image-match/python/py/Lib/site-packages/transformers/models/aria/modeling_aria.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/aria/modular_aria.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_aria.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2024 The Rhymes-AI Teams Authors and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections.abc import Callable
from dataclasses import dataclass
from typing import Optional

import torch
from torch import nn

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
from ...masking_utils import create_causal_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutputWithPast,
    BaseModelOutputWithPooling,
    CausalLMOutputWithPast,
    ModelOutput,
)
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, torch_compilable_check
from ...utils.generic import maybe_autocast, merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from ..auto import AutoModel
from .configuration_aria import AriaConfig, AriaTextConfig


@use_kernel_forward_from_hub("RMSNorm")
class AriaTextRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        AriaTextRMSNorm 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:
        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}"


class AriaProjectorMLP(nn.Module):
    """
    Feed-Forward Network module for the Aria Projector.

    Args:
        in_features (`int`):
            Input embedding dimension.
        hidden_features (`int`):
            Hidden dimension of the feed-forward network.
        output_dim (`int`):
            Output dimension.
    """

    def __init__(self, in_features, hidden_features, output_dim):
        super().__init__()
        self.linear_in = nn.Linear(in_features, hidden_features, bias=False)
        self.linear_out = nn.Linear(hidden_features, output_dim, bias=False)
        self.act = ACT2FN["gelu_new"]

    def forward(self, hidden_states):
        hidden_states = self.act(self.linear_in(hidden_states))
        hidden_states = self.linear_out(hidden_states)
        return hidden_states


class AriaCrossAttention(nn.Module):
    """
    Aria Cross-Attention module.

    Args:
        config (`AriaConfig`):
            The configuration to use.
    """

    def __init__(self, config: AriaConfig, dropout_rate: float = 0):
        super().__init__()
        hidden_size = config.vision_config.hidden_size
        num_heads = config.vision_config.num_attention_heads
        self.num_heads = num_heads
        self.q_proj = nn.Linear(hidden_size, hidden_size, bias=False)
        self.k_proj = nn.Linear(hidden_size, hidden_size, bias=False)
        self.v_proj = nn.Linear(hidden_size, hidden_size, bias=False)

        # Original code here: https://github.com/rhymes-ai/Aria/blob/719ff4e52b727443cba3793b0e27fe64e0244fe1/aria/model/projector.py#L48
        self.multihead_attn = nn.MultiheadAttention(hidden_size, num_heads, batch_first=True)
        self.linear = nn.Linear(hidden_size, hidden_size)
        self.dropout = nn.Dropout(dropout_rate)

        self.layer_norm = nn.LayerNorm(hidden_size)
        self.layer_norm_kv = nn.LayerNorm(hidden_size)

    def forward(self, key_value_states, hidden_states, attn_mask=None):
        """
        Forward pass of the AriaCrossAttention module.

        Args:
            key_value_states (`torch.Tensor`):
                Input tensor for key and value.
            hidden_states (`torch.Tensor`):
                Input tensor for query.
            attn_mask (`torch.Tensor`, *optional*, defaults to None):
                Attention mask.

        Returns:
            torch.Tensor:
                Output tensor after cross-attention.
        """
        query = self.q_proj(self.layer_norm(hidden_states))

        key_value_states = self.layer_norm_kv(key_value_states)
        key = self.k_proj(key_value_states)
        value = self.v_proj(key_value_states)

        attn_output, _ = self.multihead_attn(query, key, value, attn_mask=attn_mask)

        attn_output = self.dropout(self.linear(attn_output))

        return attn_output


class AriaProjector(nn.Module):
    """
    Aria Projector module.

    This module projects vision features into the language model's embedding space, enabling interaction between vision and language components.

    Args:
        config (`AriaConfig`):
            Configuration object for the model.
    """

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

        self.patch_to_query_dict = config.projector_patch_to_query_dict
        self.in_features = config.vision_config.hidden_size
        self.num_heads = config.vision_config.num_attention_heads
        self.kv_dim = config.vision_config.hidden_size
        self.hidden_features = config.text_config.hidden_size
        self.output_dim = config.text_config.hidden_size

        self.query = nn.Parameter(torch.zeros(config.max_value_projector_patch_to_query_dict, self.in_features))

        self.cross_attn = AriaCrossAttention(config)

        self.layer_norm = nn.LayerNorm(self.in_features)
        self.feed_forward = AriaProjectorMLP(self.in_features, self.hidden_features, self.output_dim)

    def forward(self, key_value_states: torch.Tensor, attn_mask: torch.Tensor | None = None):
        """
        Forward pass of the Projector module.

        Args:
            key_value_states (`torch.Tensor`):
                Input tensor of shape (batch_size, num_patches, kv_dim).
            attn_mask (`torch.Tensor`, *optional*, default is None):
                Attention mask.

        Returns:
            `torch.Tensor`: Output tensor of shape (batch_size, query_number, output_dim).
        """
        batch_size, num_patches = key_value_states.shape[0], key_value_states.shape[1]

        if num_patches not in self.patch_to_query_dict:
            raise KeyError(
                f"Number of patches {num_patches} not found in patch_to_query_dict amongst possible values {self.patch_to_query_dict.keys()}."
            )
        query_num = self.patch_to_query_dict[num_patches]

        queries = self.query[:query_num].unsqueeze(0).repeat(batch_size, 1, 1)

        if attn_mask is not None:
            attn_mask = attn_mask.repeat_interleave(self.num_heads, 0)
            attn_mask = attn_mask.unsqueeze(1).expand(-1, queries.size(1), -1)

        attention_out = self.cross_attn(key_value_states, queries, attn_mask=attn_mask)

        out = self.feed_forward(self.layer_norm(attention_out))

        return out


class AriaSharedExpertsMLP(nn.Module):
    """
    Shared Expert MLP for shared experts.

    Unlike routed experts, shared experts process all tokens without routing.
    This class reconfigures the intermediate size in comparison to the LlamaMLP.

    Args:
        config (`AriaTextConfig`): Configuration object for the Aria language model.
    """

    def __init__(self, config: AriaTextConfig):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size * config.moe_num_shared_experts
        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


def sequential_experts_gemm(token_states, expert_weights, tokens_per_expert):
    """
    Compute the matrix multiplication (GEMM) for each expert sequentially. This approach is computationally inefficient, especially when dealing with a large number of experts.

    Args:
        token_states (torch.Tensor): Input tensor of shape (num_tokens, in_features).
        expert_weights (torch.Tensor): Weight tensor of shape (num_experts, in_features, out_features).
        tokens_per_expert (torch.Tensor): Number of tokens assigned to each expert.

    Returns:
        torch.Tensor: Output tensor of shape (num_tokens, out_features).
    """
    num_tokens = token_states.shape[0]
    out_features = expert_weights.shape[-1]
    output = torch.zeros(num_tokens, out_features, dtype=token_states.dtype, device=token_states.device)

    cumsum_num_tokens = torch.cumsum(tokens_per_expert, dim=0)
    # Insert zero at the beginning for offset index's convenience
    zero_tensor = torch.zeros(1, dtype=torch.long, device=cumsum_num_tokens.device)
    cumsum_num_tokens = torch.cat((zero_tensor, cumsum_num_tokens))

    for expert_num in range(expert_weights.shape[0]):
        start = cumsum_num_tokens[expert_num]
        end = cumsum_num_tokens[expert_num + 1]
        tokens = token_states[start:end]

        out = torch.matmul(tokens, expert_weights[expert_num])
        output[start:end] = out
    return output


class AriaGroupedExpertsGemm(nn.Module):
    """
    Grouped GEMM (General Matrix Multiplication) module for efficient expert computation.
    This module utilizes the grouped_gemm library (https://github.com/fanshiqing/grouped_gemm)
    for optimized performance. If the grouped_gemm library is not installed, it gracefully
    falls back to a sequential GEMM implementation, which may be slower but ensures
    functionality.

    Args:
        in_features (`int`):
            Number of input features.
        out_features (`int`):
            Number of output features.
        groups (`int`):
            Number of expert groups.
    """

    def __init__(self, in_features, out_features, groups):
        super().__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.groups = groups
        self.weight = nn.Parameter(torch.empty(groups, in_features, out_features))

    def forward(self, input, tokens_per_expert):
        """
        Perform grouped matrix multiplication.

        Args:
            input (`torch.Tensor`):
                Input tensor of shape (num_tokens, in_features).
            tokens_per_expert (`torch.Tensor`):
                Number of tokens assigned to each expert.

        Returns:
            torch.Tensor: Output tensor of shape (num_tokens, out_features).
        """
        return sequential_experts_gemm(
            input,
            self.weight,
            tokens_per_expert.cpu(),
        )


class AriaExperts(nn.Module):
    def __init__(self, config: AriaTextConfig) -> None:
        super().__init__()
        self.config = config
        self.fc1 = AriaGroupedExpertsGemm(config.hidden_size, config.intermediate_size * 2, config.moe_num_experts)
        self.fc2 = AriaGroupedExpertsGemm(config.intermediate_size, config.hidden_size, config.moe_num_experts)

    def route_tokens_to_experts(self, router_logits):
        top_logits, top_indices = torch.topk(router_logits, k=self.config.moe_topk, dim=1)
        scores = nn.functional.softmax(top_logits, dim=-1)
        return top_indices, scores

    def forward(self, hidden_states, router_logits) -> torch.Tensor:
        top_k_index, top_k_weights = self.route_tokens_to_experts(router_logits)
        original_dtype = top_k_index.dtype
        tokens_per_expert = torch.histc(
            top_k_index.flatten().to(torch.float32),
            bins=self.config.moe_num_experts,
            min=0,
            max=self.config.moe_num_experts - 1,
        ).to(original_dtype)
        indices = top_k_index

        flatten_indices = indices.view(-1)
        sorted_indices = torch.argsort(flatten_indices)
        permuted_tokens = hidden_states.index_select(0, sorted_indices // self.config.moe_topk)

        fc1_output = self.fc1(permuted_tokens, tokens_per_expert)
        projection, gate = torch.chunk(fc1_output, 2, dim=-1)
        fc1_output = nn.functional.silu(projection) * gate
        expert_output = self.fc2(fc1_output, tokens_per_expert)

        unpermuted_tokens = torch.zeros(
            (top_k_weights.shape[0] * self.config.moe_topk, expert_output.size(1)),
            dtype=expert_output.dtype,
            device=expert_output.device,
        )
        unpermuted_tokens.index_copy_(0, sorted_indices, expert_output)
        unpermuted_tokens = unpermuted_tokens.view(-1, self.config.moe_topk, expert_output.size(1))

        output = (unpermuted_tokens * top_k_weights.unsqueeze(-1)).sum(dim=1)
        return output


class AriaTextMoELayer(nn.Module):
    def __init__(self, config: AriaTextConfig):
        super().__init__()
        self.router = nn.Linear(config.hidden_size, config.moe_num_experts, bias=False)
        self.experts = AriaExperts(config)
        self.shared_experts = AriaSharedExpertsMLP(config)
        self.config = config

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        original_shape = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_states.size(-1))
        router_logits = self.router(hidden_states)
        expert_output = self.experts(hidden_states, router_logits).view(original_shape)
        shared_expert_output = self.shared_experts(hidden_states.view(original_shape))
        return expert_output + shared_expert_output


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)


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

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    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: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    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:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


@use_kernelized_func(apply_rotary_pos_emb)
class AriaTextAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: AriaTextConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        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 = self.head_dim**-0.5
        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.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_mask: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 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)

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

        if past_key_values is not None:
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

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

        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(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class AriaTextDecoderLayer(GradientCheckpointingLayer):
    """
    Aria Text Decoder Layer.

    This class defines a single decoder layer in the language model, incorporating self-attention and Mixture of Experts (MoE) feed-forward network.

    Args:
        config (`AriaTextConfig`):
            Configuration object for the text component of the model.
        layer_idx (`int`):
            Index of the layer.
    """

    def __init__(self, config: AriaTextConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = AriaTextAttention(config=config, layer_idx=layer_idx)
        self.mlp = AriaTextMoELayer(config)
        self.input_layernorm = AriaTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = AriaTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        # Self Attention
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            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
        return hidden_states


@auto_docstring
class AriaTextPreTrainedModel(PreTrainedModel):
    config: AriaTextConfig
    base_model_prefix = "model"
    input_modalities = ("image", "text")
    _no_split_modules = ["AriaTextDecoderLayer", "AriaGroupedExpertsGemm"]
    supports_gradient_checkpointing = True
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True

    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": AriaTextDecoderLayer,
        "attentions": AriaTextAttention,
    }

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, AriaGroupedExpertsGemm):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)


@auto_docstring
class AriaPreTrainedModel(PreTrainedModel):
    config: AriaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["AriaDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = False  # MoE models don't work with torch.compile (dynamic slicing)
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": AriaTextDecoderLayer,
        "attentions": AriaTextAttention,
    }

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, AriaProjector):
            init.trunc_normal_(module.query, std=self.config.initializer_range)


class AriaTextRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: AriaTextConfig, device=None):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config

        self.rope_type = self.config.rope_parameters["rope_type"]
        rope_init_fn: Callable = self.compute_default_rope_parameters
        if self.rope_type != "default":
            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)

        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)

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

        attention_factor = 1.0  # Unused in this type of RoPE

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

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        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 maybe_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)


@auto_docstring
class AriaTextModel(AriaTextPreTrainedModel):
    def __init__(self, config: AriaTextConfig):
        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(
            [AriaTextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = AriaTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = AriaTextRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

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

        if inputs_embeds is None:
            inputs_embeds: torch.Tensor = self.embed_tokens(input_ids)

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

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

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

        hidden_states = inputs_embeds
        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_embeddings=position_embeddings,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


@auto_docstring
class AriaTextForCausalLM(AriaTextPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

    def __init__(self, config: AriaTextConfig):
        super().__init__(config)
        self.model = AriaTextModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from transformers import AutoTokenizer, AriaTextForCausalLM

        >>> model = AriaTextForCausalLM.from_pretrained("meta-aria_text/AriaText-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-aria_text/AriaText-2-7b-hf")

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        outputs: BaseModelOutputWithPast = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

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

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


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Aria causal language model (or autoregressive) outputs.
    """
)
class AriaCausalLMOutputWithPast(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Language modeling loss (for next-token prediction).
    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    image_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    """

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    past_key_values: Cache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None
    image_hidden_states: torch.FloatTensor | None = None


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Aria outputs, with hidden states and attentions.
    """
)
class AriaModelOutputWithPast(BaseModelOutputWithPast):
    r"""
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    image_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    """

    image_hidden_states: torch.FloatTensor | None = None


@auto_docstring(
    custom_intro="""
    The Aria model which consists of a vision backbone and a language model, without a language modeling head.
    """
)
class AriaModel(AriaPreTrainedModel):
    def __init__(self, config: AriaConfig):
        super().__init__(config)
        self.vision_tower = AutoModel.from_config(config.vision_config)
        self.multi_modal_projector = AriaProjector(config)
        self.language_model = AutoModel.from_config(config.text_config)
        self.post_init()

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring(
        custom_intro="Obtains image last hidden states from the vision tower and apply multimodal projection."
    )
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        pixel_mask: torch.FloatTensor | None = None,
        vision_feature_layer: int | list[int] = -1,
        output_hidden_states: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        patch_attention_mask = self._create_patch_attention_mask(pixel_mask)
        image_outputs = self.vision_tower(
            pixel_values,
            patch_attention_mask=patch_attention_mask,
            output_hidden_states=True,  # Ignore arg on purpose
            return_dict=True,
            **kwargs,
        )
        image_attn_mask = None
        if patch_attention_mask is not None:
            flattened_mask = patch_attention_mask.flatten(1)
            image_attn_mask = torch.logical_not(flattened_mask)

        selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
        image_outputs.pooler_output = self.multi_modal_projector(selected_image_feature, attn_mask=image_attn_mask)

        return image_outputs

    def get_placeholder_mask(
        self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor
    ):
        """
        Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
        equal to the length of multimodal features. If the lengths are different, an error is raised.
        """
        if input_ids is None:
            special_image_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_image_mask = special_image_mask.all(-1)
        else:
            special_image_mask = input_ids == self.config.image_token_id

        n_image_tokens = special_image_mask.sum()
        n_image_features = image_features.shape[0] * image_features.shape[1]
        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        torch_compilable_check(
            inputs_embeds[special_image_mask].numel() == image_features.numel(),
            f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {n_image_features}",
        )
        return special_image_mask

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_mask: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple | AriaModelOutputWithPast:
        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

        # 2. Merge text and images
        if pixel_values is not None and inputs_embeds.shape[1] != 1:
            image_features = self.get_image_features(
                pixel_values=pixel_values,
                pixel_mask=pixel_mask,
                vision_feature_layer=self.config.vision_feature_layer,
                return_dict=True,
            ).pooler_output
            image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
            special_image_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_features
            )
            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

        outputs = self.language_model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

        return AriaModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values if use_cache else None,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features if pixel_values is not None else None,
        )

    def _create_patch_attention_mask(self, pixel_mask):
        if pixel_mask is None:
            return None

        patches_subgrid = pixel_mask.unfold(
            dimension=1,
            size=self.vision_tower.config.patch_size,
            step=self.vision_tower.config.patch_size,
        )
        patches_subgrid = patches_subgrid.unfold(
            dimension=2,
            size=self.vision_tower.config.patch_size,
            step=self.vision_tower.config.patch_size,
        )
        return (patches_subgrid.sum(dim=(-1, -2)) > 0).bool()


@auto_docstring(
    custom_intro="""
    Aria model for conditional generation tasks.

    This model combines a vision tower, a multi-modal projector, and a language model
    to perform tasks that involve both image and text inputs.
    """
)
class AriaForConditionalGeneration(AriaPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}

    def __init__(self, config: AriaConfig):
        super().__init__(config)
        self.model = AriaModel(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.post_init()

    def get_input_embeddings(self):
        return self.model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.model.set_input_embeddings(value)

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

    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        pixel_mask: torch.FloatTensor | None = None,
        vision_feature_layer: int | list[int] = -1,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        return self.model.get_image_features(
            pixel_values=pixel_values,
            pixel_mask=pixel_mask,
            vision_feature_layer=vision_feature_layer,
            **kwargs,
        )

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_mask: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | AriaCausalLMOutputWithPast:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or `model.image_token_id` (where `model` is your instance of `AriaForConditionalGeneration`).
            Tokens with indices set to `model.image_token_id` are ignored (masked), the loss is only
            computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Example:

        ```python
        >>> import httpx
        >>> from io import BytesIO
        >>> import torch
        >>> from PIL import Image
        >>> from io import BytesIO

        >>> from transformers import AutoProcessor, AutoModel
        >>> from transformers.image_utils import load_image

        >>> # Note that passing the image urls (instead of the actual pil images) to the processor is also possible
        >>> image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
        >>> image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg")
        >>> image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg")

        >>> processor = AutoProcessor.from_pretrained("Rhymes-AI/Aria")
        >>> model = AutoModel.from_pretrained("Rhymes-AI/Aria", dtype=torch.bfloat16, device_map="auto")

        >>> # Create inputs
        >>> messages = [
        ...     {
        ...         "role": "user",
        ...         "content": [
        ...             {"type": "image"},
        ...             {"type": "text", "text": "In this image, we can see the city of New York, and more specifically the Statue of Liberty."},
        ...             {"type": "image"},
        ...             {"type": "text", "text": "What can we see in this image?"},
        ...         ]
        ...     },
        ...     {
        ...         "role": "user",
        ...         "content": [
        ...             {"type": "image"},
        ...             {"type": "text", "text": "In which city is that bridge located?"},
        ...         ]
        ...     }
        ... ]

        >>> prompts = [processor.apply_chat_template([message], add_generation_prompt=True) for message in messages]
        >>> images = [[image1, image2], [image3]]
        >>> inputs = processor(text=prompts, images=images, padding=True, return_tensors="pt").to(model.device)

        >>> # Generate
        >>> generated_ids = model.generate(**inputs, max_new_tokens=256)
        >>> generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)

        >>> print(generated_texts[0])
        Assistant: There are buildings, trees, lights, and water visible in this image.

        >>> print(generated_texts[1])
        Assistant: The bridge is in San Francisco.
        ```"""
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            pixel_mask=pixel_mask,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            **kwargs,
        )

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

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

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

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        pixel_mask=None,
        attention_mask=None,
        logits_to_keep=None,
        is_first_iteration=False,
        **kwargs,
    ):
        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            logits_to_keep=logits_to_keep,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        if is_first_iteration or not kwargs.get("use_cache", True):
            # Pixel values are used only in the first iteration if available
            # In subsequent iterations, they are already merged with text and cached
            # NOTE: first iteration doesn't have to be prefill, it can be the first
            # iteration with a question and cached system prompt (continue generate from cache)
            model_inputs["pixel_values"] = pixel_values
            model_inputs["pixel_mask"] = pixel_mask

        return model_inputs


__all__ = [
    "AriaForConditionalGeneration",
    "AriaPreTrainedModel",
    "AriaTextPreTrainedModel",
    "AriaTextModel",
    "AriaModel",
    "AriaTextForCausalLM",
]