webapp/seed/Lib/site-packages/transformers/models/llama4/modeling_llama4.py

# Copyright 2025 The LLAMA4 and 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.
import math
from collections.abc import Callable
from dataclasses import dataclass
from typing import Optional

import torch
import torch.nn as nn
import torch.nn.functional as F

from transformers.models.llama4.configuration_llama4 import Llama4VisionConfig

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
from ...masking_utils import create_causal_mask, create_chunked_causal_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutput,
    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, logging, torch_compilable_check
from ...utils.generic import check_model_inputs, maybe_autocast
from .configuration_llama4 import Llama4Config, Llama4TextConfig


logger = logging.get_logger(__name__)


class Llama4TextExperts(nn.Module):
    def __init__(self, config: Llama4TextConfig):
        super().__init__()
        self.num_experts = config.num_local_experts
        self.intermediate_size = config.intermediate_size
        self.hidden_size = config.hidden_size
        self.expert_dim = self.intermediate_size
        self.gate_up_proj = nn.Parameter(torch.zeros(self.num_experts, self.hidden_size, 2 * self.expert_dim))
        self.down_proj = nn.Parameter(torch.empty((self.num_experts, self.expert_dim, self.hidden_size)))
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        """
        This should really not be run on a single machine, as we are reaching compute bound:
        - the inputs are expected to be "sorted" per expert already.
        - the weights are viewed with another dim, to match num_expert, 1, shape * num_tokens, shape

        Args:
            hidden_states (torch.Tensor): (batch_size * token_num, hidden_size)
            selected_experts (torch.Tensor): (batch_size * token_num, top_k)
            routing_weights (torch.Tensor): (batch_size * token_num, top_k)
        Returns:
            torch.Tensor
        """
        hidden_states = hidden_states.view(self.gate_up_proj.shape[0], -1, self.hidden_size)
        gate_up = torch.bmm(hidden_states, self.gate_up_proj)
        gate, up = gate_up.chunk(2, dim=-1)  # not supported for DTensors
        next_states = torch.bmm((up * self.act_fn(gate)), self.down_proj)
        next_states = next_states.view(-1, self.hidden_size)
        return next_states


# Phi3MLP
class Llama4TextMLP(nn.Module):
    def __init__(self, config, intermediate_size=None):
        super().__init__()

        if intermediate_size is None:
            intermediate_size = config.intermediate_size

        self.config = config
        self.gate_proj = nn.Linear(config.hidden_size, intermediate_size, bias=False)
        self.up_proj = nn.Linear(config.hidden_size, intermediate_size, bias=False)
        self.down_proj = nn.Linear(intermediate_size, config.hidden_size, bias=False)
        self.activation_fn = ACT2FN[config.hidden_act]

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


class Llama4TextL2Norm(torch.nn.Module):
    def __init__(self, eps: float = 1e-6):
        super().__init__()
        self.eps = eps

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        return self._norm(x.float()).type_as(x)

    def extra_repr(self):
        return f"eps={self.eps}"


class Llama4TextRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-5):
        """
        Llama4RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(hidden_size))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        output = self._norm(x.float()).type_as(x)
        return output * self.weight

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


class Llama4Router(nn.Linear):
    def __init__(self, config):
        super().__init__(config.hidden_size, config.num_local_experts, bias=False)
        self.num_experts = config.num_local_experts
        self.top_k = config.num_experts_per_tok

    def forward(self, hidden_states):
        router_logits = super().forward(hidden_states)
        router_top_value, router_indices = torch.topk(router_logits, self.top_k, dim=1)
        router_scores = torch.full_like(router_logits, float("-inf")).scatter_(1, router_indices, router_top_value)
        router_scores = torch.nn.functional.sigmoid(router_scores.float()).to(router_scores.dtype)
        return router_scores, router_logits


@use_kernel_forward_from_hub("Llama4TextMoe")
class Llama4TextMoe(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.hidden_dim = config.hidden_size
        self.num_experts = config.num_local_experts
        self.experts = Llama4TextExperts(config)
        self.router = Llama4Router(config)
        self.shared_expert = Llama4TextMLP(config)

    def forward(self, hidden_states):
        hidden_states = hidden_states.reshape(-1, self.hidden_dim)
        router_scores, router_logits = self.router(hidden_states)
        routed_in = hidden_states.repeat(router_scores.shape[1], 1)
        routed_in = routed_in * router_scores.transpose(0, 1).reshape(-1, 1)
        routed_out = self.experts(routed_in)
        out = self.shared_expert(hidden_states)
        out.add_(routed_out.reshape(router_scores.shape[1], -1, routed_out.shape[-1]).sum(dim=0))
        return out, router_logits


# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Llama4Text
class Llama4TextRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    # Ignore copy
    def __init__(self, config: Llama4TextConfig, 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: Llama4TextConfig | 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

    # Ignore copy
    @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)
        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.to(x.device) @ position_ids_expanded).transpose(1, 2)
            freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # Convert to complex representation
            freqs_cis = freqs_cis * self.attention_scaling

        return freqs_cis


def apply_rotary_emb(
    xq: torch.Tensor,
    xk: torch.Tensor,
    freqs_cis: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
    xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
    xq_out = torch.view_as_real(xq_ * freqs_cis[:, :, None, :]).flatten(3)
    xk_out = torch.view_as_real(xk_ * freqs_cis[:, :, None, :]).flatten(3)
    return xq_out.type_as(xq), xk_out.type_as(xk)


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)


# Adapted from transformers.models.llama.modeling_llama.eager_attention_forward -> llama4 doesn't cast attn weights to fp32
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,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

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

    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
    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


# Adapted from transformers.models.llama.modeling_llama.eager_attention_forward -> llama4 doesn't cast attn weights to fp32
def vision_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,
):
    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)) * module.head_dim**-0.5
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

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

    return attn_output, attn_weights


class Llama4TextAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Llama4TextConfig, layer_idx):
        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_attention_heads = config.num_attention_heads
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attn_scale = config.attn_scale
        self.floor_scale = config.floor_scale
        self.attn_temperature_tuning = config.attn_temperature_tuning
        self.attention_dropout = config.attention_dropout
        self.is_causal = True
        self.use_rope = config.no_rope_layers[layer_idx]
        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
        )
        if self.config.use_qk_norm and self.use_rope:
            self.qk_norm = Llama4TextL2Norm(config.rms_norm_eps)

    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,
        cache_position: torch.LongTensor | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

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

        if self.use_rope:  # the 16E model skips rope for long context on certain layers
            query_states, key_states = apply_rotary_emb(
                query_states, key_states, position_embeddings.to(query_states.device)
            )

        if hasattr(self, "qk_norm"):  # the 128E model does not use qk_norm
            query_states = self.qk_norm(query_states)
            key_states = self.qk_norm(key_states)

        # Use temperature tuning from https://huggingface.co/papers/2501.19399) to NoROPE layers
        if self.attn_temperature_tuning and not self.use_rope:
            attn_scales = (
                torch.log1p(torch.floor((cache_position.float() + 1.0) / self.floor_scale)) * self.attn_scale + 1.0
            )
            attn_scales = attn_scales.view((1, input_shape[-1], 1, 1)).expand((*input_shape, 1, 1))  # batch size > 1
            query_states = (query_states * attn_scales).to(query_states.dtype)

        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)

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

        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 Llama4TextDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config, layer_idx):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.layer_idx = layer_idx
        self.attention_type = config.layer_types[layer_idx]
        self.self_attn = Llama4TextAttention(config, layer_idx)
        self.is_moe_layer = layer_idx in config.moe_layers
        if self.is_moe_layer:  # the 128E model interleaves dense / sparse
            self.feed_forward = Llama4TextMoe(config)
        else:
            self.feed_forward = Llama4TextMLP(config, intermediate_size=config.intermediate_size_mlp)

        self.input_layernorm = Llama4TextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Llama4TextRMSNorm(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,
        cache_position: torch.LongTensor | None = None,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        attention_states, _ = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            cache_position=cache_position,
            **kwargs,
        )
        hidden_states = residual + attention_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.feed_forward(hidden_states)
        if self.is_moe_layer:
            hidden_states, _ = hidden_states
        hidden_states = residual + hidden_states.view(residual.shape)
        return hidden_states


@auto_docstring
class Llama4PreTrainedModel(PreTrainedModel):
    config: Llama4Config
    input_modalities = ("image", "text")
    supports_gradient_checkpointing = True
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = False
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        std = (
            self.config.initializer_range
            if hasattr(self.config, "initializer_range")
            else self.config.text_config.initializer_range
        )
        if isinstance(module, Llama4TextExperts):
            init.normal_(module.gate_up_proj, mean=0.0, std=std)
            init.normal_(module.down_proj, mean=0.0, std=std)
        elif isinstance(module, Llama4VisionModel):
            init.normal_(module.class_embedding, std=module.scale)
            init.normal_(module.positional_embedding_vlm, std=module.scale)


@auto_docstring
class Llama4TextModel(Llama4PreTrainedModel):
    _no_split_modules = ["Llama4TextDecoderLayer"]
    base_model_prefix = "model"
    input_modalities = ("text",)
    config: Llama4TextConfig
    _can_record_outputs = {
        "attentions": Llama4TextAttention,
        "hidden_states": Llama4TextDecoderLayer,
        "router_logits": Llama4TextMoe,
    }

    def __init__(self, config: Llama4TextConfig):
        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(
            [Llama4TextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Llama4TextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Llama4TextRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

    @can_return_tuple
    @check_model_inputs
    @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,
        cache_position: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | 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.to(self.embed_tokens.weight.device))

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

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

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

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            # Prepare mask arguments
            mask_kwargs = {
                "config": self.config,
                "input_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "cache_position": cache_position,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
            }
            # Create the masks
            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
                "chunked_attention": create_chunked_causal_mask(**mask_kwargs),
            }

        hidden_states = inputs_embeds

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

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[decoder_layer.attention_type],
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=freq_cis,
                **kwargs,
            )
        hidden_states = self.norm(hidden_states)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
        )


class Llama4ForCausalLM(Llama4PreTrainedModel, GenerationMixin):
    _no_split_modules = ["Llama4TextDecoderLayer"]
    base_model_prefix = "language_model"
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    config: Llama4TextConfig

    def __init__(self, config: Llama4TextConfig):
        super().__init__(config)
        self.model = Llama4TextModel(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()

    @can_return_tuple
    @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,
        cache_position: torch.LongTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | CausalLMOutputWithPast:
        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 -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Example:

        ```python
        >>> from transformers import AutoTokenizer, Llama4ForCausalLM

        >>> model = Llama4ForCausalLM.from_pretrained("meta-llama4/Llama4-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama4/Llama4-2-7b-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 = 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,
            cache_position=cache_position,
            **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.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 Llava causal language model (or autoregressive) outputs.
    """
)
class Llama4CausalLMOutputWithPast(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


class Llama4VisionMLP2(torch.nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.fc1 = nn.Linear(self.intermediate_size, config.projector_input_dim, bias=False)
        self.fc2 = nn.Linear(config.projector_output_dim, config.projector_output_dim, bias=False)
        self.activation_fn = nn.GELU()  # ACT2FN[config.hidden_act]
        self.dropout = config.projector_dropout

    def forward(self, hidden_states):
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        return self.activation_fn(self.fc2(hidden_states))


class Llama4MultiModalProjector(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.linear_1 = nn.Linear(
            config.vision_config.vision_output_dim,
            config.text_config.hidden_size,
            bias=False,
        )

    def forward(self, image_features):
        hidden_states = self.linear_1(image_features)
        return hidden_states


def pixel_shuffle(input_tensor, shuffle_ratio):
    # input_tensor: [batch_size, num_patches, channels]
    batch_size, num_patches, channels = input_tensor.shape
    patch_size = int(math.sqrt(num_patches))

    input_tensor = input_tensor.view(batch_size, patch_size, patch_size, -1)
    batch_size, height, width, channels = input_tensor.size()

    reshaped_tensor = input_tensor.view(batch_size, height, int(width * shuffle_ratio), int(channels / shuffle_ratio))
    reshaped_tensor = reshaped_tensor.permute(0, 2, 1, 3).contiguous()

    reshaped_tensor = reshaped_tensor.view(
        batch_size, int(height * shuffle_ratio), int(width * shuffle_ratio), int(channels / (shuffle_ratio**2))
    )
    reshaped_tensor = reshaped_tensor.permute(0, 2, 1, 3).contiguous()

    output_tensor = reshaped_tensor.view(batch_size, -1, reshaped_tensor.shape[-1])
    return output_tensor


class Llama4VisionPixelShuffleMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.pixel_shuffle_ratio = config.pixel_shuffle_ratio
        self.inner_dim = int(config.projector_input_dim // (self.pixel_shuffle_ratio**2))
        self.output_dim = config.projector_output_dim
        self.mlp = Llama4VisionMLP2(config)

    def forward(self, encoded_patches: torch.Tensor) -> torch.Tensor:
        encoded_patches = pixel_shuffle(encoded_patches, self.pixel_shuffle_ratio)
        return self.mlp(encoded_patches)


# TODO there is a different RoPE for vision encoder, defined as below
def reshape_for_broadcast(freqs_ci: torch.Tensor, query: torch.Tensor):
    ndim = query.ndim
    shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(query.shape)]
    return freqs_ci.view(*shape)


def vision_apply_rotary_emb(
    query: torch.Tensor,
    key: torch.Tensor,
    freqs_ci: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
    query_ = torch.view_as_complex(query.float().reshape(*query.shape[:-1], -1, 2))
    key_ = torch.view_as_complex(key.float().reshape(*key.shape[:-1], -1, 2))
    freqs_ci = reshape_for_broadcast(freqs_ci=freqs_ci, query=query_)  # freqs_ci[:,:,None,:]
    freqs_ci = freqs_ci.to(query_.device)
    query_out = torch.view_as_real(query_ * freqs_ci).flatten(3)
    key_out = torch.view_as_real(key_ * freqs_ci).flatten(3)
    return query_out.type_as(query), key_out.type_as(key)  # but this drops to 8e-3


class Llama4VisionAttention(nn.Module):
    def __init__(self, config: Llama4VisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = config.hidden_size // config.num_attention_heads
        self.num_key_value_groups = 1
        self.attention_dropout = config.attention_dropout
        self.scaling = self.head_dim**-0.5

        self.q_proj = nn.Linear(self.embed_dim, self.num_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(self.embed_dim, self.num_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(self.embed_dim, self.num_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.embed_dim, bias=True)

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

        query_states = self.q_proj(hidden_states).view(hidden_shape)
        key_states = self.k_proj(hidden_states).view(hidden_shape)
        value_states = self.v_proj(hidden_states).view(hidden_shape)

        query_states, key_states = vision_apply_rotary_emb(query_states, key_states, freqs_ci=freqs_ci)

        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

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

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            None,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=None,  # TODO Might be enforced here for TP compatibility as scaling is not just sqrt(head_dim)
            is_causal=False,  # HAS TO BE ENFORCED
            **kwargs,
        )

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


class Llama4VisionMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.activation_fn = nn.GELU()  # ACT2FN[config.hidden_act]
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size, bias=True)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size, bias=True)

    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


class Llama4VisionEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Llama4VisionConfig):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = Llama4VisionAttention(config)
        self.mlp = Llama4VisionMLP(config)

        self.input_layernorm = nn.LayerNorm(config.hidden_size)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size)

    def forward(
        self,
        hidden_state: torch.Tensor,
        freqs_ci: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        output_attentions: bool | None = None,
    ):
        # Self Attention
        residual = hidden_state

        hidden_state = self.input_layernorm(hidden_state)

        hidden_state, attn_weights = self.self_attn(
            hidden_state,
            freqs_ci=freqs_ci,
            attention_mask=attention_mask,
        )
        hidden_state = residual + hidden_state

        # Feed forward
        residual = hidden_state
        hidden_state = self.post_attention_layernorm(hidden_state)
        hidden_state = self.mlp(hidden_state)
        hidden_state = residual + hidden_state

        outputs = (hidden_state,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


class Llama4VisionEncoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`Llama4VisionEncoderLayer`].

    Args:
        config: Llama4VisionConfig
    """

    def __init__(self, config: Llama4VisionConfig):
        super().__init__()
        self.config = config
        self.layers = nn.ModuleList([Llama4VisionEncoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False
        self.config = config

    def forward(
        self,
        hidden_states: torch.Tensor,
        freqs_ci: torch.Tensor,  # TODO move this to an attribute instead of keeping it around
        attention_mask: torch.Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
    ) -> tuple | BaseModelOutput:
        r"""
        Args:
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        for encoder_layer in self.layers:
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)

            layer_outputs = encoder_layer(
                hidden_state=hidden_states,
                attention_mask=attention_mask,
                output_attentions=output_attentions,
                freqs_ci=freqs_ci,
            )

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

            hidden_states = layer_outputs[0]

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
        return BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )


class Llama4UnfoldConvolution(nn.Module):
    def __init__(self, config):
        super().__init__()
        kernel_size = config.patch_size
        if isinstance(kernel_size, int):
            kernel_size = (kernel_size, kernel_size)
        self.unfold = torch.nn.Unfold(kernel_size=kernel_size, stride=config.patch_size)
        self.linear = nn.Linear(
            config.num_channels * kernel_size[0] * kernel_size[1],
            config.hidden_size,
            bias=False,
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.unfold(hidden_states)
        hidden_states = hidden_states.permute(0, 2, 1)
        hidden_states = self.linear(hidden_states)
        return hidden_states


class Llama4VisionRotaryEmbedding(nn.Module):
    def __init__(self, config: Llama4VisionConfig):
        super().__init__()
        idx = config.image_size // config.patch_size
        img_idx = torch.arange(idx**2, dtype=torch.int32).reshape(idx**2, 1)
        img_idx = torch.cat([img_idx, img_idx[:1]], dim=0)
        img_idx[-1, -1] = -2  # ID_CLS_TOKEN
        frequencies_x = img_idx % idx  # get the coordinates of the 2d matrix along x
        frequencies_y = img_idx // idx  # get the coordinates of the 2d matrix along y
        freq_dim = config.hidden_size // config.num_attention_heads // 2
        rope_freq = 1.0 / (
            config.rope_parameters["rope_theta"]
            ** (torch.arange(0, freq_dim, 2)[: (freq_dim // 2)].float() / freq_dim)
        )
        freqs_x = ((frequencies_x + 1)[..., None] * rope_freq[None, None, :]).repeat_interleave(2, dim=-1)
        freqs_y = ((frequencies_y + 1)[..., None] * rope_freq[None, None, :]).repeat_interleave(2, dim=-1)
        freqs = torch.cat([freqs_x, freqs_y], dim=-1).float().contiguous()[..., ::2]
        freqs = freqs.masked_fill(img_idx.reshape(-1, 1, 1) < 0, 0)
        freq_cis = torch.view_as_complex(torch.stack([torch.cos(freqs), torch.sin(freqs)], dim=-1))
        self.freqs_ci = freq_cis  # idx**2, idx**2, idx * 2

    def forward(self, hidden_states):
        return self.freqs_ci.to(hidden_states.device)


class Llama4VisionModel(Llama4PreTrainedModel):
    base_model_prefix = "vision_model"
    input_modalities = ("image",)
    _no_split_modules = ["Llama4VisionEncoderLayer"]
    config: Llama4VisionConfig

    def __init__(self, config: Llama4VisionConfig):
        super().__init__(config)
        self.image_size = config.image_size
        self.patch_size = config.patch_size
        self.hidden_size = config.hidden_size
        self.num_channels = config.num_channels

        self.num_patches = (self.image_size // self.patch_size) ** 2 + 1
        self.scale = config.hidden_size**-0.5

        self.patch_embedding = Llama4UnfoldConvolution(config)

        self.class_embedding = nn.Parameter(self.scale * torch.randn(self.hidden_size))
        self.positional_embedding_vlm = nn.Parameter(self.scale * torch.randn(self.num_patches, self.hidden_size))
        self.rotary_embedding = Llama4VisionRotaryEmbedding(config)

        # layer norms
        self.layernorm_pre = nn.LayerNorm(self.hidden_size)
        self.layernorm_post = nn.LayerNorm(self.hidden_size)

        # encoders
        self.model = Llama4VisionEncoder(config)
        self.vision_adapter = Llama4VisionPixelShuffleMLP(config)
        self.post_init()

    def get_input_embeddings(self):
        """
        This function is used to fetch the first embedding layer to activate grads on inputs.
        """
        return self.patch_embedding

    def forward(
        self,
        pixel_values: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> BaseModelOutput | tuple[torch.Tensor, ...]:
        r"""

        Example:

        ```python
        >>> from PIL import Image
        >>> import httpx
        >>> from io import BytesIO
        >>> from transformers import AutoProcessor, MllamaVisionModel

        >>> checkpoint = "meta-llama/Llama-3.2-11B-Vision"
        >>> model = MllamaVisionModel.from_pretrained(checkpoint)
        >>> processor = AutoProcessor.from_pretrained(checkpoint)

        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))
        >>> inputs = processor(images=image, return_tensors="pt")

        >>> output = model(**inputs)

        >>> print(output.last_hidden_state.shape)
        torch.Size([1, 1, 4, 1025, 7680])
        ```
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # num_concurrent_media and num_chunks are both currently 1
        batch_size_times_num_tiles, num_channels, height, width = pixel_values.shape
        num_concurrent_media = 1
        num_chunks = 1
        hidden_state = self.patch_embedding(pixel_values)
        _, num_patches, hidden_dim = hidden_state.shape

        # Add cls token
        hidden_state = hidden_state.reshape(
            batch_size_times_num_tiles * num_concurrent_media * num_chunks, num_patches, hidden_dim
        )
        class_embedding = self.class_embedding.expand(hidden_state.shape[0], 1, hidden_state.shape[-1])
        hidden_state = torch.cat([hidden_state, class_embedding], dim=1)
        num_patches += 1

        # Position embeddings
        hidden_state = hidden_state.reshape(
            batch_size_times_num_tiles * num_concurrent_media, num_chunks, num_patches, hidden_dim
        )
        positional_embedding = self.positional_embedding_vlm.to(dtype=hidden_state.dtype, device=hidden_state.device)
        hidden_state = hidden_state + positional_embedding

        hidden_state = self.layernorm_pre(hidden_state)

        hidden_state = hidden_state.view(batch_size_times_num_tiles, -1, hidden_dim)
        freqs_ci = self.rotary_embedding(pixel_values)

        output = self.model(
            hidden_state,
            attention_mask=None,
            output_hidden_states=output_hidden_states,
            output_attentions=output_attentions,
            freqs_ci=freqs_ci,
        )

        hidden_state = output.last_hidden_state

        hidden_state = self.layernorm_post(hidden_state)

        hidden_state = hidden_state[:, :-1, :]

        # now, we use Llama4VisionPixelShuffle + mlp to project embeddings
        hidden_state = self.vision_adapter(hidden_state)

        hidden_states = output.hidden_states if output_hidden_states else None

        if output_attentions:
            attentions = output[2]
        else:
            attentions = None

        if not return_dict:
            return tuple(v for v in [hidden_state, hidden_states, attentions] if v is not None)

        return BaseModelOutput(
            last_hidden_state=hidden_state,
            hidden_states=hidden_states,
            attentions=attentions,
        )


class Llama4ForConditionalGeneration(Llama4PreTrainedModel, GenerationMixin):
    _no_split_modules = ["Llama4TextDecoderLayer", "Llama4VisionEncoderLayer"]
    _tp_plan = {}
    base_model_prefix = "model"
    config: Llama4Config

    def __init__(self, config: Llama4Config):
        super().__init__(config)
        self.vision_model = Llama4VisionModel(config.vision_config)

        self.multi_modal_projector = Llama4MultiModalProjector(config)
        self.language_model = Llama4ForCausalLM(config.text_config)
        self.vocab_size = config.text_config.vocab_size
        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1

        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)

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

    def set_output_embeddings(self, new_embeddings):
        self.language_model.set_output_embeddings(new_embeddings)

    def set_decoder(self, decoder):
        self.language_model.set_decoder(decoder)

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

    @check_model_inputs(tie_last_hidden_states=False)
    @auto_docstring(custom_intro="Obtains image last hidden states from the vision tower and apply al projection.")
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        vision_feature_select_strategy: str,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`)
            The tensors corresponding to the input images.
        vision_feature_select_strategy (`str`):
            The feature selection strategy used to select the vision feature from the vision backbone.
            Can be one of `"default"` or `"full"`
        """
        kwargs = {k: v for k, v in kwargs.items() if v is not None}
        return self.vision_model(pixel_values, **kwargs)

    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()
        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: {image_features.shape[0]}",
        )
        return special_image_mask

    @check_model_inputs(tie_last_hidden_states=False)
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | 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,
        vision_feature_select_strategy: str | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        cache_position: torch.LongTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Llama4CausalLMOutputWithPast:
        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 -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Example:

        ```python
        >>> from PIL import Image
        >>> import httpx
        >>> from io import BytesIO
        >>> from transformers import AutoProcessor, LlavaForConditionalGeneration

        >>> model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf")
        >>> processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")

        >>> prompt = "USER: <image>\nWhat's the content of the image? ASSISTANT:"
        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))

        >>> inputs = processor(images=image, text=prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(**inputs, max_new_tokens=15)
        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "USER:  \nWhat's the content of the image? ASSISTANT: The image features a busy city street with a stop sign prominently displayed"
        ```"""

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if pixel_values is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
            )

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

        if pixel_values is not None:
            image_features = self.get_image_features(
                pixel_values=pixel_values,
                vision_feature_select_strategy=vision_feature_select_strategy,
                return_dict=True,
            ).pooler_output

            vision_flat = image_features.view(-1, image_features.size(-1))
            projected_vision_flat = self.multi_modal_projector(vision_flat).to(
                inputs_embeds.device, inputs_embeds.dtype
            )
            special_image_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=projected_vision_flat
            )
            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, projected_vision_flat)

        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,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
            logits_to_keep=logits_to_keep,
            **kwargs,
        )

        logits = outputs[0]

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            if attention_mask is not None:
                # we use the input attention mask to shift the logits and labels, because it is 2D.
                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
                shift_attention_mask = attention_mask[:, -(logits.shape[1] - 1) :].to(logits.device)
                shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
                shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
            else:
                shift_logits = logits[..., :-1, :].contiguous()
                shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(
                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device)
            )

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return Llama4CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features if pixel_values is not None else None,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        attention_mask=None,
        cache_position=None,
        logits_to_keep=None,
        is_first_iteration=False,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        model_inputs = self.language_model.prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            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 subsquent 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

        return model_inputs


__all__ = [
    "Llama4PreTrainedModel",
    "Llama4TextModel",
    "Llama4VisionModel",
    "Llama4ForCausalLM",
    "Llama4ForConditionalGeneration",
]