#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/minimax_m3_vl/modular_minimax_m3_vl.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_minimax_m3_vl.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 the MiniMax AI Team and HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from collections.abc import Callable
from dataclasses import dataclass
from typing import Optional

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

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache, DynamicLayer, StaticLayer
from ...configuration_utils import PreTrainedConfig
from ...generation import GenerationMixin
from ...integrations import use_experts_implementation
from ...masking_utils import create_causal_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutputWithPast,
    BaseModelOutputWithPooling,
    MoeCausalLMOutputWithPast,
    MoeModelOutputWithPast,
)
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 ModelOutput, TransformersKwargs, auto_docstring, torch_compilable_check
from ...utils.generic import can_return_tuple, maybe_autocast, merge_with_config_defaults
from ...utils.import_utils import is_torchdynamo_compiling
from ...utils.output_capturing import OutputRecorder, capture_outputs
from .configuration_minimax_m3_vl import MiniMaxM3VLConfig, MiniMaxM3VLTextConfig, MiniMaxM3VLVisionConfig


class MiniMaxM3VLSparseCacheLayer(DynamicLayer):
    layer_type = "minimax_m3_sparse"

    def __init__(self, config: PreTrainedConfig | None = None):
        super().__init__(config)
        self.idx_keys: torch.Tensor | None = None

    def update_index(self, idx_k: torch.Tensor) -> torch.Tensor:
        """Append the new token's `idx_k` to the cache and return the full history."""
        self.idx_keys = idx_k if self.idx_keys is None else torch.cat([self.idx_keys, idx_k], dim=-2)
        return self.idx_keys

    def reorder_cache(self, beam_idx: torch.LongTensor) -> None:
        super().reorder_cache(beam_idx)
        if self.idx_keys is not None:
            self.idx_keys = self.idx_keys.index_select(0, beam_idx.to(self.idx_keys.device))

    def batch_repeat_interleave(self, repeats: int) -> None:
        super().batch_repeat_interleave(repeats)
        if self.idx_keys is not None:
            self.idx_keys = self.idx_keys.repeat_interleave(repeats, dim=0)

    def batch_select_indices(self, indices: torch.Tensor) -> None:
        super().batch_select_indices(indices)
        if self.idx_keys is not None:
            self.idx_keys = self.idx_keys[indices, ...]

    def crop(self, max_length: int) -> None:
        super().crop(max_length)
        if max_length < 0:
            max_length = self.get_seq_length() - abs(max_length)
        if self.idx_keys is not None and self.idx_keys.shape[-2] > max_length:
            self.idx_keys = self.idx_keys[..., :max_length, :]


class MiniMaxM3VLSparseStaticCacheLayer(StaticLayer):
    layer_type = "minimax_m3_sparse"

    def __init__(self, max_cache_len: int):
        super().__init__(max_cache_len)
        self.idx_keys: torch.Tensor | None = None
        # Tensor (not int) so it can be marked as a static address for cudagraphs, like `cumulative_length`.
        self.idx_cumulative_length = torch.tensor([0], dtype=int)

    def update_index(self, idx_k: torch.Tensor) -> torch.Tensor:
        """Write the new token's `idx_k` into the static buffer in place and return the whole buffer.

        The buffer's unfilled tail holds zeros, but those slots sit at key positions ahead of every
        current query, so the indexer's block- and token-level causal masking discards them — the
        returned `[B, 1, max_cache_len, D]` history is therefore safe to score against directly.
        """
        if self.idx_keys is None:
            self.idx_keys = torch.zeros(
                (idx_k.shape[0], idx_k.shape[1], self.max_cache_len, idx_k.shape[-1]),
                dtype=idx_k.dtype,
                device=idx_k.device,
            )
            self.idx_cumulative_length = self.idx_cumulative_length.to(idx_k.device)
            if not is_torchdynamo_compiling():
                torch._dynamo.mark_static_address(self.idx_keys)
                torch._dynamo.mark_static_address(self.idx_cumulative_length)

        kv_len = idx_k.shape[-2]
        cache_position = torch.arange(kv_len, device=self.idx_keys.device) + self.idx_cumulative_length
        self.idx_cumulative_length.add_(kv_len)
        try:
            self.idx_keys.index_copy_(2, cache_position, idx_k)
        except NotImplementedError:
            # Fallback for devices like MPS where index_copy_ might not be supported.
            self.idx_keys[:, :, cache_position] = idx_k
        return self.idx_keys

    def reset(self) -> None:
        super().reset()
        if self.idx_keys is not None:
            self.idx_keys.zero_()
        self.idx_cumulative_length.zero_()

    def reorder_cache(self, beam_idx: torch.LongTensor) -> None:
        super().reorder_cache(beam_idx)
        if self.idx_keys is not None:
            self.idx_keys = self.idx_keys.index_select(0, beam_idx.to(self.idx_keys.device))


class MiniMaxM3VLRMSNorm(nn.Module):
    """Gemma-style RMSNorm: normalizes in fp32 and scales by `weight + 1`."""

    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.zeros(dim))

    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())
        # Llama does x.to(float16) * w whilst MiniMaxM3VL is (x * w).to(float16)
        # See https://github.com/huggingface/transformers/pull/29402
        output = output * (1.0 + self.weight.float())
        return output.type_as(x)

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


class MiniMaxM3VLDenseMLP(nn.Module):
    def __init__(self, config: MiniMaxM3VLTextConfig, intermediate_size: int | None = None):
        super().__init__()
        inter = intermediate_size if intermediate_size is not None else config.dense_intermediate_size
        self.swiglu_alpha = config.swiglu_alpha
        self.swiglu_limit = config.swiglu_limit
        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * inter, bias=False)
        self.down_proj = nn.Linear(inter, config.hidden_size, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        gate_up = self.gate_up_proj(hidden_states)
        gate, up = gate_up.chunk(2, dim=-1)
        gate = gate.clamp(max=self.swiglu_limit)
        up = up.clamp(min=-self.swiglu_limit, max=self.swiglu_limit)
        glu = gate * torch.sigmoid(gate * self.swiglu_alpha)
        return self.down_proj((up + 1.0) * glu)


@use_experts_implementation
class MiniMaxM3VLExperts(nn.Module):
    """Collection of expert weights stored as 3D tensors."""

    def __init__(self, config: MiniMaxM3VLTextConfig):
        super().__init__()
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.intermediate_dim = config.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.limit = config.swiglu_limit
        self.swiglu_alpha = config.swiglu_alpha
        self.swiglu_limit = config.swiglu_limit

    def forward(
        self, hidden_states: torch.Tensor, top_k_index: torch.Tensor, top_k_weights: torch.Tensor
    ) -> torch.Tensor:
        final = torch.zeros_like(hidden_states)
        with torch.no_grad():
            mask = F.one_hot(top_k_index, num_classes=self.num_experts).permute(2, 1, 0)
            hit = torch.greater(mask.sum(dim=(-1, -2)), 0).nonzero()
        for expert_idx in hit:
            expert_idx = expert_idx[0]
            if expert_idx == self.num_experts:
                continue
            top_k_pos, token_idx = torch.where(mask[expert_idx])
            current = self._apply_gate(F.linear(hidden_states[token_idx], self.gate_up_proj[expert_idx]))
            current = F.linear(current, self.down_proj[expert_idx]) * top_k_weights[token_idx, top_k_pos, None]
            final.index_add_(0, token_idx, current.to(final.dtype))
        return final

    def _apply_gate(self, gate_up: torch.Tensor) -> torch.Tensor:
        # same as GPT OSS, but the weights are not interleaved
        gate, up = gate_up.chunk(2, dim=-1)
        gate = gate.clamp(max=self.swiglu_limit)
        up = up.clamp(min=-self.swiglu_limit, max=self.swiglu_limit)
        glu = gate * torch.sigmoid(gate * self.swiglu_alpha)
        return (up + 1.0) * glu


class MiniMaxM3VLTopKRouter(nn.Module):
    def __init__(self, config: MiniMaxM3VLTextConfig):
        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.empty(self.num_experts, self.hidden_dim))
        self.register_buffer("e_score_correction_bias", torch.zeros(config.num_local_experts))

    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        hidden_states = hidden_states.reshape(-1, self.hidden_dim)
        router_logits = F.linear(hidden_states.to(self.weight.dtype), self.weight)
        # Sigmoid scoring (not softmax), as in M2.
        routing_weights = F.sigmoid(router_logits.float())
        scores_for_choice = routing_weights + self.e_score_correction_bias
        _, top_k_index = torch.topk(scores_for_choice, self.top_k, dim=-1, sorted=False)
        top_k_weights = routing_weights.gather(1, top_k_index)
        top_k_weights /= top_k_weights.sum(dim=-1, keepdim=True)
        return router_logits, top_k_weights, top_k_index


class MiniMaxM3VLSparseMoeBlock(nn.Module):
    def __init__(self, config: MiniMaxM3VLTextConfig):
        super().__init__()
        self.gate = MiniMaxM3VLTopKRouter(config)
        self.experts = MiniMaxM3VLExperts(config)
        self.routed_scaling_factor = config.routed_scaling_factor
        self.shared_experts = MiniMaxM3VLDenseMLP(config, intermediate_size=config.shared_intermediate_size)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        shared_output = self.shared_experts(hidden_states)

        _, routing_weights, selected_experts = self.gate(hidden_states)
        hidden_states = self.experts(hidden_states, selected_experts, routing_weights)
        # Additional scaling
        hidden_states = hidden_states * self.routed_scaling_factor
        hidden_states = hidden_states + shared_output

        hidden_states = hidden_states.reshape(batch_size, sequence_length, hidden_dim)
        return hidden_states


class MiniMaxM3VLRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: MiniMaxM3VLConfig, 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: MiniMaxM3VLConfig | 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"]
        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
        dim = int(head_dim * partial_rotary_factor)

        attention_factor = 1.0  # Unused in this type of RoPE

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

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

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

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


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

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)

    # Keep half or full tensor for later concatenation
    rotary_dim = cos.shape[-1]
    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]

    # Apply rotary embeddings on the first half or full tensor
    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)

    # Concatenate back to full shape
    q_embed = torch.cat([q_embed, q_pass], dim=-1)
    k_embed = torch.cat([k_embed, k_pass], dim=-1)
    return q_embed, k_embed


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


class MiniMaxM3VLAttention(nn.Module):
    """
    M3 attention: per-head Gemma QK-norm + partial RoPE, optionally sparse indexer selection which require position IDs.
    """

    def __init__(self, config: MiniMaxM3VLTextConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True
        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
        self.q_norm = MiniMaxM3VLRMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.k_norm = MiniMaxM3VLRMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.indexer = (
            MiniMaxM3VLIndexer(config, layer_idx) if config.layer_types[layer_idx] == "minimax_m3_sparse" else None
        )

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

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

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

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

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )
        block_indices = None
        if self.indexer is not None:
            position_ids = kwargs.get("position_ids")
            if position_ids is None:
                position_ids = torch.arange(
                    key_states.shape[2] - query_states.shape[2], key_states.shape[2], device=query_states.device
                )
            position_ids = (position_ids if position_ids.ndim > 1 else position_ids.unsqueeze(0)).expand(
                query_states.shape[0], -1
            )
            block_indices = self.indexer(hidden_states, position_embeddings, past_key_values, position_ids)
            if self.config._attn_implementation in ("eager", "sdpa"):
                attention_mask = self.indexer.build_block_mask(
                    block_indices,
                    attention_mask,
                    key_states.shape[2],
                    query_states.dtype,
                    query_states.device,
                    position_ids,
                )

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


class MiniMaxM3VLIndexer(nn.Module):
    r"""Lightning Indexer for MiniMax M3 sparse attention.

    Scores each query against every key with a small `index_n_heads`-head
    dot-product branch, then max-pools those per-key scores into *blocks* of
    `index_block_size` keys and keeps, per query, the top-`index_topk_blocks`
    key blocks plus the `index_local_blocks` blocks immediately preceding the
    query (always visible). Selection therefore happens at the granularity of a
    *block of keys* rather than individual keys: the expensive main attention
    only has to attend the handful of selected key blocks, which is what makes
    it block-sparse (and cheaper) on long sequences.

    The `index_local_blocks` boosting their score so they always win key slots, the
    same way the deployment block-sparse kernel (MiniMax `topk_sparse`) does it.

    `forward` returns the per-query selected key-block indices
    `[B, S_q, index_topk_blocks]`. Valid indices are left-packed and `-1`
    right-pads the unused slots (future/empty blocks), and the local boost makes
    selections deduplicated -- the exact contract the block-sparse attention
    kernel consumes (it counts the valid entries, then reads them sequentially
    and would double-count a repeated block). The eager/SDPA path instead calls
    `build_block_mask`, which expands the indices into the dense
    `[B, 1, S_q, S_k]` additive mask the standard attention interface expects
    (`0` at every allowed (query, key) pair, `-inf` elsewhere).

    Like DeepSeek-V4's indexer this is purely a *selection* branch: it has no
    value projection and produces no residual output of its own (the upstream
    checkpoint disables the index-value path on every sparse layer).

    TODO: blocks are anchored to absolute key *slots* (the contiguous reshape in
    `forward` and `q_block = slot // block_size`), so left-padding shifts the block
    boundaries and the selection diverges from an unpadded run -- only right-padding
    is equivalent (same limitation as DeepSeek-V4; see `test_right_padding_does_not_leak`
    / the skipped `test_left_padding_compatibility`). For *true* left-padding equivalence
    we'd make blocking content-relative instead of slot-relative:
      1. derive block ids from `position_ids` (content positions, 0 at each row's first
         real token) rather than from absolute slots, and
      2. replace the contiguous `view(..., num_key_blocks, block_size).amax(-1)` key pool
         with a per-row position-binned pool (e.g. `scatter_reduce` over `key_position //
         block_size`), so pad never shifts the boundaries, and
      3. mask padded keys' scores to `-inf` before the pool so a pad key can't win a block
         a top-k slot.
    """

    def __init__(self, config: MiniMaxM3VLTextConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = config.index_head_dim
        self.num_heads = config.index_n_heads
        self.block_size = config.index_block_size
        self.topk_blocks = config.index_topk_blocks
        self.local_blocks = config.index_local_blocks
        self.q_proj = nn.Linear(config.hidden_size, config.index_n_heads * config.index_head_dim, bias=False)
        self.k_proj = nn.Linear(config.hidden_size, config.index_head_dim, bias=False)
        self.q_norm = MiniMaxM3VLRMSNorm(config.index_head_dim, eps=config.rms_norm_eps)
        self.k_norm = MiniMaxM3VLRMSNorm(config.index_head_dim, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        past_key_values: Cache | None,
        position_ids: torch.Tensor,
    ) -> torch.Tensor:
        batch, q_len, _ = hidden_states.shape
        idx_q = self.q_proj(hidden_states).view(batch, q_len, -1, self.head_dim)
        idx_q = self.q_norm(idx_q).transpose(1, 2)  # [B, H_idx, Sq, D]
        idx_k = self.k_proj(hidden_states).view(batch, q_len, 1, self.head_dim)
        idx_k = self.k_norm(idx_k).transpose(1, 2)  # [B, 1, Sq, D]
        cos, sin = position_embeddings
        idx_q, idx_k = apply_rotary_pos_emb(idx_q, idx_k, cos[..., : self.head_dim], sin[..., : self.head_dim])

        if past_key_values is not None:
            idx_k = past_key_values.layers[self.layer_idx].update_index(idx_k)

        k_len = idx_k.shape[2]
        num_key_blocks = -(-k_len // self.block_size)  # ceil-div
        pad = num_key_blocks * self.block_size - k_len

        scores = torch.matmul(idx_q.float(), idx_k.float().transpose(-1, -2))
        k_positions = torch.arange(k_len, device=idx_q.device)
        token_future = k_positions[None, None, None, :] > position_ids[:, None, :, None]  # [B, 1, S_q, S_k]
        scores = scores.masked_fill(token_future, float("-inf"))
        if pad:
            scores = F.pad(scores, (0, pad), value=float("-inf"))
        scores = scores.view(batch, self.num_heads, q_len, num_key_blocks, self.block_size)
        block_scores = scores.amax(dim=-1).amax(dim=1)  # -> [B, S_q, num_key_blocks]

        q_block = position_ids // self.block_size  # [B, S_q]

        if self.local_blocks > 0:
            local = torch.arange(self.local_blocks, device=idx_q.device)
            local_idx = (q_block[..., None] - local.view(1, 1, -1)).clamp(min=0)  # [B, S_q, local]
            block_scores.scatter_(-1, local_idx, float("inf"))

        # Slots that fall on a future/empty block keep their `-inf`
        # score, which top-k sorts to the end, so tagging them `-1` yields left-packed block indices
        # with `-1` right-padding which is the format expect by block-sparse attention kernel.
        topk = min(self.topk_blocks, num_key_blocks)
        topk_scores, topk_indices = block_scores.topk(topk, dim=-1)  # [B, S_q, topk]
        return topk_indices.masked_fill(topk_scores == float("-inf"), -1)

    def build_block_mask(
        self,
        block_indices: torch.Tensor,
        attention_mask: torch.Tensor | None,
        key_length: int,
        dtype: torch.dtype,
        device: torch.device,
        position_ids: torch.Tensor,
    ) -> torch.Tensor:
        """
        We build the full 4D attention mask (Batch, query, key, head)
        """
        batch, q_len, _ = block_indices.shape
        num_key_blocks = -(-key_length // self.block_size)

        # Scatter the kept blocks to `0`; `-1` slots land in a throwaway column we drop afterwards.
        safe = block_indices.masked_fill(block_indices < 0, num_key_blocks)
        bias = block_indices.new_full((batch, q_len, num_key_blocks + 1), float("-inf"), dtype=dtype)
        bias.scatter_(-1, safe, 0.0)
        bias = bias[..., :num_key_blocks]

        # Broadcast the per-block keep/drop verdict back onto every key (block granularity), add head axis.
        block_keep = (bias == 0.0).repeat_interleave(self.block_size, dim=-1)[..., :key_length].unsqueeze(1)

        # Compose block-selection with the existing mask, then emit a single additive float mask.
        if attention_mask is not None:
            padding_mask = attention_mask if attention_mask.dtype == torch.bool else attention_mask == 0
            keep = block_keep & padding_mask
        else:
            k_positions = torch.arange(key_length, device=device)
            token_future = k_positions[None, None, None, :] > position_ids[:, None, :, None]  # [B, 1, S_q, S_k]
            keep = block_keep & ~token_future
        min_dtype = torch.finfo(dtype).min
        return torch.zeros(keep.shape, dtype=dtype, device=device).masked_fill(~keep, min_dtype)


class MiniMaxM3VLDecoderLayer(GradientCheckpointingLayer):
    """M3 decoder layer: per-layer dense/MoE MLP and dense/sparse attention."""

    def __init__(self, config: MiniMaxM3VLTextConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = MiniMaxM3VLAttention(config, layer_idx)
        self.mlp = (
            MiniMaxM3VLSparseMoeBlock(config)
            if config.mlp_layer_types[layer_idx] == "sparse"
            else MiniMaxM3VLDenseMLP(config, intermediate_size=config.dense_intermediate_size)
        )
        self.input_layernorm = MiniMaxM3VLRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = MiniMaxM3VLRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

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


@auto_docstring
class MiniMaxM3VLPreTrainedModel(PreTrainedModel):
    config: MiniMaxM3VLConfig | MiniMaxM3VLTextConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["MiniMaxM3VLDecoderLayer", "MiniMaxM3VLVisionEncoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = False
    _supports_sdpa = True
    _supports_flex_attn = False
    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "router_logits": OutputRecorder(MiniMaxM3VLTopKRouter, index=0),
        "hidden_states": MiniMaxM3VLDecoderLayer,
        "attentions": MiniMaxM3VLAttention,
    }
    input_modalities = ("image", "video", "text")
    _keys_to_ignore_on_load_unexpected = [r"(^|\.)mtp\..*"]
    _compatible_flash_implementations = ["MiniMaxAI/msa"]

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        std = getattr(self.config, "initializer_range", 0.02)
        if isinstance(module, MiniMaxM3VLExperts):
            init.normal_(module.gate_up_proj, mean=0.0, std=std)
            init.normal_(module.down_proj, mean=0.0, std=std)
        elif isinstance(module, MiniMaxM3VLTopKRouter):
            init.normal_(module.weight, mean=0.0, std=std)
            init.zeros_(module.e_score_correction_bias)
        elif isinstance(module, MiniMaxM3VLRMSNorm):
            init.zeros_(module.weight)


@auto_docstring
class MiniMaxM3VLTextModel(MiniMaxM3VLPreTrainedModel):
    config: MiniMaxM3VLTextConfig

    def __init__(self, config: MiniMaxM3VLTextConfig):
        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([MiniMaxM3VLDecoderLayer(config, i) for i in range(config.num_hidden_layers)])
        self.norm = MiniMaxM3VLRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = MiniMaxM3VLRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> 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 use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

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

        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)

        if isinstance(attention_mask, dict):
            causal_mask = next(iter(attention_mask.values()))
        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,
            )

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

        # `position_ids` is threaded to every layer so the sparse layers' lightning indexer can anchor
        # block selection to each query's content position (see `MiniMaxM3VLIndexer`).
        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)

        return MoeModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


def load_balancing_loss_func(
    gate_logits: torch.Tensor | tuple[torch.Tensor] | None,
    num_experts: int | None = None,
    top_k=2,
    attention_mask: torch.Tensor | None = None,
) -> torch.Tensor | int:
    r"""
    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.

    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
    experts is too unbalanced.

    Args:
        gate_logits:
            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
            shape [batch_size X sequence_length, num_experts].
        num_experts:
            Number of experts
        top_k:
            The number of experts to route per-token, can be also interpreted as the `top-k` routing
            parameter.
        attention_mask (`torch.Tensor`, *optional*):
            The attention_mask used in forward function
            shape [batch_size X sequence_length] if not None.

    Returns:
        The auxiliary loss.
    """
    if gate_logits is None or not isinstance(gate_logits, tuple):
        return 0

    if isinstance(gate_logits, tuple):
        compute_device = gate_logits[0].device
        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)

    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)

    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)

    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)

    if attention_mask is None:
        # Compute the percentage of tokens routed to each experts
        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)

        # Compute the average probability of routing to these experts
        router_prob_per_expert = torch.mean(routing_weights, dim=0)
    else:
        batch_size, sequence_length = attention_mask.shape
        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)

        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
        expert_attention_mask = (
            attention_mask[None, :, :, None, None]
            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
            .reshape(-1, top_k, num_experts)
            .to(compute_device)
        )

        # Compute the percentage of tokens routed to each experts
        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
            expert_attention_mask, dim=0
        )

        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
        router_per_expert_attention_mask = (
            attention_mask[None, :, :, None]
            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
            .reshape(-1, num_experts)
            .to(compute_device)
        )

        # Compute the average probability of routing to these experts
        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
            router_per_expert_attention_mask, dim=0
        )

    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
    return overall_loss * num_experts


@auto_docstring
class MiniMaxM3VLForCausalLM(MiniMaxM3VLPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
    config: MiniMaxM3VLTextConfig

    def __init__(self, config: MiniMaxM3VLTextConfig):
        super().__init__(config)
        self.model = MiniMaxM3VLTextModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.router_aux_loss_coef = config.router_aux_loss_coef
        self.num_experts = config.num_local_experts
        self.num_experts_per_tok = config.num_experts_per_tok

        # 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,
        output_router_logits: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> MoeCausalLMOutputWithPast:
        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, MiniMaxM3VLForCausalLM

        >>> model = MiniMaxM3VLForCausalLM.from_pretrained("mistralai/MiniMaxM3VL-8x7B-v0.1")
        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/MiniMaxM3VL-8x7B-v0.1")

        >>> 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."
        ```"""

        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs: MoeModelOutputWithPast = 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,
            output_router_logits=output_router_logits,
            **kwargs,
        )

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

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

        aux_loss = None
        if output_router_logits:
            aux_loss = load_balancing_loss_func(
                outputs.router_logits,
                self.num_experts,
                self.num_experts_per_tok,
                attention_mask,
            )
            if labels is not None:
                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device

        return MoeCausalLMOutputWithPast(
            loss=loss,
            aux_loss=aux_loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            router_logits=outputs.router_logits,
        )


class MiniMaxM3VLVisionEmbeddings(nn.Module):
    """Patch embedding, identical to [`Qwen2_5_VisionPatchEmbed`] (reads its dims from the vision
    config). The upstream checkpoint stores the conv as `patch_embedding`, renamed to the
    inherited `proj` in the conversion mapping."""

    def __init__(self, config) -> None:
        super().__init__()
        self.patch_size = config.patch_size
        self.temporal_patch_size = config.temporal_patch_size
        self.in_channels = config.num_channels
        self.embed_dim = config.hidden_size

        kernel_size = [self.temporal_patch_size, self.patch_size, self.patch_size]
        self.proj = nn.Conv3d(
            self.in_channels, self.embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        target_dtype = self.proj.weight.dtype
        hidden_states = hidden_states.view(
            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
        )
        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
        return hidden_states


class MiniMaxM3VL3DRotaryEmbedding(nn.Module):
    r"""3D RoPE for the vision tower: each patch is rotated by its `(T, H, W)` grid position.

    `2 * (head_dim // 2)` rotary dims are split evenly across the three axes (each rounded
    down to a multiple of 2), giving `axis_dim` dims per axis and `axis_dim // 2` frequencies::

        |<------------------ rotated (3 * axis_dim) ------------------>|<- pass ->|
        +--------------------+--------------------+--------------------+----------+
        |     T  (frames)    |      H  (rows)     |      W  (cols)     |          |
        |      axis_dim      |      axis_dim      |      axis_dim      |          |
        +--------------------+--------------------+--------------------+----------+

    Each axis' coordinate scales its own band of frequencies; the bands are concatenated as
    `T|H|W` and duplicated via `cat([f, f])` to pair with the half-rotation in
    `apply_rotary_pos_emb_vision`. Any head dims past `3 * axis_dim` are left unrotated.
    """

    def __init__(self, head_dim: int, theta: float = 10000.0, spatial_merge_size: int = 1):
        super().__init__()
        # `2 * (head_dim // 2)` rotary dims are split evenly across T/H/W, each axis rounded
        # down to a multiple of 2. With head_dim=80 that is 26 dims/axis (39 freqs total); the
        # remaining `head_dim - 3 * axis_dim` dims are never rotated (they pass through).
        rope_dims = 2 * (head_dim // 2)
        self.axis_dim = 2 * ((rope_dims // 3) // 2)
        self.spatial_merge_size = spatial_merge_size
        self.theta = theta

    def forward(
        self, grid_thw: torch.Tensor, device: torch.device, dtype: torch.dtype
    ) -> tuple[torch.Tensor, torch.Tensor]:
        m = self.spatial_merge_size
        coords = []
        for t, h, w in grid_thw.tolist():
            hi = torch.arange(h).unsqueeze(1).expand(-1, w)
            hi = hi.reshape(h // m, m, w // m, m).permute(0, 2, 1, 3).flatten()
            wi = torch.arange(w).unsqueeze(0).expand(h, -1)
            wi = wi.reshape(h // m, m, w // m, m).permute(0, 2, 1, 3).flatten()
            ti = torch.arange(t).repeat_interleave(h * w)
            coords.append(torch.stack([ti, hi.repeat(t), wi.repeat(t)], dim=-1))
        coords = torch.cat(coords).to(device=device, dtype=torch.float32)

        # meta device init was having trouble when it was registered. TODO standardize?
        inv_freq = 1.0 / (
            self.theta ** (torch.arange(0, self.axis_dim, 2, dtype=torch.float32, device=device) / self.axis_dim)
        )
        freqs = torch.cat([coords[:, i : i + 1] * inv_freq for i in range(3)], dim=-1)
        emb = torch.cat([freqs, freqs], dim=-1)
        return emb.cos().to(dtype), emb.sin().to(dtype)


def apply_rotary_pos_emb_vision(
    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
    # Only the first `rot_dim` head dims carry 3D RoPE; the tail passes through untouched.
    rot_dim = cos.shape[-1]
    cos, sin = cos[None, :, None, :], sin[None, :, None, :]
    q_rot, q_pass = q[..., :rot_dim], q[..., rot_dim:]
    k_rot, k_pass = k[..., :rot_dim], k[..., rot_dim:]
    q_rot = q_rot * cos + rotate_half(q_rot) * sin
    k_rot = k_rot * cos + rotate_half(k_rot) * sin
    return torch.cat([q_rot, q_pass], dim=-1), torch.cat([k_rot, k_pass], dim=-1)


class MiniMaxM3VLVisionAttention(nn.Module):
    """CLIP-style vision attention; the only difference from [`CLIPAttention`] is
    that queries and keys are rotated by the tower's 3D RoPE before the
    (interface-dispatched) scaled dot-product attention."""

    def __init__(self, config: MiniMaxM3VLVisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_heads
        self.scale = self.head_dim**-0.5
        self.dropout = config.attention_dropout
        self.is_causal = False

        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
        # The vision tower has no grouped-query attention; the shared eager kernel
        # still expects this attribute to drive its (no-op) `repeat_kv`.
        self.num_key_value_groups = 1

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        """Input shape: Batch x Time x Channel"""
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        queries = self.q_proj(hidden_states).view(hidden_shape)
        keys = self.k_proj(hidden_states).view(hidden_shape)
        values = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        queries, keys = apply_rotary_pos_emb_vision(queries, keys, cos, sin)
        queries, keys = queries.transpose(1, 2), keys.transpose(1, 2)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )
        attn_output, attn_weights = attention_interface(
            self,
            queries,
            keys,
            values,
            attention_mask,
            scaling=self.scale,
            dropout=0.0 if not self.training else self.dropout,
            **kwargs,
        )
        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        return self.out_proj(attn_output), attn_weights


class MiniMaxM3VLVisionMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.activation_fn = ACT2FN[config.hidden_act]
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)

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


class MiniMaxM3VLVisionEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: MiniMaxM3VLVisionConfig):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.self_attn = MiniMaxM3VLVisionAttention(config)
        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
        self.mlp = MiniMaxM3VLVisionMLP(config)
        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)

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

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

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

        return hidden_states


@auto_docstring
class MiniMaxM3VLVisionModel(MiniMaxM3VLPreTrainedModel):
    """CLIP-like vision tower with Conv3d patch embed + 3D RoPE."""

    config: MiniMaxM3VLVisionConfig
    main_input_name = "pixel_values"
    _can_record_outputs = {
        "hidden_states": MiniMaxM3VLVisionEncoderLayer,
        "attentions": MiniMaxM3VLVisionAttention,
    }

    def __init__(self, config: MiniMaxM3VLVisionConfig):
        super().__init__(config)
        self.embeddings = MiniMaxM3VLVisionEmbeddings(config)
        self.pre_layrnorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.layers = nn.ModuleList([MiniMaxM3VLVisionEncoderLayer(config) for _ in range(config.num_hidden_layers)])
        head_dim = config.hidden_size // config.num_attention_heads
        self.rotary_emb = MiniMaxM3VL3DRotaryEmbedding(
            head_dim, theta=config.rope_parameters["rope_theta"], spatial_merge_size=config.spatial_merge_size
        )
        self.post_init()

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self, pixel_values: torch.Tensor, image_grid_thw: torch.Tensor, **kwargs: Unpack[TransformersKwargs]
    ) -> BaseModelOutputWithPooling:
        r"""
        image_grid_thw (`torch.Tensor` of shape `(num_images, 3)`):
            The temporal, height and width of feature shape of each image.
        """
        embeds = self.embeddings(pixel_values).to(self.pre_layrnorm.weight.dtype)
        cos, sin = self.rotary_emb(image_grid_thw, device=embeds.device, dtype=embeds.dtype)
        hidden_states = self.pre_layrnorm(embeds).unsqueeze(0)
        for layer in self.layers:
            hidden_states = layer(hidden_states, attention_mask=None, position_embeddings=(cos, sin), **kwargs)
        return BaseModelOutputWithPooling(last_hidden_state=hidden_states, pooler_output=hidden_states[:, 0])


class MiniMaxM3VLMultiModalProjector(nn.Module):
    """Projects each vision patch from `vision_config.hidden_size` to `text_config.hidden_size`
    (GELU MLP), then groups `spatial_merge_size**2` neighbouring patches into the channel dim and
    fuses them back to a single `text_config.hidden_size` token with a second GELU MLP."""

    def __init__(self, config: MiniMaxM3VLConfig):
        super().__init__()
        text_hidden = config.text_config.hidden_size
        self.spatial_merge_size = config.vision_config.spatial_merge_size
        self.linear_1 = nn.Linear(config.vision_config.hidden_size, config.projector_hidden_size, bias=True)
        self.act = ACT2FN["gelu"]
        self.linear_2 = nn.Linear(config.projector_hidden_size, text_hidden, bias=True)
        self.merge_linear_1 = nn.Linear(config.merged_hidden_size, config.projector_hidden_size, bias=True)
        self.merge_act = ACT2FN["gelu"]
        self.merge_linear_2 = nn.Linear(config.projector_hidden_size, text_hidden, bias=True)

    def forward(self, image_features: torch.Tensor) -> torch.Tensor:
        hidden_states = self.linear_2(self.act(self.linear_1(image_features)))
        hidden_states = hidden_states.reshape(hidden_states.shape[0] // (self.spatial_merge_size**2), -1)
        return self.merge_linear_2(self.merge_act(self.merge_linear_1(hidden_states)))


@auto_docstring(
    custom_intro="""
    Base class for MiniMaxM3VL outputs, with hidden states and attentions.
    """
)
@dataclass
class MiniMaxM3VLModelOutputWithPast(BaseModelOutputWithPast):
    r"""
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

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

    image_hidden_states: torch.FloatTensor | None = None

    video_hidden_states: torch.FloatTensor | None = None


@auto_docstring(
    custom_intro="""
    Base class for MiniMaxM3VL causal language model (or autoregressive) outputs.
    """
)
@dataclass
class MiniMaxM3VLCausalLMOutputWithPast(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 `(num_image_patches, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    video_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(num_video_patches, hidden_size)`.
        video_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

    video_hidden_states: torch.FloatTensor | None = None


@auto_docstring(custom_intro="MiniMax M3 VL backbone (vision + projector + text), without LM head.")
class MiniMaxM3VLModel(MiniMaxM3VLPreTrainedModel):
    config: MiniMaxM3VLConfig

    def __init__(self, config: MiniMaxM3VLConfig):
        super().__init__(config)
        self.vision_tower = MiniMaxM3VLVisionModel(config.vision_config)
        self.multi_modal_projector = MiniMaxM3VLMultiModalProjector(config)
        self.language_model = MiniMaxM3VLTextModel(config.text_config)
        self.post_init()

    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring(
        custom_intro="Obtains image last hidden states from the vision tower and apply multimodal projection."
    )
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_grid_thw: torch.Tensor,
        **kwargs,
    ) -> BaseModelOutputWithPooling:
        r"""
        image_grid_thw (`torch.Tensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of each image's feature grid, used to build the vision 3D RoPE
            and to merge patch features.
        """
        # Return the raw vision-tower output (so callers can inspect hidden states /
        # attentions) while stashing the projected + spatially-merged features —
        # ready to scatter into the text embeddings — in `pooler_output`.
        vision_outputs = self.vision_tower(pixel_values=pixel_values, image_grid_thw=image_grid_thw, **kwargs)
        vision_outputs.pooler_output = self.multi_modal_projector(vision_outputs.last_hidden_state.squeeze(0))
        return vision_outputs

    def get_placeholder_mask(
        self,
        input_ids: torch.LongTensor,
        inputs_embeds: torch.FloatTensor,
        image_features: torch.FloatTensor | None = None,
        video_features: torch.FloatTensor | None = None,
    ):
        """
        Obtains the image/video placeholder masks from `input_ids` or `inputs_embeds`, and checks that the
        placeholder token count matches the multimodal feature length. Raises if they differ.
        """
        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)
            special_video_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_video_mask = special_video_mask.all(-1)
        else:
            special_image_mask = input_ids == self.config.image_token_id
            special_video_mask = input_ids == self.config.video_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)
        if image_features is not None:
            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]}",
            )

        n_video_tokens = special_video_mask.sum()
        special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if video_features is not None:
            torch_compilable_check(
                inputs_embeds[special_video_mask].numel() == video_features.numel(),
                f"Video features and video tokens do not match, tokens: {n_video_tokens}, features: {video_features.shape[0]}",
            )
        return special_image_mask, special_video_mask

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_values_videos: torch.FloatTensor | None = None,
        image_grid_thw: torch.Tensor | None = None,
        video_grid_thw: 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.FloatTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | MiniMaxM3VLModelOutputWithPast:
        r"""
        image_grid_thw (`torch.Tensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of each image's feature grid, used to build the vision 3D RoPE
            and to merge patch features.
        video_grid_thw (`torch.Tensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of each video's feature grid, used to build the vision 3D RoPE
            and to merge patch features.
        """
        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.get_input_embeddings()(input_ids)

        image_features = None
        if pixel_values is not None:
            image_features = self.get_image_features(
                pixel_values=pixel_values, image_grid_thw=image_grid_thw
            ).pooler_output.to(inputs_embeds.device, inputs_embeds.dtype)

        video_features = None
        if pixel_values_videos is not None:
            video_features = self.get_video_features(
                pixel_values_videos=pixel_values_videos, video_grid_thw=video_grid_thw
            ).pooler_output.to(inputs_embeds.device, inputs_embeds.dtype)

        image_mask, video_mask = self.get_placeholder_mask(
            input_ids, inputs_embeds, image_features=image_features, video_features=video_features
        )
        if image_features is not None:
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_features)
        if video_features is not None:
            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_features)

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

        return MiniMaxM3VLModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=getattr(outputs, "hidden_states", None),
            attentions=getattr(outputs, "attentions", None),
            image_hidden_states=image_features,
            video_hidden_states=video_features,
        )

    @merge_with_config_defaults
    @can_return_tuple
    @auto_docstring(
        custom_intro="Obtains video last hidden states from the vision tower and apply multimodal projection."
    )
    def get_video_features(
        self,
        pixel_values_videos: torch.FloatTensor,
        video_grid_thw: torch.Tensor,
        **kwargs,
    ) -> BaseModelOutputWithPooling:
        r"""
        pixel_values_videos (`torch.FloatTensor`):
            The tensors corresponding to the input video frames.
        video_grid_thw (`torch.Tensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of each video's feature grid, used to build the vision 3D RoPE
            and to merge patch features.
        """
        # Video frames flow through the same vision pipeline as images (the tower is
        # grid-agnostic); only the placeholder token they scatter into differs.
        vision_outputs = self.vision_tower(pixel_values=pixel_values_videos, image_grid_thw=video_grid_thw, **kwargs)
        vision_outputs.pooler_output = self.multi_modal_projector(vision_outputs.last_hidden_state.squeeze(0))
        return vision_outputs


@auto_docstring(custom_intro="MiniMax M3 VL full model with LM head (text + vision).")
class MiniMaxM3SparseForConditionalGeneration(MiniMaxM3VLPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
    config: MiniMaxM3VLConfig

    def __init__(self, config: MiniMaxM3VLConfig):
        super().__init__(config)
        self.model = MiniMaxM3VLModel(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)
        self.post_init()

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

    @auto_docstring
    def get_image_features(self, pixel_values, image_grid_thw, **kwargs) -> tuple | BaseModelOutputWithPooling:
        r"""
        image_grid_thw (`torch.Tensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of each image's feature grid, used to build the vision 3D RoPE
            and to merge patch features.
        """
        return self.model.get_image_features(pixel_values, image_grid_thw, **kwargs)

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        pixel_values: torch.FloatTensor | None = None,
        pixel_values_videos: torch.FloatTensor | None = None,
        image_grid_thw: torch.Tensor | None = None,
        video_grid_thw: 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.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | MiniMaxM3VLCausalLMOutputWithPast:
        r"""
        image_grid_thw (`torch.Tensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of each image's feature grid, used to build the vision 3D RoPE
            and to merge patch features.
        video_grid_thw (`torch.Tensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of each video's feature grid, used to build the vision 3D RoPE
            and to merge patch features.
        """
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )
        hidden_states = outputs.last_hidden_state
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

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

        return MiniMaxM3VLCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=outputs.image_hidden_states,
            video_hidden_states=outputs.video_hidden_states,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        pixel_values_videos=None,
        attention_mask=None,
        logits_to_keep=None,
        is_first_iteration=False,
        **kwargs,
    ):
        # Overwritten -- pixel inputs are merged into the cache on the first step, so we
        # only forward them once (image and video alike).
        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            logits_to_keep=logits_to_keep,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        if is_first_iteration or not kwargs.get("use_cache", True):
            model_inputs["pixel_values"] = pixel_values
            model_inputs["pixel_values_videos"] = pixel_values_videos

        return model_inputs

    def get_video_features(self, pixel_values_videos, video_grid_thw, **kwargs):
        r"""
        pixel_values_videos (`torch.FloatTensor`):
            The tensors corresponding to the input video frames.
        video_grid_thw (`torch.Tensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of each video's feature grid, used to build the vision 3D RoPE
            and to merge patch features.
        """
        return self.model.get_video_features(pixel_values_videos, video_grid_thw, **kwargs)


__all__ = [
    "MiniMaxM3VLForCausalLM",
    "MiniMaxM3SparseForConditionalGeneration",
    "MiniMaxM3VLModel",
    "MiniMaxM3VLPreTrainedModel",
    "MiniMaxM3VLTextModel",
    "MiniMaxM3VLVisionModel",
]