#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/molmo2/modular_molmo2.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_molmo2.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 The HuggingFace 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

import torch
from torch import nn

from ... import initialization as init
from ...activations import ACT2FN
from ...backbone_utils import filter_output_hidden_states
from ...cache_utils import Cache, DynamicCache
from ...configuration_utils import PreTrainedConfig
from ...generation import GenerationMixin
from ...integrations import use_kernel_forward_from_hub, use_kernelized_func
from ...masking_utils import create_bidirectional_mask, create_causal_mask, create_masks_for_generate
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPast, BaseModelOutputWithPooling, 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 merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from .configuration_molmo2 import Molmo2AdapterConfig, Molmo2Config, Molmo2TextConfig, Molmo2VisionConfig


@auto_docstring(
    custom_intro="""
    Base class for Molmo2 causal language model (or autoregressive) outputs.
    """
)
@dataclass
class Molmo2CausalLMOutputWithPast(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


@auto_docstring(
    custom_intro="""
    Base class for Molmo2 outputs, with hidden states and attentions.
    """
)
@dataclass
class Molmo2ModelOutputWithPast(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


class Molmo2RotaryEmbedding(nn.Module):
    """RoPE from the flat `rope_parameters`; `scaled=False` keeps the theta but skips the scaling, for the
    layers outside `rope_scaling_layers`."""

    def __init__(self, config: Molmo2TextConfig, scaled: bool = True):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config

        # `rope_type = "default"` also keeps the generic `_init_weights` rotary re-init on the unscaled table.
        self.rope_type = self.config.rope_parameters["rope_type"] if scaled else "default"
        rope_init_fn: Callable = (
            self.compute_default_rope_parameters
            if self.rope_type == "default"
            else ROPE_INIT_FUNCTIONS[self.rope_type]
        )
        inv_freq, self.attention_scaling = rope_init_fn(self.config)

        self.inv_freq = nn.Buffer(inv_freq, persistent=False)
        self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False)

    @staticmethod
    def compute_default_rope_parameters(config: Molmo2Config, **kwargs) -> tuple[torch.Tensor, float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
        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.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):
        freqs = position_ids[..., None].float() * self.inv_freq.to(device=x.device, dtype=torch.float)
        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)


@use_kernel_forward_from_hub("RMSNorm")
class Molmo2RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        Molmo2RMSNorm 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}"


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_kernel_forward_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 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_kernelized_func(apply_rotary_pos_emb)
class Molmo2Attention(nn.Module):
    """Olmo2 attention whose q/k RMSNorm runs either over the full projection (`olmo`) or per head (`qwen3`)."""

    def __init__(self, config: Molmo2TextConfig, layer_idx: int) -> None:
        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=False)
        self.q_norm = Molmo2RMSNorm(
            self.head_dim if config.qk_norm_type == "qwen3" else config.num_attention_heads * self.head_dim,
            eps=config.rms_norm_eps,
        )
        self.k_norm = Molmo2RMSNorm(
            self.head_dim if config.qk_norm_type == "qwen3" else config.num_key_value_heads * self.head_dim,
            eps=config.rms_norm_eps,
        )
        self.qk_norm_type = config.qk_norm_type

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

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

        if self.qk_norm_type == "olmo":
            query_states = self.q_norm(query_states)
            key_states = self.k_norm(key_states)

        query_states = query_states.view(hidden_shape)
        key_states = key_states.view(hidden_shape)

        if self.qk_norm_type == "qwen3":
            query_states = self.q_norm(query_states)
            key_states = self.k_norm(key_states)

        query_states = query_states.transpose(1, 2)
        key_states = key_states.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 Molmo2MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

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


class Molmo2DecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Molmo2TextConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = Molmo2Attention(config=config, layer_idx=layer_idx)

        self.mlp = Molmo2MLP(config)
        self.input_layernorm = Molmo2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Molmo2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.norm_after = config.norm_after

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, 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,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        # `norm_after=True` normalizes each sublayer's output (post-norm) instead of its input (pre-norm)
        residual = hidden_states
        if not self.norm_after:
            hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            **kwargs,
        )
        if self.norm_after:
            hidden_states = self.input_layernorm(hidden_states)
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        if not self.norm_after:
            hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        if self.norm_after:
            hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states


@auto_docstring
class Molmo2PreTrainedModel(PreTrainedModel):
    config: Molmo2Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Molmo2DecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": Molmo2DecoderLayer,
        "attentions": Molmo2Attention,
    }
    input_modalities = ("image", "video", "text")

    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, Molmo2VisionModel):
            init.normal_(module.positional_embedding, mean=0.0, std=self.config.initializer_range)


class Molmo2VisionMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.activation_fn = ACT2FN[config.hidden_act]
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states = self.fc2(hidden_states)
        return hidden_states


# Modular automatically inherits RoPE, hence no inheritance for now
class Molmo2VisionAttention(nn.Module):
    def __init__(self, config: Molmo2VisionConfig | Molmo2AdapterConfig, hidden_size: int | None = None):
        super().__init__()
        self.config = config
        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 = False
        self.q_proj = nn.Linear(
            hidden_size or config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            hidden_size or config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            hidden_size or 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,
        encoder_hidden_states: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        encoder_hidden_states = hidden_states if encoder_hidden_states is None else encoder_hidden_states

        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)
        key_value_shape = (*encoder_hidden_states.shape[:-1], -1, self.head_dim)

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

        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,
            is_causal=self.is_causal,
            scaling=self.scaling,
            dropout=0.0 if not self.training else self.attention_dropout,
            **kwargs,
        )

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

        return attn_output, attn_weights


class Molmo2VisionEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Molmo2VisionConfig):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.self_attn = Molmo2VisionAttention(config)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.mlp = Molmo2VisionMLP(config)

    @auto_docstring
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.FloatTensor:
        residual = hidden_states

        hidden_states = self.layer_norm1(hidden_states)
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.layer_norm2(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states


@auto_docstring
class Molmo2VisionModel(Molmo2PreTrainedModel):
    config_class = Molmo2VisionConfig
    main_input_name = "pixel_values"
    input_modalities = ("image",)
    _no_split_modules = ["Molmo2VisionEncoderLayer"]
    _can_record_outputs = {
        "hidden_states": Molmo2VisionEncoderLayer,
        "attentions": Molmo2VisionAttention,
    }

    def __init__(self, config: Molmo2VisionConfig):
        super().__init__(config)
        self.positional_embedding = nn.Parameter(torch.zeros(config.num_position_embeddings, config.hidden_size))
        self.patch_embedding = nn.Linear(
            config.patch_size * config.patch_size * config.num_channels, config.hidden_size
        )
        # trf-ignore: TRF034 (false positive: Siglip2EncoderLayer subclasses GradientCheckpointingLayer)
        self.layers = nn.ModuleList([Molmo2VisionEncoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.post_init()

    @merge_with_config_defaults
    @capture_outputs(tie_last_hidden_states=False)
    @auto_docstring
    def forward(self, pixel_values: torch.Tensor, **kwargs: Unpack[TransformersKwargs]) -> BaseModelOutputWithPooling:
        hidden_states = self.patch_embedding(pixel_values.to(dtype=self.dtype))
        # patch count == num_position_embeddings, locked by config; only retraining with a different grid breaks this.
        hidden_states = hidden_states + self.positional_embedding[None, :, :].to(hidden_states.dtype)

        for encoder_layer in self.layers:
            hidden_states = encoder_layer(hidden_states, None, **kwargs)

        return BaseModelOutputWithPooling(
            last_hidden_state=hidden_states,
            pooler_output=hidden_states.mean(dim=1),
        )


class Molmo2ImageProjectorMLP(nn.Module):
    def __init__(self, config: Molmo2AdapterConfig):
        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(config.intermediate_size, config.text_hidden_size, bias=False)
        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


@auto_docstring(
    custom_intro="""
    The Molmo2 vision adapter: pools ViT patch features with a cross-attention layer and projects them into the
    language model's embedding space.
    """
)
class Molmo2Adapter(Molmo2PreTrainedModel):
    config_class = Molmo2AdapterConfig
    input_modalities = ("image",)
    _no_split_modules = ["Molmo2VisionAttention"]
    _can_record_outputs = {"attentions": Molmo2VisionAttention}

    def __init__(self, config: Molmo2AdapterConfig):
        super().__init__(config)
        pooling_input_dim = config.hidden_size * len(config.vision_feature_layer)
        self.image_pooling_2d = Molmo2VisionAttention(config, hidden_size=pooling_input_dim)
        self.image_projector = Molmo2ImageProjectorMLP(config)
        self.post_init()

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        image_features: torch.Tensor,
        pooled_patches_idx: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutput:
        r"""
        image_features (`torch.Tensor` of shape `(num_crops, num_patches, hidden_size * len(vision_feature_layer))`):
            Concatenated intermediate ViT features of every crop.
        pooled_patches_idx (`torch.Tensor` of shape `(num_tokens, pool_h * pool_w)`):
            Indices into the flattened patch sequence pooled by each output token; `-1` marks padding slots.
        """
        flat_features = image_features.reshape(-1, image_features.shape[-1])

        valid_mask = pooled_patches_idx >= 0
        valid_token_mask = torch.any(valid_mask, -1)

        patches_to_pool = flat_features[torch.clip(pooled_patches_idx, 0)]
        patches_to_pool = patches_to_pool * valid_mask.to(patches_to_pool.dtype)[..., None]

        num_valid_patches = valid_mask.float().sum(-1)
        num_valid_patches = torch.where(num_valid_patches == 0, 1, num_valid_patches)
        query = patches_to_pool.sum(-2, keepdim=True) / num_valid_patches[:, None, None].to(patches_to_pool.dtype)

        attention_mask = create_bidirectional_mask(
            config=self.config, inputs_embeds=query, attention_mask=valid_mask, encoder_hidden_states=patches_to_pool
        )
        pooled_features, _ = self.image_pooling_2d(query, patches_to_pool, attention_mask=attention_mask, **kwargs)
        pooled_features = pooled_features.squeeze(1)
        pooled_features = self.image_projector(pooled_features)
        return BaseModelOutput(last_hidden_state=pooled_features[valid_token_mask])


@auto_docstring
class Molmo2TextModel(Molmo2PreTrainedModel):
    config: Molmo2TextConfig

    def __init__(self, config: Molmo2TextConfig):
        super().__init__(config)
        self.vocab_size = config.vocab_size
        # The checkpoint's extra-vocabulary table is concatenated onto the base one at load time
        # (see `conversion_mapping.py`), so the embedding covers `vocab_size + additional_vocab_size`.
        self.embed_tokens = nn.Embedding(config.vocab_size + config.additional_vocab_size, config.hidden_size)
        # trf-ignore: TRF034 (false positive: LlamaDecoderLayer subclasses GradientCheckpointingLayer)
        self.layers = nn.ModuleList(
            [Molmo2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Molmo2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Molmo2RotaryEmbedding(config)
        self.rotary_emb_unscaled = Molmo2RotaryEmbedding(config, scaled=False)
        self.rope_types = [
            "scaled" if layer_idx in config.rope_scaling_layers else "unscaled"
            for layer_idx in range(config.num_hidden_layers)
        ]
        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 = 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 = {
            "scaled": self.rotary_emb(hidden_states, position_ids=position_ids),
            "unscaled": self.rotary_emb_unscaled(hidden_states, position_ids=position_ids),
        }

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

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


@auto_docstring(
    custom_intro="""
    The Molmo2 model which consists of a vision backbone, a pooling adapter and a language model, without a language
    modeling head.
    """
)
class Molmo2Model(Molmo2PreTrainedModel):
    config: Molmo2Config

    def __init__(self, config: Molmo2Config):
        super().__init__(config)
        self.vision_tower = Molmo2VisionModel(config.vision_config)
        self.multi_modal_projector = Molmo2Adapter(config.adapter_config)
        self.language_model = Molmo2TextModel(config.text_config)
        self.post_init()

    @can_return_tuple
    @filter_output_hidden_states
    @auto_docstring(
        custom_intro="Obtains pooled image features from the vision tower and the adapter: the `pooler_output` is a "
        "tuple with one `(num_image_tokens, hidden_size)` tensor per image."
    )
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_token_pooling: torch.Tensor,
        image_grids: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        image_shape = None
        if pixel_values.dim() == 4:
            image_shape = pixel_values.shape[:3]
            pixel_values = pixel_values.reshape(-1, pixel_values.shape[-2], pixel_values.shape[-1])

        image_outputs: BaseModelOutputWithPooling = self.vision_tower(pixel_values, **kwargs)
        image_features = torch.cat(
            [
                image_outputs.hidden_states[layer + 1 if layer >= 0 else layer]
                for layer in self.config.adapter_config.vision_feature_layer
            ],
            dim=-1,
        )

        if image_shape is not None:
            image_features = image_features.reshape(*image_shape, -1)
        pooled_features = self.multi_modal_projector(image_features, image_token_pooling).last_hidden_state
        split_sizes = (image_grids[:, 0] * image_grids[:, 1] + image_grids[:, 2] * image_grids[:, 3]).tolist()
        image_outputs.pooler_output = torch.split(pooled_features, split_sizes)
        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.full((), 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).to(inputs_embeds.device)
        torch_compilable_check(
            n_image_tokens * inputs_embeds.shape[-1] == 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,
        image_token_pooling: torch.Tensor | None = None,
        image_grids: torch.Tensor | None = None,
        image_num_crops: torch.Tensor | None = None,
        pixel_values_videos: torch.Tensor | None = None,
        video_token_pooling: torch.Tensor | None = None,
        video_grids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        mm_token_type_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        mm_encoder_outputs: dict[str, BaseModelOutputWithPooling] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Molmo2ModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
        if pixel_values is not None and pixel_values_videos is not None:
            raise ValueError("pixel_values and pixel_values_videos are provided at the same time")
        if (pixel_values is not None or pixel_values_videos is not None) and mm_encoder_outputs is not None:
            raise ValueError(
                "You cannot specify both pixel_values/pixel_values_videos and mm_encoder_outputs at the same time"
            )
        if self.training and mm_token_type_ids is None:
            raise ValueError("`mm_token_type_ids` is required as a model input when training")

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

        mm_encoder_outputs = mm_encoder_outputs if mm_encoder_outputs is not None else {}
        if mm_encoder_outputs.get("image") is None and pixel_values is not None:
            mm_encoder_outputs["image"] = self.get_image_features(pixel_values, image_token_pooling, image_grids)
        if mm_encoder_outputs.get("video") is None and pixel_values_videos is not None:
            mm_encoder_outputs["video"] = self.get_video_features(
                pixel_values_videos, video_token_pooling, video_grids
            )

        image_features: torch.FloatTensor | None = None
        if mm_encoder_outputs.get("image") is not None:
            image_features = mm_encoder_outputs["image"].pooler_output
        elif mm_encoder_outputs.get("video") is not None:
            image_features = mm_encoder_outputs["video"].pooler_output

        if image_features is not None:
            image_features = torch.cat(image_features, dim=0)
            # `get_placeholder_mask` returns a [batch, seq, 1] mask; Molmo2 *adds* the image features onto
            # the placeholder-token embeddings (residual), which means we index `inputs_embeds` directly.
            # Boolean indexing does not broadcast, so the mask must be expanded to the hidden dim first.
            special_image_mask = self.get_placeholder_mask(input_ids, inputs_embeds, image_features)
            special_image_mask = special_image_mask.expand_as(inputs_embeds)
            inputs_embeds = inputs_embeds.masked_scatter(
                special_image_mask,
                inputs_embeds[special_image_mask] + image_features.reshape(-1),
            )

        # An already prepared 4D mask (e.g. from `generate`) is returned as is by `create_causal_mask`
        mask_kwargs = {
            "config": self.config.get_text_config(),
            "inputs_embeds": inputs_embeds,
            "attention_mask": attention_mask,
            "past_key_values": past_key_values,
            "position_ids": position_ids,
        }
        is_prefill = past_key_values is None or not past_key_values.is_initialized or image_features is not None
        if mm_token_type_ids is not None and is_prefill:
            # every image or frame token attends to every other one, so they all share a single block
            mask_kwargs["block_sequence_ids"] = torch.where(mm_token_type_ids.to(inputs_embeds.device) == 1, 0, -1)
        causal_mask = create_causal_mask(**mask_kwargs)

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

        return Molmo2ModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features,
        )

    @can_return_tuple
    @auto_docstring(
        custom_intro="Obtains pooled video features from the vision tower and the adapter: the `pooler_output` is a "
        "tuple with one `(num_video_tokens, hidden_size)` tensor per video."
    )
    def get_video_features(
        self,
        pixel_values_videos: torch.FloatTensor,
        video_token_pooling: torch.Tensor,
        video_grids: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        # Video frames only go through the low-res path: `num_frames` frames of `rows x cols` tokens, no high-res crops.
        num_frames, rows, cols = video_grids.unbind(-1)
        image_grids = torch.stack([num_frames * rows, cols, torch.zeros_like(cols), torch.zeros_like(cols)], dim=-1)
        return self.get_image_features(pixel_values_videos, video_token_pooling, image_grids, **kwargs)


@auto_docstring(
    custom_intro="""
    The Molmo2 model which consists of a vision backbone, a pooling adapter and a language model.
    """
)
class Molmo2ForConditionalGeneration(Molmo2PreTrainedModel, GenerationMixin):
    # the loss must not be normalized by `num_items_in_batch`: logits/labels are filtered before computing it
    accepts_loss_kwargs = False
    config: Molmo2Config

    def __init__(self, config: Molmo2Config):
        super().__init__(config)

        self.model = Molmo2Model(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.vocab_size = config.text_config.vocab_size

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

    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_token_pooling: torch.Tensor,
        image_grids: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        return self.model.get_image_features(pixel_values, image_token_pooling, image_grids, **kwargs)

    @auto_docstring
    def get_video_features(
        self,
        pixel_values_videos: torch.FloatTensor,
        video_token_pooling: torch.Tensor,
        video_grids: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        return self.model.get_video_features(pixel_values_videos, video_token_pooling, video_grids, **kwargs)

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.Tensor | None = None,
        image_token_pooling: torch.Tensor | None = None,
        image_grids: torch.Tensor | None = None,
        image_num_crops: torch.Tensor | None = None,
        pixel_values_videos: torch.Tensor | None = None,
        video_token_pooling: torch.Tensor | None = None,
        video_grids: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: list[torch.FloatTensor] | None = None,
        mm_token_type_ids: torch.LongTensor | 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,
        mm_encoder_outputs: dict[str, BaseModelOutputWithPooling] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Molmo2CausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, Molmo2ForConditionalGeneration

        >>> model = Molmo2ForConditionalGeneration.from_pretrained("allenai/Molmo2-8B")
        >>> processor = AutoProcessor.from_pretrained("allenai/Molmo2-8B")

        >>> prompt = "What's the content of the image?"
        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> messages = [{"role": "user", "content": [{"type": "text", "text": prompt}, {"type": "image", "image": image}]}]

        >>> inputs = processor.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt", return_dict=True)

        >>> # Generate
        >>> generated_ids = model.generate(**inputs, max_new_tokens=15)
        >>> generated_tokens = generated_ids[:, inputs['input_ids'].size(1):]
        >>> processor.post_process_image_text_to_text(generated_tokens, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "The image shows a bustling street scene in what appears to be a Chinatown area. There's ..."
        ```"""
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            image_token_pooling=image_token_pooling,
            image_grids=image_grids,
            image_num_crops=image_num_crops,
            pixel_values_videos=pixel_values_videos,
            video_token_pooling=video_token_pooling,
            video_grids=video_grids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            mm_token_type_ids=mm_token_type_ids,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            mm_encoder_outputs=mm_encoder_outputs,
            **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.vocab_size, **kwargs)

        return Molmo2CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=outputs.image_hidden_states,
        )

    @staticmethod
    def create_masks_for_generate(
        config: PreTrainedConfig,
        inputs_embeds: torch.Tensor,
        attention_mask: torch.Tensor | None,
        past_key_values: Cache | None,
        position_ids: torch.Tensor | None,
        mm_token_type_ids: torch.Tensor | None = None,
        **kwargs,
    ) -> dict:
        mask_kwargs = {
            "config": config.get_text_config(),
            "inputs_embeds": inputs_embeds,
            "attention_mask": attention_mask,
            "past_key_values": past_key_values,
            "position_ids": position_ids,
        }
        if mm_token_type_ids is not None and inputs_embeds.shape[1] != 1:
            mask_kwargs["block_sequence_ids"] = torch.where(mm_token_type_ids.to(inputs_embeds.device) == 1, 0, -1)

        return create_masks_for_generate(**mask_kwargs)


__all__ = [
    "Molmo2Adapter",
    "Molmo2ForConditionalGeneration",
    "Molmo2Model",
    "Molmo2PreTrainedModel",
    "Molmo2TextModel",
    "Molmo2VisionModel",
]