = None,
    output_hidden_states: Optional[bool] = None,
    return_dict: Optional[bool] = None,
) -> Union[Tuple[mindspore.Tensor], LEDSeq2SeqSequenceClassifierOutput]:
    r"""
    Args:
        labels (`mindspore.Tensor` 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 classification loss is computed (Cross-Entropy).
    """
    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
    if labels is not None:
        use_cache = False

    if input_ids is None and inputs_embeds is not None:
        raise NotImplementedError(
            f"Passing input embeddings is currently not supported for {self.__class__.__name__}"
        )

    outputs = self.led(
        input_ids,
        attention_mask=attention_mask,
        decoder_input_ids=decoder_input_ids,
        decoder_attention_mask=decoder_attention_mask,
        global_attention_mask=global_attention_mask,
        head_mask=head_mask,
        decoder_head_mask=decoder_head_mask,
        cross_attn_head_mask=cross_attn_head_mask,
        encoder_outputs=encoder_outputs,
        inputs_embeds=inputs_embeds,
        decoder_inputs_embeds=decoder_inputs_embeds,
        use_cache=use_cache,
        output_attentions=output_attentions,
        output_hidden_states=output_hidden_states,
        return_dict=return_dict,
    )
    hidden_states = outputs[0]  # last hidden state

    eos_mask = input_ids.eq(self.config.eos_token_id)
    if len(ops.unique_consecutive(eos_mask.sum(1))) > 1:
        raise ValueError("All examples must have the same number of <eos> tokens.")
    sentence_representation = hidden_states[eos_mask].view(hidden_states.shape[0], -1, hidden_states.shape[-1])[:, -1, :]
    logits = self.classification_head(sentence_representation)

    loss = None
    if labels is not None:
        if self.config.problem_type is None:
            if self.config.num_labels == 1:
                self.config.problem_type = "regression"
            elif self.config.num_labels > 1 and labels.dtype in (mindspore.int64, mindspore.int32):
                self.config.problem_type = "single_label_classification"
            else:
                self.config.problem_type = "multi_label_classification"

        if self.config.problem_type == "regression":
            if self.config.num_labels == 1:
                loss = ops.mse_loss(logits.squeeze(), labels.squeeze())
            else:
                loss = ops.mse_loss(logits, labels)
        elif self.config.problem_type == "single_label_classification":
            loss = ops.cross_entropy(logits.view(-1, self.config.num_labels), labels.view(-1))
        elif self.config.problem_type == "multi_label_classification":
            loss = ops.binary_cross_entropy_with_logits(logits, labels)
    if not return_dict:
        output = (logits,) + outputs[1:]
        return ((loss,) + output) if loss is not None else output

    return LEDSeq2SeqSequenceClassifierOutput(
        loss=loss,
        logits=logits,
        past_key_values=outputs.past_key_values,
        decoder_hidden_states=outputs.decoder_hidden_states,
        decoder_attentions=outputs.decoder_attentions,
        cross_attentions=outputs.cross_attentions,
        encoder_last_hidden_state=outputs.encoder_last_hidden_state,
        encoder_hidden_states=outputs.encoder_hidden_states,
        encoder_attentions=outputs.encoder_attentions,
        encoder_global_attentions=outputs.encoder_global_attentions,
    )

mindnlp.transformers.models.led.modeling_led.LEDLearnedPositionalEmbedding

Bases: Embedding

This module learns positional embeddings up to a fixed maximum size.

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDLearnedPositionalEmbedding(nn.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size.
    """

    def __init__(self, num_embeddings: int, embedding_dim: int):
        super().__init__(num_embeddings, embedding_dim)

    def forward(self, input_ids_shape, past_key_values_length: int = 0):
        """`input_ids_shape` is expected to be [bsz x seqlen]."""
        bsz, seq_len = input_ids_shape[:2]
        positions = ops.arange(
            past_key_values_length, past_key_values_length + seq_len, dtype=mindspore.int64
        )
        return super().forward(positions)

mindnlp.transformers.models.led.modeling_led.LEDLearnedPositionalEmbedding.forward(input_ids_shape, past_key_values_length=0)

input_ids_shape is expected to be [bsz x seqlen].

Source code in mindnlp/transformers/models/led/modeling_led.py

def forward(self, input_ids_shape, past_key_values_length: int = 0):
    """`input_ids_shape` is expected to be [bsz x seqlen]."""
    bsz, seq_len = input_ids_shape[:2]
    positions = ops.arange(
        past_key_values_length, past_key_values_length + seq_len, dtype=mindspore.int64
    )
    return super().forward(positions)

mindnlp.transformers.models.led.modeling_led.LEDPreTrainedModel

Bases: PreTrainedModel

Source code in mindnlp/transformers/models/led/modeling_led.py

class LEDPreTrainedModel(PreTrainedModel):
    config_class = LEDConfig
    base_model_prefix = "led"
    supports_gradient_checkpointing = True

    def _init_weights(self, cell):
        """Initialize the weights"""
        std = self.config.init_std
        if isinstance(cell, nn.Linear):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            cell.weight.set_data(initializer(Normal(std), cell.weight.shape, cell.weight.dtype))
            if cell.bias:
                cell.bias.set_data(initializer('zeros', cell.bias.shape, cell.bias.dtype))
        elif isinstance(cell, nn.Embedding):
            weight = np.random.normal(0.0, std, cell.weight.shape)
            if cell.padding_idx:
                weight[cell.padding_idx] = 0

            cell.weight.set_data(Tensor(weight, cell.weight.dtype))

    @property
    def dummy_inputs(self):
        pad_token = self.config.pad_token_id
        input_ids = mindspore.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]])
        dummy_inputs = {
            "attention_mask": input_ids.ne(pad_token),
            "input_ids": input_ids,
        }
        return dummy_inputs

mindnlp.transformers.models.led.modeling_led.LEDSeq2SeqLMOutput dataclass

Bases: ModelOutput

Base class for sequence-to-sequence language models outputs.

PARAMETER	DESCRIPTION
loss	Language modeling loss. TYPE: `mindspore.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided DEFAULT: None
logits	Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)` DEFAULT: None
encoder_last_hidden_state	Sequence of hidden-states at the output of the last layer of the encoder of the model. TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional* DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

@dataclass
class LEDSeq2SeqLMOutput(ModelOutput):
    """
    Base class for sequence-to-sequence language models outputs.

    Args:
        loss (`mindspore.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss.
        logits (`mindspore.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        past_key_values (`List[mindspore.Tensor]`, *optional*, returned when `use_cache=True`
            is passed or when `config.use_cache=True`):
            List of `mindspore.Tensor` of length `config.n_layers`, with each tensor of shape `(2, batch_size,
            num_heads, sequence_length, embed_size_per_head)`).

            Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
            used (see `past_key_values` input) to speed up sequential decoding.
        decoder_hidden_states (`tuple(mindspore.Tensor)`, *optional*, returned when `output_hidden_states=True`
            is passed or when `config.output_hidden_states=True`):
            Tuple of `mindspore.Tensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
        decoder_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
            self-attention heads.
        cross_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
            weighted average in the cross-attention heads.
        encoder_last_hidden_state (`mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Sequence of hidden-states at the output of the last layer of the encoder of the model.
        encoder_hidden_states (`tuple(mindspore.Tensor)`, *optional*, returned when `output_hidden_states=True`
            is passed or when `config.output_hidden_states=True`):
            Tuple of `mindspore.Tensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
        encoder_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
            self-attention heads.
        encoder_global_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x)`,
            where `x` is the number of tokens with global attention mask.

            Global attentions weights after the attention softmax, used to compute the weighted average in the
            self-attention heads. Those are the attention weights from every token with global attention to every token
            in the sequence.
    """

    loss: Optional[mindspore.Tensor] = None
    logits: mindspore.Tensor = None
    past_key_values: Optional[List[mindspore.Tensor]] = None
    decoder_hidden_states: Optional[Tuple[mindspore.Tensor, ...]] = None
    decoder_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    cross_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_last_hidden_state: Optional[mindspore.Tensor] = None
    encoder_hidden_states: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_global_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None

mindnlp.transformers.models.led.modeling_led.LEDSeq2SeqModelOutput dataclass

Bases: ModelOutput

Base class for model encoder's outputs that also contains : pre-computed hidden states that can speed up sequential decoding.

PARAMETER	DESCRIPTION
last_hidden_state	Sequence of hidden-states at the output of the last layer of the decoder of the model. If past_key_values is used only the last hidden-state of the sequences of shape (batch_size, 1, hidden_size) is output. TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)` DEFAULT: None
encoder_last_hidden_state	Sequence of hidden-states at the output of the last layer of the encoder of the model. TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional* DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

@dataclass
class LEDSeq2SeqModelOutput(ModelOutput):
    """
    Base class for model encoder's outputs that also contains : pre-computed hidden states that can speed up sequential
    decoding.

    Args:
        last_hidden_state (`mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the decoder of the model.

            If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
            hidden_size)` is output.
        past_key_values (`List[mindspore.Tensor]`, *optional*, returned when `use_cache=True` is passed
            or when `config.use_cache=True`):
            List of `mindspore.Tensor` of length `config.n_layers`, with each tensor of shape `(2, batch_size,
            num_heads, sequence_length, embed_size_per_head)`).

            Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
            used (see `past_key_values` input) to speed up sequential decoding.
        decoder_hidden_states (`tuple(mindspore.Tensor)`, *optional*, returned when `output_hidden_states=True`
            is passed or when `config.output_hidden_states=True`):
            Tuple of `mindspore.Tensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
        decoder_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
            self-attention heads.
        cross_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
            weighted average in the cross-attention heads.
        encoder_last_hidden_state (`mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Sequence of hidden-states at the output of the last layer of the encoder of the model.
        encoder_hidden_states (`tuple(mindspore.Tensor)`, *optional*, returned when `output_hidden_states=True`
            is passed or when `config.output_hidden_states=True`):
            Tuple of `mindspore.Tensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
        encoder_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
            self-attention heads.
        encoder_global_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x)`,
            where `x` is the number of tokens with global attention mask.

            Global attentions weights after the attention softmax, used to compute the weighted average in the
            self-attention heads. Those are the attention weights from every token with global attention to every token
            in the sequence.
    """

    last_hidden_state: mindspore.Tensor = None
    past_key_values: Optional[List[mindspore.Tensor]] = None
    decoder_hidden_states: Optional[Tuple[mindspore.Tensor, ...]] = None
    decoder_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    cross_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_last_hidden_state: Optional[mindspore.Tensor] = None
    encoder_hidden_states: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_global_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None

mindnlp.transformers.models.led.modeling_led.LEDSeq2SeqQuestionAnsweringModelOutput dataclass

Bases: ModelOutput

Base class for outputs of sequence-to-sequence question answering models.

PARAMETER	DESCRIPTION
loss	Total span extraction loss is the sum of a Cross-Entropy for the start and end positions. TYPE: `mindspore.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided DEFAULT: None
start_logits	Span-start scores (before SoftMax). TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length)` DEFAULT: None
end_logits	Span-end scores (before SoftMax). TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length)` DEFAULT: None
encoder_last_hidden_state	Sequence of hidden-states at the output of the last layer of the encoder of the model. TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional* DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

@dataclass
class LEDSeq2SeqQuestionAnsweringModelOutput(ModelOutput):
    """
    Base class for outputs of sequence-to-sequence question answering models.

    Args:
        loss (`mindspore.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
        start_logits (`mindspore.Tensor` of shape `(batch_size, sequence_length)`):
            Span-start scores (before SoftMax).
        end_logits (`mindspore.Tensor` of shape `(batch_size, sequence_length)`):
            Span-end scores (before SoftMax).
        past_key_values (`List[mindspore.Tensor]`, *optional*, returned when `use_cache=True`
            is passed or when `config.use_cache=True`):
            List of `mindspore.Tensor` of length `config.n_layers`, with each tensor of shape `(2, batch_size,
            num_heads, sequence_length, embed_size_per_head)`).

            Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
            used (see `past_key_values` input) to speed up sequential decoding.
        decoder_hidden_states (`tuple(mindspore.Tensor)`, *optional*, returned when `output_hidden_states=True`
            is passed or when `config.output_hidden_states=True`):
            Tuple of `mindspore.Tensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
        decoder_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
            self-attention heads.
        cross_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
            weighted average in the cross-attention heads.
        encoder_last_hidden_state (`mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Sequence of hidden-states at the output of the last layer of the encoder of the model.
        encoder_hidden_states (`tuple(mindspore.Tensor)`, *optional*, returned when `output_hidden_states=True`
            is passed or when `config.output_hidden_states=True`):
            Tuple of `mindspore.Tensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
        encoder_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
            self-attention heads.
        encoder_global_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x)`,
            where `x` is the number of tokens with global attention mask.

            Global attentions weights after the attention softmax, used to compute the weighted average in the
            self-attention heads. Those are the attention weights from every token with global attention to every token
            in the sequence.
    """

    loss: Optional[mindspore.Tensor] = None
    start_logits: mindspore.Tensor = None
    end_logits: mindspore.Tensor = None
    past_key_values: Optional[List[mindspore.Tensor]] = None
    decoder_hidden_states: Optional[Tuple[mindspore.Tensor, ...]] = None
    decoder_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    cross_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_last_hidden_state: Optional[mindspore.Tensor] = None
    encoder_hidden_states: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_global_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None

mindnlp.transformers.models.led.modeling_led.LEDSeq2SeqSequenceClassifierOutput dataclass

Bases: ModelOutput

Base class for outputs of sequence-to-sequence sentence classification models.

PARAMETER	DESCRIPTION
loss	Classification (or regression if config.num_labels==1) loss. TYPE: `mindspore.Tensor` of shape `(1,)`, *optional*, returned when `label` is provided DEFAULT: None
logits	Classification (or regression if config.num_labels==1) scores (before SoftMax). TYPE: `mindspore.Tensor` of shape `(batch_size, config.num_labels)` DEFAULT: None
encoder_last_hidden_state	Sequence of hidden-states at the output of the last layer of the encoder of the model. TYPE: `mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional* DEFAULT: None

Source code in mindnlp/transformers/models/led/modeling_led.py

@dataclass
class LEDSeq2SeqSequenceClassifierOutput(ModelOutput):
    """
    Base class for outputs of sequence-to-sequence sentence classification models.

    Args:
        loss (`mindspore.Tensor` of shape `(1,)`, *optional*, returned when `label` is provided):
            Classification (or regression if config.num_labels==1) loss.
        logits (`mindspore.Tensor` of shape `(batch_size, config.num_labels)`):
            Classification (or regression if config.num_labels==1) scores (before SoftMax).
        past_key_values (`List[mindspore.Tensor]`, *optional*, returned when `use_cache=True`
            is passed or when `config.use_cache=True`):
            List of `mindspore.Tensor` of length `config.n_layers`, with each tensor of shape `(2, batch_size,
            num_heads, sequence_length, embed_size_per_head)`).

            Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
            used (see `past_key_values` input) to speed up sequential decoding.
        decoder_hidden_states (`tuple(mindspore.Tensor)`, *optional*, returned when `output_hidden_states=True`
            is passed or when `config.output_hidden_states=True`):
            Tuple of `mindspore.Tensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
        decoder_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
            self-attention heads.
        cross_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
            weighted average in the cross-attention heads.
        encoder_last_hidden_state (`mindspore.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Sequence of hidden-states at the output of the last layer of the encoder of the model.
        encoder_hidden_states (`tuple(mindspore.Tensor)`, *optional*, returned when `output_hidden_states=True`
            is passed or when `config.output_hidden_states=True`):
            Tuple of `mindspore.Tensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
        encoder_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.

            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
            self-attention heads.
        encoder_global_attentions (`tuple(mindspore.Tensor)`, *optional*, returned when `output_attentions=True`
            is passed or when `config.output_attentions=True`):
            Tuple of `mindspore.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x)`,
            where `x` is the number of tokens with global attention mask.

            Global attentions weights after the attention softmax, used to compute the weighted average in the
            self-attention heads. Those are the attention weights from every token with global attention to every token
            in the sequence.
    """

    loss: Optional[mindspore.Tensor] = None
    logits: mindspore.Tensor = None
    past_key_values: Optional[List[mindspore.Tensor]] = None
    decoder_hidden_states: Optional[Tuple[mindspore.Tensor, ...]] = None
    decoder_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    cross_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_last_hidden_state: Optional[mindspore.Tensor] = None
    encoder_hidden_states: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None
    encoder_global_attentions: Optional[Tuple[mindspore.Tensor, ...]] = None

mindnlp.transformers.models.led.modeling_led.scalar_div(input, other, *, rounding_mode='trunc')

scalar div since ops.div do not support scalar

Source code in mindnlp/transformers/models/led/modeling_led.py

def scalar_div(input, other, *, rounding_mode="trunc"):
    """scalar div since ops.div do not support scalar"""
    if rounding_mode == 'trunc':
        res = input // other
        if res < 0:
            res = res + 1
        return res
    if rounding_mode == 'floor':
        return input // other
    return input / other

mindnlp.transformers.models.led.modeling_led.shift_tokens_right(input_ids, pad_token_id, decoder_start_token_id)

Shift input ids one token to the right.

Source code in mindnlp/transformers/models/led/modeling_led.py

def shift_tokens_right(input_ids: mindspore.Tensor, pad_token_id: int, decoder_start_token_id: int):
    """
    Shift input ids one token to the right.
    """
    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
    shifted_input_ids[:, 1:] = input_ids[:, :-1].copy()
    shifted_input_ids[:, 0] = decoder_start_token_id

    if pad_token_id is None:
        raise ValueError("config.pad_token_id has to be defined.")
    # replace possible -100 values in labels by `pad_token_id`
    shifted_input_ids = shifted_input_ids.masked_fill(shifted_input_ids == -100, pad_token_id)

    return shifted_input_ids

mindnlp.transformers.models.led.tokenization_led

Tokenization classes for LED.

mindnlp.transformers.models.led.tokenization_led.LEDTokenizer

Bases: PreTrainedTokenizer

Constructs a LED tokenizer, which is smilar to the ROBERTa tokenizer, using byte-level Byte-Pair-Encoding.

This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will be encoded differently whether it is at the beginning of the sentence (without space) or not:

Example

>>> from transformers import LEDTokenizer
...
>>> tokenizer = LEDTokenizer.from_pretrained("allenai/led-base-16384")
>>> tokenizer("Hello world")["input_ids"]
[0, 31414, 232, 2]
>>> tokenizer(" Hello world")["input_ids"]
[0, 20920, 232, 2]

You can get around that behavior by passing add_prefix_space=True when instantiating this tokenizer or when you call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance.

When used with is_split_into_words=True, this tokenizer will add a space before each word (even the first one).

This tokenizer inherits from [PreTrainedTokenizer] which contains most of the main methods. Users should refer to this superclass for more information regarding those methods.

PARAMETER	DESCRIPTION
vocab_file	Path to the vocabulary file. TYPE: `str`
merges_file	Path to the merges file. TYPE: `str`
errors	Paradigm to follow when decoding bytes to UTF-8. See bytes.decode for more information. TYPE: `str`, *optional*, defaults to `"replace"` DEFAULT: 'replace'
bos_token	The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token. When building a sequence using special tokens, this is not the token that is used for the beginning of sequence. The token used is the cls_token. TYPE: `str`, *optional*, defaults to `"<s>"` DEFAULT: '<s>'
eos_token	The end of sequence token. When building a sequence using special tokens, this is not the token that is used for the end of sequence. The token used is the sep_token. TYPE: `str`, *optional*, defaults to `"</s>"` DEFAULT: '</s>'
sep_token	The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for sequence classification or for a text and a question for question answering. It is also used as the last token of a sequence built with special tokens. TYPE: `str`, *optional*, defaults to `"</s>"` DEFAULT: '</s>'
cls_token	The classifier token which is used when doing sequence classification (classification of the whole sequence instead of per-token classification). It is the first token of the sequence when built with special tokens. TYPE: `str`, *optional*, defaults to `"<s>"` DEFAULT: '<s>'
unk_token	The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this token instead. TYPE: `str`, *optional*, defaults to `"<unk>"` DEFAULT: '<unk>'
pad_token	The token used for padding, for example when batching sequences of different lengths. TYPE: `str`, *optional*, defaults to `"<pad>"` DEFAULT: '<pad>'
mask_token	The token used for masking values. This is the token used when training this model with masked language modeling. This is the token which the model will try to predict. TYPE: `str`, *optional*, defaults to `"<mask>"` DEFAULT: '<mask>'
add_prefix_space	Whether or not to add an initial space to the input. This allows to treat the leading word just as any other word. (BART tokenizer detect beginning of words by the preceding space). TYPE: `bool`, *optional*, defaults to `False` DEFAULT: False

Source code in mindnlp/transformers/models/led/tokenization_led.py

class LEDTokenizer(PreTrainedTokenizer):
    r"""
    Constructs a LED tokenizer, which is smilar to the ROBERTa tokenizer, using byte-level Byte-Pair-Encoding.

    This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will
    be encoded differently whether it is at the beginning of the sentence (without space) or not:

    Example:
        ```python
        >>> from transformers import LEDTokenizer
        ...
        >>> tokenizer = LEDTokenizer.from_pretrained("allenai/led-base-16384")
        >>> tokenizer("Hello world")["input_ids"]
        [0, 31414, 232, 2]
        >>> tokenizer(" Hello world")["input_ids"]
        [0, 20920, 232, 2]
        ```

    You can get around that behavior by passing `add_prefix_space=True` when instantiating this tokenizer or when you
    call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance.

    <Tip>

    When used with `is_split_into_words=True`, this tokenizer will add a space before each word (even the first one).

    </Tip>

    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
    this superclass for more information regarding those methods.

    Args:
        vocab_file (`str`):
            Path to the vocabulary file.
        merges_file (`str`):
            Path to the merges file.
        errors (`str`, *optional*, defaults to `"replace"`):
            Paradigm to follow when decoding bytes to UTF-8. See
            [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
        bos_token (`str`, *optional*, defaults to `"<s>"`):
            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.

            <Tip>

            When building a sequence using special tokens, this is not the token that is used for the beginning of
            sequence. The token used is the `cls_token`.

            </Tip>

        eos_token (`str`, *optional*, defaults to `"</s>"`):
            The end of sequence token.

            <Tip>

            When building a sequence using special tokens, this is not the token that is used for the end of sequence.
            The token used is the `sep_token`.

            </Tip>

        sep_token (`str`, *optional*, defaults to `"</s>"`):
            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
            sequence classification or for a text and a question for question answering. It is also used as the last
            token of a sequence built with special tokens.
        cls_token (`str`, *optional*, defaults to `"<s>"`):
            The classifier token which is used when doing sequence classification (classification of the whole sequence
            instead of per-token classification). It is the first token of the sequence when built with special tokens.
        unk_token (`str`, *optional*, defaults to `"<unk>"`):
            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
            token instead.
        pad_token (`str`, *optional*, defaults to `"<pad>"`):
            The token used for padding, for example when batching sequences of different lengths.
        mask_token (`str`, *optional*, defaults to `"<mask>"`):
            The token used for masking values. This is the token used when training this model with masked language
            modeling. This is the token which the model will try to predict.
        add_prefix_space (`bool`, *optional*, defaults to `False`):
            Whether or not to add an initial space to the input. This allows to treat the leading word just as any
            other word. (BART tokenizer detect beginning of words by the preceding space).
    """

    vocab_files_names = VOCAB_FILES_NAMES
    model_input_names = ["input_ids", "attention_mask"]

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.__init__
    def __init__(
        self,
        vocab_file,
        merges_file,
        errors="replace",
        bos_token="<s>",
        eos_token="</s>",
        sep_token="</s>",
        cls_token="<s>",
        unk_token="<unk>",
        pad_token="<pad>",
        mask_token="<mask>",
        add_prefix_space=False,
        **kwargs,
    ):
        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
        sep_token = AddedToken(sep_token, lstrip=False, rstrip=False) if isinstance(sep_token, str) else sep_token
        cls_token = AddedToken(cls_token, lstrip=False, rstrip=False) if isinstance(cls_token, str) else cls_token
        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token

        # Mask token behave like a normal word, i.e. include the space before it
        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token

        with open(vocab_file, encoding="utf-8") as vocab_handle:
            self.encoder = json.load(vocab_handle)
        self.decoder = {v: k for k, v in self.encoder.items()}
        self.errors = errors  # how to handle errors in decoding
        self.byte_encoder = bytes_to_unicode()
        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
        with open(merges_file, encoding="utf-8") as merges_handle:
            bpe_merges = merges_handle.read().split("\n")[1:-1]
        bpe_merges = [tuple(merge.split()) for merge in bpe_merges]
        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
        self.cache = {}
        self.add_prefix_space = add_prefix_space

        # Should have added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions
        self.pat = re.compile(r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""")

        super().__init__(
            errors=errors,
            bos_token=bos_token,
            eos_token=eos_token,
            unk_token=unk_token,
            sep_token=sep_token,
            cls_token=cls_token,
            pad_token=pad_token,
            mask_token=mask_token,
            add_prefix_space=add_prefix_space,
            **kwargs,
        )

    @property
    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.vocab_size
    def vocab_size(self):
        return len(self.encoder)

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.get_vocab
    def get_vocab(self):
        return dict(self.encoder, **self.added_tokens_encoder)

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.bpe
    def bpe(self, token):
        if token in self.cache:
            return self.cache[token]
        word = tuple(token)
        pairs = get_pairs(word)

        if not pairs:
            return token

        while True:
            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
            if bigram not in self.bpe_ranks:
                break
            first, second = bigram
            new_word = []
            i = 0
            while i < len(word):
                try:
                    j = word.index(first, i)
                except ValueError:
                    new_word.extend(word[i:])
                    break
                else:
                    new_word.extend(word[i:j])
                    i = j

                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
                    new_word.append(first + second)
                    i += 2
                else:
                    new_word.append(word[i])
                    i += 1
            new_word = tuple(new_word)
            word = new_word
            if len(word) == 1:
                break
            else:
                pairs = get_pairs(word)
        word = " ".join(word)
        self.cache[token] = word
        return word

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer._tokenize
    def _tokenize(self, text):
        """Tokenize a string."""
        bpe_tokens = []
        for token in re.findall(self.pat, text):
            token = "".join(
                self.byte_encoder[b] for b in token.encode("utf-8")
            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
        return bpe_tokens

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer._convert_token_to_id
    def _convert_token_to_id(self, token):
        """Converts a token (str) in an id using the vocab."""
        return self.encoder.get(token, self.encoder.get(self.unk_token))

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer._convert_id_to_token
    def _convert_id_to_token(self, index):
        """Converts an index (integer) in a token (str) using the vocab."""
        return self.decoder.get(index)

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.convert_tokens_to_string
    def convert_tokens_to_string(self, tokens):
        """Converts a sequence of tokens (string) in a single string."""
        text = "".join(tokens)
        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
        return text

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.save_vocabulary
    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
        if not os.path.isdir(save_directory):
            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
            return
        vocab_file = os.path.join(
            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
        )
        merge_file = os.path.join(
            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
        )

        with open(vocab_file, "w", encoding="utf-8") as f:
            f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n")

        index = 0
        with open(merge_file, "w", encoding="utf-8") as writer:
            writer.write("#version: 0.2\n")
            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
                if index != token_index:
                    logger.warning(
                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
                        " Please check that the tokenizer is not corrupted!"
                    )
                    index = token_index
                writer.write(" ".join(bpe_tokens) + "\n")
                index += 1

        return vocab_file, merge_file

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.build_inputs_with_special_tokens with BART->LED
    def build_inputs_with_special_tokens(
        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
    ) -> List[int]:
        """
        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
        adding special tokens. A LED sequence has the following format:

        - single sequence: `<s> X </s>`
        - pair of sequences: `<s> A </s></s> B </s>`

        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 IDs](../glossary#input-ids) 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

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.get_special_tokens_mask
    def get_special_tokens_mask(
        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
    ) -> List[int]:
        """
        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
        special tokens using the tokenizer `prepare_for_model` method.

        Args:
            token_ids_0 (`List[int]`):
                List of IDs.
            token_ids_1 (`List[int]`, *optional*):
                Optional second list of IDs for sequence pairs.
            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
                Whether or not the token list is already formatted with special tokens for the model.

        Returns:
            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
        """
        if already_has_special_tokens:
            return super().get_special_tokens_mask(
                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
            )

        if token_ids_1 is None:
            return [1] + ([0] * len(token_ids_0)) + [1]
        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.create_token_type_ids_from_sequences with BART->LED
    def create_token_type_ids_from_sequences(
        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
    ) -> List[int]:
        """
        Create a mask from the two sequences passed to be used in a sequence-pair classification task. LED does not
        make use of token type ids, therefore a list of zeros is returned.

        Args:
            token_ids_0 (`List[int]`):
                List of IDs.
            token_ids_1 (`List[int]`, *optional*):
                Optional second list of IDs for sequence pairs.

        Returns:
            `List[int]`: List of zeros.
        """
        sep = [self.sep_token_id]
        cls = [self.cls_token_id]

        if token_ids_1 is None:
            return len(cls + token_ids_0 + sep) * [0]
        return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]

    # Copied from transformers.models.bart.tokenization_bart.BartTokenizer.prepare_for_tokenization
    def prepare_for_tokenization(self, text, is_split_into_words=False, **kwargs):
        add_prefix_space = kwargs.pop("add_prefix_space", self.add_prefix_space)
        if (is_split_into_words or add_prefix_space) and (len(text) > 0 and not text[0].isspace()):
            text = " " + text
        return (text, kwargs)

    def _pad(
        self,
        encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding],
        max_length: Optional[int] = None,
        padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
        pad_to_multiple_of: Optional[int] = None,
        return_attention_mask: Optional[bool] = None,
    ) -> dict:
        encoded_inputs = super()._pad(
            encoded_inputs=encoded_inputs,
            max_length=max_length,
            padding_strategy=padding_strategy,
            pad_to_multiple_of=pad_to_multiple_of,
            return_attention_mask=return_attention_mask,
        )

        # Load from model defaults
        if return_attention_mask is None:
            return_attention_mask = "attention_mask" in self.model_input_names

        if return_attention_mask and "global_attention_mask" in encoded_inputs:
            required_input = encoded_inputs[self.model_input_names[0]]
            # `global_attention_mask` need to have the same length as other (sequential) inputs.
            needs_to_be_padded = len(encoded_inputs["global_attention_mask"]) != len(required_input)

            if needs_to_be_padded:
                difference = len(required_input) - len(encoded_inputs["global_attention_mask"])

                if self.padding_side == "right":
                    # Use `-1` since `0` in `global_attention_mask` means `local attention` instead of `not to attend`
                    encoded_inputs["global_attention_mask"] = (
                        encoded_inputs["global_attention_mask"] + [-1] * difference
                    )
                elif self.padding_side == "left":
                    encoded_inputs["global_attention_mask"] = [-1] * difference + encoded_inputs[
                        "global_attention_mask"
                    ]
                else:
                    raise ValueError("Invalid padding strategy:" + str(self.padding_side))

        return encoded_inputs

mindnlp.transformers.models.led.tokenization_led.LEDTokenizer.build_inputs_with_special_tokens(token_ids_0, token_ids_1=None)

Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and adding special tokens. A LED sequence has the following format:

• single sequence: <s> X </s>
• pair of sequences: <s> A </s></s> B </s>

PARAMETER	DESCRIPTION
token_ids_0	List of IDs to which the special tokens will be added. TYPE: `List[int]`
token_ids_1	Optional second list of IDs for sequence pairs. TYPE: `List[int]`, *optional* DEFAULT: None

RETURNS	DESCRIPTION
List[int]	List[int]: List of input IDs with the appropriate special tokens.

Source code in mindnlp/transformers/models/led/tokenization_led.py

def build_inputs_with_special_tokens(
    self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
    """
    Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
    adding special tokens. A LED sequence has the following format:

    - single sequence: `<s> X </s>`
    - pair of sequences: `<s> A </s></s> B </s>`

    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 IDs](../glossary#input-ids) 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

mindnlp.transformers.models.led.tokenization_led.LEDTokenizer.convert_tokens_to_string(tokens)

Converts a sequence of tokens (string) in a single string.

Source code in mindnlp/transformers/models/led/tokenization_led.py

def convert_tokens_to_string(self, tokens):
    """Converts a sequence of tokens (string) in a single string."""
    text = "".join(tokens)
    text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
    return text

mindnlp.transformers.models.led.tokenization_led.LEDTokenizer.create_token_type_ids_from_sequences(token_ids_0, token_ids_1=None)

Create a mask from the two sequences passed to be used in a sequence-pair classification task. LED does not make use of token type ids, therefore a list of zeros is returned.

PARAMETER	DESCRIPTION
token_ids_0	List of IDs. TYPE: `List[int]`
token_ids_1	Optional second list of IDs for sequence pairs. TYPE: `List[int]`, *optional* DEFAULT: None

RETURNS	DESCRIPTION
List[int]	List[int]: List of zeros.

Source code in mindnlp/transformers/models/led/tokenization_led.py

def create_token_type_ids_from_sequences(
    self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
    """
    Create a mask from the two sequences passed to be used in a sequence-pair classification task. LED does not
    make use of token type ids, therefore a list of zeros is returned.

    Args:
        token_ids_0 (`List[int]`):
            List of IDs.
        token_ids_1 (`List[int]`, *optional*):
            Optional second list of IDs for sequence pairs.

    Returns:
        `List[int]`: List of zeros.
    """
    sep = [self.sep_token_id]
    cls = [self.cls_token_id]

    if token_ids_1 is None:
        return len(cls + token_ids_0 + sep) * [0]
    return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]

mindnlp.transformers.models.led.tokenization_led.LEDTokenizer.get_special_tokens_mask(token_ids_0, token_ids_1=None, already_has_special_tokens=False)

Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding special tokens using the tokenizer prepare_for_model method.

PARAMETER	DESCRIPTION
token_ids_0	List of IDs. TYPE: `List[int]`
token_ids_1	Optional second list of IDs for sequence pairs. TYPE: `List[int]`, *optional* DEFAULT: None
already_has_special_tokens	Whether or not the token list is already formatted with special tokens for the model. TYPE: `bool`, *optional*, defaults to `False` DEFAULT: False

RETURNS	DESCRIPTION
List[int]	List[int]: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.

Source code in mindnlp/transformers/models/led/tokenization_led.py

def get_special_tokens_mask(
    self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
) -> List[int]:
    """
    Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
    special tokens using the tokenizer `prepare_for_model` method.

    Args:
        token_ids_0 (`List[int]`):
            List of IDs.
        token_ids_1 (`List[int]`, *optional*):
            Optional second list of IDs for sequence pairs.
        already_has_special_tokens (`bool`, *optional*, defaults to `False`):
            Whether or not the token list is already formatted with special tokens for the model.

    Returns:
        `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
    """
    if already_has_special_tokens:
        return super().get_special_tokens_mask(
            token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
        )

    if token_ids_1 is None:
        return [1] + ([0] * len(token_ids_0)) + [1]
    return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]

mindnlp.transformers.models.led.tokenization_led.bytes_to_unicode() cached

Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control characters the bpe code barfs on.

The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup tables between utf-8 bytes and unicode strings.

Source code in mindnlp/transformers/models/led/tokenization_led.py

@lru_cache()
# Copied from transformers.models.bart.tokenization_bart.bytes_to_unicode
def bytes_to_unicode():
    """
    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
    characters the bpe code barfs on.

    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
    tables between utf-8 bytes and unicode strings.
    """
    bs = (
        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
    )
    cs = bs[:]
    n = 0
    for b in range(2**8):
        if b not in bs:
            bs.append(b)
            cs.append(2**8 + n)
            n += 1
    cs = [chr(n) for n in cs]
    return dict(zip(bs, cs))

mindnlp.transformers.models.led.tokenization_led.get_pairs(word)

Return set of symbol pair