#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/mistral_large4/modular_mistral_large4.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_mistral_large4.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 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

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

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache, KVSharingDict
from ...generation import GenerationMixin
from ...integrations import use_experts_implementation, use_kernel_forward_from_hub
from ...masking_utils import create_causal_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutputWithPast,
    BaseModelOutputWithPooling,
    ModelOutput,
    MoeModelOutputWithPast,
)
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel, slice_logits_to_keep
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 OutputRecorder, capture_outputs, select_hidden_states
from ..auto import AutoModel
from .configuration_mistral_large4 import MistralLarge4Config, MistralLarge4TextConfig


@use_kernel_forward_from_hub("RMSNorm")
class MistralLarge4RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        MistralLarge4RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

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

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


class MistralLarge4TextRotaryEmbedding(nn.Module):
    """LlamaRotaryEmbedding with a learned tensor for scaling the inverse frequencies."""

    def __init__(self, config: MistralLarge4TextConfig) -> 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)

        self.inv_freq = nn.Buffer(inv_freq, persistent=False)
        self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False)
        self.alpha = nn.Buffer(torch.zeros_like(self.inv_freq))  # init to 0 to avoid NaNs in alpha.exp()

    @staticmethod
    def compute_default_rope_parameters(config: MistralLarge4TextConfig, **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):
        scaled_inv_freq = (
            self.inv_freq.to(device=x.device, dtype=torch.float)
            * self.alpha.to(device=x.device, dtype=torch.float).exp()
        )
        freqs = position_ids[..., None].float() * scaled_inv_freq
        emb = torch.cat((freqs, freqs), dim=-1)
        cos = emb.cos() * self.attention_scaling
        sin = emb.sin() * self.attention_scaling

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


class MistralLarge4TextSoftmaxScaling(nn.Module):
    def __init__(self, num_heads: int):
        super().__init__()
        self.scale = nn.Parameter(torch.empty(num_heads))
        self.bias = nn.Parameter(torch.empty(num_heads))

    def forward(self, query_states: torch.Tensor, position_ids: torch.Tensor, head_dim_first: bool) -> torch.Tensor:
        log_pos = (position_ids + 1).float().log()  # convert to float to avoid overflow
        log_pos, scale, bias = (x.to(query_states.dtype) for x in (log_pos, self.scale, self.bias))
        if head_dim_first:
            scale = scale[None, :, None] * log_pos[:, None, :] + bias[None, :, None]
        else:
            scale = scale[None, None, :] * log_pos[:, :, None] + bias[None, None, :]
        return query_states * scale.unsqueeze(-1)


def apply_rotary_pos_emb_interleave(
    q: torch.Tensor, k: torch.Tensor | None, cos: torch.Tensor, sin: torch.Tensor, unsqueeze_dim: int = 1
) -> tuple[torch.Tensor, torch.Tensor | None]:
    r"""
    Applies interleaved Rotary Position Embedding to the query and (optionally) the key tensors.

    DeepSeek lays the rotary dimensions out in interleaved pairs `(x0, x1), (x2, x3), ...`, each rotated by a
    single frequency. We compute that rotation directly on the even/odd slices instead of de-interleaving with a
    `view`/`transpose`/`reshape`; the output is bit-identical to the de-interleaved `rotate_half` formulation while
    avoiding the extra contiguous copy.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`, *optional*): The key tensor if there is one, otherwise `None`.
        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 (optionally) the key tensors rotated using the Rotary Position
        Embedding.
    """
    # `cos`/`sin` are `cat(freqs, freqs)`; the first half holds the per-pair angle.
    cos = cos[..., : cos.shape[-1] // 2].unsqueeze(unsqueeze_dim)
    sin = sin[..., : sin.shape[-1] // 2].unsqueeze(unsqueeze_dim)

    q1, q2 = q[..., 0::2], q[..., 1::2]
    q_embed = torch.cat([q1 * cos - q2 * sin, q2 * cos + q1 * sin], dim=-1)

    if k is not None:
        k1, k2 = k[..., 0::2], k[..., 1::2]
        k_embed = torch.cat([k1 * cos - k2 * sin, k2 * cos + k1 * sin], dim=-1)
    else:
        k_embed = None

    return q_embed, k_embed


class MistralLarge4TextIndexer(nn.Module):
    """Indexer from glm_moe_dsa, but since there is no Q-LoRA in the attention, the indexer queries are derived from
    hidden states, using norm -> proj -> softmax scaling. Also, for layers with no RoPE, the indexer does not use RoPE
    as well."""

    def __init__(self, config: MistralLarge4TextConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx

        self.hidden_size: int = config.hidden_size
        self.n_heads: int = config.index_n_heads
        self.head_dim: int = config.index_head_dim
        self.qk_rope_head_dim: int = config.qk_rope_head_dim
        self.index_topk: int = config.index_topk
        self.wq_b = nn.Linear(config.hidden_size, self.n_heads * self.head_dim, bias=False)
        self.wk = nn.Linear(self.hidden_size, self.head_dim, bias=False)
        self.k_norm = nn.LayerNorm(self.head_dim, eps=1e-6)
        self.weights_proj = nn.Linear(self.hidden_size, self.n_heads, bias=False)
        self.softmax_scale = self.head_dim**-0.5
        self.q_input_norm = nn.LayerNorm(config.hidden_size, eps=1e-6)
        self.softmax_scaling = MistralLarge4TextSoftmaxScaling(self.n_heads)
        self.use_rope = config.no_rope_layers[layer_idx]

    @torch.no_grad()
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor,
        position_ids: torch.Tensor,
        past_key_values: Cache | None = None,
    ) -> torch.Tensor:
        """
        Selects the top-k tokens per query for DeepSeek Sparse Attention (DSA).

        Same as [`DeepseekV32Indexer.forward`], but the indexer applies **interleaved** RoPE
        rather than the non-interleaved half-split RoPE used by DeepSeek-V3.2.

        Args:
            hidden_states: Input hidden states `[B, S, hidden_size]`.
            position_embeddings: `(cos, sin)` from RotaryEmbedding.
            attention_mask: Causal mask, broadcastable to `[B, S, T]`.
            past_key_values: Cache object containing the indexer key cache for this layer.

        Returns:
            `torch.Tensor`: the `int32` top-k token indices of shape `[B, S, topk]`. The eager / SDPA paths
                turn these into an additive sparse mask; the `flash-mla` kernel consumes them directly.
        """
        batch_size, seq_len, _ = hidden_states.shape
        cos, sin = position_embeddings

        # The indexer queries are derived directly from the hidden states
        q = self.q_input_norm(hidden_states)
        q = self.wq_b(q)  # [B, S, H*D]
        q = q.view(batch_size, seq_len, self.n_heads, self.head_dim)  # [B, S, H, D]
        q = self.softmax_scaling(q, position_ids, head_dim_first=False)

        # Same for the keys
        k = self.k_norm(self.wk(hidden_states)).unsqueeze(2)  # [B, S, 1, D]

        # If there is RoPE, apply it (interleaved)
        if self.use_rope:
            q_rot, q_pass = torch.split(q, [self.qk_rope_head_dim, self.head_dim - self.qk_rope_head_dim], dim=-1)
            k_rot, k_pass = torch.split(k, [self.qk_rope_head_dim, self.head_dim - self.qk_rope_head_dim], dim=-1)
            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin, unsqueeze_dim=2)
            q = torch.cat([q_rot, q_pass], dim=-1)  # [B, S, H, D]
            k = torch.cat([k_rot, k_pass], dim=-1)  # [B, S, 1, D]
        k = k.squeeze(2)

        if past_key_values is not None:
            k = past_key_values.update_indexer(k, self.layer_idx)

        scores = torch.matmul(q.float(), k.transpose(-1, -2).float().unsqueeze(1)) * self.softmax_scale
        scores = F.relu(scores)

        # Weight per head and sum across heads: [B, S, 1, H] @ [B, S, H, T] → [B, S, T]
        weights = self.weights_proj(hidden_states.to(self.weights_proj.weight.dtype)).float() * (self.n_heads**-0.5)
        index_scores = torch.matmul(weights.unsqueeze(-2), scores).squeeze(-2)

        # Causality needs to be taken into account when computing scores so padding tokens don't affect computation
        if attention_mask.dtype == torch.bool:
            index_scores = index_scores.masked_fill(~attention_mask, float("-inf"))
        else:
            index_scores = index_scores + attention_mask

        topk = min(self.index_topk, index_scores.shape[-1])
        return index_scores.topk(topk, dim=-1).indices.to(torch.int32)  # [B, S, topk]


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


def yarn_get_mscale(scale=1, mscale=1):
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


def yarn_apply_mscale(rope_parameters, scaling):
    if rope_parameters.get("rope_type", "default") != "default":
        mscale_all_dim = rope_parameters.get("mscale_all_dim", 0)
        if mscale_all_dim:
            mscale = yarn_get_mscale(rope_parameters["factor"], mscale_all_dim)
            scaling = scaling * mscale * mscale
    return scaling


class MistralLarge4TextAttention(nn.Module):
    """DSA attention from GlmMoeDsa, with no Q-LoRA and KV sharing."""

    def __init__(self, config: MistralLarge4TextConfig, layer_idx: int):
        super().__init__()
        self.use_rope = config.no_rope_layers[layer_idx]
        self.is_kv_shared_layer = config.kv_sharing_modes[layer_idx] in ("reuse", "reuse_but_indexer")
        self.store_full_length_kv = config.kv_sharing_modes[layer_idx] == "store"
        self.config = config
        self.layer_idx = layer_idx
        self.attention_dropout = config.attention_dropout
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.qk_rope_head_dim = config.qk_rope_head_dim if self.use_rope else 0
        self.kv_lora_rank = config.kv_lora_rank
        self.v_head_dim = config.v_head_dim
        self.qk_nope_head_dim = config.qk_nope_head_dim + (0 if self.use_rope else config.qk_rope_head_dim)
        self.qk_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads

        self.is_causal = True
        self.kv_a_proj_with_mqa = (
            nn.Linear(self.hidden_size, config.kv_lora_rank + self.qk_rope_head_dim, bias=config.attention_bias)
            if not self.is_kv_shared_layer
            else None
        )
        self.kv_a_layernorm = (
            MistralLarge4RMSNorm(config.kv_lora_rank, eps=config.rms_norm_eps) if not self.is_kv_shared_layer else None
        )
        self.kv_b_proj = nn.Linear(
            config.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
        )

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

        self.scaling = yarn_apply_mscale(config.rope_parameters, self.qk_head_dim ** (-0.5))
        # Refer: https://arxiv.org/abs/2603.12201 for more details.
        self.skip_topk = config.indexer_types[layer_idx] == "shared"
        self.indexer = None if self.skip_topk else MistralLarge4TextIndexer(config, layer_idx)

        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
        self.softmax_scaling = MistralLarge4TextSoftmaxScaling(self.num_heads)

    def expand_kv(self, kv_nope: torch.Tensor, k_rot: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        """Expands the compressed latents into key and value states. Args:
            - kv_nope: key + value without positional encoding, shape [batch_size, 1, seqlen, self.kv_lora_rank]
            - k_rot: shared key with positional encoding, shape [batch_size, 1, seqlen, self.qk_rope_head_dim]
        Returns the key and value states, two tensors of shape [batch, num_heads, seq, (k or v)_head_dim].
        """
        batch_size, _, seq_length, _ = kv_nope.shape
        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)

        kv_nope = self.kv_b_proj(kv_nope).view(key_shape).transpose(1, 2)
        k_nope, value_states = torch.split(kv_nope, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
        k_rot = k_rot.expand(-1, k_nope.shape[1], -1, -1)  # does not affect the underlying storage

        # Concatenate k_nope and k_rot in a new tensor
        key_states = kv_nope.new_empty(*kv_nope.shape[:-1], self.qk_nope_head_dim + self.qk_rope_head_dim)
        key_states[..., : self.qk_nope_head_dim].copy_(k_nope)
        key_states[..., self.qk_nope_head_dim :].copy_(k_rot)
        return key_states, value_states

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor,
        shared_kv_states: KVSharingDict,
        position_ids: torch.Tensor,
        past_key_values: Cache | None = None,
        prev_topk_indices: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None, torch.Tensor | None]:
        batch_size, seq_length = hidden_states.shape[:-1]
        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)

        q_states = self.q_proj(hidden_states).view(query_shape).transpose(1, 2)
        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)

        # Either the layer computes it's own cache or it loads it from shared_kv_states
        if self.is_kv_shared_layer:
            k_pass, k_rot = shared_kv_states["rope" if self.use_rope else "nope"]
        else:
            compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
            kv_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
            # Both latents are viewed as single-head, 4D tensors, as expected by `expand_kv`
            k_pass = self.kv_a_layernorm(kv_pass).view(batch_size, 1, seq_length, self.kv_lora_rank)
            k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)

        # Apply LeRoPE to the queries and the keys, except if this is a no_rope layer
        if self.use_rope:
            cos, sin = position_embeddings
            # For layers with cache sharing, only rotate the queries: the keys were already rotated by the source layer
            if self.is_kv_shared_layer:
                q_rot, _ = apply_rotary_pos_emb_interleave(q_rot, None, cos, sin)
            else:
                q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)

        # Update the cache while latent KV is still compressed
        if past_key_values is not None and not self.is_kv_shared_layer:
            k_pass, k_rot = past_key_values.update(k_pass, k_rot, self.layer_idx)
        # If the layer takes part in cache sharing and computes its own cache, store a copy in shared_kv_states
        if self.store_full_length_kv:
            shared_kv_states["rope" if self.use_rope else "nope"] = (k_pass, k_rot)

        # Recompose and apply softmax-scaling to the queries
        query_states = torch.cat((q_pass, q_rot), dim=-1)
        query_states = self.softmax_scaling(query_states, position_ids, head_dim_first=True)

        key_states, value_states = self.expand_kv(k_pass, k_rot)

        # DSA: select this layer's top-k tokens, or reuse the previous full layer's on `"shared"` layers.
        if self.indexer is not None:
            topk_indices = self.indexer(
                hidden_states,
                position_embeddings,
                attention_mask[:, 0, :, :],
                position_ids,  # Kept for BC
                past_key_values=past_key_values,
            )  # [B, S, topk]
        else:
            if prev_topk_indices is None:
                raise ValueError("Shared DSA layers require top-k indices from a previous full indexer layer.")
            topk_indices = prev_topk_indices

        sparse_indices = None
        if self.config._attn_implementation in ("eager", "sdpa"):
            index_mask = (
                topk_indices.new_ones((batch_size, seq_length, key_states.shape[2]), dtype=torch.bool)
                .scatter(-1, topk_indices.long(), False)
                .unsqueeze(1)
            )

            if attention_mask.dtype == torch.bool:
                attention_mask = attention_mask & ~index_mask
            else:
                attention_mask = attention_mask.masked_fill(index_mask, torch.finfo(hidden_states.dtype).min)
        else:
            sparse_indices = topk_indices

        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,
            indices=sparse_indices,  # consumed by flash_mla_with_kvcache; ignored by eager / SDPA
            position_ids=position_ids,
            **kwargs,
        )

        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights, topk_indices


class MistralLarge4TextMLP(nn.Module):
    def __init__(self, config, intermediate_size=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size if intermediate_size is None else 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 MistralLarge4TextTopkRouter(nn.Module):
    """Same as Mistral4TopkRouter, with fp32 precision for routing logits and norming of topk weights is always on."""

    def __init__(self, config):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))
        self.routed_scaling_factor = config.routed_scaling_factor
        self.num_group = config.n_group
        self.topk_group = config.topk_group

    def forward(self, hidden_states):
        hidden_states = hidden_states.view(-1, self.hidden_dim)
        # weight.float() should be a no-op because the weight is in _keep_in_fp32_modules_strict, but if a user changed
        # the dtype of the model after loading, it will prevent an error
        router_logits = F.linear(hidden_states.float(), self.weight.float())
        scores = router_logits.softmax(-1)
        group_scores = (
            scores.view(-1, self.num_group, self.num_experts // self.num_group).topk(2, dim=-1)[0].sum(dim=-1)
        )
        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
        group_mask = torch.zeros_like(group_scores)
        group_mask.scatter_(1, group_idx, 1)
        score_mask = (
            group_mask.unsqueeze(-1)
            .expand(-1, self.num_group, self.num_experts // self.num_group)
            .reshape(-1, self.num_experts)
        )
        scores_for_choice = scores.masked_fill(~score_mask.bool(), 0.0)
        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
        topk_weights = scores.gather(1, topk_indices)
        denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
        topk_weights /= denominator
        topk_weights = topk_weights * self.routed_scaling_factor
        return router_logits, topk_weights, topk_indices


@use_experts_implementation
class MistralLarge4TextExperts(nn.Module):
    """Collection of expert weights stored as 3D tensors."""

    def __init__(self, config):
        super().__init__()
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.intermediate_dim = config.moe_intermediate_size
        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(
        self,
        hidden_states: torch.Tensor,
        top_k_index: torch.Tensor,
        top_k_weights: torch.Tensor,
    ) -> torch.Tensor:
        final_hidden_states = torch.zeros_like(hidden_states)
        with torch.no_grad():
            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts + 1)
            expert_mask = expert_mask.permute(2, 1, 0)
            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()

        for expert_idx in expert_hit:
            expert_idx = expert_idx[0]
            if expert_idx == self.num_experts:
                continue
            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
            current_state = hidden_states[token_idx]
            gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
            current_hidden_states = self.act_fn(gate) * up
            current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))

        return final_hidden_states


class MistralLarge4TextMoE(nn.Module):
    """
    A mixed expert module containing shared experts.
    """

    def __init__(self, config: MistralLarge4TextConfig):
        super().__init__()
        self.config = config
        self.experts = MistralLarge4TextExperts(config)
        self.gate = MistralLarge4TextTopkRouter(config)
        self.shared_experts = MistralLarge4TextMLP(
            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        residuals = hidden_states
        orig_shape = hidden_states.shape
        _, topk_weights, topk_indices = self.gate(hidden_states)
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape)
        hidden_states = hidden_states + self.shared_experts(residuals)
        return hidden_states


class MistralLarge4TextDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: MistralLarge4TextConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = MistralLarge4TextAttention(config, layer_idx)

        self.mlp = (
            MistralLarge4TextMoE(config)
            if config.mlp_layer_types[layer_idx] == "sparse"
            else MistralLarge4TextMLP(config)
        )
        self.input_layernorm = MistralLarge4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = MistralLarge4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        shared_kv_states: KVSharingDict,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        prev_topk_indices: torch.Tensor | None = None,  # MAIN DIFF with DSV3.2
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        # Self Attention
        hidden_states, _, topk_indices = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            shared_kv_states=shared_kv_states,  # the forward is overloaded to pass this args
            position_ids=position_ids,
            past_key_values=past_key_values,
            prev_topk_indices=prev_topk_indices,
            use_cache=use_cache,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states, topk_indices


@auto_docstring
class MistralLarge4PreTrainedModel(PreTrainedModel):
    config: MistralLarge4Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["MistralLarge4TextDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = False  # flash-mla kernels need a bit more work in the way we enable them!
    _supports_sdpa = True
    _supports_flex_attn = False

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": MistralLarge4TextDecoderLayer,
        "attentions": MistralLarge4TextAttention,
        "router_logits": OutputRecorder(MistralLarge4TextTopkRouter, index=0),
    }
    _keep_in_fp32_modules_strict = ["rotary_emb", "mlp.gate.weight"]
    # The checkpoint ships a multi-token-prediction draft head that this model does not implement yet # TODO
    _keys_to_ignore_on_load_unexpected = [r"mtp_dspark\..*"]
    _keep_in_fp32_modules = ["indexer.weights_proj"]
    input_modalities = ("image", "text")

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, MistralLarge4TextTopkRouter):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, MistralLarge4TextExperts):
            init.normal_(module.gate_up_proj, mean=0.0, std=self.config.initializer_range)
            init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, MistralLarge4TextRotaryEmbedding):
            init.zeros_(module.alpha)  # exp(0) = 1: LeRoPE has no effect until `alpha` is loaded
        elif isinstance(module, MistralLarge4TextSoftmaxScaling):
            # 0 * log(pos + 1) + 1 = 1: no query scaling until `scale` / `bias` are loaded
            init.zeros_(module.scale)
            init.ones_(module.bias)


@auto_docstring
class MistralLarge4TextModel(MistralLarge4PreTrainedModel):
    config: MistralLarge4TextConfig

    def __init__(self, config: MistralLarge4TextConfig):
        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(
            [MistralLarge4TextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = MistralLarge4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = MistralLarge4TextRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

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

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

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

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

        # It may already have been prepared by e.g. `generate`
        if isinstance(attention_mask, dict):
            causal_mask = attention_mask["indexed_attention"]
        else:
            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,
                allow_is_causal_skip=False,  # Always force creation to account for causality in the indexer
            )

        # Initialize as empty storage and it will be filled in the right layers.
        shared_kv_states = KVSharingDict()

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

        topk_indices = None  # MAIN DIFF with DSV3.2
        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states, topk_indices = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_embeddings=position_embeddings,
                position_ids=position_ids,
                past_key_values=past_key_values,
                shared_kv_states=shared_kv_states,
                use_cache=use_cache,
                prev_topk_indices=topk_indices,  # MAIN DIFF with DSV3.2
                **kwargs,
            )

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


@auto_docstring(
    custom_intro="""
    Base class for MistralLarge4 outputs, with hidden states and attentions.
    """
)
@dataclass
class MistralLarge4ModelOutputWithPast(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.Tensor`, *optional*):
        A `torch.Tensor` 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.
    router_logits (`tuple(torch.Tensor)`, *optional*, returned when `output_router_logits=True` is passed):
        Tuple of `torch.Tensor` (one for each MoE layer) of shape `(batch_size * sequence_length, num_experts)`.
    """

    image_hidden_states: torch.Tensor | None = None

    router_logits: tuple[torch.Tensor] | None = None


class MistralLarge4PatchMerger(nn.Module):
    """
    Learned merging of spatial_merge_size ** 2 patches
    """

    def __init__(self, config: MistralLarge4Config):
        super().__init__()
        self.config = config

        hidden_size = config.vision_config.hidden_size
        self.spatial_merge_size = config.spatial_merge_size
        self.patch_size = self.config.vision_config.patch_size
        self.merging_layer = nn.Linear(hidden_size * self.spatial_merge_size**2, hidden_size, bias=False)

    def forward(self, image_features: torch.Tensor, image_sizes: torch.Tensor) -> torch.Tensor:
        image_sizes = [
            (image_size[0] // self.patch_size, image_size[1] // self.patch_size) for image_size in image_sizes
        ]

        tokens_per_image = [h * w for h, w in image_sizes]
        d = image_features.shape[-1]

        permuted_tensor = []
        for image_index, image_tokens in enumerate(image_features.split(tokens_per_image)):
            # Reshape image_tokens into a 2D grid
            h, w = image_sizes[image_index]
            image_grid = image_tokens.view(h, w, d).permute(2, 0, 1).unsqueeze(0)
            grid = torch.nn.functional.unfold(
                image_grid, kernel_size=self.spatial_merge_size, stride=self.spatial_merge_size
            )
            grid = grid.view(d * self.spatial_merge_size**2, -1).t()
            permuted_tensor.append(grid)

        image_features = torch.cat(permuted_tensor, dim=0)
        image_features = self.merging_layer(image_features)
        return image_features


class MistralLarge4MultiModalProjector(nn.Module):
    def __init__(self, config: MistralLarge4Config):
        super().__init__()
        self.norm = MistralLarge4RMSNorm(config.vision_config.hidden_size, eps=config.text_config.rms_norm_eps)
        self.patch_merger = MistralLarge4PatchMerger(config)
        # We have hidden_size * the number of vision feature layers
        self.num_feature_layers = (
            1 if isinstance(config.vision_feature_layer, int) else len(config.vision_feature_layer)
        )
        self.linear_1 = nn.Linear(
            config.vision_config.hidden_size * self.num_feature_layers,
            config.text_config.hidden_size,
            bias=config.multimodal_projector_bias,
        )
        self.act = ACT2FN[config.projector_hidden_act]
        self.linear_2 = nn.Linear(
            config.text_config.hidden_size, config.text_config.hidden_size, bias=config.multimodal_projector_bias
        )

    def forward(self, image_features: torch.Tensor, image_sizes: torch.Tensor):
        image_features = self.norm(image_features)
        image_features = self.patch_merger(image_features, image_sizes)
        hidden_states = self.linear_1(image_features)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        return hidden_states


@auto_docstring(
    custom_intro="""
    The MistralLarge4 model which consists of a vision backbone and a language model, without a language modeling head.
    """
)
class MistralLarge4Model(MistralLarge4PreTrainedModel):
    def __init__(self, config: MistralLarge4Config):
        super().__init__(config)
        self.vision_tower = AutoModel.from_config(config.vision_config)

        self.multi_modal_projector = MistralLarge4MultiModalProjector(config)
        self.language_model = AutoModel.from_config(config.text_config)
        self.post_init()

    @merge_with_config_defaults
    @can_return_tuple
    @select_hidden_states("vision_feature_layer")
    @auto_docstring(
        custom_intro="Obtains image last hidden states from the vision tower and apply multimodal projection."
    )
    def get_image_features(
        self,
        pixel_values: torch.Tensor,
        image_sizes: torch.Tensor,
        vision_feature_layer: int | list[int] | list[int] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        image_outputs = self.vision_tower(
            pixel_values,
            image_sizes=image_sizes,
            return_dict=True,
            **kwargs,
        )
        # If we have one vision feature layer, return the corresponding hidden states,
        # otherwise, select the hidden states of each feature layer and concatenate them
        if isinstance(vision_feature_layer, int):
            selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
        else:
            hs_pool = [image_outputs.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
            selected_image_feature = torch.cat(hs_pool, dim=-1)

        image_features = self.multi_modal_projector(selected_image_feature.squeeze(0), image_sizes)
        downsample_ratio = self.vision_tower.patch_size * self.config.spatial_merge_size
        split_sizes = (
            (torch.as_tensor(image_sizes, device=image_features.device) // downsample_ratio).prod(dim=-1).tolist()
        )
        image_outputs.pooler_output = torch.split(image_features.squeeze(0), split_sizes)
        return image_outputs

    def get_placeholder_mask(
        self, input_ids: torch.LongTensor, inputs_embeds: torch.Tensor, image_features: torch.Tensor
    ):
        """
        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

    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.Tensor | None = None,
        vision_feature_layer: int | list[int] | list[int] | None = None,
        use_cache: bool | None = None,
        image_sizes: torch.Tensor | None = None,
        mm_encoder_outputs: dict[str, BaseModelOutputWithPooling] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | MistralLarge4ModelOutputWithPast:
        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 mm_encoder_outputs is not None:
            raise ValueError("You cannot specify both pixel_values and mm_encoder_outputs at the same time")

        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=pixel_values,
                vision_feature_layer=vision_feature_layer,
                image_sizes=image_sizes,
                return_dict=True,
            )

        if mm_encoder_outputs.get("image") is not None:
            image_features = torch.cat(mm_encoder_outputs["image"].pooler_output, dim=0).to(
                inputs_embeds.device, inputs_embeds.dtype
            )
            special_image_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_features
            )
            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

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

        return MistralLarge4ModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            router_logits=outputs.router_logits,  # only diff with Mistral3Model: we also return the router logits
            image_hidden_states=image_features if mm_encoder_outputs.get("image") is not None else None,
        )


@auto_docstring(
    custom_intro="""
    Base class for MistralLarge4 causal language model (or autoregressive) outputs.
    """
)
@dataclass
class MistralLarge4CausalLMOutputWithPast(ModelOutput):
    r"""
    loss (`torch.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Language modeling loss (for next-token prediction).
    logits (`torch.Tensor` 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.Tensor`, *optional*):
        A `torch.Tensor` 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.Tensor | None = None
    logits: torch.Tensor | None = None
    past_key_values: Cache | None = None
    hidden_states: tuple[torch.Tensor] | None = None
    attentions: tuple[torch.Tensor] | None = None
    image_hidden_states: torch.Tensor | None = None
    router_logits: tuple[torch.Tensor] | None = None


@auto_docstring(
    custom_intro="""
    The MISTRAL_LARGE4 model which consists of a vision backbone and a language model.
    """
)
class MistralLarge4ForConditionalGeneration(MistralLarge4PreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}

    def __init__(self, config: MistralLarge4Config):
        super().__init__(config)
        self.model = MistralLarge4Model(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.final_logit_softcapping = self.config.get_text_config().final_logit_softcapping
        self.post_init()

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

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

    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.Tensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        image_sizes: torch.Tensor | None = None,
        mm_encoder_outputs: dict[str, BaseModelOutputWithPooling] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | MistralLarge4CausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from PIL import Image
        >>> from huggingface_hub.utils import httpx
        >>> from io import BytesIO
        >>> from transformers import AutoProcessor, MistralLarge4ForConditionalGeneration

        >>> model = MistralLarge4ForConditionalGeneration.from_pretrained("mistralai/Mistral-Small-3.1-24B-Instruct-2503")
        >>> processor = AutoProcessor.from_pretrained("mistralai/Mistral-Small-3.1-24B-Instruct-2503")

        >>> prompt = "<s>[INST][IMG]What is the image?[/INST]"
        >>> url = "http://images.cocodataset.org/val2017/000000039769.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]
        "What is the image?The image depicts two cats lying on a pink blanket."
        ```"""
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            image_sizes=image_sizes,
            mm_encoder_outputs=mm_encoder_outputs,
            **kwargs,
        )

        hidden_states = outputs[0]
        hidden_states = slice_logits_to_keep(hidden_states, logits_to_keep)
        logits = self.lm_head(hidden_states)

        # Logits softcapping
        logits = logits / self.final_logit_softcapping
        logits = torch.tanh(logits)
        logits = logits * self.final_logit_softcapping

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

        return MistralLarge4CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            router_logits=outputs.router_logits,
            image_hidden_states=outputs.image_hidden_states,
        )


__all__ = [
    "MistralLarge4PreTrainedModel",
    "MistralLarge4TextModel",
    "MistralLarge4Model",
    "MistralLarge4ForConditionalGeneration",
]
