# Copyright 2023 HuggingFace Inc. team and MosaicML NLP team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch MPT model."""

import math

import torch
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
from torch.nn import functional as F

from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutputWithPastAndCrossAttentions,
    CausalLMOutputWithCrossAttentions,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
from ...modeling_utils import PreTrainedModel
from ...utils import auto_docstring, logging
from .configuration_mpt import MptConfig


logger = logging.get_logger(__name__)


def build_mpt_alibi_tensor(num_heads, sequence_length, alibi_bias_max=8, device=None):
    r"""
    Link to paper: https://huggingface.co/papers/2108.12409 - Alibi tensor is not causal as the original paper mentions, it
    relies on a translation invariance of softmax for quick implementation. This implementation has been copied from
    the alibi implementation of MPT source code that led to slightly different results than the Bloom alibi:
    https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L292
    """
    alibi = torch.arange(1 - sequence_length, 1, dtype=torch.int32, device=device).view(1, 1, 1, sequence_length)
    num_heads_power_of_2 = 2 ** math.ceil(math.log2(num_heads))

    base = torch.arange(1, num_heads_power_of_2 + 1, dtype=torch.int64, device=device).float()
    base = base * (alibi_bias_max / num_heads_power_of_2)

    slopes = 1.0 / torch.pow(2, base)
    slopes = slopes.view(1, num_heads_power_of_2, 1, 1)

    if num_heads_power_of_2 != num_heads:
        slopes = torch.concat([slopes[:, 1::2, ...], slopes[:, ::2, ...]], dim=1)[:, :num_heads, ...]

    alibi = alibi * slopes
    return alibi.squeeze(0)


class MptAttention(nn.Module):
    """Multi-head self attention.
    Using torch or triton attention implementation enables user to also use additive bias.
    """

    def __init__(self, config: MptConfig, layer_idx: int | None = None):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.n_heads = config.n_heads
        self.max_seq_length = config.max_seq_len
        self.head_dim = self.hidden_size // self.n_heads
        self.softmax_scale = config.attn_config.softmax_scale
        if self.softmax_scale is None:
            self.softmax_scale = 1 / math.sqrt(self.hidden_size / self.n_heads)

        self.attn_dropout_p = config.attn_config.attn_pdrop
        self.clip_qkv = config.attn_config.clip_qkv
        self.Wqkv = nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
        self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
        self.layer_idx = layer_idx

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_bias: torch.Tensor,
        past_key_values: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        cache_position: torch.Tensor | None = None,
    ):
        batch_size, seq_length = hidden_states.shape[:2]

        mixed_qkv = self.Wqkv(hidden_states)
        if self.clip_qkv:
            mixed_qkv = mixed_qkv.clamp(min=-self.clip_qkv, max=self.clip_qkv)

        query_states, key_states, value_states = mixed_qkv.chunk(3, dim=2)
        query_states = query_states.reshape(batch_size, seq_length, self.n_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.reshape(batch_size, seq_length, self.n_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.reshape(batch_size, seq_length, self.n_heads, self.head_dim).transpose(1, 2)

        if past_key_values is not None:
            cache_kwargs = {"cache_position": cache_position}
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_scores = torch.matmul(query_states, key_states.transpose(-1, -2)) * self.softmax_scale
        query_length = seq_length if past_key_values is None else seq_length + past_key_values.get_seq_length()

        if position_bias is not None:
            if len(position_bias.shape) != 3:
                raise ValueError(f"Expecting position_bias shape to be 3 dimensions, got {len(position_bias.shape)}")
            key_length = key_states.shape[-2]

            position_bias_query_index = max(0, position_bias.size(1) - query_length)
            position_bias_key_index = max(0, position_bias.size(2) - key_length)

            position_bias = position_bias[:, position_bias_query_index:, position_bias_key_index:]

            attention_scores = attention_scores + position_bias

        if attention_mask is not None:
            attention_scores = attention_scores.masked_fill(attention_mask, torch.finfo(query_states.dtype).min)

        # (batch_size, n_heads, seq_length, key_length)
        attn_weights = nn.functional.softmax(attention_scores.float(), dim=-1).to(value_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attn_dropout_p, training=self.training)

        context_states = torch.matmul(attn_weights, value_states)
        context_states = context_states.permute(0, 2, 1, 3).contiguous().view(batch_size, seq_length, -1)
        attn_output = self.out_proj(context_states)

        return attn_output, attn_weights


class MptMLP(nn.Module):
    def __init__(self, config: MptConfig):
        super().__init__()
        hidden_size = config.hidden_size

        self.up_proj = nn.Linear(hidden_size, 4 * hidden_size, bias=False)
        self.act = nn.GELU(approximate="none")
        self.down_proj = nn.Linear(4 * hidden_size, hidden_size, bias=False)
        self.hidden_dropout = config.attn_config.attn_pdrop

    def forward(self, hidden_states: torch.Tensor, residual: torch.Tensor) -> torch.Tensor:
        hidden_states = self.act(self.up_proj(hidden_states))

        intermediate_output = self.down_proj(hidden_states)

        output = F.dropout(intermediate_output, p=self.hidden_dropout, training=self.training)
        output = output + residual

        return output


class MptBlock(GradientCheckpointingLayer):
    def __init__(self, config: MptConfig, layer_idx: int | None = None):
        super().__init__()
        hidden_size = config.hidden_size

        self.norm_1 = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        # backward compatibility with weights on the Hub
        self.norm_1.bias = None

        self.num_heads = config.n_heads
        self.attn = MptAttention(config, layer_idx)

        self.norm_2 = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        # backward compatibility with weights on the Hub
        self.norm_2.bias = None

        self.ffn = MptMLP(config)

        self.dropout_rate = config.attn_config.attn_pdrop
        self.resid_attn_dropout = nn.Dropout(self.dropout_rate)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_bias: torch.Tensor,
        attention_mask: torch.Tensor,
        layer_past: Cache | None = None,
        use_cache: bool = False,
        output_attentions: bool = False,
        cache_position: torch.Tensor | None = None,
    ):
        # hidden_states: [batch_size, seq_length, hidden_size]
        # Layer norm at the beginning of the transformer layer.
        layernorm_output = self.norm_1(hidden_states)

        residual = hidden_states

        # Self attention.
        attn_outputs, attn_weights = self.attn(
            layernorm_output,
            position_bias=position_bias,
            attention_mask=attention_mask,
            past_key_values=layer_past,
            cache_position=cache_position,
        )

        hidden_states = self.resid_attn_dropout(attn_outputs) + residual

        layernorm_output = self.norm_2(hidden_states)

        # Get residual
        residual = hidden_states

        # MLP.
        output = self.ffn(layernorm_output, residual)
        return output, attn_weights


@auto_docstring
class MptPreTrainedModel(PreTrainedModel):
    config: MptConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = True
    _no_split_modules = ["MptBlock"]


@auto_docstring
class MptModel(MptPreTrainedModel):
    def __init__(self, config: MptConfig):
        super().__init__(config)

        self.hidden_size = config.hidden_size
        self.num_heads = config.n_heads

        # Embedding + LN Embedding
        self.wte = nn.Embedding(config.vocab_size, self.hidden_size)

        # Transformer blocks
        self.blocks = nn.ModuleList([MptBlock(config, layer_idx=i) for i in range(config.n_layers)])

        # Final Layer Norm
        self.norm_f = LayerNorm(self.hidden_size, eps=config.layer_norm_epsilon)
        # backward compatibility with weights on the Hub
        self.norm_f.bias = None

        self.gradient_checkpointing = False

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

    def get_input_embeddings(self):
        return self.wte

    def build_mpt_alibi_tensor(self, num_heads, sequence_length, alibi_bias_max=8, device=None):
        return build_mpt_alibi_tensor(num_heads, sequence_length, alibi_bias_max, device)

    def set_input_embeddings(self, new_embeddings: torch.Tensor):
        self.wte = new_embeddings

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        inputs_embeds: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        cache_position: torch.Tensor | None = None,
        **kwargs,  # NOOP kwargs, for now
    ) -> tuple[torch.Tensor, ...] | BaseModelOutputWithPastAndCrossAttentions:
        r"""
        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values.get_seq_length()`
            (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary.

            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
            `input_ids`.

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

            [What are input IDs?](../glossary#input-ids)
        """
        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

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        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

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

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

        hidden_states = inputs_embeds

        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None

        # Compute alibi tensor: check build_alibi_tensor documentation
        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
        seq_length_with_past = seq_length + past_key_values_length
        if attention_mask is None:
            attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
        else:
            attention_mask = attention_mask.to(hidden_states.device)

        alibi = self.build_mpt_alibi_tensor(self.num_heads, self.config.max_seq_len, device=hidden_states.device)

        causal_mask = _prepare_4d_causal_attention_mask(
            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
        )
        causal_mask = causal_mask.bool()

        for block in self.blocks:
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            outputs = block(
                hidden_states,
                layer_past=past_key_values,
                attention_mask=causal_mask,
                use_cache=use_cache,
                output_attentions=output_attentions,
                position_bias=alibi,
                cache_position=cache_position,
            )

            hidden_states = outputs[0]
            if output_attentions:
                all_self_attentions = all_self_attentions + (outputs[1],)

        # Add last hidden state
        hidden_states = self.norm_f(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v for v in [hidden_states, past_key_values, all_hidden_states, all_self_attentions] if v is not None
            )

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )


@auto_docstring(
    custom_intro="""
    The MPT Model transformer with a language modeling head on top (linear layer with weights tied to the input
    embeddings).
    """
)
class MptForCausalLM(MptPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "transformer.wte.weight"}

    def __init__(self, config: MptConfig):
        super().__init__(config)
        self.transformer = MptModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        cache_position: torch.Tensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs,
    ) -> tuple[torch.Tensor] | CausalLMOutputWithCrossAttentions:
        r"""
        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values.get_seq_length()`
            (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary.

            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
            `input_ids`.

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

            [What are input IDs?](../glossary#input-ids)
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        hidden_states = transformer_outputs[0]
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

        if not return_dict:
            output = (logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return CausalLMOutputWithCrossAttentions(
            loss=loss,
            logits=logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )


@auto_docstring(
    custom_intro="""
    The MPT Model transformer with a sequence classification head on top (linear layer).

    [`MptForSequenceClassification`] uses the last token in order to do the classification, as other causal models
    (e.g. GPT-1) do.

    Since it does classification on the last token, it requires to know the position of the last token. If a
    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
    each row of the batch).
    """
)
class MptForSequenceClassification(MptPreTrainedModel):
    def __init__(self, config: MptConfig):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.transformer = MptModel(config)
        self.score = nn.Linear(config.hidden_size, config.num_labels, bias=False)

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

    def set_output_embeddings(self, new_embeddings: torch.Tensor):
        self.score = new_embeddings

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor] | SequenceClassifierOutputWithPast:
        r"""
        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values.get_seq_length()`
            (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary.

            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
            `input_ids`.

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

            [What are input IDs?](../glossary#input-ids)
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = transformer_outputs[0]
        logits = self.score(hidden_states)

        if input_ids is not None:
            batch_size = input_ids.shape[0]
        else:
            batch_size = inputs_embeds.shape[0]

        if self.config.pad_token_id is None and batch_size != 1:
            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
        if self.config.pad_token_id is None:
            last_non_pad_token = -1
        elif input_ids is not None:
            # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
            non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32)
            token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32)
            last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
        else:
            last_non_pad_token = -1
            logger.warning_once(
                f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
                "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
            )

        pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token]

        loss = None
        if labels is not None:
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(pooled_logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(pooled_logits, labels)
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(pooled_logits, labels)
        if not return_dict:
            output = (pooled_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutputWithPast(
            loss=loss,
            logits=pooled_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )


@auto_docstring
class MptForTokenClassification(MptPreTrainedModel):
    def __init__(self, config: MptConfig):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.transformer = MptModel(config)
        if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
            classifier_dropout = config.classifier_dropout
        elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
            classifier_dropout = config.hidden_dropout
        else:
            classifier_dropout = 0.1
        self.dropout = nn.Dropout(classifier_dropout)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        attention_mask: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **deprecated_arguments,
    ) -> tuple[torch.Tensor] | TokenClassifierOutput:
        r"""
        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values.get_seq_length()`
            (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary.

            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
            `input_ids`.

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

            [What are input IDs?](../glossary#input-ids)
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = transformer_outputs[0]
        hidden_states = self.dropout(hidden_states)
        logits = self.classifier(hidden_states)

        loss = None
        if labels is not None:
            # move labels to correct device
            labels = labels.to(logits.device)
            batch_size, seq_length = labels.shape
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(
                logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
            )

        if not return_dict:
            output = (logits,) + transformer_outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )


@auto_docstring
class MptForQuestionAnswering(MptPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.transformer = MptModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, 2)

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        start_positions: torch.LongTensor | None = None,
        end_positions: torch.LongTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> tuple | QuestionAnsweringModelOutput:
        r"""
        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values.get_seq_length()`
            (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary.

            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
            `input_ids`.

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

            [What are input IDs?](../glossary#input-ids)
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

        if not return_dict:
            output = (start_logits, end_logits) + outputs[2:]
            return ((total_loss,) + output) if total_loss is not None else output

        return QuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


__all__ = [
    "MptForCausalLM",
    "MptModel",
    "MptPreTrainedModel",
    "MptForSequenceClassification",
    "MptForTokenClassification",
    "MptForQuestionAnswering",
]