modeling_gigarembed.py

import copy
import logging
from typing import Callable, List, Optional, Tuple, Union, Mapping

import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F

from einops import rearrange, repeat
from transformers import AutoModel, AutoTokenizer

from transformers.cache_utils import Cache
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS

from transformers.activations import ACT2FN
from transformers.cache_utils import DynamicCache, StaticCache
from transformers.generation import GenerationMixin
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
from transformers.modeling_utils import PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, replace_return_docstrings

from .configuration_gigarembed import GigarConfig, GigarEmbedConfig, LatentAttentionConfig


logger = logging.getLogger(__name__)
_CONFIG_FOR_DOC = "GigarEmbedConfig"


class GigarMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


class GigarRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        GigarRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, 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.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        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)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_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_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class GigarLatentAttention(nn.Module):
    """
    Multi-headed Latent Attention (MLA)

    Check out the original paper: https://arxiv.org/pdf/2405.04434,
    and the reference implementation: https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
    """

    def __init__(self, config: GigarConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)

        assert config.num_attention_heads == config.num_key_value_heads, (
            "GQA for MLA is not supported (does it even make sense?)"
        )
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads

        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.apply_qk_norm = config.apply_qk_norm
        self.attention_dropout = config.attention_dropout

        assert config.mla_config is not None
        self.qk_nope_head_dim = config.mla_config["qk_nope_head_dim"]
        self.qk_rope_head_dim = config.mla_config["qk_rope_head_dim"]
        self.v_head_dim = config.mla_config["v_head_dim"]  # V has no rope part
        self.kv_lora_rank = config.mla_config["kv_lora_rank"]
        self.q_lora_rank = config.mla_config["q_lora_rank"]

        self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim

        self.scaling = self.qk_head_dim**-0.5

        if self.q_lora_rank == 0:
            self.q_proj = nn.Linear(
                self.hidden_size,
                self.num_heads * self.qk_head_dim,
                bias=config.attention_bias,
            )
        else:
            self.dq_proj = nn.Linear(
                self.hidden_size,
                self.q_lora_rank,
                bias=config.attention_bias,
            )
            self.q_norm = GigarRMSNorm(self.q_lora_rank)
            self.uq_proj = nn.Linear(
                self.q_lora_rank,
                self.num_heads * self.qk_head_dim,
                bias=config.attention_bias,
            )

        self.kv_norm = GigarRMSNorm(self.kv_lora_rank)
        self.dkv_proj = nn.Linear(
            self.hidden_size,
            self.kv_lora_rank,
            bias=config.attention_bias,
        )
        self.uk_proj = nn.Linear(
            config.kv_lora_rank,
            self.num_heads * self.qk_nope_head_dim,
            bias=config.attention_bias,
        )
        self.uv_proj = nn.Linear(
            config.kv_lora_rank,
            self.num_heads * self.v_head_dim,
            bias=config.attention_bias,
        )
        self.kr_proj = nn.Linear(
            self.hidden_size,
            self.num_heads * self.qk_rope_head_dim,
            bias=config.attention_bias,
        )

        self.o_proj = nn.Linear(
            self.num_heads * self.v_head_dim,
            self.hidden_size,
            bias=config.attention_bias,
        )

        if self.apply_qk_norm:
            self.qk_q_norm = nn.LayerNorm(self.num_heads * self.qk_head_dim, bias=False)
            self.qk_k_norm = nn.LayerNorm(self.num_heads * self.qk_head_dim, bias=False)

        config_for_rope = copy.copy(self.config)
        config_for_rope.head_dim = self.config.qk_rope_head_dim

        self.is_causal = False

    def _compute_qkv(
        self,
        hidden_states: torch.Tensor,
    ):
        """Compute query, key, and value tensors from hidden states."""
        bsz, seq_len, _ = hidden_states.size()

        if self.q_lora_rank == 0:
            query = self.q_proj(hidden_states)
        else:
            query = self.uq_proj(self.q_norm(self.dq_proj(hidden_states)))

        latent = self.dkv_proj(hidden_states)
        latent = self.kv_norm(latent)
        k_rope = self.kr_proj(hidden_states)

        k_nope = self.uk_proj(latent)
        value = self.uv_proj(latent)

        if self.apply_qk_norm:
            query = self.qk_q_norm(query).to(query.dtype)
            key = self.qk_k_norm(torch.cat([k_nope, k_rope], dim=-1)).to(k_nope.dtype)
            k_nope, k_rope = torch.split(key, [k_nope.shape[-1], k_rope.shape[-1]], dim=-1)

        # Reshape tensors
        query = query.view(bsz, seq_len, self.num_heads, self.qk_head_dim).transpose(1, 2)
        k_nope = k_nope.view(bsz, seq_len, self.num_heads, self.qk_nope_head_dim).transpose(1, 2)
        k_rope = k_rope.view(bsz, seq_len, self.num_heads, self.qk_rope_head_dim).transpose(1, 2)
        value = value.view(bsz, seq_len, self.num_heads, self.v_head_dim).transpose(1, 2)

        q_nope, q_rope = torch.split(query, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)

        return q_nope, q_rope, k_nope, k_rope, value

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        """
        hidden_states: [bsz, seq_len, hidden_size]
        attention_mask: [bsz, seq_len]
        """
        batch_size, seq_len, _ = hidden_states.size()

        q_nope, q_rope, k_nope, k_rope, value_states = self._compute_qkv(hidden_states)

        # cos, sin = self.rotary_emb(q_rope, seq_len=seq_len)
        cos, sin = position_embeddings
        q_rope, k_rope = apply_rotary_pos_emb(q_rope, k_rope, cos, sin)
        query_states = torch.cat([q_nope, q_rope], dim=-1)
        key_states = torch.cat([k_nope, k_rope], dim=-1)

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

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

        attn_output = attn_output.reshape(batch_size, seq_len, -1).contiguous()
        attn_output = self.o_proj(attn_output)

        return attn_output, attn_weights


class GigarDecoderLayer(nn.Module):
    def __init__(self, config: GigarConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = GigarLatentAttention(config, layer_idx)
        self.mlp = GigarMLP(config)
        self.input_layernorm = GigarRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = GigarRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)
        if output_attentions:
            outputs += (self_attn_weights,)

        return outputs


class GigarRotaryEmbedding(nn.Module):
    def __init__(self, config: GigarConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()

        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
        cos = cos * self.attention_scaling
        sin = sin * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


GIGAR_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`GigarConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    "The bare Gigar Model outputting raw hidden-states without any specific head on top.",
    GIGAR_START_DOCSTRING,
)
class GigarPreTrainedModel(PreTrainedModel):
    config_class = GigarConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["GigarDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_cache_class = True
    _supports_quantized_cache = True
    _supports_static_cache = True

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()


GIGAR_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` 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 (`torch.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)

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

            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
            `past_key_values`).

            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
            information on the default strategy.

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`.

            [What are position IDs?](../glossary#position-ids)
        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.

            Two formats are allowed:
            - a [`~cache_utils.Cache`] instance, see our
            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
            cache format.

            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
            legacy cache format will be returned.

            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
            of shape `(batch_size, sequence_length)`.
        inputs_embeds (`torch.FloatTensor` 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.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        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.
        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
            the complete sequence length.
"""

@add_start_docstrings(
    "The bare Gigar Model outputting raw hidden-states without any specific head on top.",
    GIGAR_START_DOCSTRING,
)
class GigarModel(GigarPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`GigarDecoderLayer`]

    Args:
        config: GigarConfig
    """

    def __init__(self, config: GigarConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [GigarDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = GigarRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = GigarRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

    def get_input_embeddings(self):
        return self.embed_tokens

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

    @add_start_docstrings_to_model_forward(GIGAR_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        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 None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if self.gradient_checkpointing and self.training and 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.embed_tokens(input_ids)

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache()

        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)
        
        attention_mask = self._update_encoder_mask(attention_mask, inputs_embeds)

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

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

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states,
                    attention_mask,  # causal_mask
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                    cache_position,
                    position_embeddings,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,  # causal_mask
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    cache_position=cache_position,
                    position_embeddings=position_embeddings,
                    **flash_attn_kwargs,
                )

            hidden_states = layer_outputs[0]

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

        hidden_states = self.norm(hidden_states)

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

        output = BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )
        return output if return_dict else output.to_tuple()
    
    def _update_encoder_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
    ):
        # Для flash_attention_2 возвращаем исходную маску
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and (attention_mask == 0).any():
                return attention_mask
            return None

        dtype, device = input_tensor.dtype, input_tensor.device
        batch_size, sequence_length = input_tensor.shape[:2]

        # 1. Создаём базовую маску без ограничений (все токены видят друг друга)
        encoder_mask = torch.full(
            (batch_size, 1, sequence_length, sequence_length),
            fill_value=1.0,
            dtype=dtype,
            device=device
        )

        # 2. Применяем padding-маску если есть
        if attention_mask is not None:
            # Создаём 4D padding-маску [batch, 1, 1, seq_len]
            padding_mask = attention_mask[:, None, None, :].to(dtype=dtype)

            # Комбинируем: обнуляем позиции где padding_mask == 0
            encoder_mask = encoder_mask * padding_mask

            # Конвертируем в формат для softmax (0 = -inf)
            min_dtype = torch.finfo(dtype).min
            encoder_mask = encoder_mask.masked_fill(encoder_mask == 0.0, min_dtype)

        return encoder_mask

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and (attention_mask == 0.0).any():
                return attention_mask
            return None

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        sequence_length = input_tensor.shape[1]
        if using_static_cache:
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type == "cuda"
            and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            min_dtype = torch.finfo(dtype).min
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        device: torch.device,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            device (`torch.device`):
                The device to plcae the 4D attention mask on.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask


class FeedForward(nn.Module):
    def __init__(self, dim, mult = 4):
        super().__init__()
        self.hidden_size = dim
        self.intermediate_size = dim * mult
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = nn.SiLU()

    def forward(self, x):
        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))


class Attention(nn.Module):
    def __init__(self, query_dimension, context_dimension=None, num_heads=8, head_dim=64):
        super().__init__()
        inner_dimension = head_dim * num_heads
        context_dimension = context_dimension if context_dimension is not None else query_dimension

        self.scaling_factor = head_dim ** -0.5
        self.num_heads = num_heads

        self.to_q = nn.Linear(query_dimension, inner_dimension, bias=False)
        self.to_kv = nn.Linear(context_dimension, inner_dimension * 2, bias=False)
        self.to_out = nn.Linear(inner_dimension, query_dimension, bias=False)

    def forward(self, input_tensor, context=None, attention_mask=None):
        batch_size, seq_len, _ = input_tensor.shape
        num_heads = self.num_heads

        # Project input to query
        query = self.to_q(input_tensor)

        # Use input as context if not provided
        context = input_tensor if context is None else context
        key, value = self.to_kv(context).chunk(2, dim=-1)

        # Rearrange for multi-head attention
        query = rearrange(query, 'b n (h d) -> (b h) n d', h=num_heads)
        key = rearrange(key, 'b n (h d) -> (b h) n d', h=num_heads)
        value = rearrange(value, 'b n (h d) -> (b h) n d', h=num_heads)

        # Compute scaled dot-product attention
        with torch.backends.cuda.sdp_kernel(
            enable_flash=True, 
            enable_math=True, 
            enable_mem_efficient=True
        ):
            attention_output = F.scaled_dot_product_attention(query, key, value)

        # Rearrange back to original shape
        attention_output = rearrange(attention_output, '(b h) n d -> b n (h d)', h=num_heads)

        return self.to_out(attention_output)


class LatentAttentionModel(PreTrainedModel):
    config_class = LatentAttentionConfig

    def __init__(self, configuration: LatentAttentionConfig):
        super().__init__(configuration)

        # Extract configuration parameters
        num_latents = configuration.num_latents_value
        latent_dimension = configuration.latent_dim
        cross_attention_heads = configuration.num_cross_heads
        cross_head_dimension = configuration.cross_dim_head
        hidden_dimension = configuration.hidden_dim

        # Initialize cross-attention components
        self.cross_attend_blocks = nn.ModuleList([
            Attention(
                query_dimension=latent_dimension,
                context_dimension=hidden_dimension,
                num_heads=cross_attention_heads,
                head_dim=cross_head_dimension
            ),
            FeedForward(latent_dimension)
        ])

        # Register learnable latents as model parameter
        self.latents = nn.Parameter(torch.randn(num_latents, latent_dimension))

    def forward(self, hidden_states, attention_mask: Optional[torch.Tensor] = None):
        cross_attention, feed_forward = self.cross_attend_blocks
        
        batch_size, device = hidden_states.size(0), hidden_states.device
        
        # Expand latents to match batch size
        expanded_latents = self.latents.repeat(batch_size, 1, 1)
        
        # Apply cross-attention with residual connection
        attended_output = cross_attention(
            hidden_states, context=expanded_latents, attention_mask=attention_mask) + hidden_states
        
        # Apply feed-forward with residual connection
        processed_output = feed_forward(attended_output) + attended_output
        
        return processed_output


class GigarEmbedModel(PreTrainedModel):
    config_class = GigarEmbedConfig
    _supports_flash_attn_2 = True
    _no_split_modules = ["GigarDecoderLayer", "LatentAttentionModel"]
    
    def __init__(self, configuration: GigarEmbedConfig):
        super().__init__(configuration)
        
        # Initialize latent attention model
        self.latent_attention_model = AutoModel.from_config(
            configuration.latent_attention_config
        )

        self.tokenizer, self.text_encoder = None, None
        if configuration.text_config is not None:
            # Initialize text model if provided in config
            self.model = AutoModel.from_config(configuration.text_config)

            # Initialize tokenizer if text config is available
            self.tokenizer = AutoTokenizer.from_pretrained(
                configuration.text_config.name_or_path
            )
        
        # Set configuration parameters
        self.padding_side = configuration.padding_side
        self.add_eos = configuration.add_eos
        self.mask_type = configuration.mask_type
        
        # Add padding token if configured
        if configuration.add_pad_token and self.tokenizer is not None:
            self.add_pad_token()

    def add_pad_token(self):
        self.tokenizer.pad_token_id = 0
        self.tokenizer.padding_side = self.padding_side

    def gradient_checkpointing_enable(self, *args, **kwargs):
        self.model.gradient_checkpointing_enable(*args, **kwargs)

    def forward(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, 
                return_embeddings: bool = False, **kwargs):
        kwargs.pop('token_type_ids', None)

        with torch.autocast('cuda', dtype=torch.bfloat16):
            outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, **kwargs)

            last_hidden = self.latent_attention_model(outputs.last_hidden_state, attention_mask)

        if return_embeddings:
            return self.mean_pool(last_hidden, attention_mask)

        return BaseModelOutputWithPast(last_hidden_state=last_hidden)

    def mean_pool(self, last_hidden: torch.Tensor, attention_mask: torch.Tensor):
        last_hidden = last_hidden.masked_fill(~attention_mask[..., None].bool(), 0.0)
        embeddings = last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
        return F.normalize(embeddings, p=2, dim=-1)


## AutoModel Register
AutoModel.register(GigarConfig, GigarModel)
AutoModel.register(GigarEmbedConfig, GigarEmbedModel)
AutoModel.register(LatentAttentionConfig, LatentAttentionModel)

## Register for auto class
GigarModel.register_for_auto_class("AutoModel")
GigarEmbedModel.register_for_auto_class("AutoModel")
LatentAttentionModel.register_for_auto_class("AutoModel")