Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2025 MiniMaxAI and HuggingFace Inc. teams. All rights reserved.3#4#5# Licensed under the Apache License, Version 2.0 (the "License");6# you may not use this file except in compliance with the License.7# You may obtain a copy of the License at8#9#     http://www.apache.org/licenses/LICENSE-2.010#11# Unless required by applicable law or agreed to in writing, software12# distributed under the License is distributed on an "AS IS" BASIS,13# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.14# See the License for the specific language governing permissions and15# limitations under the License.16"""PyTorch MiniMax model."""17 18from typing import Optional19 20import torch21import torch.nn.functional as F22from torch import nn23 24from ...activations import ACT2FN25from ...cache_utils import Cache, DynamicCache26from ...configuration_utils import layer_type_validation27from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask28from ...modeling_flash_attention_utils import FlashAttentionKwargs29from ...modeling_layers import GradientCheckpointingLayer30from ...modeling_outputs import MoeModelOutputWithPast31from ...processing_utils import Unpack32from ...utils import TransformersKwargs, logging33from ...utils.deprecation import deprecate_kwarg34from ...utils.generic import OutputRecorder, check_model_inputs35from ..mixtral.configuration_mixtral import MixtralConfig36from ..mixtral.modeling_mixtral import (37    MixtralAttention,38    MixtralDecoderLayer,39    MixtralForCausalLM,40    MixtralForQuestionAnswering,41    MixtralForSequenceClassification,42    MixtralForTokenClassification,43    MixtralModel,44    MixtralPreTrainedModel,45    MixtralRMSNorm,46    MixtralSparseMoeBlock,47)48 49 50logger = logging.get_logger(__name__)51 52 53class MiniMaxConfig(MixtralConfig):54    r"""55    This is the configuration class to store the configuration of a [`MiniMaxModel`]. It is used to instantiate an56    MiniMax model according to the specified arguments, defining the model architecture. Instantiating a configuration57    with the defaults will yield a similar configuration to that of the MiniMax.58 59    [MiniMaxAI/MiniMax-Text-01-hf](https://huggingface.co/MiniMaxAI/MiniMax-Text-01-hf)60 61    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the62    documentation from [`PretrainedConfig`] for more information.63 64 65    Args:66        vocab_size (`int`, *optional*, defaults to 32000):67            Vocabulary size of the MiniMax model. Defines the number of different tokens that can be represented by the68            `inputs_ids` passed when calling [`MiniMaxModel`]69        hidden_size (`int`, *optional*, defaults to 4096):70            Dimension of the hidden representations.71        intermediate_size (`int`, *optional*, defaults to 14336):72            Dimension of the MLP representations.73        num_hidden_layers (`int`, *optional*, defaults to 32):74            Number of hidden layers in the Transformer encoder.75        num_attention_heads (`int`, *optional*, defaults to 32):76            Number of attention heads for each attention layer in the Transformer encoder.77        num_key_value_heads (`int`, *optional*, defaults to 8):78            This is the number of key_value heads that should be used to implement Grouped Query Attention. If79            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if80            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When81            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed82            by meanpooling all the original heads within that group. For more details, check out [this83            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `8`.84        head_dim (`int`, *optional*, defaults to `hidden_size // num_attention_heads`):85            The attention head dimension.86        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):87            The non-linear activation function (function or string) in the decoder.88        max_position_embeddings (`int`, *optional*, defaults to `4096*32`):89            The maximum sequence length that this model might ever be used with. MiniMax's sliding window attention90            allows sequence of up to 4096*32 tokens.91        initializer_range (`float`, *optional*, defaults to 0.02):92            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.93        rms_norm_eps (`float`, *optional*, defaults to 1e-05):94            The epsilon used by the rms normalization layers.95        use_cache (`bool`, *optional*, defaults to `True`):96            Whether or not the model should return the last key/values attentions (not used by all models). Only97            relevant if `config.is_decoder=True`.98        pad_token_id (`int`, *optional*):99            The id of the padding token.100        bos_token_id (`int`, *optional*, defaults to 1):101            The id of the "beginning-of-sequence" token.102        eos_token_id (`int`, *optional*, defaults to 2):103            The id of the "end-of-sequence" token.104        tie_word_embeddings (`bool`, *optional*, defaults to `False`):105            Whether the model's input and output word embeddings should be tied.106        rope_theta (`float`, *optional*, defaults to 1000000.0):107            The base period of the RoPE embeddings.108        sliding_window (`int`, *optional*):109            Sliding window attention window size. If not specified, will default to `4096`.110        attention_dropout (`float`, *optional*, defaults to 0.0):111            The dropout ratio for the attention probabilities.112        num_experts_per_tok (`int`, *optional*, defaults to 2):113            The number of experts to route per-token, can be also interpreted as the `top-k` routing114            parameter115        num_local_experts (`int`, *optional*, defaults to 8):116            Number of experts per Sparse MLP layer.117        output_router_logits (`bool`, *optional*, defaults to `False`):118            Whether or not the router logits should be returned by the model. Enabling this will also119            allow the model to output the auxiliary loss. See [here]() for more details120        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):121            The aux loss factor for the total loss.122        router_jitter_noise (`float`, *optional*, defaults to 0.0):123            Amount of noise to add to the router.124        layer_types (`list`, *optional*):125            Attention pattern for each layer.126        block_size (`int`, *optional*, defaults to 256):127            The length of each attention block, determining how queries, keys, and values128            are grouped and processed for intra- and inter-block attention.129        full_attn_alpha_factor (`float`, *optional*, defaults to 1):130            Weight for residual value in residual connection after normal attention.131        full_attn_beta_factor (`float`, *optional*, defaults to 1):132            Weight for hidden state value in residual connection after normal attention.133        linear_attn_alpha_factor (`float`, *optional*, defaults to 1):134            Weight for residual value in residual connection after lightning attention.135        linear_attn_beta_factor (`float`, *optional*, defaults to 1):136            Weight for hidden state value in residual connection after lightning attention.137        mlp_alpha_factor (`float`, *optional*, defaults to 1):138            Weight for residual value in residual connection after MLP.139        mlp_beta_factor (`float`, *optional*, defaults to 1):140            Weight for hidden state value in residual connection after MLP.141 142    ```python143    >>> from transformers import MiniMaxModel, MiniMaxConfig144 145    >>> # Initializing a MiniMax style configuration146    >>> configuration = MiniMaxConfig()147 148    >>> # Initializing a model from the MiniMax style configuration149    >>> model = MiniMaxModel(configuration)150 151    >>> # Accessing the model configuration152    >>> configuration = model.config153    ```"""154 155    def __init__(156        self,157        layer_types=None,158        block_size=256,159        full_attn_alpha_factor=1,160        full_attn_beta_factor=1,161        linear_attn_alpha_factor=1,162        linear_attn_beta_factor=1,163        mlp_alpha_factor=1,164        mlp_beta_factor=1,165        **super_kwargs,166    ):167        super().__init__(**super_kwargs)168        self.layer_types = layer_types169        self.block_size = block_size170        self.full_attn_alpha_factor = full_attn_alpha_factor171        self.full_attn_beta_factor = full_attn_beta_factor172        self.linear_attn_alpha_factor = linear_attn_alpha_factor173        self.linear_attn_beta_factor = linear_attn_beta_factor174        self.mlp_alpha_factor = mlp_alpha_factor175        self.mlp_beta_factor = mlp_beta_factor176 177        if self.layer_types is None:178            self.layer_types = [179                "full_attention" if bool((i + 1) % 2) else "linear_attention" for i in range(self.num_hidden_layers)180            ]181        layer_type_validation(self.layer_types, self.num_hidden_layers)182 183 184class MiniMaxRMSNorm(MixtralRMSNorm):185    pass186 187 188class MiniMaxCache(DynamicCache):189    def __init__(self):190        super().__init__()191        self.linear_cache: list[torch.Tensor] = []192 193    def set_linear_cache(self, layer_idx, linear_cache):194        # There may be skipped layers, fill them with empty lists195        for _ in range(len(self.linear_cache), layer_idx + 1):196            self.linear_cache.append([])197        self.linear_cache[layer_idx] = linear_cache198 199    def get_linear_cache(self, layer_idx: int):200        if layer_idx < len(self):201            return self.linear_cache[layer_idx]202        return None203 204    def __len__(self):205        return max(super().__len__(), len(self.linear_cache))206 207    def __getitem__(self, layer_idx: int):208        if layer_idx < len(self.linear_cache) and self.linear_cache[layer_idx] != []:209            return (self.linear_cache[layer_idx],)210        return super().__getitem__(layer_idx)211 212    def __iter__(self):213        for layer_idx in range(len(self)):214            yield self[layer_idx]215 216    def batch_repeat_interleave(self, repeats: int):217        for layer_idx in range(len(self)):218            if self.linear_cache[layer_idx] != []:219                self.linear_cache[layer_idx] = self.linear_cache[layer_idx].repeat_interleave(repeats, dim=0)220            else:221                self.layers[layer_idx].batch_repeat_interleave(repeats)222 223    def batch_select_indices(self, indices: torch.Tensor):224        for layer_idx in range(len(self)):225            if self.linear_cache[layer_idx] != []:226                self.linear_cache[layer_idx] = self.linear_cache[layer_idx][indices, ...]227            else:228                self.layers[layer_idx].batch_select_indices(indices)229 230    def crop(self, max_length: int):231        raise RuntimeError("MiniMaxCache doesnot support `crop` method")232 233 234class MiniMaxLightningAttention(nn.Module):235    def __init__(self, config: MiniMaxConfig, layer_idx: int):236        super().__init__()237        self.layer_idx = layer_idx238        self.head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads239        self.num_attention_heads = config.num_attention_heads240        self.num_hidden_layers = config.num_hidden_layers241        self.block_size = config.block_size242 243        self.act_fn = ACT2FN[config.hidden_act]244        self.norm = MiniMaxRMSNorm(self.head_dim * self.num_attention_heads)245        self.qkv_proj = nn.Linear(config.hidden_size, self.num_attention_heads * self.head_dim * 3, bias=False)246        self.out_proj = nn.Linear(self.num_attention_heads * self.head_dim, config.hidden_size, bias=False)247        self.output_gate = nn.Linear(config.hidden_size, self.num_attention_heads * self.head_dim, bias=False)248 249        slope_rate = self.get_slope_rate()250        query_decay, key_decay, diagonal_decay = self.decay_factors(slope_rate)251 252        self.register_buffer("slope_rate", slope_rate)253        self.register_buffer("query_decay", query_decay)254        self.register_buffer("key_decay", key_decay)255        self.register_buffer("diagonal_decay", diagonal_decay)256 257    def get_slope_rate(self):258        base = 1 / (2 ** (8 / self.num_attention_heads))259        exponent = torch.arange(self.num_attention_heads) + 1260        factor = 1 - self.layer_idx / (self.num_hidden_layers - 1 + 1e-5) + 1e-5261 262        rate = base**exponent263        rate = rate * factor264        rate = rate[:, None, None]265 266        return rate267 268    def decay_factors(self, slope_rate):269        block_size_range = torch.arange(self.block_size) + 1270 271        query_decay = torch.exp(-slope_rate * block_size_range[:, None])272        key_decay = torch.exp(-slope_rate * (self.block_size - block_size_range[:, None]))273 274        diagonal_decay = block_size_range[:, None] - block_size_range[None, :]275        diagonal_decay = diagonal_decay[None, None, :, :]276        diagonal_decay = slope_rate * diagonal_decay277        diagonal_decay = torch.where(diagonal_decay >= 0, -diagonal_decay, float("-inf"))278        diagonal_decay = torch.exp(diagonal_decay)279 280        return query_decay, key_decay, diagonal_decay281 282    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")283    def forward(284        self,285        hidden_states: torch.Tensor,286        position_embeddings: tuple[torch.Tensor, torch.Tensor],287        attention_mask: Optional[torch.Tensor],288        past_key_values: Optional[Cache] = None,289        cache_position: Optional[torch.LongTensor] = None,290        **kwargs: Unpack[FlashAttentionKwargs],291    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:292        batch_size, seq_len, hidden_size = hidden_states.shape293        num_blocks = (seq_len + self.block_size - 1) // self.block_size294 295        qkv_states = self.act_fn(self.qkv_proj(hidden_states))296        qkv_states = qkv_states.reshape(batch_size, seq_len, self.num_attention_heads, 3 * self.head_dim)297 298        query_states, key_states, value_states = torch.split(qkv_states, self.head_dim, dim=3)299 300        query_states = query_states.transpose(1, 2)301        key_states = key_states.transpose(1, 2)302        value_states = value_states.transpose(1, 2)303 304        # calculated (K.T @ V) and saved as cache305        attn_weights_inter = None306        if past_key_values is not None:307            attn_weights_inter = past_key_values.get_linear_cache(self.layer_idx)308 309        if attn_weights_inter is None:310            attn_weights_inter = torch.zeros(batch_size, self.num_attention_heads, self.head_dim, self.head_dim).to(311                value_states312            )313 314            # apply attention_mask315            if attention_mask is not None:316                attention_mask = attention_mask.to(dtype=torch.bool)  # Ensure it's a boolean tensor317                value_states = value_states.masked_fill(~attention_mask.unsqueeze(1).unsqueeze(-1), 0)318 319            attn_output = []320            for i in range(num_blocks):321                start_idx = i * self.block_size322                end_idx = min(start_idx + self.block_size, seq_len)323                current_block_size = end_idx - start_idx324 325                current_query_states = query_states[:, :, start_idx:end_idx]326                current_key_states = key_states[:, :, start_idx:end_idx]327                current_value_states = value_states[:, :, start_idx:end_idx]328 329                current_query_decay = self.query_decay[:, :current_block_size]330                current_key_decay = self.key_decay[:, -current_block_size:]331                current_diagonal_decay = self.diagonal_decay[:, :, :current_block_size, :current_block_size]332                block_decay = torch.exp(-self.slope_rate * current_block_size)333 334                # intra: ( Q @ K.T ) @ V -> QK * V335                attn_weights_intra = torch.matmul(current_query_states, current_key_states.transpose(-1, -2))336                attn_output_intra = torch.matmul(attn_weights_intra * current_diagonal_decay, current_value_states)337 338                # inter: Q @ ( K.T @ V ) -> Q * KV339                attn_output_inter = torch.matmul(current_query_states * current_query_decay, attn_weights_inter)340 341                # final attention output342                current_attn_output = attn_output_inter + attn_output_intra343                attn_output.append(current_attn_output)344 345                # calculate attn_weights_inter for next block or cache346                next_attn_weights_inter = torch.matmul(347                    (current_key_states * current_key_decay).transpose(-1, -2), current_value_states348                )349                attn_weights_inter = attn_weights_inter * block_decay + next_attn_weights_inter350 351        else:352            ratio = torch.exp(-self.slope_rate)353            attn_output = []354            for i in range(seq_len):355                current_query_states = query_states[:, :, i : i + 1]356                current_key_states = key_states[:, :, i : i + 1]357                current_value_states = value_states[:, :, i : i + 1]358 359                current_attn_weights_inter = torch.matmul(current_key_states.transpose(-1, -2), current_value_states)360                attn_weights_inter = ratio * attn_weights_inter + current_attn_weights_inter361                current_attn_output = torch.matmul(current_query_states, attn_weights_inter)362 363                attn_output.append(current_attn_output)364 365        # concatenate attention outputs over all blocks366        attn_output = torch.cat(attn_output, dim=-2)367 368        # final output projection369        attn_output = attn_output.transpose(1, 2)370        attn_output = attn_output.reshape(batch_size, seq_len, self.num_attention_heads * self.head_dim)371        attn_output = self.norm(attn_output)372        attn_output = F.sigmoid(self.output_gate(hidden_states)) * attn_output373        attn_output = self.out_proj(attn_output)374 375        # update cache376        if past_key_values is not None:377            past_key_values.set_linear_cache(self.layer_idx, attn_weights_inter)378 379        return attn_output, attn_weights_inter380 381 382class MiniMaxAttention(MixtralAttention):383    pass384 385 386class MiniMaxSparseMoeBlock(MixtralSparseMoeBlock):387    pass388 389 390class MiniMaxDecoderLayer(MixtralDecoderLayer, GradientCheckpointingLayer):391    def __init__(self, config: MiniMaxConfig, layer_idx: int):392        super().__init__(config, layer_idx)393 394        self.layer_idx = layer_idx395        self.layer_type = config.layer_types[layer_idx]396        self.mlp_alpha_factor = config.mlp_alpha_factor397        self.mlp_beta_factor = config.mlp_beta_factor398 399        if self.layer_type == "linear_attention":400            self.self_attn = MiniMaxLightningAttention(config, layer_idx)401            self.attn_alpha_factor = config.linear_attn_alpha_factor402            self.attn_beta_factor = config.linear_attn_beta_factor403        else:404            self.self_attn = MiniMaxAttention(config, layer_idx)405            self.attn_alpha_factor = config.full_attn_alpha_factor406            self.attn_beta_factor = config.full_attn_beta_factor407 408    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")409    def forward(410        self,411        hidden_states: torch.Tensor,412        position_embeddings: tuple[torch.Tensor, torch.Tensor],413        attention_mask: Optional[torch.Tensor] = None,414        position_ids: Optional[torch.LongTensor] = None,415        past_key_values: Optional[Cache] = None,416        output_attentions: Optional[bool] = False,417        output_router_logits: Optional[bool] = False,418        use_cache: Optional[bool] = False,419        cache_position: Optional[torch.LongTensor] = None,420        **kwargs: Unpack[FlashAttentionKwargs],421    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:422        """423        Args:424            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`425            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`):426                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,427                with `head_dim` being the embedding dimension of each attention head.428            attention_mask (`torch.Tensor`, *optional*): attention mask of size429                `(batch, sequence_length)` where padding elements are indicated by 0.430            past_key_values (`Cache`, *optional*): cached past key and value projection states431            output_attentions (`bool`, *optional*):432                Whether or not to return the attentions tensors of all attention layers. See `attentions` under433                returned tensors for more detail.434            output_router_logits (`bool`, *optional*):435                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and436                should not be returned during inference.437            use_cache (`bool`, *optional*):438                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding439                (see `past_key_values`).440            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):441                Indices depicting the position of the input sequence tokens in the sequence.442            kwargs (`dict`, *optional*):443                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code444                into the model445        """446 447        hidden_states = self.input_layernorm(hidden_states)448        residual = hidden_states449 450        # Self Attention451        hidden_states, _ = self.self_attn(452            hidden_states=hidden_states,453            position_embeddings=position_embeddings,454            attention_mask=attention_mask,455            position_ids=position_ids,456            past_key_values=past_key_values,457            output_attentions=output_attentions,458            use_cache=use_cache,459            cache_position=cache_position,460            **kwargs,461        )462        hidden_states = residual * self.attn_alpha_factor + hidden_states * self.attn_beta_factor463 464        # Fully Connected465        hidden_states = self.post_attention_layernorm(hidden_states)466        residual = hidden_states467        hidden_states, _ = self.block_sparse_moe(hidden_states)468        hidden_states = residual * self.mlp_alpha_factor + hidden_states * self.mlp_beta_factor469 470        return hidden_states471 472 473class MiniMaxPreTrainedModel(MixtralPreTrainedModel):474    _can_compile_fullgraph = False475    _can_record_outputs = {476        "router_logits": OutputRecorder(MiniMaxSparseMoeBlock, index=1),477        "hidden_states": MiniMaxDecoderLayer,478        "attentions": [MiniMaxAttention, MiniMaxLightningAttention],479    }480 481 482class MiniMaxModel(MixtralModel):483    @check_model_inputs()484    def forward(485        self,486        input_ids: Optional[torch.LongTensor] = None,487        attention_mask: Optional[torch.Tensor] = None,488        position_ids: Optional[torch.LongTensor] = None,489        past_key_values: Optional[MiniMaxCache] = None,490        inputs_embeds: Optional[torch.FloatTensor] = None,491        use_cache: Optional[bool] = None,492        output_attentions: Optional[bool] = None,493        cache_position: Optional[torch.LongTensor] = None,494        **kwargs: Unpack[TransformersKwargs],495    ) -> MoeModelOutputWithPast:496        if (input_ids is None) ^ (inputs_embeds is not None):497            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")498 499        if use_cache and past_key_values is None:500            past_key_values = MiniMaxCache()501        elif use_cache and not isinstance(past_key_values, MiniMaxCache):502            raise ValueError(503                f"MiniMax uses cache of its own and is not compatible with `past_key_values` of type {type(past_key_values)}."504            )505 506        if inputs_embeds is None:507            inputs_embeds = self.embed_tokens(input_ids)508 509        if cache_position is None:510            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0511            cache_position = torch.arange(512                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device513            )514        if position_ids is None:515            position_ids = cache_position.unsqueeze(0)516 517        mask_function = create_causal_mask if self.config.sliding_window is None else create_sliding_window_causal_mask518        causal_mask = mask_function(519            config=self.config,520            input_embeds=inputs_embeds,521            attention_mask=attention_mask,522            cache_position=cache_position,523            past_key_values=past_key_values,524            position_ids=position_ids,525        )526 527        hidden_states = inputs_embeds528 529        # create position embeddings to be shared across the decoder layers530        position_embeddings = self.rotary_emb(hidden_states, position_ids)531 532        for decoder_layer in self.layers:533            if decoder_layer.layer_type == "full_attention":534                input_attention_mask = causal_mask535            else:536                # lightning attention uses original attention_mask, and uses it only for the first step537                input_attention_mask = attention_mask538 539            hidden_states = decoder_layer(540                hidden_states,541                position_embeddings=position_embeddings,542                attention_mask=input_attention_mask,543                position_ids=position_ids,544                past_key_values=past_key_values,545                use_cache=use_cache,546                cache_position=cache_position,547                **kwargs,548            )549 550        hidden_states = self.norm(hidden_states)551 552        return MoeModelOutputWithPast(553            last_hidden_state=hidden_states,554            past_key_values=past_key_values,555        )556 557 558class MiniMaxForCausalLM(MixtralForCausalLM):559    def forward(self, **super_kwargs):560        r"""561        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):562            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,563            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored564            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.565 566        Example:567 568        ```python569        >>> from transformers import AutoTokenizer, MiniMaxForCausalLM570 571        >>> model = MiniMaxForCausalLM.from_pretrained("MiniMaxAI/MiniMax-Text-01-hf")572        >>> tokenizer = AutoTokenizer.from_pretrained("MiniMaxAI/MiniMax-Text-01-hf")573 574        >>> prompt = "Hey, are you conscious? Can you talk to me?"575        >>> inputs = tokenizer(prompt, return_tensors="pt")576 577        >>> # Generate578        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)579        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]580        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."581        ```"""582        return super().forward(**super_kwargs)583 584 585class MiniMaxForSequenceClassification(MixtralForSequenceClassification):586    pass587 588 589class MiniMaxForTokenClassification(MixtralForTokenClassification):590    pass591 592 593class MiniMaxForQuestionAnswering(MixtralForQuestionAnswering):594    pass595 596 597__all__ = [598    "MiniMaxConfig",599    "MiniMaxPreTrainedModel",600    "MiniMaxModel",601    "MiniMaxForCausalLM",602    "MiniMaxForSequenceClassification",603    "MiniMaxForTokenClassification",604    "MiniMaxForQuestionAnswering",605]606