xhyi/CodeGen-2B-Multi

1# coding=utf-82# Copyright 2021 The EleutherAI and HuggingFace Teams. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15 16# Modified forward-pass implementation based on https://github.com/huggingface/transformers/blob/main/src/transformers/models/gptj/modeling_gptj.py17 18from typing import Tuple19 20import numpy as np21 22import torch23import torch.utils.checkpoint24from torch import nn25from torch.nn import CrossEntropyLoss26 27from transformers.activations import ACT2FN28from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast29from transformers.modeling_utils import PreTrainedModel30from transformers.utils import logging31from transformers.utils.model_parallel_utils import assert_device_map, get_device_map32from .configuration_codegen import CodeGenConfig33 34 35logger = logging.get_logger(__name__)36 37 38def fixed_pos_embedding(x, seq_dim=1, seq_len=None):39    dim = x.shape[-1]40    if seq_len is None:41        seq_len = x.shape[seq_dim]42    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2) / dim))43    sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(seq_len), inv_freq).to(x.device).float()44    return torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)45 46 47def rotate_every_two(x):48    x1 = x[:, :, :, ::2]49    x2 = x[:, :, :, 1::2]50    x = torch.stack((-x2, x1), axis=-1)51    return x.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')52 53 54def apply_rotary_pos_emb(x, sincos, offset=0):55    sin, cos = map(lambda t: t[None, offset : x.shape[1] + offset, None, :].repeat_interleave(2, 3), sincos)56    # einsum notation for lambda t: repeat(t[offset:x.shape[1]+offset,:], "n d -> () n () (d j)", j=2)57    return (x * cos) + (rotate_every_two(x) * sin)58 59 60class CodeGenAttention(nn.Module):61    def __init__(self, config):62        super().__init__()63 64        max_positions = config.max_position_embeddings65        self.register_buffer(66            "bias",67            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(68                1, 1, max_positions, max_positions69            ),70        )71        self.register_buffer("masked_bias", torch.tensor(-1e9))72 73        self.attn_dropout = nn.Dropout(config.attn_pdrop)74        self.resid_dropout = nn.Dropout(config.resid_pdrop)75 76        self.embed_dim = config.hidden_size77        self.num_attention_heads = config.num_attention_heads78        self.head_dim = self.embed_dim // self.num_attention_heads79        if self.head_dim * self.num_attention_heads != self.embed_dim:80            raise ValueError(81                f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and `num_attention_heads`: {self.num_attention_heads})."82            )83        self.scale_attn = torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype())84        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)85 86        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)87        self.rotary_dim = None88        if config.rotary_dim is not None:89            self.rotary_dim = config.rotary_dim90 91    def _split_heads(self, x, n_head, dim_head, mp_num):92        reshaped = x.reshape(x.shape[:-1] + (n_head//mp_num, dim_head))93        reshaped = reshaped.reshape(x.shape[:-2] + (-1, ) + reshaped.shape[-1:])94        return reshaped95 96    def _merge_heads(self, tensor, num_attention_heads, attn_head_size):97        """98        Merges attn_head_size dim and num_attn_heads dim into n_ctx99        """100        if len(tensor.shape) == 5:101            tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()102        elif len(tensor.shape) == 4:103            tensor = tensor.permute(0, 2, 1, 3).contiguous()104        else:105            raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")106        new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)107        return tensor.view(new_shape)108 109    def _attn(110        self,111        query,112        key,113        value,114        attention_mask=None,115        head_mask=None,116    ):117 118        # compute causal mask from causal mask buffer119        query_length, key_length = query.size(-2), key.size(-2)120        causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]121 122        # Keep the attention weights computation in fp32 to avoid overflow issues123        query = query.to(torch.float32)124        key = key.to(torch.float32)125 126        attn_weights = torch.matmul(query, key.transpose(-1, -2))127 128        attn_weights = attn_weights / self.scale_attn129        attn_weights = torch.where(causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype))130 131        if attention_mask is not None:132            # Apply the attention mask133            attn_weights = attn_weights + attention_mask134 135        attn_weights = nn.Softmax(dim=-1)(attn_weights)136        attn_weights = attn_weights.to(value.dtype)137        attn_weights = self.attn_dropout(attn_weights)138 139        # Mask heads if we want to140        if head_mask is not None:141            attn_weights = attn_weights * head_mask142 143        attn_output = torch.matmul(attn_weights, value)144 145        return attn_output, attn_weights146 147    def forward(148        self,149        hidden_states,150        attention_mask=None,151        layer_past=None,152        head_mask=None,153        use_cache=False,154        output_attentions=False,155    ):156 157        qkv = self.qkv_proj(hidden_states)158        # TODO(enijkamp): factor out number of logical TPU-v4 cores or make forward pass agnostic159        mp_num = 4160        qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))161 162        local_dim = self.head_dim * self.num_attention_heads // mp_num163        query, value, key = torch.split(qkv_split, local_dim, dim=-1)164        query = self._split_heads(query, self.num_attention_heads, self.head_dim, mp_num=mp_num)165        key = self._split_heads(key, self.num_attention_heads, self.head_dim, mp_num=mp_num)166 167        value = self._split_heads(value, self.num_attention_heads, self.head_dim, mp_num=mp_num)168        value = value.permute(0, 2, 1, 3)169 170        seq_len = key.shape[1]171        offset = 0172 173        if layer_past is not None:174            offset = layer_past[0].shape[-2]175            seq_len += offset176 177        if self.rotary_dim is not None:178            k_rot = key[:, :, :, : self.rotary_dim]179            k_pass = key[:, :, :, self.rotary_dim :]180 181            q_rot = query[:, :, :, : self.rotary_dim]182            q_pass = query[:, :, :, self.rotary_dim :]183 184            sincos = fixed_pos_embedding(k_rot, 1, seq_len=seq_len)185            k_rot = apply_rotary_pos_emb(k_rot, sincos, offset=offset)186            q_rot = apply_rotary_pos_emb(q_rot, sincos, offset=offset)187 188            key = torch.cat([k_rot, k_pass], dim=-1)189            query = torch.cat([q_rot, q_pass], dim=-1)190        else:191            sincos = fixed_pos_embedding(key, 1, seq_len=seq_len)192            key = apply_rotary_pos_emb(key, sincos, offset=offset)193            query = apply_rotary_pos_emb(query, sincos, offset=offset)194 195        key = key.permute(0, 2, 1, 3)196        query = query.permute(0, 2, 1, 3)197 198        if layer_past is not None:199            past_key = layer_past[0]200            past_value = layer_past[1]201            key = torch.cat((past_key, key), dim=-2)202            value = torch.cat((past_value, value), dim=-2)203 204        if use_cache is True:205            present = (key, value)206        else:207            present = None208 209        # compute self-attention: V x Softmax(QK^T)210        attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)211 212        attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)213 214        attn_output = self.out_proj(attn_output)215        attn_output = self.resid_dropout(attn_output)216 217        outputs = (attn_output, present)218        if output_attentions:219            outputs += (attn_weights,)220 221        return outputs  # a, present, (attentions)222 223 224class CodeGenMLP(nn.Module):225    def __init__(self, intermediate_size, config):  # in MLP: intermediate_size= 4 * embed_dim226        super().__init__()227        embed_dim = config.n_embd228 229        self.fc_in = nn.Linear(embed_dim, intermediate_size)230        self.fc_out = nn.Linear(intermediate_size, embed_dim)231 232        self.act = ACT2FN[config.activation_function]233        self.dropout = nn.Dropout(config.resid_pdrop)234 235    def forward(self, hidden_states):236        hidden_states = self.fc_in(hidden_states)237        hidden_states = self.act(hidden_states)238        hidden_states = self.fc_out(hidden_states)239        hidden_states = self.dropout(hidden_states)240        return hidden_states241 242 243class CodeGenBlock(nn.Module):244    def __init__(self, config):245        super().__init__()246        inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd247        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)248        self.attn = CodeGenAttention(config)249        self.mlp = CodeGenMLP(inner_dim, config)250 251    def forward(252        self,253        hidden_states,254        layer_past=None,255        attention_mask=None,256        head_mask=None,257        use_cache=False,258        output_attentions=False,259    ):260        residual = hidden_states261        hidden_states = self.ln_1(hidden_states)262        attn_outputs = self.attn(263            hidden_states,264            layer_past=layer_past,265            attention_mask=attention_mask,266            head_mask=head_mask,267            use_cache=use_cache,268            output_attentions=output_attentions,269        )270        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)271        outputs = attn_outputs[1:]272 273        feed_forward_hidden_states = self.mlp(hidden_states)274        hidden_states = attn_output + feed_forward_hidden_states + residual275 276        if use_cache:277            outputs = (hidden_states,) + outputs278        else:279            outputs = (hidden_states,) + outputs[1:]280 281        return outputs  # hidden_states, present, (attentions)282 283 284class CodeGenPreTrainedModel(PreTrainedModel):285    """286    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained287    models.288    """289 290    config_class = CodeGenConfig291    base_model_prefix = "transformer"292    is_parallelizable = True293 294    def __init__(self, *inputs, **kwargs):295        super().__init__(*inputs, **kwargs)296 297    def _init_weights(self, module):298        """Initialize the weights."""299        if isinstance(module, (nn.Linear,)):300            # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization301            # cf https://github.com/pytorch/pytorch/pull/5617302            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)303            if module.bias is not None:304                module.bias.data.zero_()305        elif isinstance(module, nn.Embedding):306            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)307            if module.padding_idx is not None:308                module.weight.data[module.padding_idx].zero_()309        elif isinstance(module, nn.LayerNorm):310            module.bias.data.zero_()311            module.weight.data.fill_(1.0)312 313 314class CodeGenModel(CodeGenPreTrainedModel):315    def __init__(self, config):316        super().__init__(config)317 318        self.embed_dim = config.n_embd319        self.vocab_size = config.vocab_size320        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)321        self.drop = nn.Dropout(config.embd_pdrop)322        self.h = nn.ModuleList([CodeGenBlock(config) for _ in range(config.n_layer)])323        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)324        self.rotary_dim = min(config.rotary_dim, config.n_ctx // config.num_attention_heads)325        self.init_weights()326 327        # Model parallel328        self.model_parallel = False329        self.device_map = None330 331 332    def parallelize(self, device_map=None):333        # Check validity of device_map334        self.device_map = (335            get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map336        )337        assert_device_map(self.device_map, len(self.h))338        self.model_parallel = True339        self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))340        self.last_device = "cuda:" + str(max(self.device_map.keys()))341        self.wte = self.wte.to(self.first_device)342        # Load onto devices343        for k, v in self.device_map.items():344            for block in v:345                cuda_device = "cuda:" + str(k)346                self.h[block] = self.h[block].to(cuda_device)347        # ln_f to last348        self.ln_f = self.ln_f.to(self.last_device)349 350 351    def deparallelize(self):352        self.model_parallel = False353        self.device_map = None354        self.first_device = "cpu"355        self.last_device = "cpu"356        self.wte = self.wte.to("cpu")357        for index in range(len(self.h)):358            self.h[index] = self.h[index].to("cpu")359        self.ln_f = self.ln_f.to("cpu")360        torch.cuda.empty_cache()361 362    def get_input_embeddings(self):363        return self.wte364 365    def set_input_embeddings(self, new_embeddings):366        self.wte = new_embeddings367 368    def forward(369        self,370        input_ids=None,371        past_key_values=None,372        attention_mask=None,373        token_type_ids=None,374        position_ids=None,375        head_mask=None,376        inputs_embeds=None,377        use_cache=None,378        output_attentions=None,379        output_hidden_states=None,380        return_dict=None,381    ):382        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions383        output_hidden_states = (384            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states385        )386        use_cache = use_cache if use_cache is not None else self.config.use_cache387        return_dict = return_dict if return_dict is not None else self.config.use_return_dict388 389        if input_ids is not None and inputs_embeds is not None:390            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")391        elif input_ids is not None:392            input_shape = input_ids.size()393            input_ids = input_ids.view(-1, input_shape[-1])394            batch_size = input_ids.shape[0]395        elif inputs_embeds is not None:396            input_shape = inputs_embeds.size()[:-1]397            batch_size = inputs_embeds.shape[0]398        else:399            raise ValueError("You have to specify either input_ids or inputs_embeds")400 401        device = input_ids.device if input_ids is not None else inputs_embeds.device402 403        if token_type_ids is not None:404            token_type_ids = token_type_ids.view(-1, input_shape[-1])405 406        if position_ids is not None:407            position_ids = position_ids.view(-1, input_shape[-1])408 409        if past_key_values is None:410            past_length = 0411            past_key_values = tuple([None] * len(self.h))412        else:413            past_length = past_key_values[0][0].size(-2)414 415        if position_ids is None:416            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)417            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])418 419        # Attention mask.420        if attention_mask is not None:421            assert batch_size > 0, "batch_size has to be defined and > 0"422            attention_mask = attention_mask.view(batch_size, -1)423            # We create a 3D attention mask from a 2D tensor mask.424            # Sizes are [batch_size, 1, 1, to_seq_length]425            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]426            # this attention mask is more simple than the triangular masking of causal attention427            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.428            attention_mask = attention_mask[:, None, None, :]429 430            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for431            # masked positions, this operation will create a tensor which is 0.0 for432            # positions we want to attend and -10000.0 for masked positions.433            # Since we are adding it to the raw scores before the softmax, this is434            # effectively the same as removing these entirely.435            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility436            attention_mask = (1.0 - attention_mask) * -10000.0437 438        # Prepare head mask if needed439        # 1.0 in head_mask indicate we keep the head440        # attention_probs has shape bsz x num_attention_heads x N x N441        # head_mask has shape n_layer x batch x num_attention_heads x N x N442        head_mask = self.get_head_mask(head_mask, self.config.n_layer)443 444        if inputs_embeds is None:445            inputs_embeds = self.wte(input_ids)446 447        hidden_states = inputs_embeds448 449        if token_type_ids is not None:450            token_type_embeds = self.wte(token_type_ids)451            hidden_states = hidden_states + token_type_embeds452 453        hidden_states = self.drop(hidden_states)454 455        output_shape = input_shape + (hidden_states.size(-1),)456 457        presents = () if use_cache else None458        all_self_attentions = () if output_attentions else None459        all_hidden_states = () if output_hidden_states else None460        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):461 462            # Model parallel463            if self.model_parallel:464                torch.cuda.set_device(hidden_states.device)465                # Ensure layer_past is on same device as hidden_states (might not be correct)466                if layer_past is not None:467                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)468                # Ensure that attention_mask is always on the same device as hidden_states469                if attention_mask is not None:470                    attention_mask = attention_mask.to(hidden_states.device)471                if isinstance(head_mask, torch.Tensor):472                    head_mask = head_mask.to(hidden_states.device)473            if output_hidden_states:474                all_hidden_states = all_hidden_states + (hidden_states,)475 476            if getattr(self.config, "gradient_checkpointing", False) and self.training:477 478                if use_cache:479                    logger.warning(480                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "481                        "`use_cache=False`..."482                    )483                    use_cache = False484 485                def create_custom_forward(module):486                    def custom_forward(*inputs):487                        # None for past_key_value488                        return module(*inputs, use_cache, output_attentions)489 490                    return custom_forward491 492                outputs = torch.utils.checkpoint.checkpoint(493                    create_custom_forward(block),494                    hidden_states,495                    None,496                    attention_mask,497                    head_mask[i],498                )499            else:500                outputs = block(501                    hidden_states,502                    layer_past=layer_past,503                    attention_mask=attention_mask,504                    head_mask=head_mask[i],505                    use_cache=use_cache,506                    output_attentions=output_attentions,507                )508 509            hidden_states = outputs[0]510            if use_cache is True:511                presents = presents + (outputs[1],)512 513            if output_attentions:514                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)515 516            # Model Parallel: If it's the last layer for that device, put things on the next device517            if self.model_parallel:518                for k, v in self.device_map.items():519                    if i == v[-1] and "cuda:" + str(k) != self.last_device:520                        hidden_states = hidden_states.to("cuda:" + str(k + 1))521 522        hidden_states = self.ln_f(hidden_states)523 524        hidden_states = hidden_states.view(*output_shape)525        # Add last hidden state526        if output_hidden_states:527            all_hidden_states = all_hidden_states + (hidden_states,)528 529        if not return_dict:530            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)531 532        return BaseModelOutputWithPast(533            last_hidden_state=hidden_states,534            past_key_values=presents,535            hidden_states=all_hidden_states,536            attentions=all_self_attentions,537        )538 539 540class CodeGenForCausalLM(CodeGenPreTrainedModel):541    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias", r"lm_head\.weight"]542 543    def __init__(self, config):544        super().__init__(config)545        self.transformer = CodeGenModel(config)546        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)547        self.init_weights()548 549        # Model parallel550        self.model_parallel = False551        self.device_map = None552 553    def parallelize(self, device_map=None):554        self.device_map = (555            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))556            if device_map is None557            else device_map558        )559        assert_device_map(self.device_map, len(self.transformer.h))560        self.transformer.parallelize(self.device_map)561        self.lm_head = self.lm_head.to(self.transformer.first_device)562        self.model_parallel = True563 564    def deparallelize(self):565        self.transformer.deparallelize()566        self.transformer = self.transformer.to("cpu")567        self.lm_head = self.lm_head.to("cpu")568        self.model_parallel = False569        torch.cuda.empty_cache()570 571    def get_output_embeddings(self):572        return None573 574    def set_output_embeddings(self, new_embeddings):575        return576 577    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):578        token_type_ids = kwargs.get("token_type_ids", None)579        # only last token for inputs_ids if past is defined in kwargs580        if past:581            input_ids = input_ids[:, -1].unsqueeze(-1)582            if token_type_ids is not None:583                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)584 585        attention_mask = kwargs.get("attention_mask", None)586        position_ids = kwargs.get("position_ids", None)587 588        if attention_mask is not None and position_ids is None:589            # create position_ids on the fly for batch generation590            position_ids = attention_mask.long().cumsum(-1) - 1591            position_ids.masked_fill_(attention_mask == 0, 1)592            if past:593                position_ids = position_ids[:, -1].unsqueeze(-1)594        else:595            position_ids = None596        return {597            "input_ids": input_ids,598            "past_key_values": past,599            "use_cache": kwargs.get("use_cache"),600            "position_ids": position_ids,601            "attention_mask": attention_mask,602            "token_type_ids": token_type_ids,603        }604 605    def forward(606        self,607        input_ids=None,608        past_key_values=None,609        attention_mask=None,610        token_type_ids=None,611        position_ids=None,612        head_mask=None,613        inputs_embeds=None,614        labels=None,615        use_cache=None,616        output_attentions=None,617        output_hidden_states=None,618        return_dict=None,619    ):620        r"""621        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):622            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set623            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size]`` All labels set to624            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``625        """626        return_dict = return_dict if return_dict is not None else self.config.use_return_dict627 628        transformer_outputs = self.transformer(629            input_ids,630            past_key_values=past_key_values,631            attention_mask=attention_mask,632            token_type_ids=token_type_ids,633            position_ids=position_ids,634            head_mask=head_mask,635            inputs_embeds=inputs_embeds,636            use_cache=use_cache,637            output_attentions=output_attentions,638            output_hidden_states=output_hidden_states,639            return_dict=return_dict,640        )641        hidden_states = transformer_outputs[0]642 643        # Set device for model parallelism644        if self.model_parallel:645            torch.cuda.set_device(self.transformer.first_device)646            hidden_states = hidden_states.to(self.lm_head.weight.device)647 648        # make sure sampling in fp16 works correctly and649        # compute loss in fp32 to match with mesh-tf version650        # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179651        lm_logits = self.lm_head(hidden_states).to(torch.float32)652 653        loss = None654        if labels is not None:655            # Shift so that tokens < n predict n656            shift_logits = lm_logits[..., :-1, :].contiguous()657            shift_labels = labels[..., 1:].contiguous()658            # Flatten the tokens659            loss_fct = CrossEntropyLoss()660            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))661 662            loss = loss.to(hidden_states.dtype)663 664        if not return_dict:665            output = (lm_logits,) + transformer_outputs[1:]666            return ((loss,) + output) if loss is not None else output667 668        return CausalLMOutputWithPast(669            loss=loss,670            logits=lm_logits,671            past_key_values=transformer_outputs.past_key_values,672            hidden_states=transformer_outputs.hidden_states,673            attentions=transformer_outputs.attentions,674        )675 676    @staticmethod677    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:678        """679        This function is used to re-order the :obj:`past_key_values` cache if680        :meth:`~transformers.PretrainedModel.beam_search` or :meth:`~transformers.PretrainedModel.beam_sample` is681        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.682        """683        return tuple(684            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)685            for layer_past in past686        )687