language_helium1_casa.py

# ADAPTED FROM https://github.com/huggingface/transformers/blob/main/src/transformers/models/helium/modeling_helium.py
# GIT HASH 1b222903c3e1cfd9492d75e4b2548aa8bd458674
import logging
import math
from dataclasses import dataclass
from functools import partial
from typing import Any, Callable, Literal, Optional
from typing import cast as type_cast

import torch
from torch import nn
from transformers import (
    ROPE_INIT_FUNCTIONS,  # pyright: ignore[reportPrivateImportUsage]
    dynamic_rope_update,  # pyright: ignore[reportPrivateImportUsage]
)
from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.configuration_utils import PretrainedConfig
from transformers.generation.utils import GenerationMixin
from transformers.loss.loss_utils import ForCausalLMLoss
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_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils.generic import LossKwargs, can_return_tuple
from transformers.utils.import_utils import is_torch_flex_attn_available

from .casa_attention import CASAAttention, CASAAttentionHandler, insert_image_tokens
from .configuration_helium1_casa import Helium1CASAConfig

logger = logging.getLogger(__name__)

if is_torch_flex_attn_available():
    from transformers.integrations.flex_attention import make_flex_block_causal_mask


def remove_image_tokens(
    inputs_embeds: torch.Tensor,
    image_tokens_mask: torch.Tensor,
) -> torch.Tensor:
    """Remove the image tokens from inputs_embeds as indicated by image_tokens_mask

    :param inputs_embeds: Tokens of shape (Batch, Seqlen, Dims) containing image tokens
    :param image_tokens_mask: 1-0 mask indicating where image tokens are; (Batch, Seqlen)

    :return: Tokens tensor of shape (Batch, S' < Seqlen, Dims)
    """
    image_seq_lengths = torch.sum(image_tokens_mask, dim=1)[:, 0]
    image_seq_length = int(image_seq_lengths[0].item())
    assert torch.all(image_seq_lengths == image_seq_length)
    new_shape = (
        inputs_embeds.shape[0],
        inputs_embeds.shape[1] - image_seq_length,
        inputs_embeds.shape[-1],
    )
    tokens = torch.masked_select(
        inputs_embeds,
        torch.logical_not(image_tokens_mask).expand((-1, -1, inputs_embeds.shape[-1])),
    )
    return tokens.reshape(new_shape)


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: "HeliumAttention",
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: None | torch.Tensor,
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Any,
):
    del kwargs  # unused
    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


# Different Attention Classes
class HeliumAttention(torch.nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Helium1CASAConfig, layer_idx: None | int = None):
        super().__init__()
        self.config = config
        assert layer_idx is not None
        self.layer_idx: int = layer_idx

        self.apply_rotary_fn = ApplyRotaryPosEmbHelium1()
        self.head_dim = getattr(
            config, "head_dim", config.hidden_size // config.num_attention_heads
        )
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = 1 / math.sqrt(self.head_dim)
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size,
            config.num_attention_heads * self.head_dim,
            bias=config.attention_bias,
        )
        self.k_proj = nn.Linear(
            config.hidden_size,
            config.num_key_value_heads * self.head_dim,
            bias=config.attention_bias,
        )
        self.v_proj = nn.Linear(
            config.hidden_size,
            config.num_key_value_heads * self.head_dim,
            bias=config.attention_bias,
        )
        self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: None | torch.Tensor,
        past_key_values: None | Cache = None,
        cache_position: None | torch.LongTensor = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        # del (cache_position, past_key_value)  # we use our own generate/caching
        bs, seq_len, _ = hidden_states.shape
        # Get QKV
        hidden_shape = (bs, seq_len, -1, self.head_dim)

        # Embed Queries
        # Shape: (batch_size, num_heads, seq_len, head_dim)
        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        num_queries = query_states.shape[2]

        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        # Applies rotation
        cos, sin = position_embeddings
        query_states, key_states = self.apply_rotary_fn(
            query_states, key_states, cos, sin, num_queries=num_queries
        )
        assert key_states is not None and query_states is not None

        attention_interface: Callable = eager_attention_forward

        if self.config._attn_implementation != "eager":
            if self.config._attn_implementation == "sdpa" and kwargs.get(
                "output_attentions", False
            ):
                print(
                    "`torch.nn.functional.scaled_dot_product_attention` does not support"
                    " `output_attentions=True`. Falling back to "
                    'eager attention. This warning can be removed using the argument"\
                    " `attn_implementation="eager"` when loading the model.'
                )
            else:
                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        if past_key_values 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_values.update(
                key_states, value_states, self.layer_idx, cache_kwargs
            )
        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(bs, num_queries, -1).contiguous()
        attn_output = self.o_proj(attn_output)

        assert isinstance(attn_output, torch.Tensor)
        return attn_output, attn_weights


class ApplyRotaryPosEmbHelium1:
    @staticmethod
    def rotate_half(x: torch.Tensor) -> torch.Tensor:
        """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)

    @staticmethod
    def __call__(
        q: torch.Tensor,
        k: torch.Tensor,
        cos: torch.Tensor,
        sin: torch.Tensor,
        position_ids: torch.Tensor | None = None,
        unsqueeze_dim: int = 1,
        num_queries: int | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """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.
        """
        del position_ids
        cos = cos.unsqueeze(unsqueeze_dim)
        sin = sin.unsqueeze(unsqueeze_dim)
        if num_queries is None:
            offset = 0
        else:
            offset = -num_queries

        q_embed = (q * cos[:, :, offset:]) + (
            ApplyRotaryPosEmbHelium1.rotate_half(q) * sin[:, :, offset:]
        )
        k_embed = (k * cos) + (ApplyRotaryPosEmbHelium1.rotate_half(k) * sin)

        return q_embed, k_embed


class HeliumRotaryEmbedding(nn.Module):
    def __init__(self, config: Helium1CASAConfig, device: None | torch.device | str = None):
        super().__init__()
        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
        assert self.rope_type in ROPE_INIT_FUNCTIONS, (
            f"Invalid rope type {self.rope_type}. Supported types are: {list(ROPE_INIT_FUNCTIONS.keys())}"
        )
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
        inv_freq, self.attention_scaling = self.rope_init_fn(config, device=device)
        self.inv_freq: torch.Tensor  # only defined for typing
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(
        self, x: torch.Tensor, position_ids: torch.Tensor
    ) -> tuple[torch.Tensor, torch.Tensor]:
        inv_freq_expanded = (
            self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        )
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = (
            x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        )
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

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


class Helium1CASAAttention(CASAAttention):
    """A CASA Attention layer compatible with Qwen"""

    def __init__(
        self,
        config: Helium1CASAConfig,
        layer_idx: int | None,
        self_attn: torch.nn.Module | None = None,
        input_layernorm_fn: Callable[[torch.Tensor], torch.Tensor] | None = None,
    ):
        # Only adding  this init for typing purposes for the config
        super().__init__(config, layer_idx, self_attn, input_layernorm_fn)  # pyright: ignore[reportArgumentType]

    @staticmethod
    def rotate_half(x: torch.Tensor) -> torch.Tensor:
        """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_position_embeddings(
        self,
        key: Literal["q", "kv"],
        x: torch.Tensor,  # (batch, seq_len, num_heads, head_dim)
        casa_handler: CASAAttentionHandler | None,
        num_queries: int = 0,
        unsqueeze_dim: int = 1,
    ) -> torch.Tensor:  # (batch, seq_len, num_heads, head_dim)
        """Apply position embeddings to query and key states"""
        if casa_handler is not None:
            posemb = casa_handler.get_position_embedding(key, num_queries=num_queries)

            if posemb is not None:
                x = x.transpose(1, 2).to(torch.float32)
                x = (x * posemb[0].unsqueeze(dim=unsqueeze_dim)) + (
                    self.rotate_half(x) * posemb[1].unsqueeze(dim=unsqueeze_dim)
                )
                return x.transpose(1, 2)
        return x

    def init_from_config_proj(
        self, key: Literal["q", "o", "k", "v"], config: PretrainedConfig
    ) -> torch.nn.Linear:
        """Initialize the Linear proj in this module"""
        num_heads = config.num_key_value_heads if key in {"k", "v"} else config.num_attention_heads
        return torch.nn.Linear(
            config.hidden_size,
            num_heads * config.head_dim,
            bias=config.attention_bias if key != "o" else False,
        )


# NORMALISATION LAYER
def __rms_norm_forward__(
    hidden_states: torch.Tensor, weight: torch.Tensor, variance_epsilon: float = 1e-6
) -> torch.Tensor:
    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 + variance_epsilon)
    return weight * hidden_states.to(input_dtype)


class Helium1RMSNorm(nn.Module):
    def __init__(self, hidden_size: int, eps: float = 1e-6) -> None:
        """
        Helium1RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        return __rms_norm_forward__(hidden_states, self.weight, self.variance_epsilon)

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


def delta_w_factory_rms_norm(
    org_lin: Helium1RMSNorm, new_lin: Helium1RMSNorm
) -> Callable[[torch.Tensor], torch.Tensor]:
    """Factory for building rms norm where the weights are the sum of two layers' weights"""

    def _delta_w_fwd(input: torch.Tensor) -> torch.Tensor:
        nonlocal org_lin, new_lin
        return __rms_norm_forward__(
            input, org_lin.weight + new_lin.weight, new_lin.variance_epsilon
        )

    return _delta_w_fwd


# FULL CONNECTED LAYER


class HeliumMLP(nn.Module):
    def __init__(self, config: Helium1CASAConfig) -> None:
        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: torch.Tensor) -> torch.Tensor:
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


class HeliumDecoderLayer(nn.Module):
    def __init__(self, config: Helium1CASAConfig, layer_idx: None | int = None):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.config = config
        self.mlp = HeliumMLP(config)
        self.input_layernorm = Helium1RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Helium1RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        # Self-attention
        self.self_attn = HeliumAttention(config=config, layer_idx=layer_idx)

        # Setup norm for fusion mechanisms; Note that this norm is on the text tokens
        is_xa_layer = layer_idx is None or not config.xa_layers or layer_idx in config.xa_layers
        self.norm_cross: None | Helium1RMSNorm = None
        self.override_norm_cross: Callable[[torch.Tensor], torch.Tensor] | None = None
        if is_xa_layer and config.casa_attention:
            # Custom normalization layer for the extra fusion module
            if self.config.xa_custom_norm:
                self.norm_cross = Helium1RMSNorm(config.hidden_size)
                if config.casa_delta_w:
                    self.override_norm_cross = delta_w_factory_rms_norm(
                        self.input_layernorm, self.norm_cross
                    )
                    with torch.no_grad():
                        torch.nn.init.ones_(self.norm_cross.weight)

        # Setup additional norm for images tokens which is set in each individual mechansims
        norm_on_images_fn = (
            None
            if not self.config.xa_norm_on_images
            else self.override_norm_cross
            if self.override_norm_cross is not None
            else self.norm_cross.forward
            if self.norm_cross is not None
            else self.input_layernorm.forward
        )

        # CASA
        self.casa_attn: Helium1CASAAttention | None = None
        if config.casa_attention and is_xa_layer:
            self.casa_attn = Helium1CASAAttention(
                config, layer_idx, self_attn=self.self_attn, input_layernorm_fn=norm_on_images_fn
            )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: None | torch.Tensor = None,
        position_ids: None | torch.LongTensor = None,
        past_key_values: None | Cache = None,
        output_attentions: None | bool = False,
        use_cache: None | bool = False,
        cache_position: None | torch.LongTensor = None,
        position_embeddings: None
        | tuple[torch.Tensor, torch.Tensor] = None,  # necessary, but kept here for BC
        # CASA
        casa_handler: CASAAttentionHandler | None = None,
        cu_seqlens: torch.Tensor | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor] | tuple[torch.Tensor]:
        # Image fusion mechanisms
        apply_ca = self.casa_attn is not None
        ca_update: torch.Tensor | None = None
        if (
            self.config.xa_order
            in {
                "parallel",
                "ca_first",
                "instead",
            }
            and apply_ca
        ):
            # Apply layer norm
            assert self.norm_cross is not None
            ca_input = (
                self.override_norm_cross
                if self.override_norm_cross is not None
                else self.norm_cross
            )(hidden_states)
            # CASA
            if self.casa_attn is not None:
                ca_update = self.casa_attn(ca_input, casa_handler=casa_handler)

            # If we're here, it's because we had proper inputs (no text-only samples)
            # so the output better be not None !
            if ca_update is not None:
                # `instead`: directly return the output of the CA module as residual
                if self.config.xa_order == "instead":
                    outputs = (hidden_states + ca_update,)
                    if output_attentions:
                        outputs += (
                            torch.zeros((), device=ca_update.device, dtype=ca_update.dtype),
                        )
                    return outputs

                # `ca_first`: update then continue with normal self-attention
                if self.config.xa_order == "ca_first":
                    hidden_states = hidden_states + ca_update
                    ca_update = None

        # Self Attention with initial input layer norm
        residual = hidden_states
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=self.input_layernorm(hidden_states),
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            output_attentions=output_attentions,
            use_cache=use_cache,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            cu_seqlens=cu_seqlens,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # parallel - residual update
        if self.config.xa_order == "parallel" and apply_ca and ca_update is not None:
            hidden_states = hidden_states + ca_update

        # Fully Connected layer
        residual = hidden_states
        # MLP updates for image embeddings
        if (
            self.config.xa_update_image_embeds
            and self.casa_attn is not None
            and casa_handler is not None
            and casa_handler.image_embeds is not None
        ):
            # Text flattening
            hs = self.post_attention_layernorm(hidden_states).reshape(-1, hidden_states.shape[-1])
            # Image flattening
            img_seq_lengths = [_x.shape[0] for _x in casa_handler.image_embeds]
            img_residual = torch.cat(list(casa_handler.image_embeds), dim=0)
            update = self.mlp(torch.cat([hs, self.post_attention_layernorm(img_residual)], dim=0))
            # update text
            hidden_states = hidden_states + update[: hs.shape[0]].reshape(hidden_states.shape)
            casa_handler.image_embeds = list(
                torch.split(img_residual + update[hs.shape[0] :], img_seq_lengths)
            )
        else:
            hidden_states = self.mlp(self.post_attention_layernorm(hidden_states))
            hidden_states = residual + hidden_states

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

        return outputs


# FULL HELIUM MODEL


@dataclass
class CausalHeliumOutput(CausalLMOutputWithPast):
    attention_mask: Optional[torch.Tensor] = None
    num_image_tokens_log: Optional[torch.Tensor] = None
    num_text_tokens_log: Optional[torch.Tensor] = None


class Helium1PreTrainedModel(PreTrainedModel):
    config_class = Helium1CASAConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["HeliumDecoderLayer"]
    _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
    _supports_attention_backend = True

    def _init_weights(self, module: torch.nn.Module) -> None:
        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_()
        elif isinstance(module, Helium1RMSNorm):
            module.weight.data.fill_(1.0)


class Helium1Model(Helium1PreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]

    Args:
        config: Helium1CASAConfig
    """

    def __init__(self, config: Helium1CASAConfig):
        Helium1PreTrainedModel.__init__(self, config)
        self.training: bool
        self._gradient_checkpointing_func: Callable
        self.config = 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(
            [HeliumDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Helium1RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = HeliumRotaryEmbedding(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: nn.Module) -> None:
        self.embed_tokens = value

    @can_return_tuple
    def forward(
        self,
        input_ids: None | torch.LongTensor = None,
        attention_mask: None | torch.Tensor = None,
        position_ids: None | torch.Tensor = None,
        past_key_values: None | DynamicCache = None,
        inputs_embeds: None | torch.Tensor = None,
        use_cache: None | bool = None,
        output_attentions: None | bool = None,
        output_hidden_states: None | bool = None,
        cache_position: None | torch.Tensor = None,
        # Insertion
        image_tokens_mask: torch.Tensor | None = None,
        # CASA
        casa_handler: CASAAttentionHandler | None = None,
        cu_seqlens: torch.Tensor | None = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> 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 = not self.training and (
            use_cache if use_cache is not None else self.config.use_cache
        )

        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:
            print(
                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
            )
            use_cache = False

        # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
        if not isinstance(past_key_values, (type(None), Cache)):
            raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)
        assert inputs_embeds is not None

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

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

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

        # Get attention mask
        causal_mask: None | torch.Tensor = self._update_causal_mask(
            attention_mask,
            inputs_embeds,
            cache_position,
            past_key_values,
            output_attentions,
            force_mask=False,
        )

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

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

        for decoder_layer_idx, decoder_layer in enumerate(
            self.layers[: self.config.num_hidden_layers]
        ):
            is_xa_layer = not self.config.xa_layers or decoder_layer_idx in self.config.xa_layers
            if output_hidden_states is not None:
                if all_hidden_states is None:
                    all_hidden_states = ()
                all_hidden_states += (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                    hidden_states,
                    causal_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                    cache_position,
                    position_embeddings,
                    casa_handler if is_xa_layer else None,
                    cu_seqlens,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=causal_mask,
                    position_ids=position_ids,
                    past_key_values=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    cache_position=cache_position,
                    position_embeddings=position_embeddings,
                    casa_handler=casa_handler if is_xa_layer else None,
                    cu_seqlens=cu_seqlens,
                    **flash_attn_kwargs,
                )

            hidden_states = layer_outputs[0]

            if output_attentions:
                if all_self_attns is None:
                    all_self_attns = ()
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            if all_hidden_states is None:
                all_hidden_states = ()
            all_hidden_states += (hidden_states,)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,  # pyright: ignore[reportArgumentType]
            hidden_states=all_hidden_states,  # pyright: ignore[reportArgumentType]
            attentions=all_self_attns,
        )

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor | None,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: None | DynamicCache | Cache,
        output_attentions: bool = False,
        force_mask: bool = False,
    ) -> torch.Tensor | None:
        if self.config._attn_implementation == "flex_attention":
            if isinstance(attention_mask, torch.Tensor):
                attention_mask = make_flex_block_causal_mask(attention_mask)  # type: ignore
            return attention_mask

        assert attention_mask is None or isinstance(attention_mask, torch.Tensor)
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and (force_mask or (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_compilable_cache = (
            past_key_values.is_compileable if past_key_values is not None else False
        )

        # 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_compilable_cache
            and not output_attentions
        ):
            if not force_mask and 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 = input_tensor.dtype
        sequence_length = input_tensor.shape[1]
        if using_compilable_cache and past_key_values is not None:
            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
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        assert target_length is not None
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            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 in ["cuda", "xpu", "npu"]
            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(
                type_cast(torch.FloatTensor, causal_mask), min_dtype
            )

        return causal_mask

    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor | None,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs: Any,
    ):
        """
        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.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        del kwargs
        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=cache_position.device,
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(
                target_length, device=cache_position.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, :
                ].to(causal_mask.device)
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask


class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...


class Helium1ForCausalLM(Helium1PreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

    def __init__(self, config: Helium1CASAConfig, **kwargs: Any) -> None:
        del kwargs
        super().__init__(config)
        self.model: Helium1Model
        self.model = Helium1Model(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self._loss_function = ForCausalLMLoss

    def get_input_embeddings(self) -> nn.Module:
        return self.model.embed_tokens

    def set_input_embeddings(self, value: nn.Module) -> None:
        self.model.embed_tokens = value

    def get_output_embeddings(self) -> nn.Module:
        return self.lm_head

    def set_output_embeddings(self, new_embeddings: nn.Module) -> None:
        self.lm_head = new_embeddings

    def set_decoder(self, decoder: Helium1Model) -> None:
        self.model = decoder

    def get_decoder(self) -> Helium1Model:
        return self.model

    @can_return_tuple
    def forward(
        self,
        input_ids: None | torch.LongTensor = None,
        attention_mask: None | torch.Tensor = None,
        position_ids: None | torch.LongTensor = None,
        past_key_values: None | Cache = None,
        inputs_embeds: None | torch.Tensor = None,
        image_embeds: None | torch.Tensor | list[torch.Tensor] = None,
        image_embeds_insertion_points: None | list[torch.Tensor] = None,
        labels: None | torch.LongTensor = None,
        use_cache: None | bool = None,
        output_attentions: None | bool = None,
        output_hidden_states: None | bool = None,
        cache_position: None | torch.LongTensor = None,
        logits_to_keep: int | torch.Tensor = 0,
        # CASA
        casa_windows_info: None | dict = None,
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> CausalHeliumOutput:
        r"""
        Helium1 augmented with CASA layers
        """
        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
        )
        if input_ids is not None:
            assert inputs_embeds is None, (
                "Need to provide only one of `input_ids` or `inputs_embeds`."
            )
            inputs_embeds = self.model.embed_tokens(input_ids)
        assert inputs_embeds is not None

        # Setup image + text token fusion
        bs, og_seq_len, _ = inputs_embeds.shape
        image_tokens_mask: torch.Tensor | None = None
        casa_handler: CASAAttentionHandler | None = None

        num_image_tokens = -1
        if image_embeds is not None:
            num_image_tokens = sum(_x.shape[0] for _x in image_embeds)
            assert image_embeds_insertion_points is not None, (
                "Missing image embeddings insertion points"
            )
            # B1. CASA layers: We need to init the shared Handler
            if self.model.config.casa_attention:
                casa_handler = CASAAttentionHandler(
                    # for text tokens, we don't need the actual values
                    inputs_embeds=torch.zeros_like(inputs_embeds),
                    # for image embeddings, we put real inputs as this will be fixed
                    image_embeds=image_embeds,
                    image_embeds_insertion_points=image_embeds_insertion_points,
                    # attention mask is only needed at inference / left padding
                    attention_mask=None if self.training else attention_mask,
                    rope_fn=self.model.rotary_emb,
                    windows=self.model.config.casa_windows,
                    use_asymetric_q_kv=self.model.config.casa_use_asymetric_qkv,
                    # further params are fed to the funtion computing attention
                    casa_windows_info=casa_windows_info,
                )
            # B2. Direct image insertion
            else:
                inputs_embeds, _, attention_mask, image_tokens_mask = insert_image_tokens(
                    inputs_embeds=inputs_embeds,
                    image_embeds=image_embeds,
                    image_embeds_insertion_points=image_embeds_insertion_points,
                    attention_mask=attention_mask,
                    padding_side="right" if self.training else "left",
                    recover_batch_dim=True,
                )

        del image_embeds
        del input_ids
        outputs: BaseModelOutputWithPast = self.model(
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            cache_position=cache_position,
            image_tokens_mask=image_tokens_mask,
            casa_handler=casa_handler,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        assert hidden_states is not None
        if image_tokens_mask is not None:
            hidden_states = remove_image_tokens(hidden_states, image_tokens_mask)
        # 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,
            )
        out = CausalHeliumOutput(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            num_image_tokens_log=torch.tensor(num_image_tokens).to(logits.device).to(torch.float),
            num_text_tokens_log=torch.tensor(og_seq_len).to(logits.device).to(torch.float),
        )
        return out