from collections.abc import Callable
from typing import Optional

import torch
from torch import nn

from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...masking_utils import create_causal_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
from ..llama.modeling_llama import LlamaModel, LlamaPreTrainedModel, LlamaRotaryEmbedding, rotate_half
from .configuration_gpt_neox import GPTNeoXConfig


logger = logging.get_logger(__name__)


class GPTNeoXMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size)
        self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size)
        self.act = ACT2FN[config.hidden_act]

    def forward(self, hidden_states):
        hidden_states = self.dense_h_to_4h(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.dense_4h_to_h(hidden_states)
        return hidden_states


class GPTNeoXRotaryEmbedding(LlamaRotaryEmbedding):
    @staticmethod
    def compute_default_rope_parameters(
        config: GPTNeoXConfig | None = None,
        device: Optional["torch.device"] = None,
        seq_len: int | None = None,
    ) -> tuple["torch.Tensor", float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
        dim = int(head_dim * partial_rotary_factor)

        attention_factor = 1.0  # Unused in this type of RoPE

        # Compute the inverse frequencies
        inv_freq = 1.0 / (
            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
        )
        return inv_freq, attention_factor


def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)

    # Keep half or full tensor for later concatenation
    rotary_dim = cos.shape[-1]
    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]

    # Apply rotary embeddings on the first half or full tensor
    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)

    # Concatenate back to full shape
    q_embed = torch.cat([q_embed, q_pass], dim=-1)
    k_embed = torch.cat([k_embed, k_pass], dim=-1)
    return q_embed, k_embed


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor,
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling

    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)

    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value)

    # Reshape outputs
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class GPTNeoXAttention(nn.Module):
    def __init__(self, config, layer_idx=None):
        super().__init__()
        self.config = config
        self.head_size = config.hidden_size // config.num_attention_heads
        self.attention_dropout = config.attention_dropout
        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
        self.rotary_ndims = int(self.head_size * partial_rotary_factor)
        self.scaling = self.head_size**-0.5
        self.is_causal = True
        self.layer_idx = layer_idx

        self.query_key_value = nn.Linear(config.hidden_size, 3 * config.hidden_size, bias=config.attention_bias)
        self.dense = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attention_bias)

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        attention_mask: torch.FloatTensor,
        layer_past: Cache | None = None,
        cache_position: torch.LongTensor | None = None,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ):
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, 3 * self.head_size)

        qkv = self.query_key_value(hidden_states).view(hidden_shape).transpose(1, 2)
        query_states, key_states, value_states = qkv.chunk(3, dim=-1)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        # Cache QKV values
        if layer_past is not None:
            cache_kwargs = {
                "sin": sin,
                "cos": cos,
                "partial_rotation_size": self.rotary_ndims,
                "cache_position": cache_position,
            }
            key_states, value_states = layer_past.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        # Compute attention
        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            scaling=self.scaling,
            dropout=0.0 if not self.training else self.attention_dropout,
            **kwargs,
        )

        # Reshape outputs and final projection
        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.dense(attn_output)

        return attn_output, attn_weights


class GPTNeoXLayer(GradientCheckpointingLayer):
    def __init__(self, config, layer_idx):
        super().__init__()
        self.use_parallel_residual = config.use_parallel_residual
        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.post_attention_dropout = nn.Dropout(config.hidden_dropout)
        self.post_mlp_dropout = nn.Dropout(config.hidden_dropout)
        self.attention = GPTNeoXAttention(config, layer_idx)
        self.mlp = GPTNeoXMLP(config)

    def forward(
        self,
        hidden_states: torch.FloatTensor | None,
        attention_mask: torch.FloatTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        use_cache: bool | None = False,
        layer_past: Cache | None = None,
        cache_position: torch.LongTensor | None = None,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ):
        attn_output, _ = self.attention(
            self.input_layernorm(hidden_states),
            attention_mask=attention_mask,
            position_ids=position_ids,
            layer_past=layer_past,
            use_cache=use_cache,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        attn_output = self.post_attention_dropout(attn_output)

        if self.use_parallel_residual:
            # pseudocode:
            # x = x + attn(ln1(x)) + mlp(ln2(x))
            mlp_output = self.mlp(self.post_attention_layernorm(hidden_states))
            mlp_output = self.post_mlp_dropout(mlp_output)
            hidden_states = mlp_output + attn_output + hidden_states
        else:
            # pseudocode:
            # x = x + attn(ln1(x))
            # x = x + mlp(ln2(x))
            attn_output = attn_output + hidden_states
            mlp_output = self.mlp(self.post_attention_layernorm(attn_output))
            mlp_output = self.post_mlp_dropout(mlp_output)
            hidden_states = mlp_output + attn_output

        return hidden_states


class GPTNeoXPreTrainedModel(LlamaPreTrainedModel):
    base_model_prefix = "gpt_neox"
    _no_split_modules = ["GPTNeoXLayer"]
    _keys_to_ignore_on_load_unexpected = [r"attention.bias", r"attention.masked_bias"]
    _can_record_outputs = {
        "hidden_states": GPTNeoXLayer,
        "attentions": GPTNeoXAttention,
    }


GPT_NEOX_START_DOCSTRING = None  # Will be picked up by modular
GPT_NEOX_INPUTS_DOCSTRING = None  # Will be picked up by modular


class GPTNeoXModel(LlamaModel):
    def __init__(self, config):
        PreTrainedModel.__init__(self, config)
        self.config = config

        self.embed_in = nn.Embedding(config.vocab_size, config.hidden_size)
        self.emb_dropout = nn.Dropout(config.hidden_dropout)
        self.layers = nn.ModuleList([GPTNeoXLayer(config, i) for i in range(config.num_hidden_layers)])
        self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.rotary_emb = GPTNeoXRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

    def get_input_embeddings(self):
        return self.embed_in

    def set_input_embeddings(self, value):
        self.embed_in = value

    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

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

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        causal_mask = create_causal_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )

        hidden_states = self.emb_dropout(inputs_embeds)
        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
        for layer in self.layers:
            hidden_states = layer(
                hidden_states,
                attention_mask=causal_mask,
                position_ids=position_ids,
                layer_past=past_key_values,
                use_cache=use_cache,
                position_embeddings=position_embeddings,
                cache_position=cache_position,
                **kwargs,
            )

        hidden_states = self.final_layer_norm(hidden_states)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


@auto_docstring(
    custom_intro="""
    GPTNeoX Model with a `language modeling` head on top for CLM fine-tuning.
    """
)
class GPTNeoXForCausalLM(GPTNeoXPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"embed_out.weight": "gpt_neox.embed_in.weight"}
    _tp_plan = {"embed_out": "colwise_gather_output"}
    _pp_plan = {"embed_out": (["hidden_states"], ["logits"])}

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

        self.gpt_neox = GPTNeoXModel(config)
        self.embed_out = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

    def get_output_embeddings(self):
        return self.embed_out

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        past_key_values: Cache | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | CausalLMOutputWithPast:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
            `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are
            ignored (masked), the loss is only computed for the tokens with labels n `[0, ..., config.vocab_size]`.

        Example:

        ```python
        >>> from transformers import AutoTokenizer, GPTNeoXForCausalLM, GPTNeoXConfig
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
        >>> config = GPTNeoXConfig.from_pretrained("EleutherAI/gpt-neox-20b")
        >>> config.is_decoder = True
        >>> model = GPTNeoXForCausalLM.from_pretrained("EleutherAI/gpt-neox-20b", config=config)

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> outputs = model(**inputs)

        >>> prediction_logits = outputs.logits
        ```"""

        outputs: BaseModelOutputWithPast = self.gpt_neox(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            past_key_values=past_key_values,
            use_cache=use_cache,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        # 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.embed_out(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)

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


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

    [`GPTNeoXForSequenceClassification`] 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 GPTNeoXForSequenceClassification(GPTNeoXPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.gpt_neox = GPTNeoXModel(config)
        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        past_key_values: Cache | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        **kwargs,
    ) -> SequenceClassifierOutputWithPast:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """

        outputs: BaseModelOutputWithPast = self.gpt_neox(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            past_key_values=past_key_values,
            use_cache=use_cache,
            **kwargs,
        )
        hidden_states = outputs.last_hidden_state
        logits = self.score(hidden_states)

        batch_size = logits.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:
            loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config)

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


class GPTNeoXForTokenClassification(GPTNeoXPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.gpt_neox = GPTNeoXModel(config)
        self.dropout = nn.Dropout(config.classifier_dropout)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        **kwargs,
    ) -> TokenClassifierOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *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).
        """

        outputs: BaseModelOutputWithPast = self.gpt_neox(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        hidden_states = self.dropout(hidden_states)
        logits = self.classifier(hidden_states)

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

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


@auto_docstring
class GPTNeoXForQuestionAnswering(GPTNeoXPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.gpt_neox = GPTNeoXModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, 2)

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        start_positions: torch.LongTensor | None = None,
        end_positions: torch.LongTensor | None = None,
        **kwargs,
    ) -> QuestionAnsweringModelOutput:
        outputs: BaseModelOutputWithPast = self.gpt_neox(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

        sequence_output = outputs.last_hidden_state

        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()

        loss = None
        if start_positions is not None and end_positions is not None:
            loss = self.loss_function(start_logits, end_logits, start_positions, end_positions)

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


__all__ = [
    "GPTNeoXForCausalLM",
    "GPTNeoXForQuestionAnswering",
    "GPTNeoXForSequenceClassification",
    "GPTNeoXForTokenClassification",
    "GPTNeoXLayer",
    "GPTNeoXModel",
    "GPTNeoXPreTrainedModel",
]