"""
LLaMA intermediate layer
Based on https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
See commit history for authorship.
"""
from typing import Optional, Tuple

import torch
from transformers.models.llama.modeling_llama import LlamaDecoderLayer, LlamaModel


class WrappedLlamaBlock(LlamaDecoderLayer):
    def forward(
        self,
        hidden_states: torch.Tensor,
        *args,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        use_cache: bool = False,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        batch_size, seq_length, _ = hidden_states.shape

        seq_length_with_past = seq_length
        past_key_values_length = 0

        past_key_value = layer_past
        if past_key_value is not None:
            past_key_values_length = past_key_value[0].shape[2]
            seq_length_with_past = seq_length_with_past + past_key_values_length
            past_key_value = self._reorder_cache_from_bloom_to_llama(past_key_value, batch_size, past_key_values_length)

        if position_ids is None:
            device = hidden_states.device
            position_ids = torch.arange(
                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
            )
            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
        else:
            position_ids = position_ids.view(-1, seq_length).long()

        # embed positions
        if attention_mask is None:
            attention_mask = torch.ones(
                (batch_size, seq_length_with_past), dtype=torch.bool, device=hidden_states.device
            )
        attention_mask = LlamaModel._prepare_decoder_attention_mask(
            None, attention_mask, (batch_size, seq_length), hidden_states, past_key_values_length
        )

        outputs = super().forward(
            hidden_states,
            *args,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            use_cache=use_cache,
            **kwargs,
        )

        if use_cache:
            present_key_value = outputs[-1]
            present_key_value = self._reorder_cache_from_llama_to_bloom(
                present_key_value, batch_size, seq_length_with_past
            )
            outputs = outputs[:-1] + (present_key_value,)

        return outputs

    def _reorder_cache_from_bloom_to_llama(
        self, key_value: Tuple[torch.Tensor], batch_size: int, seq_length: int
    ) -> Tuple[torch.Tensor]:
        key_states, value_states = key_value
        key_states = key_states.permute(0, 2, 1)
        key_states = key_states.view(
            batch_size, self.self_attn.num_key_value_heads, seq_length, self.self_attn.head_dim
        )
        value_states = value_states.view(*key_states.shape)
        return (key_states, value_states)

    def _reorder_cache_from_llama_to_bloom(
        self, key_value: Tuple[torch.Tensor], batch_size: int, seq_length: int
    ) -> Tuple[torch.Tensor]:
        key_states, value_states = key_value
        value_states = value_states.view(
            batch_size * self.self_attn.num_key_value_heads, seq_length, self.self_attn.head_dim
        )
        key_states = key_states.view(*value_states.shape)
        key_states = key_states.permute(0, 2, 1)
        return (key_states, value_states)