inference-optimization/MiniMax-M2.5-BF16

1#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨2#           This file was automatically generated from src/transformers/models/minimax_m2/modular_minimax_m2.py.3#               Do NOT edit this file manually as any edits will be overwritten by the generation of4#             the file from the modular. If any change should be done, please apply the change to the5#                          modular_minimax_m2.py file directly. One of our CI enforces this.6#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨7# coding=utf-88# Copyright 2025 the HuggingFace Team. All rights reserved.9#10# Licensed under the Apache License, Version 2.0 (the "License");11# you may not use this file except in compliance with the License.12# You may obtain a copy of the License at13#14#     http://www.apache.org/licenses/LICENSE-2.015#16# Unless required by applicable law or agreed to in writing, software17# distributed under the License is distributed on an "AS IS" BASIS,18# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.19# See the License for the specific language governing permissions and20# limitations under the License.21 22 23from collections.abc import Callable24from typing import Optional, Union, Unpack25 26import torch27from torch import nn28 29from transformers.activations import ACT2FN30from transformers.cache_utils import Cache, DynamicCache31from transformers.generation import GenerationMixin32from transformers.integrations import use_kernel_forward_from_hub33from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask34from transformers.modeling_flash_attention_utils import FlashAttentionKwargs35from transformers.modeling_layers import (36    GenericForQuestionAnswering,37    GenericForSequenceClassification,38    GenericForTokenClassification,39    GradientCheckpointingLayer,40)41from transformers.modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast42from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update43from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel44from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple45from transformers.utils.deprecation import deprecate_kwarg46from transformers.utils.generic import OutputRecorder, check_model_inputs47from .configuration_minimax_m2 import MiniMaxM2Config48 49 50class MiniMaxM2MLP(nn.Module):51    def __init__(self, config: MiniMaxM2Config):52        super().__init__()53        self.ffn_dim = config.intermediate_size54        self.hidden_dim = config.hidden_size55 56        self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)57        self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)58        self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)59 60        self.act_fn = ACT2FN[config.hidden_act]61 62    def forward(self, hidden_states):63        current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)64        current_hidden_states = self.w2(current_hidden_states)65        return current_hidden_states66 67 68class MiniMaxM2Experts(nn.ModuleList):69    """70    ModuleList of experts.71    """72 73    def __init__(self, config: MiniMaxM2Config):74        super().__init__()75        self.top_k = config.num_experts_per_tok76        self.num_experts = config.num_local_experts77        for _ in range(self.num_experts):78            self.append(MiniMaxM2MLP(config))79 80    def forward(81        self, hidden_states: torch.Tensor, top_k_index: torch.Tensor, top_k_weights: torch.Tensor82    ) -> torch.Tensor:83        """84        Args:85            hidden_states: (batch_size * sequence_length, hidden_dim)86            selected_experts: (batch_size * sequence_length, top_k)87            routing_weights: (batch_size * sequence_length, top_k)88        Returns:89            (batch_size * sequence_length, hidden_dim)90        """91        final_hidden_states = torch.zeros_like(hidden_states)92        expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts).permute(2, 1, 0)93 94        expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()95        for expert_idx in expert_hit:96            idx, top_x = torch.where(expert_mask[expert_idx].squeeze(0))97            current_state = hidden_states[None, top_x].reshape(-1, hidden_states.shape[-1])98            current_hidden_states = self[expert_idx](current_state) * top_k_weights[top_x, idx, None]99            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))100        return final_hidden_states101 102 103class MiniMaxM2SparseMoeBlock(nn.Module):104    def __init__(self, config):105        super().__init__()106        self.top_k = config.num_experts_per_tok107        self.jitter_noise = config.router_jitter_noise108        self.gate = nn.Linear(config.hidden_size, config.num_local_experts, bias=False)109        self.experts = MiniMaxM2Experts(config)110        self.register_buffer("e_score_correction_bias", torch.zeros(config.num_local_experts))111 112    def route_tokens_to_experts(self, router_logits):113        routing_weights = torch.nn.functional.sigmoid(router_logits.float())114        scores_for_choice = routing_weights + self.e_score_correction_bias115        _, top_k_index = torch.topk(scores_for_choice, self.top_k, dim=-1, sorted=False)116        top_k_weights = routing_weights.gather(1, top_k_index)117        top_k_weights /= top_k_weights.sum(dim=-1, keepdim=True)118        return top_k_index, top_k_weights.to(router_logits.dtype)119 120    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:121        batch_size, sequence_length, hidden_dim = hidden_states.shape122        if self.training and self.jitter_noise > 0:123            hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise)124        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])125        router_logits = self.gate(hidden_states)126        top_k_index, top_k_weights = self.route_tokens_to_experts(router_logits)127        hidden_states = self.experts(hidden_states, top_k_index, top_k_weights.to(hidden_states.dtype))128        hidden_states = hidden_states.reshape(batch_size, sequence_length, hidden_dim)129        return hidden_states, router_logits130 131 132@use_kernel_forward_from_hub("RMSNorm")133class MiniMaxM2RMSNorm(nn.Module):134    def __init__(self, hidden_size, eps=1e-6):135        """136        MiniMaxM2RMSNorm is equivalent to T5LayerNorm137        """138        super().__init__()139        self.weight = nn.Parameter(torch.ones(hidden_size))140        self.variance_epsilon = eps141 142    def forward(self, hidden_states):143        input_dtype = hidden_states.dtype144        hidden_states = hidden_states.to(torch.float32)145        variance = hidden_states.pow(2).mean(-1, keepdim=True)146        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)147        return self.weight * hidden_states.to(input_dtype)148 149    def extra_repr(self):150        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"151 152 153def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:154    """155    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,156    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)157    """158    batch, num_key_value_heads, slen, head_dim = hidden_states.shape159    if n_rep == 1:160        return hidden_states161    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)162    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)163 164 165def eager_attention_forward(166    module: nn.Module,167    query: torch.Tensor,168    key: torch.Tensor,169    value: torch.Tensor,170    attention_mask: Optional[torch.Tensor],171    scaling: float,172    dropout: float = 0.0,173    **kwargs: Unpack[TransformersKwargs],174):175    key_states = repeat_kv(key, module.num_key_value_groups)176    value_states = repeat_kv(value, module.num_key_value_groups)177 178    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling179    if attention_mask is not None:180        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]181        attn_weights = attn_weights + causal_mask182 183    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)184    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)185    attn_output = torch.matmul(attn_weights, value_states)186    attn_output = attn_output.transpose(1, 2).contiguous()187 188    return attn_output, attn_weights189 190 191def rotate_half(x):192    """Rotates half the hidden dims of the input."""193    x1 = x[..., : x.shape[-1] // 2]194    x2 = x[..., x.shape[-1] // 2 :]195    return torch.cat((-x2, x1), dim=-1)196 197 198def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):199    """Applies Rotary Position Embedding to the query and key tensors.200 201    Args:202        q (`torch.Tensor`): The query tensor.203        k (`torch.Tensor`): The key tensor.204        cos (`torch.Tensor`): The cosine part of the rotary embedding.205        sin (`torch.Tensor`): The sine part of the rotary embedding.206        position_ids (`torch.Tensor`, *optional*):207            Deprecated and unused.208        unsqueeze_dim (`int`, *optional*, defaults to 1):209            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and210            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note211            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and212            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes213            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have214            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.215    Returns:216        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.217    """218    cos = cos.unsqueeze(unsqueeze_dim)219    sin = sin.unsqueeze(unsqueeze_dim)220 221    # Keep half or full tensor for later concatenation222    rotary_dim = cos.shape[-1]223    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]224    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]225 226    # Apply rotary embeddings on the first half or full tensor227    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)228    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)229 230    # Concatenate back to full shape231    q_embed = torch.cat([q_embed, q_pass], dim=-1)232    k_embed = torch.cat([k_embed, k_pass], dim=-1)233    return q_embed, k_embed234 235 236class MiniMaxM2Attention(nn.Module):237    """Multi-headed attention from 'Attention Is All You Need' paper"""238 239    def __init__(self, config: MiniMaxM2Config, layer_idx: int):240        super().__init__()241        self.config = config242        self.layer_idx = layer_idx243        self.head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads244        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads245        self.scaling = self.head_dim**-0.5246        self.attention_dropout = config.attention_dropout247        self.is_causal = True248        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=False)249        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)250        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)251        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)252 253        self.use_qk_norm = config.use_qk_norm254        if self.use_qk_norm:255            self.q_norm = MiniMaxM2RMSNorm(self.head_dim * config.num_attention_heads, eps=config.rms_norm_eps)256            self.k_norm = MiniMaxM2RMSNorm(self.head_dim * config.num_key_value_heads, eps=config.rms_norm_eps)257 258    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")259    def forward(260        self,261        hidden_states: torch.Tensor,262        position_embeddings: tuple[torch.Tensor, torch.Tensor],263        attention_mask: Optional[torch.Tensor],264        past_key_values: Optional[Cache] = None,265        cache_position: Optional[torch.LongTensor] = None,266        **kwargs: Unpack[FlashAttentionKwargs],267    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:268        input_shape = hidden_states.shape[:-1]269        hidden_shape = (*input_shape, -1, self.head_dim)270 271        query_states = self.q_proj(hidden_states)272        key_states = self.k_proj(hidden_states)273        value_states = self.v_proj(hidden_states)274 275        if self.use_qk_norm:  # main diff from Llama276            query_states = self.q_norm(query_states)277            key_states = self.k_norm(key_states)278 279        key_states = key_states.view(hidden_shape)280        query_states = query_states.view(hidden_shape)281        value_states = value_states.view(hidden_shape)282 283        query_states = query_states.transpose(1, 2)284        key_states = key_states.transpose(1, 2)285        value_states = value_states.transpose(1, 2)286 287        cos, sin = position_embeddings288        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)289 290        if past_key_values is not None:291            # sin and cos are specific to RoPE models; position_ids needed for the static cache292            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}293            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)294 295        attention_interface: Callable = eager_attention_forward296        if self.config._attn_implementation != "eager":297            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]298 299        attn_output, attn_weights = attention_interface(300            self,301            query_states,302            key_states,303            value_states,304            attention_mask,305            dropout=0.0 if not self.training else self.attention_dropout,306            scaling=self.scaling,307            **kwargs,308        )309 310        attn_output = attn_output.reshape(*input_shape, -1).contiguous()311        attn_output = self.o_proj(attn_output)312        return attn_output, attn_weights313 314 315class MiniMaxM2DecoderLayer(GradientCheckpointingLayer):316    def __init__(self, config: MiniMaxM2Config, layer_idx: int):317        super().__init__()318        self.hidden_size = config.hidden_size319 320        self.self_attn = MiniMaxM2Attention(config, layer_idx)321 322        self.block_sparse_moe = MiniMaxM2SparseMoeBlock(config)323        self.input_layernorm = MiniMaxM2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)324        self.post_attention_layernorm = MiniMaxM2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)325 326    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")327    def forward(328        self,329        hidden_states: torch.Tensor,330        position_embeddings: tuple[torch.Tensor, torch.Tensor],331        attention_mask: Optional[torch.Tensor] = None,332        position_ids: Optional[torch.LongTensor] = None,333        past_key_values: Optional[Cache] = None,334        cache_position: Optional[torch.LongTensor] = None,335        **kwargs: Unpack[TransformersKwargs],336    ) -> torch.FloatTensor:337        residual = hidden_states338 339        hidden_states = self.input_layernorm(hidden_states)340 341        # Self Attention342        hidden_states, _ = self.self_attn(343            hidden_states=hidden_states,344            position_embeddings=position_embeddings,345            attention_mask=attention_mask,346            position_ids=position_ids,347            past_key_values=past_key_values,348            cache_position=cache_position,349            **kwargs,350        )351        hidden_states = residual + hidden_states352 353        # Fully Connected354        residual = hidden_states355        hidden_states = self.post_attention_layernorm(hidden_states)356        hidden_states, _ = self.block_sparse_moe(hidden_states)357        hidden_states = residual + hidden_states358 359        return hidden_states360 361 362class MiniMaxM2RotaryEmbedding(nn.Module):363    inv_freq: torch.Tensor  # fix linting for `register_buffer`364 365    def __init__(self, config: MiniMaxM2Config, device=None):366        super().__init__()367        # BC: "rope_type" was originally "type"368        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):369            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))370        else:371            self.rope_type = "default"372        self.max_seq_len_cached = config.max_position_embeddings373        self.original_max_seq_len = config.max_position_embeddings374 375        self.config = config376        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]377 378        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)379        self.register_buffer("inv_freq", inv_freq, persistent=False)380        self.original_inv_freq = self.inv_freq381 382    @torch.no_grad()383    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)384    def forward(self, x, position_ids):385        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)386        position_ids_expanded = position_ids[:, None, :].float()387 388        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"389        with torch.autocast(device_type=device_type, enabled=False):  # Force float32390            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)391            emb = torch.cat((freqs, freqs), dim=-1)392            cos = emb.cos() * self.attention_scaling393            sin = emb.sin() * self.attention_scaling394 395        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)396 397 398@auto_docstring399class MiniMaxM2PreTrainedModel(PreTrainedModel):400    config: MiniMaxM2Config401    base_model_prefix = "model"402    supports_gradient_checkpointing = True403    _no_split_modules = ["MiniMaxM2DecoderLayer"]404    _skip_keys_device_placement = ["past_key_values"]405    _supports_flash_attn = True406    _supports_sdpa = True407    _supports_flex_attn = True408    _can_compile_fullgraph = False  # MoE models don't work with torch.compile (`torch.where(condition)` not supported)409    _supports_attention_backend = True410    _can_record_outputs = {411        "router_logits": OutputRecorder(MiniMaxM2SparseMoeBlock, index=1),412        "hidden_states": MiniMaxM2DecoderLayer,413        "attentions": MiniMaxM2Attention,414    }415 416 417@auto_docstring418class MiniMaxM2Model(MiniMaxM2PreTrainedModel):419    def __init__(self, config: MiniMaxM2Config):420        super().__init__(config)421        self.padding_idx = config.pad_token_id422        self.vocab_size = config.vocab_size423 424        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)425        self.layers = nn.ModuleList(426            [MiniMaxM2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]427        )428        self.norm = MiniMaxM2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)429        self.rotary_emb = MiniMaxM2RotaryEmbedding(config=config)430        self.gradient_checkpointing = False431 432        # Initialize weights and apply final processing433        self.post_init()434 435    @check_model_inputs436    @auto_docstring437    def forward(438        self,439        input_ids: Optional[torch.LongTensor] = None,440        attention_mask: Optional[torch.Tensor] = None,441        position_ids: Optional[torch.LongTensor] = None,442        past_key_values: Optional[Cache] = None,443        inputs_embeds: Optional[torch.FloatTensor] = None,444        use_cache: Optional[bool] = None,445        cache_position: Optional[torch.LongTensor] = None,446        **kwargs: Unpack[TransformersKwargs],447    ) -> MoeModelOutputWithPast:448        if (input_ids is None) ^ (inputs_embeds is not None):449            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")450 451        if use_cache and past_key_values is None:452            past_key_values = DynamicCache(config=self.config)453 454        if inputs_embeds is None:455            inputs_embeds = self.embed_tokens(input_ids)456 457        if cache_position is None:458            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0459            cache_position = torch.arange(460                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device461            )462        if position_ids is None:463            position_ids = cache_position.unsqueeze(0)464 465        mask_function = create_causal_mask if self.config.sliding_window is None else create_sliding_window_causal_mask466        causal_mask = mask_function(467            config=self.config,468            input_embeds=inputs_embeds,469            attention_mask=attention_mask,470            cache_position=cache_position,471            past_key_values=past_key_values,472            position_ids=position_ids,473        )474 475        hidden_states = inputs_embeds476 477        # create position embeddings to be shared across the decoder layers478        position_embeddings = self.rotary_emb(hidden_states, position_ids)479 480        for decoder_layer in self.layers[: self.config.num_hidden_layers]:481            hidden_states = decoder_layer(482                hidden_states,483                position_embeddings=position_embeddings,484                attention_mask=causal_mask,485                position_ids=position_ids,486                past_key_values=past_key_values,487                use_cache=use_cache,488                cache_position=cache_position,489                **kwargs,490            )491 492        hidden_states = self.norm(hidden_states)493 494        return MoeModelOutputWithPast(  # only diff with Mistral is the output type, we need MoE495            last_hidden_state=hidden_states,496            past_key_values=past_key_values,497        )498 499 500def load_balancing_loss_func(501    gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None],502    num_experts: Optional[int] = None,503    top_k=2,504    attention_mask: Optional[torch.Tensor] = None,505) -> Union[torch.Tensor, int]:506    r"""507    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.508 509    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss510    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between511    experts is too unbalanced.512 513    Args:514        gate_logits:515            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of516            shape [batch_size X sequence_length, num_experts].517        num_experts:518            Number of experts519        top_k:520            The number of experts to route per-token, can be also interpreted as the `top-k` routing521            parameter.522        attention_mask (`torch.Tensor`, *optional*):523            The attention_mask used in forward function524            shape [batch_size X sequence_length] if not None.525 526    Returns:527        The auxiliary loss.528    """529    if gate_logits is None or not isinstance(gate_logits, tuple):530        return 0531 532    if isinstance(gate_logits, tuple):533        compute_device = gate_logits[0].device534        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)535 536    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)537 538    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)539 540    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)541 542    if attention_mask is None:543        # Compute the percentage of tokens routed to each experts544        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)545 546        # Compute the average probability of routing to these experts547        router_prob_per_expert = torch.mean(routing_weights, dim=0)548    else:549        batch_size, sequence_length = attention_mask.shape550        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)551 552        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask553        expert_attention_mask = (554            attention_mask[None, :, :, None, None]555            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))556            .reshape(-1, top_k, num_experts)557            .to(compute_device)558        )559 560        # Compute the percentage of tokens routed to each experts561        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(562            expert_attention_mask, dim=0563        )564 565        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert566        router_per_expert_attention_mask = (567            attention_mask[None, :, :, None]568            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))569            .reshape(-1, num_experts)570            .to(compute_device)571        )572 573        # Compute the average probability of routing to these experts574        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(575            router_per_expert_attention_mask, dim=0576        )577 578    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))579    return overall_loss * num_experts580 581 582@auto_docstring583class MiniMaxM2ForCausalLM(MiniMaxM2PreTrainedModel, GenerationMixin):584    _tied_weights_keys = ["lm_head.weight"]585    _tp_plan = {"lm_head": "colwise_rep"}586    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}587 588    def __init__(self, config):589        super().__init__(config)590        self.model = MiniMaxM2Model(config)591        self.vocab_size = config.vocab_size592        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)593        self.router_aux_loss_coef = config.router_aux_loss_coef594        self.num_experts = config.num_local_experts595        self.num_experts_per_tok = config.num_experts_per_tok596 597        # Initialize weights and apply final processing598        self.post_init()599 600    @can_return_tuple601    @auto_docstring602    def forward(603        self,604        input_ids: Optional[torch.LongTensor] = None,605        attention_mask: Optional[torch.Tensor] = None,606        position_ids: Optional[torch.LongTensor] = None,607        past_key_values: Optional[Cache] = None,608        inputs_embeds: Optional[torch.FloatTensor] = None,609        labels: Optional[torch.LongTensor] = None,610        use_cache: Optional[bool] = None,611        output_router_logits: Optional[bool] = None,612        cache_position: Optional[torch.LongTensor] = None,613        logits_to_keep: Union[int, torch.Tensor] = 0,614        **kwargs: Unpack[TransformersKwargs],615    ) -> MoeCausalLMOutputWithPast:616        r"""617        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):618            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,619            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored620            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.621 622        Example:623 624        ```python625        >>> from transformers import AutoTokenizer, MiniMaxM2ForCausalLM626 627        >>> model = MiniMaxM2ForCausalLM.from_pretrained("mistralai/MiniMaxM2-8x7B-v0.1")628        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/MiniMaxM2-8x7B-v0.1")629 630        >>> prompt = "Hey, are you conscious? Can you talk to me?"631        >>> inputs = tokenizer(prompt, return_tensors="pt")632 633        >>> # Generate634        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)635        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]636        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."637        ```"""638 639        output_router_logits = (640            output_router_logits if output_router_logits is not None else self.config.output_router_logits641        )642 643        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)644        outputs: MoeModelOutputWithPast = self.model(645            input_ids=input_ids,646            attention_mask=attention_mask,647            position_ids=position_ids,648            past_key_values=past_key_values,649            inputs_embeds=inputs_embeds,650            use_cache=use_cache,651            output_router_logits=output_router_logits,652            cache_position=cache_position,653            **kwargs,654        )655 656        hidden_states = outputs.last_hidden_state657        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss658        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep659        logits = self.lm_head(hidden_states[:, slice_indices, :])660 661        loss = None662        if labels is not None:663            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)664 665        aux_loss = None666        if output_router_logits:667            aux_loss = load_balancing_loss_func(668                outputs.router_logits,669                self.num_experts,670                self.num_experts_per_tok,671                attention_mask,672            )673            if labels is not None:674                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device675 676        return MoeCausalLMOutputWithPast(677            loss=loss,678            aux_loss=aux_loss,679            logits=logits,680            past_key_values=outputs.past_key_values,681            hidden_states=outputs.hidden_states,682            attentions=outputs.attentions,683            router_logits=outputs.router_logits,684        )685 686 687class MiniMaxM2ForSequenceClassification(GenericForSequenceClassification, MiniMaxM2PreTrainedModel):688    pass689 690 691class MiniMaxM2ForTokenClassification(GenericForTokenClassification, MiniMaxM2PreTrainedModel):692    pass693 694 695class MiniMaxM2ForQuestionAnswering(GenericForQuestionAnswering, MiniMaxM2PreTrainedModel):696    pass697 698 699__all__ = [700    "MiniMaxM2ForCausalLM",701    "MiniMaxM2ForQuestionAnswering",702    "MiniMaxM2Model",703    "MiniMaxM2PreTrainedModel",704    "MiniMaxM2ForSequenceClassification",705    "MiniMaxM2ForTokenClassification",706]707