transformers/pr_48314/en/_app/immutable/nodes/123.DRLN9KIn.js

import"../chunks/DsnmJJEf.js";import{i as H,h as $,C as Q,H as s,D as a,E as Y,s as K,a as P}from"../chunks/CD2rhSaz.js";import{p as ee,o as oe,s as e,f,a as r,b as ne,c as i,d as c,r as d,n as U}from"../chunks/DmjbnfDo.js";import{T as g}from"../chunks/B2suExpn.js";import{E as R}from"../chunks/B5qO6y4i.js";const te='{"title":"DeepSeek-V2","local":"deepseek-v2","sections":[{"title":"Overview","local":"overview","sections":[{"title":"Usage tips","local":"usage-tips","sections":[],"depth":3}],"depth":2},{"title":"DeepseekV2Config","local":"transformers.DeepseekV2Config","sections":[],"depth":2},{"title":"DeepseekV2Model","local":"transformers.DeepseekV2Model","sections":[],"depth":2},{"title":"DeepseekV2ForCausalLM","local":"transformers.DeepseekV2ForCausalLM","sections":[],"depth":2},{"title":"DeepseekV2ForSequenceClassification","local":"transformers.DeepseekV2ForSequenceClassification","sections":[],"depth":2}],"depth":1}';var se=c('<meta name="hf:doc:metadata"/>'),Z=c("<p>Example:</p> <!>",1),_=c(`<p>Although the recipe for forward pass needs to be defined within this function, one should call the <code>Module</code> instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.</p>`),ae=c(`<p></p> <p><em>This model was published in HF papers on 2024-05-07 and contributed to Hugging Face Transformers on 2025-07-09.</em></p> <!> <!> <!> <p>The DeepSeek-V2 model was proposed in <a href="https://huggingface.co/papers/2405.04434" rel="nofollow">DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model</a> by DeepSeek-AI Team.</p> <p>The abstract from the paper is the following:
We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE. MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times. We pretrain DeepSeek-V2 on a high-quality and multi-source corpus consisting of 8.1T tokens, and further perform Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unlock its potential. Evaluation results show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models.</p> <p>This model was contributed by <a href="https://github.com/VladOS95-cyber" rel="nofollow">VladOS95-cyber</a>.
The original code can be found <a href="https://huggingface.co/deepseek-ai/DeepSeek-V2" rel="nofollow">here</a>.</p> <!> <p>The model uses Multi-head Latent Attention (MLA) and DeepSeekMoE architectures for efficient inference and cost-effective training. It employs an auxiliary-loss-free strategy for load balancing and multi-token prediction training objective. The model can be used for various language tasks after being pre-trained on 14.8 trillion tokens and going through Supervised Fine-Tuning and Reinforcement Learning stages.</p> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>This is the configuration class to store the configuration of a DeepseekV2Model. It is used to instantiate a Deepseek V2
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the <a href="https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite" rel="nofollow">deepseek-ai/DeepSeek-V2-Lite</a></p> <p>Configuration objects inherit from <a href="/docs/transformers/pr_48314/en/main_classes/configuration#transformers.PreTrainedConfig">PreTrainedConfig</a> and can be used to control the model outputs. Read the
documentation from <a href="/docs/transformers/pr_48314/en/main_classes/configuration#transformers.PreTrainedConfig">PreTrainedConfig</a> for more information.</p> <!></div> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>The bare Deepseek V2 Model outputting raw hidden-states without any specific head on top.</p> <p>This model inherits from <a href="/docs/transformers/pr_48314/en/main_classes/model#transformers.PreTrainedModel">PreTrainedModel</a>. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)</p> <p>This model is also a PyTorch <a href="https://pytorch.org/docs/stable/nn.html#torch.nn.Module" rel="nofollow">torch.nn.Module</a> subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.</p> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>The <a href="/docs/transformers/pr_48314/en/model_doc/deepseek_v2#transformers.DeepseekV2Model">DeepseekV2Model</a> forward method, overrides the <code>__call__</code> special method.</p> <!> <ul><li><p><strong>last_hidden_state</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, sequence_length, hidden_size)</code>) — Sequence of hidden-states at the output of the last layer of the model.</p> <p>If <code>past_key_values</code> is used only the last hidden-state of the sequences of shape <code>(batch_size, 1, hidden_size)</code> is output.</p></li> <li><p><strong>past_key_values</strong> (<code>Cache</code>, <em>optional</em>, returned when <code>use_cache=True</code> is passed or when <code>config.use_cache=True</code>) — It is a <a href="/docs/transformers/pr_48314/en/internal/generation_utils#transformers.Cache">Cache</a> instance. For more details, see our <a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>.</p> <p>Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if <code>config.is_encoder_decoder=True</code> in the cross-attention blocks) that can be used (see <code>past_key_values</code> input) to speed up sequential decoding.</p></li> <li><p><strong>hidden_states</strong> (<code>tuple(torch.FloatTensor)</code>, <em>optional</em>, returned when <code>output_hidden_states=True</code> is passed or when <code>config.output_hidden_states=True</code>) — Tuple of <code>torch.FloatTensor</code> (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape <code>(batch_size, sequence_length, hidden_size)</code>.</p> <p>Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.</p></li> <li><p><strong>attentions</strong> (<code>tuple(torch.FloatTensor)</code>, <em>optional</em>, returned when <code>output_attentions=True</code> is passed or when <code>config.output_attentions=True</code>) — Tuple of <code>torch.FloatTensor</code> (one for each layer) of shape <code>(batch_size, num_heads, sequence_length, sequence_length)</code>.</p> <p>Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.</p></li></ul></div></div> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>The Deepseek V2 Model for causal language modeling.</p> <p>This model inherits from <a href="/docs/transformers/pr_48314/en/main_classes/model#transformers.PreTrainedModel">PreTrainedModel</a>. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)</p> <p>This model is also a PyTorch <a href="https://pytorch.org/docs/stable/nn.html#torch.nn.Module" rel="nofollow">torch.nn.Module</a> subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.</p> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>The <a href="/docs/transformers/pr_48314/en/model_doc/deepseek_v2#transformers.DeepseekV2ForCausalLM">DeepseekV2ForCausalLM</a> forward method, overrides the <code>__call__</code> special method.</p> <!> <ul><li><p><strong>loss</strong> (<code>torch.FloatTensor</code> of shape <code>(1,)</code>, <em>optional</em>, returned when <code>labels</code> is provided) — Language modeling loss (for next-token prediction).</p></li> <li><p><strong>logits</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, sequence_length, config.vocab_size)</code>) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).</p></li> <li><p><strong>past_key_values</strong> (<code>Cache</code>, <em>optional</em>, returned when <code>use_cache=True</code> is passed or when <code>config.use_cache=True</code>) — It is a <a href="/docs/transformers/pr_48314/en/internal/generation_utils#transformers.Cache">Cache</a> instance. For more details, see our <a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>.</p> <p>Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see <code>past_key_values</code> input) to speed up sequential decoding.</p></li> <li><p><strong>hidden_states</strong> (<code>tuple(torch.FloatTensor)</code>, <em>optional</em>, returned when <code>output_hidden_states=True</code> is passed or when <code>config.output_hidden_states=True</code>) — Tuple of <code>torch.FloatTensor</code> (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape <code>(batch_size, sequence_length, hidden_size)</code>.</p> <p>Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.</p></li> <li><p><strong>attentions</strong> (<code>tuple(torch.FloatTensor)</code>, <em>optional</em>, returned when <code>output_attentions=True</code> is passed or when <code>config.output_attentions=True</code>) — Tuple of <code>torch.FloatTensor</code> (one for each layer) of shape <code>(batch_size, num_heads, sequence_length, sequence_length)</code>.</p> <p>Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.</p></li></ul> <!></div></div> <!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"><!> <p>The <code>GenericForSequenceClassification</code> forward method, overrides the <code>__call__</code> special method.</p> <!></div></div> <!> <p></p>`,1);function pe(J,B){ee(B,!1),oe(()=>{new URLSearchParams(window.location.search).get("fw")}),H();var k=ae();$("exfver",o=>{var t=se();K(t,"content",te),r(o,t)});var b=e(f(k),4);Q(b,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var v=e(b,2);s(v,{title:"DeepSeek-V2",local:"deepseek-v2",headingTag:"h1"});var y=e(v,2);s(y,{title:"Overview",local:"overview",headingTag:"h2"});var T=e(y,8);s(T,{title:"Usage tips",local:"usage-tips",headingTag:"h3"});var w=e(T,4);s(w,{title:"DeepseekV2Config",local:"transformers.DeepseekV2Config",headingTag:"h2"});var l=e(w,2),V=i(l);a(V,{name:"class transformers.DeepseekV2Config",anchor:"transformers.DeepseekV2Config",source:"https://github.com/huggingface/transformers/blob/vr_48314/src/transformers/models/deepseek_v2/configuration_deepseek_v2.py#L30",parameters:[{name:"transformers_version",val:": str | None = None"},{name:"architectures",val:": list[str] | None = None"},{name:"output_hidden_states",val:": bool | None = False"},{name:"return_dict",val:": bool | None = True"},{name:"dtype",val:": typing.Union[str, ForwardRef('torch.dtype'), NoneType] = None"},{name:"chunk_size_feed_forward",val:": int = 0"},{name:"is_encoder_decoder",val:": bool = False"},{name:"id2label",val:": dict[int, str] | dict[str, str] | None = None"},{name:"label2id",val:": dict[str, int] | dict[str, str] | None = None"},{name:"problem_type",val:": typing.Optional[typing.Literal['regression', 'single_label_classification', 'multi_label_classification']] = None"},{name:"vocab_size",val:": int = 102400"},{name:"hidden_size",val:": int = 4096"},{name:"intermediate_size",val:": int = 11008"},{name:"num_hidden_layers",val:": int = 32"},{name:"num_attention_heads",val:": int = 32"},{name:"num_key_value_heads",val:": int | None = None"},{name:"hidden_act",val:": str = 'silu'"},{name:"max_position_embeddings",val:": int = 2048"},{name:"initializer_range",val:": float = 0.02"},{name:"rms_norm_eps",val:": float = 1e-06"},{name:"use_cache",val:": bool = True"},{name:"pad_token_id",val:": int | None = None"},{name:"bos_token_id",val:": int | None = 1"},{name:"eos_token_id",val:": int | list[int] | None = 2"},{name:"pretraining_tp",val:": int | None = 1"},{name:"tie_word_embeddings",val:": bool = False"},{name:"rope_parameters",val:": transformers.modeling_rope_utils.RopeParameters | dict | None = None"},{name:"attention_bias",val:": bool = False"},{name:"attention_dropout",val:": float | None = 0.0"},{name:"mlp_bias",val:": bool = False"},{name:"head_dim",val:": int | None = None"},{name:"first_k_dense_replace",val:": int = 0"},{name:"kv_lora_rank",val:": int = 512"},{name:"q_lora_rank",val:": int | None = 1536"},{name:"n_group",val:": int | None = None"},{name:"n_routed_experts",val:": int = 64"},{name:"n_shared_experts",val:": int = 2"},{name:"qk_nope_head_dim",val:": int = 128"},{name:"qk_rope_head_dim",val:": int = 64"},{name:"routed_scaling_factor",val:": float = 1.0"},{name:"topk_group",val:": int | None = None"},{name:"topk_method",val:": str | None = 'greedy'"},{name:"norm_topk_prob",val:": bool | None = False"},{name:"v_head_dim",val:": int = 128"},{name:"num_experts_per_tok",val:": int | None = None"},{name:"moe_intermediate_size",val:": int = 1407"}],parametersDescription:[{anchor:"transformers.DeepseekV2Config.vocab_size",description:`<strong>vocab_size</strong> (<code>int</code>, <em>optional</em>, defaults to <code>102400</code>) &#x2014;
Vocabulary size of the model. Defines the number of different tokens that can be represented by the <code>input_ids</code>.`,name:"vocab_size"},{anchor:"transformers.DeepseekV2Config.hidden_size",description:`<strong>hidden_size</strong> (<code>int</code>, <em>optional</em>, defaults to <code>4096</code>) &#x2014;
Dimension of the hidden representations.`,name:"hidden_size"},{anchor:"transformers.DeepseekV2Config.intermediate_size",description:`<strong>intermediate_size</strong> (<code>int</code>, <em>optional</em>, defaults to <code>11008</code>) &#x2014;
Dimension of the MLP representations.`,name:"intermediate_size"},{anchor:"transformers.DeepseekV2Config.num_hidden_layers",description:`<strong>num_hidden_layers</strong> (<code>int</code>, <em>optional</em>, defaults to <code>32</code>) &#x2014;
Number of hidden layers in the Transformer decoder.`,name:"num_hidden_layers"},{anchor:"transformers.DeepseekV2Config.num_attention_heads",description:`<strong>num_attention_heads</strong> (<code>int</code>, <em>optional</em>, defaults to <code>32</code>) &#x2014;
Number of attention heads for each attention layer in the Transformer decoder.`,name:"num_attention_heads"},{anchor:"transformers.DeepseekV2Config.num_key_value_heads",description:`<strong>num_key_value_heads</strong> (<code>int</code>, <em>optional</em>) &#x2014;
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
<code>num_key_value_heads=num_attention_heads</code>, the model will use Multi Head Attention (MHA), if
<code>num_key_value_heads=1</code> the model will use Multi Query Attention (MQA) otherwise GQA is used. When
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
by meanpooling all the original heads within that group. For more details, check out <a href="https://huggingface.co/papers/2305.13245" rel="nofollow">this
paper</a>. If it is not specified, will default to
<code>num_attention_heads</code>.`,name:"num_key_value_heads"},{anchor:"transformers.DeepseekV2Config.hidden_act",description:`<strong>hidden_act</strong> (<code>str</code>, <em>optional</em>, defaults to <code>silu</code>) &#x2014;
The non-linear activation function (function or string) in the decoder. For example, <code>&quot;gelu&quot;</code>,
<code>&quot;relu&quot;</code>, <code>&quot;silu&quot;</code>, etc.`,name:"hidden_act"},{anchor:"transformers.DeepseekV2Config.max_position_embeddings",description:`<strong>max_position_embeddings</strong> (<code>int</code>, <em>optional</em>, defaults to <code>2048</code>) &#x2014;
The maximum sequence length that this model might ever be used with.`,name:"max_position_embeddings"},{anchor:"transformers.DeepseekV2Config.initializer_range",description:`<strong>initializer_range</strong> (<code>float</code>, <em>optional</em>, defaults to <code>0.02</code>) &#x2014;
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.`,name:"initializer_range"},{anchor:"transformers.DeepseekV2Config.rms_norm_eps",description:`<strong>rms_norm_eps</strong> (<code>float</code>, <em>optional</em>, defaults to <code>1e-06</code>) &#x2014;
The epsilon used by the rms normalization layers.`,name:"rms_norm_eps"},{anchor:"transformers.DeepseekV2Config.use_cache",description:`<strong>use_cache</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) &#x2014;
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if <code>config.is_decoder=True</code> or when the model is a decoder-only generative model.`,name:"use_cache"},{anchor:"transformers.DeepseekV2Config.pad_token_id",description:`<strong>pad_token_id</strong> (<code>int</code>, <em>optional</em>) &#x2014;
Token id used for padding in the vocabulary.`,name:"pad_token_id"},{anchor:"transformers.DeepseekV2Config.bos_token_id",description:`<strong>bos_token_id</strong> (<code>int</code>, <em>optional</em>, defaults to <code>1</code>) &#x2014;
Token id used for beginning-of-stream in the vocabulary.`,name:"bos_token_id"},{anchor:"transformers.DeepseekV2Config.eos_token_id",description:`<strong>eos_token_id</strong> (<code>Union[int, list[int]]</code>, <em>optional</em>, defaults to <code>2</code>) &#x2014;
Token id used for end-of-stream in the vocabulary.`,name:"eos_token_id"},{anchor:"transformers.DeepseekV2Config.pretraining_tp",description:`<strong>pretraining_tp</strong> (<code>int</code>, <em>optional</em>, defaults to <code>1</code>) &#x2014;
Experimental feature. Tensor parallelism rank used during pretraining. Please refer to <a href="https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism" rel="nofollow">this
document</a> to
understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
results. Please refer to <a href="https://github.com/pytorch/pytorch/issues/76232" rel="nofollow">this issue</a>.`,name:"pretraining_tp"},{anchor:"transformers.DeepseekV2Config.tie_word_embeddings",description:`<strong>tie_word_embeddings</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) &#x2014;
Whether to tie weight embeddings according to model&#x2019;s <code>tied_weights_keys</code> mapping.`,name:"tie_word_embeddings"},{anchor:"transformers.DeepseekV2Config.rope_parameters",description:`<strong>rope_parameters</strong> (<code>Union[~modeling_rope_utils.RopeParameters, dict]</code>, <em>optional</em>) &#x2014;
Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
a value for <code>rope_theta</code> and optionally parameters used for scaling in case you want to use RoPE
with longer <code>max_position_embeddings</code>.`,name:"rope_parameters"},{anchor:"transformers.DeepseekV2Config.attention_bias",description:`<strong>attention_bias</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) &#x2014;
Whether to use a bias in the query, key, value and output projection layers during self-attention.`,name:"attention_bias"},{anchor:"transformers.DeepseekV2Config.attention_dropout",description:`<strong>attention_dropout</strong> (<code>float</code>, <em>optional</em>, defaults to <code>0.0</code>) &#x2014;
The dropout ratio for the attention probabilities.`,name:"attention_dropout"},{anchor:"transformers.DeepseekV2Config.mlp_bias",description:`<strong>mlp_bias</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) &#x2014;
Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.`,name:"mlp_bias"},{anchor:"transformers.DeepseekV2Config.head_dim",description:`<strong>head_dim</strong> (<code>int</code>, <em>optional</em>) &#x2014;
The attention head dimension. If None, it will default to hidden_size // num_attention_heads`,name:"head_dim"},{anchor:"transformers.DeepseekV2Config.first_k_dense_replace",description:`<strong>first_k_dense_replace</strong> (<code>int</code>, <em>optional</em>, defaults to 0) &#x2014;
Number of dense layers in the shallow layers before switching to MoE layers.`,name:"first_k_dense_replace"},{anchor:"transformers.DeepseekV2Config.kv_lora_rank",description:`<strong>kv_lora_rank</strong> (<code>int</code>, <em>optional</em>, defaults to <code>512</code>) &#x2014;
Rank of the LoRA matrices for key and value projections.`,name:"kv_lora_rank"},{anchor:"transformers.DeepseekV2Config.q_lora_rank",description:`<strong>q_lora_rank</strong> (<code>int</code>, <em>optional</em>, defaults to <code>1536</code>) &#x2014;
Rank of the LoRA matrices for query projections.`,name:"q_lora_rank"},{anchor:"transformers.DeepseekV2Config.n_group",description:`<strong>n_group</strong> (<code>int</code>, <em>optional</em>) &#x2014;
Number of groups for routed experts.`,name:"n_group"},{anchor:"transformers.DeepseekV2Config.n_routed_experts",description:`<strong>n_routed_experts</strong> (<code>int</code>, <em>optional</em>, defaults to <code>64</code>) &#x2014;
Number of routed experts.`,name:"n_routed_experts"},{anchor:"transformers.DeepseekV2Config.n_shared_experts",description:`<strong>n_shared_experts</strong> (<code>int</code>, <em>optional</em>, defaults to <code>2</code>) &#x2014;
Number of shared experts.`,name:"n_shared_experts"},{anchor:"transformers.DeepseekV2Config.qk_nope_head_dim",description:`<strong>qk_nope_head_dim</strong> (<code>int</code>, <em>optional</em>, defaults to <code>128</code>) &#x2014;
Dimension of the query/key heads that don&#x2019;t use rotary position embeddings.`,name:"qk_nope_head_dim"},{anchor:"transformers.DeepseekV2Config.qk_rope_head_dim",description:`<strong>qk_rope_head_dim</strong> (<code>int</code>, <em>optional</em>, defaults to <code>64</code>) &#x2014;
Dimension of the query/key heads that use rotary position embeddings.`,name:"qk_rope_head_dim"},{anchor:"transformers.DeepseekV2Config.routed_scaling_factor",description:`<strong>routed_scaling_factor</strong> (<code>float</code>, <em>optional</em>, defaults to <code>1.0</code>) &#x2014;
Scaling factor or routed experts.`,name:"routed_scaling_factor"},{anchor:"transformers.DeepseekV2Config.topk_group",description:`<strong>topk_group</strong> (<code>int</code>, <em>optional</em>) &#x2014;
Number of selected groups for each token (for each token, ensuring the selected experts is only within <code>topk_group</code> groups).`,name:"topk_group"},{anchor:"transformers.DeepseekV2Config.topk_method",description:`<strong>topk_method</strong> (<code>str</code>, <em>optional</em>, defaults to <code>&quot;greedy&quot;</code>) &#x2014;
The method used for selecting top-k experts in the routed gate mechanism.`,name:"topk_method"},{anchor:"transformers.DeepseekV2Config.norm_topk_prob",description:`<strong>norm_topk_prob</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) &#x2014;
Whether to normalize the weights of the routed experts.`,name:"norm_topk_prob"},{anchor:"transformers.DeepseekV2Config.v_head_dim",description:`<strong>v_head_dim</strong> (<code>int</code>, <em>optional</em>, defaults to <code>128</code>) &#x2014;
Dimension of the value heads.`,name:"v_head_dim"},{anchor:"transformers.DeepseekV2Config.num_experts_per_tok",description:`<strong>num_experts_per_tok</strong> (<code>int</code>, <em>optional</em>) &#x2014;
Number of experts to route each token to. This is the top-k value for the token-choice routing.`,name:"num_experts_per_tok"},{anchor:"transformers.DeepseekV2Config.moe_intermediate_size",description:`<strong>moe_intermediate_size</strong> (<code>int</code>, <em>optional</em>, defaults to <code>1407</code>) &#x2014;
Intermediate size of the routed expert MLPs.`,name:"moe_intermediate_size"}]});var E=e(V,6);R(E,{anchor:"transformers.DeepseekV2Config.example",children:(o,t)=>{var n=Z(),u=e(f(n),2);P(u,{code:"ZnJvbSUyMHRyYW5zZm9ybWVycyUyMGltcG9ydCUyMERlZXBzZWVrVjJNb2RlbCUyQyUyMERlZXBzZWVrVjJDb25maWclMEElMjMlMjBJbml0aWFsaXppbmclMjBhJTIwRGVlcFNlZWstVjIlMjBzdHlsZSUyMGNvbmZpZ3VyYXRpb24lMEFjb25maWd1cmF0aW9uJTIwJTNEJTIwRGVlcHNlZWtWMkNvbmZpZygpJTBBJTIzJTIwQWNjZXNzaW5nJTIwdGhlJTIwbW9kZWwlMjBjb25maWd1cmF0aW9uJTBBbW9kZWwlMjAlM0QlMjBEZWVwc2Vla1YyTW9kZWwoY29uZmlndXJhdGlvbiklMEFwcmludChtb2RlbC5jb25maWcp",highlighted:`<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> DeepseekV2Model, DeepseekV2Config
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># Initializing a DeepSeek-V2 style configuration</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>configuration = DeepseekV2Config()
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># Accessing the model configuration</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>model = DeepseekV2Model(configuration)
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-built_in">print</span>(model.config)`,lang:"python",wrap:!1}),r(o,n)},$$slots:{default:!0}}),d(l);var D=e(l,2);s(D,{title:"DeepseekV2Model",local:"transformers.DeepseekV2Model",headingTag:"h2"});var p=e(D,2),M=i(p);a(M,{name:"class transformers.DeepseekV2Model",anchor:"transformers.DeepseekV2Model",source:"https://github.com/huggingface/transformers/blob/vr_48314/src/transformers/models/deepseek_v2/modeling_deepseek_v2.py#L502",parameters:[{name:"config",val:": DeepseekV2Config"}],parametersDescription:[{anchor:"transformers.DeepseekV2Model.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_48314/en/model_doc/deepseek_v2#transformers.DeepseekV2Config">DeepseekV2Config</a>) &#x2014;
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
<a href="/docs/transformers/pr_48314/en/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"}]});var C=e(M,8),x=i(C);a(x,{name:"forward",anchor:"transformers.DeepseekV2Model.forward",source:"https://github.com/huggingface/transformers/blob/vr_48314/src/transformers/models/deepseek_v2/modeling_deepseek_v2.py#L519",parameters:[{name:"input_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"attention_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"position_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"past_key_values",val:": transformers.cache_utils.Cache | None = None"},{name:"inputs_embeds",val:": typing.Optional[torch.FloatTensor] = None"},{name:"use_cache",val:": bool | None = None"},{name:"**kwargs",val:": Unpack"}],parametersDescription:[{anchor:"transformers.DeepseekV2Model.forward.input_ids",description:`<strong>input_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_48314/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_48314/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_48314/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.__call__">PreTrainedTokenizer.<strong>call</strong>()</a> for details.</p>
<p><a href="../glossary#input-ids">What are input IDs?</a>`,name:"input_ids"},{anchor:"transformers.DeepseekV2Model.forward.attention_mask",description:`<strong>attention_mask</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Mask to avoid performing attention on padding token indices. Mask values selected in <code>[0, 1]</code>:</p>
<ul>
<li>1 for tokens that are <strong>not masked</strong>,</li>
<li>0 for tokens that are <strong>masked</strong>.</li>
</ul>
<p><a href="../glossary#attention-mask">What are attention masks?</a>`,name:"attention_mask"},{anchor:"transformers.DeepseekV2Model.forward.position_ids",description:`<strong>position_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range <code>[0, config.n_positions - 1]</code>.</p>
<p><a href="../glossary#position-ids">What are position IDs?</a>`,name:"position_ids"},{anchor:"transformers.DeepseekV2Model.forward.past_key_values",description:`<strong>past_key_values</strong> (<code>~cache_utils.Cache</code>, <em>optional</em>) &#x2014;
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used to speed up sequential decoding. This typically consists in the <code>past_key_values</code>
returned by the model at a previous stage of decoding, when <code>use_cache=True</code> or <code>config.use_cache=True</code>.</p>
<p>Only <a href="/docs/transformers/pr_48314/en/internal/generation_utils#transformers.Cache">Cache</a> instance is allowed as input, see our <a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>.
If no <code>past_key_values</code> are passed, <a href="/docs/transformers/pr_48314/en/internal/generation_utils#transformers.DynamicCache">DynamicCache</a> will be initialized by default.</p>
<p>The model will output the same cache format that is fed as input.</p>
<p>If <code>past_key_values</code> are used, the user is expected to input only unprocessed <code>input_ids</code> (those that don&#x2019;t
have their past key value states given to this model) of shape <code>(batch_size, unprocessed_length)</code> instead of all <code>input_ids</code>
of shape <code>(batch_size, sequence_length)</code>.`,name:"past_key_values"},{anchor:"transformers.DeepseekV2Model.forward.inputs_embeds",description:`<strong>inputs_embeds</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, sequence_length, hidden_size)</code>, <em>optional</em>) &#x2014;
Optionally, instead of passing <code>input_ids</code> you can choose to directly pass an embedded representation. This
is useful if you want more control over how to convert <code>input_ids</code> indices into associated vectors than the
model&#x2019;s internal embedding lookup matrix.`,name:"inputs_embeds"},{anchor:"transformers.DeepseekV2Model.forward.use_cache",description:`<strong>use_cache</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
If set to <code>True</code>, <code>past_key_values</code> key value states are returned and can be used to speed up decoding (see
<code>past_key_values</code>).`,name:"use_cache"}],returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>A <a
  href="/docs/transformers/pr_48314/en/main_classes/output#transformers.modeling_outputs.BaseModelOutputWithPast"
>BaseModelOutputWithPast</a> or a tuple of
<code>torch.FloatTensor</code> (if <code>return_dict=False</code> is passed or when <code>config.return_dict=False</code>) comprising various
elements depending on the configuration (<a
  href="/docs/transformers/pr_48314/en/model_doc/deepseek_v2#transformers.DeepseekV2Config"
>DeepseekV2Config</a>) and inputs.</p>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><a
  href="/docs/transformers/pr_48314/en/main_classes/output#transformers.modeling_outputs.BaseModelOutputWithPast"
>BaseModelOutputWithPast</a> or <code>tuple(torch.FloatTensor)</code></p>
`});var A=e(x,4);g(A,{children:(o,t)=>{var n=_();r(o,n)},$$slots:{default:!0}}),U(2),d(C),d(p);var z=e(p,2);s(z,{title:"DeepseekV2ForCausalLM",local:"transformers.DeepseekV2ForCausalLM",headingTag:"h2"});var m=e(z,2),F=i(m);a(F,{name:"class transformers.DeepseekV2ForCausalLM",anchor:"transformers.DeepseekV2ForCausalLM",source:"https://github.com/huggingface/transformers/blob/vr_48314/src/transformers/models/deepseek_v2/modeling_deepseek_v2.py#L576",parameters:[{name:"config",val:""}],parametersDescription:[{anchor:"transformers.DeepseekV2ForCausalLM.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_48314/en/model_doc/deepseek_v2#transformers.DeepseekV2ForCausalLM">DeepseekV2ForCausalLM</a>) &#x2014;
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
<a href="/docs/transformers/pr_48314/en/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"}]});var q=e(F,8),N=i(q);a(N,{name:"forward",anchor:"transformers.DeepseekV2ForCausalLM.forward",source:"https://github.com/huggingface/transformers/blob/vr_48314/src/transformers/models/deepseek_v2/modeling_deepseek_v2.py#L591",parameters:[{name:"input_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"attention_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"position_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"past_key_values",val:": transformers.cache_utils.Cache | None = None"},{name:"inputs_embeds",val:": typing.Optional[torch.FloatTensor] = None"},{name:"labels",val:": typing.Optional[torch.LongTensor] = None"},{name:"use_cache",val:": bool | None = None"},{name:"logits_to_keep",val:": typing.Union[int, torch.Tensor] = 0"},{name:"**kwargs",val:": Unpack"}],parametersDescription:[{anchor:"transformers.DeepseekV2ForCausalLM.forward.input_ids",description:`<strong>input_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_48314/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_48314/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_48314/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.__call__">PreTrainedTokenizer.<strong>call</strong>()</a> for details.</p>
<p><a href="../glossary#input-ids">What are input IDs?</a>`,name:"input_ids"},{anchor:"transformers.DeepseekV2ForCausalLM.forward.attention_mask",description:`<strong>attention_mask</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Mask to avoid performing attention on padding token indices. Mask values selected in <code>[0, 1]</code>:</p>
<ul>
<li>1 for tokens that are <strong>not masked</strong>,</li>
<li>0 for tokens that are <strong>masked</strong>.</li>
</ul>
<p><a href="../glossary#attention-mask">What are attention masks?</a>`,name:"attention_mask"},{anchor:"transformers.DeepseekV2ForCausalLM.forward.position_ids",description:`<strong>position_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range <code>[0, config.n_positions - 1]</code>.</p>
<p><a href="../glossary#position-ids">What are position IDs?</a>`,name:"position_ids"},{anchor:"transformers.DeepseekV2ForCausalLM.forward.past_key_values",description:`<strong>past_key_values</strong> (<code>~cache_utils.Cache</code>, <em>optional</em>) &#x2014;
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used to speed up sequential decoding. This typically consists in the <code>past_key_values</code>
returned by the model at a previous stage of decoding, when <code>use_cache=True</code> or <code>config.use_cache=True</code>.</p>
<p>Only <a href="/docs/transformers/pr_48314/en/internal/generation_utils#transformers.Cache">Cache</a> instance is allowed as input, see our <a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>.
If no <code>past_key_values</code> are passed, <a href="/docs/transformers/pr_48314/en/internal/generation_utils#transformers.DynamicCache">DynamicCache</a> will be initialized by default.</p>
<p>The model will output the same cache format that is fed as input.</p>
<p>If <code>past_key_values</code> are used, the user is expected to input only unprocessed <code>input_ids</code> (those that don&#x2019;t
have their past key value states given to this model) of shape <code>(batch_size, unprocessed_length)</code> instead of all <code>input_ids</code>
of shape <code>(batch_size, sequence_length)</code>.`,name:"past_key_values"},{anchor:"transformers.DeepseekV2ForCausalLM.forward.inputs_embeds",description:`<strong>inputs_embeds</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, sequence_length, hidden_size)</code>, <em>optional</em>) &#x2014;
Optionally, instead of passing <code>input_ids</code> you can choose to directly pass an embedded representation. This
is useful if you want more control over how to convert <code>input_ids</code> indices into associated vectors than the
model&#x2019;s internal embedding lookup matrix.`,name:"inputs_embeds"},{anchor:"transformers.DeepseekV2ForCausalLM.forward.labels",description:`<strong>labels</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Labels for computing the masked language modeling loss. Indices should either be in <code>[0, ..., config.vocab_size]</code> or -100 (see <code>input_ids</code> docstring). Tokens with indices set to <code>-100</code> are ignored
(masked), the loss is only computed for the tokens with labels in <code>[0, ..., config.vocab_size]</code>.`,name:"labels"},{anchor:"transformers.DeepseekV2ForCausalLM.forward.use_cache",description:`<strong>use_cache</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
If set to <code>True</code>, <code>past_key_values</code> key value states are returned and can be used to speed up decoding (see
<code>past_key_values</code>).`,name:"use_cache"},{anchor:"transformers.DeepseekV2ForCausalLM.forward.logits_to_keep",description:`<strong>logits_to_keep</strong> (<code>Union[int, torch.Tensor]</code>, <em>optional</em>, defaults to <code>0</code>) &#x2014;
If an <code>int</code>, compute logits for the last <code>logits_to_keep</code> tokens. If <code>0</code>, calculate logits for all
<code>input_ids</code> (special case). Only last token logits are needed for generation, and calculating them only for that
token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
If a <code>torch.Tensor</code>, must be 1D corresponding to the indices to keep in the sequence length dimension.
This is useful when using packed tensor format (single dimension for batch and sequence length).`,name:"logits_to_keep"}],returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>A <a
  href="/docs/transformers/pr_48314/en/main_classes/output#transformers.modeling_outputs.CausalLMOutputWithPast"
>CausalLMOutputWithPast</a> or a tuple of
<code>torch.FloatTensor</code> (if <code>return_dict=False</code> is passed or when <code>config.return_dict=False</code>) comprising various
elements depending on the configuration (<a
  href="/docs/transformers/pr_48314/en/model_doc/deepseek_v2#transformers.DeepseekV2Config"
>DeepseekV2Config</a>) and inputs.</p>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><a
  href="/docs/transformers/pr_48314/en/main_classes/output#transformers.modeling_outputs.CausalLMOutputWithPast"
>CausalLMOutputWithPast</a> or <code>tuple(torch.FloatTensor)</code></p>
`});var W=e(N,4);g(W,{children:(o,t)=>{var n=_();r(o,n)},$$slots:{default:!0}});var O=e(W,4);R(O,{anchor:"transformers.DeepseekV2ForCausalLM.forward.example",children:(o,t)=>{var n=Z(),u=e(f(n),2);P(u,{code:"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",highlighted:`<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer, DeepseekV2ForCausalLM

<span class="hljs-meta">&gt;&gt;&gt; </span>model = DeepseekV2ForCausalLM.from_pretrained(<span class="hljs-string">&quot;meta-deepseek_v2/DeepseekV2-2-7b-hf&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;meta-deepseek_v2/DeepseekV2-2-7b-hf&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>prompt = <span class="hljs-string">&quot;Hey, are you conscious? Can you talk to me?&quot;</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>inputs = tokenizer(prompt, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># Generate</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>generate_ids = model.generate(inputs.input_ids, max_length=<span class="hljs-number">30</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer.batch_decode(generate_ids, skip_special_tokens=<span class="hljs-literal">True</span>, clean_up_tokenization_spaces=<span class="hljs-literal">False</span>)[<span class="hljs-number">0</span>]
<span class="hljs-string">&quot;Hey, are you conscious? Can you talk to me?\\nI&#x27;m not conscious, but I can talk to you.&quot;</span>`,lang:"python",wrap:!1}),r(o,n)},$$slots:{default:!0}}),d(q),d(m);var L=e(m,2);s(L,{title:"DeepseekV2ForSequenceClassification",local:"transformers.DeepseekV2ForSequenceClassification",headingTag:"h2"});var h=e(L,2),j=i(h);a(j,{name:"class transformers.DeepseekV2ForSequenceClassification",anchor:"transformers.DeepseekV2ForSequenceClassification",source:"https://github.com/huggingface/transformers/blob/vr_48314/src/transformers/models/deepseek_v2/modeling_deepseek_v2.py#L650",parameters:[{name:"config",val:""}]});var I=e(j,2),S=i(I);a(S,{name:"forward",anchor:"transformers.DeepseekV2ForSequenceClassification.forward",source:"https://github.com/huggingface/transformers/blob/vr_48314/src/transformers/modeling_layers.py#L131",parameters:[{name:"input_ids",val:": torch.LongTensor | None = None"},{name:"attention_mask",val:": torch.Tensor | None = None"},{name:"position_ids",val:": torch.LongTensor | None = None"},{name:"past_key_values",val:": Cache | None = None"},{name:"inputs_embeds",val:": torch.FloatTensor | None = None"},{name:"labels",val:": torch.LongTensor | None = None"},{name:"use_cache",val:": bool | None = None"},{name:"**kwargs",val:": Unpack[TransformersKwargs]"}],parametersDescription:[{anchor:"transformers.DeepseekV2ForSequenceClassification.forward.input_ids",description:`<strong>input_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_48314/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_48314/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_48314/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.__call__">PreTrainedTokenizer.<strong>call</strong>()</a> for details.</p>
<p><a href="../glossary#input-ids">What are input IDs?</a>`,name:"input_ids"},{anchor:"transformers.DeepseekV2ForSequenceClassification.forward.attention_mask",description:`<strong>attention_mask</strong> (<code>torch.Tensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Mask to avoid performing attention on padding token indices. Mask values selected in <code>[0, 1]</code>:</p>
<ul>
<li>1 for tokens that are <strong>not masked</strong>,</li>
<li>0 for tokens that are <strong>masked</strong>.</li>
</ul>
<p><a href="../glossary#attention-mask">What are attention masks?</a>`,name:"attention_mask"},{anchor:"transformers.DeepseekV2ForSequenceClassification.forward.position_ids",description:`<strong>position_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range <code>[0, config.n_positions - 1]</code>.</p>
<p><a href="../glossary#position-ids">What are position IDs?</a>`,name:"position_ids"},{anchor:"transformers.DeepseekV2ForSequenceClassification.forward.past_key_values",description:`<strong>past_key_values</strong> (<code>Cache</code>, <em>optional</em>) &#x2014;
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used to speed up sequential decoding. This typically consists in the <code>past_key_values</code>
returned by the model at a previous stage of decoding, when <code>use_cache=True</code> or <code>config.use_cache=True</code>.</p>
<p>Only <a href="/docs/transformers/pr_48314/en/internal/generation_utils#transformers.Cache">Cache</a> instance is allowed as input, see our <a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>.
If no <code>past_key_values</code> are passed, <a href="/docs/transformers/pr_48314/en/internal/generation_utils#transformers.DynamicCache">DynamicCache</a> will be initialized by default.</p>
<p>The model will output the same cache format that is fed as input.</p>
<p>If <code>past_key_values</code> are used, the user is expected to input only unprocessed <code>input_ids</code> (those that don&#x2019;t
have their past key value states given to this model) of shape <code>(batch_size, unprocessed_length)</code> instead of all <code>input_ids</code>
of shape <code>(batch_size, sequence_length)</code>.`,name:"past_key_values"},{anchor:"transformers.DeepseekV2ForSequenceClassification.forward.inputs_embeds",description:`<strong>inputs_embeds</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, sequence_length, hidden_size)</code>, <em>optional</em>) &#x2014;
Optionally, instead of passing <code>input_ids</code> you can choose to directly pass an embedded representation. This
is useful if you want more control over how to convert <code>input_ids</code> indices into associated vectors than the
model&#x2019;s internal embedding lookup matrix.`,name:"inputs_embeds"},{anchor:"transformers.DeepseekV2ForSequenceClassification.forward.labels",description:`<strong>labels</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Labels for computing the masked language modeling loss. Indices should either be in <code>[0, ..., config.vocab_size]</code> or -100 (see <code>input_ids</code> docstring). Tokens with indices set to <code>-100</code> are ignored
(masked), the loss is only computed for the tokens with labels in <code>[0, ..., config.vocab_size]</code>.`,name:"labels"},{anchor:"transformers.DeepseekV2ForSequenceClassification.forward.use_cache",description:`<strong>use_cache</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
If set to <code>True</code>, <code>past_key_values</code> key value states are returned and can be used to speed up decoding (see
<code>past_key_values</code>).`,name:"use_cache"}],returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><code>SequenceClassifierOutputWithPast</code></p>
`});var G=e(S,4);g(G,{children:(o,t)=>{var n=_();r(o,n)},$$slots:{default:!0}}),d(I),d(h);var X=e(h,2);Y(X,{source:"https://github.com/huggingface/transformers/blob/main/docs/source/en/model_doc/deepseek_v2.md"}),U(2),r(J,k),ne()}export{pe as component};