transformers/pr_35674/en/_app/immutable/nodes/167.ef741eeb.js

import{s as ro,o as lo,n as re}from"../chunks/scheduler.25b97de1.js";import{S as io,i as co,g as p,s as r,r as _,A as po,h as u,f as d,c as l,j as se,u as T,x as g,k as ae,l as uo,y as i,a as c,v as y,d as b,t as M,w}from"../chunks/index.d9030fc9.js";import{T as rt}from"../chunks/Tip.baa67368.js";import{D as he}from"../chunks/Docstring.ffac8efa.js";import{C as He}from"../chunks/CodeBlock.e6cd0d95.js";import{F as ho,M as ao}from"../chunks/Markdown.7217f838.js";import{E as De}from"../chunks/ExampleCodeBlock.22dfe688.js";import{H as be,E as mo}from"../chunks/EditOnGithub.91d95064.js";function fo(G){let e,h="Example:",t,s,m;return s=new He({props:{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> GPTNeoConfig, GPTNeoModel

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># Initializing a GPTNeo EleutherAI/gpt-neo-1.3B style configuration</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>configuration = GPTNeoConfig()

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># Initializing a model (with random weights) from the EleutherAI/gpt-neo-1.3B style configuration</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoModel(configuration)

<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>configuration = model.config`,wrap:!1}}),{c(){e=p("p"),e.textContent=h,t=r(),_(s.$$.fragment)},l(o){e=u(o,"P",{"data-svelte-h":!0}),g(e)!=="svelte-11lpom8"&&(e.textContent=h),t=l(o),T(s.$$.fragment,o)},m(o,$){c(o,e,$),c(o,t,$),y(s,o,$),m=!0},p:re,i(o){m||(b(s.$$.fragment,o),m=!0)},o(o){M(s.$$.fragment,o),m=!1},d(o){o&&(d(e),d(t)),w(s,o)}}}function go(G){let e,h=`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.`;return{c(){e=p("p"),e.innerHTML=h},l(t){e=u(t,"P",{"data-svelte-h":!0}),g(e)!=="svelte-fincs2"&&(e.innerHTML=h)},m(t,s){c(t,e,s)},p:re,d(t){t&&d(e)}}}function _o(G){let e,h="Example:",t,s,m;return s=new He({props:{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, GPTNeoModel
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch

<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoModel.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>inputs = tokenizer(<span class="hljs-string">&quot;Hello, my dog is cute&quot;</span>, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>outputs = model(**inputs)

<span class="hljs-meta">&gt;&gt;&gt; </span>last_hidden_states = outputs.last_hidden_state`,wrap:!1}}),{c(){e=p("p"),e.textContent=h,t=r(),_(s.$$.fragment)},l(o){e=u(o,"P",{"data-svelte-h":!0}),g(e)!=="svelte-11lpom8"&&(e.textContent=h),t=l(o),T(s.$$.fragment,o)},m(o,$){c(o,e,$),c(o,t,$),y(s,o,$),m=!0},p:re,i(o){m||(b(s.$$.fragment,o),m=!0)},o(o){M(s.$$.fragment,o),m=!1},d(o){o&&(d(e),d(t)),w(s,o)}}}function To(G){let e,h=`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.`;return{c(){e=p("p"),e.innerHTML=h},l(t){e=u(t,"P",{"data-svelte-h":!0}),g(e)!=="svelte-fincs2"&&(e.innerHTML=h)},m(t,s){c(t,e,s)},p:re,d(t){t&&d(e)}}}function yo(G){let e,h="Example:",t,s,m;return s=new He({props:{code:"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",highlighted:`<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer, GPTNeoForCausalLM

<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoForCausalLM.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>inputs = tokenizer(<span class="hljs-string">&quot;Hello, my dog is cute&quot;</span>, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>outputs = model(**inputs, labels=inputs[<span class="hljs-string">&quot;input_ids&quot;</span>])
<span class="hljs-meta">&gt;&gt;&gt; </span>loss = outputs.loss
<span class="hljs-meta">&gt;&gt;&gt; </span>logits = outputs.logits`,wrap:!1}}),{c(){e=p("p"),e.textContent=h,t=r(),_(s.$$.fragment)},l(o){e=u(o,"P",{"data-svelte-h":!0}),g(e)!=="svelte-11lpom8"&&(e.textContent=h),t=l(o),T(s.$$.fragment,o)},m(o,$){c(o,e,$),c(o,t,$),y(s,o,$),m=!0},p:re,i(o){m||(b(s.$$.fragment,o),m=!0)},o(o){M(s.$$.fragment,o),m=!1},d(o){o&&(d(e),d(t)),w(s,o)}}}function bo(G){let e,h=`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.`;return{c(){e=p("p"),e.innerHTML=h},l(t){e=u(t,"P",{"data-svelte-h":!0}),g(e)!=="svelte-fincs2"&&(e.innerHTML=h)},m(t,s){c(t,e,s)},p:re,d(t){t&&d(e)}}}function Mo(G){let e,h=`This example uses a random model as the real ones are all very big. To get proper results, you should use
EleutherAI/gpt-neo-1.3B instead of EleutherAI/gpt-neo-1.3B. If you get out-of-memory when loading that checkpoint, you can try
adding <code>device_map=&quot;auto&quot;</code> in the <code>from_pretrained</code> call.`;return{c(){e=p("p"),e.innerHTML=h},l(t){e=u(t,"P",{"data-svelte-h":!0}),g(e)!=="svelte-ayav4f"&&(e.innerHTML=h)},m(t,s){c(t,e,s)},p:re,d(t){t&&d(e)}}}function wo(G){let e,h="Example:",t,s,m;return s=new He({props:{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, GPTNeoForQuestionAnswering
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch

<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoForQuestionAnswering.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>question, text = <span class="hljs-string">&quot;Who was Jim Henson?&quot;</span>, <span class="hljs-string">&quot;Jim Henson was a nice puppet&quot;</span>

<span class="hljs-meta">&gt;&gt;&gt; </span>inputs = tokenizer(question, text, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">with</span> torch.no_grad():
<span class="hljs-meta">... </span>    outputs = model(**inputs)

<span class="hljs-meta">&gt;&gt;&gt; </span>answer_start_index = outputs.start_logits.argmax()
<span class="hljs-meta">&gt;&gt;&gt; </span>answer_end_index = outputs.end_logits.argmax()

<span class="hljs-meta">&gt;&gt;&gt; </span>predict_answer_tokens = inputs.input_ids[<span class="hljs-number">0</span>, answer_start_index : answer_end_index + <span class="hljs-number">1</span>]

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># target is &quot;nice puppet&quot;</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>target_start_index = torch.tensor([<span class="hljs-number">14</span>])
<span class="hljs-meta">&gt;&gt;&gt; </span>target_end_index = torch.tensor([<span class="hljs-number">15</span>])

<span class="hljs-meta">&gt;&gt;&gt; </span>outputs = model(**inputs, start_positions=target_start_index, end_positions=target_end_index)
<span class="hljs-meta">&gt;&gt;&gt; </span>loss = outputs.loss`,wrap:!1}}),{c(){e=p("p"),e.textContent=h,t=r(),_(s.$$.fragment)},l(o){e=u(o,"P",{"data-svelte-h":!0}),g(e)!=="svelte-11lpom8"&&(e.textContent=h),t=l(o),T(s.$$.fragment,o)},m(o,$){c(o,e,$),c(o,t,$),y(s,o,$),m=!0},p:re,i(o){m||(b(s.$$.fragment,o),m=!0)},o(o){M(s.$$.fragment,o),m=!1},d(o){o&&(d(e),d(t)),w(s,o)}}}function vo(G){let e,h=`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.`;return{c(){e=p("p"),e.innerHTML=h},l(t){e=u(t,"P",{"data-svelte-h":!0}),g(e)!=="svelte-fincs2"&&(e.innerHTML=h)},m(t,s){c(t,e,s)},p:re,d(t){t&&d(e)}}}function ko(G){let e,h="Example of single-label classification:",t,s,m;return s=new He({props:{code:"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",highlighted:`<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer, GPTNeoForSequenceClassification

<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoForSequenceClassification.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>inputs = tokenizer(<span class="hljs-string">&quot;Hello, my dog is cute&quot;</span>, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">with</span> torch.no_grad():
<span class="hljs-meta">... </span>    logits = model(**inputs).logits

<span class="hljs-meta">&gt;&gt;&gt; </span>predicted_class_id = logits.argmax().item()

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># To train a model on \`num_labels\` classes, you can pass \`num_labels=num_labels\` to \`.from_pretrained(...)\`</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>num_labels = <span class="hljs-built_in">len</span>(model.config.id2label)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoForSequenceClassification.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>, num_labels=num_labels)

<span class="hljs-meta">&gt;&gt;&gt; </span>labels = torch.tensor([<span class="hljs-number">1</span>])
<span class="hljs-meta">&gt;&gt;&gt; </span>loss = model(**inputs, labels=labels).loss`,wrap:!1}}),{c(){e=p("p"),e.textContent=h,t=r(),_(s.$$.fragment)},l(o){e=u(o,"P",{"data-svelte-h":!0}),g(e)!=="svelte-ykxpe4"&&(e.textContent=h),t=l(o),T(s.$$.fragment,o)},m(o,$){c(o,e,$),c(o,t,$),y(s,o,$),m=!0},p:re,i(o){m||(b(s.$$.fragment,o),m=!0)},o(o){M(s.$$.fragment,o),m=!1},d(o){o&&(d(e),d(t)),w(s,o)}}}function $o(G){let e,h="Example of multi-label classification:",t,s,m;return s=new He({props:{code:"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",highlighted:`<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer, GPTNeoForSequenceClassification

<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoForSequenceClassification.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>, problem_type=<span class="hljs-string">&quot;multi_label_classification&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>inputs = tokenizer(<span class="hljs-string">&quot;Hello, my dog is cute&quot;</span>, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">with</span> torch.no_grad():
<span class="hljs-meta">... </span>    logits = model(**inputs).logits

<span class="hljs-meta">&gt;&gt;&gt; </span>predicted_class_ids = torch.arange(<span class="hljs-number">0</span>, logits.shape[-<span class="hljs-number">1</span>])[torch.sigmoid(logits).squeeze(dim=<span class="hljs-number">0</span>) &gt; <span class="hljs-number">0.5</span>]

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># To train a model on \`num_labels\` classes, you can pass \`num_labels=num_labels\` to \`.from_pretrained(...)\`</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>num_labels = <span class="hljs-built_in">len</span>(model.config.id2label)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoForSequenceClassification.from_pretrained(
<span class="hljs-meta">... </span>    <span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>, num_labels=num_labels, problem_type=<span class="hljs-string">&quot;multi_label_classification&quot;</span>
<span class="hljs-meta">... </span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>labels = torch.<span class="hljs-built_in">sum</span>(
<span class="hljs-meta">... </span>    torch.nn.functional.one_hot(predicted_class_ids[<span class="hljs-literal">None</span>, :].clone(), num_classes=num_labels), dim=<span class="hljs-number">1</span>
<span class="hljs-meta">... </span>).to(torch.<span class="hljs-built_in">float</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>loss = model(**inputs, labels=labels).loss`,wrap:!1}}),{c(){e=p("p"),e.textContent=h,t=r(),_(s.$$.fragment)},l(o){e=u(o,"P",{"data-svelte-h":!0}),g(e)!=="svelte-1l8e32d"&&(e.textContent=h),t=l(o),T(s.$$.fragment,o)},m(o,$){c(o,e,$),c(o,t,$),y(s,o,$),m=!0},p:re,i(o){m||(b(s.$$.fragment,o),m=!0)},o(o){M(s.$$.fragment,o),m=!1},d(o){o&&(d(e),d(t)),w(s,o)}}}function xo(G){let e,h=`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.`;return{c(){e=p("p"),e.innerHTML=h},l(t){e=u(t,"P",{"data-svelte-h":!0}),g(e)!=="svelte-fincs2"&&(e.innerHTML=h)},m(t,s){c(t,e,s)},p:re,d(t){t&&d(e)}}}function Go(G){let e,h="Example:",t,s,m;return s=new He({props:{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, GPTNeoForTokenClassification
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch

<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-125m&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoForTokenClassification.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-125m&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>inputs = tokenizer(
<span class="hljs-meta">... </span>    <span class="hljs-string">&quot;HuggingFace is a company based in Paris and New York&quot;</span>, add_special_tokens=<span class="hljs-literal">False</span>, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>
<span class="hljs-meta">... </span>)

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">with</span> torch.no_grad():
<span class="hljs-meta">... </span>    logits = model(**inputs).logits

<span class="hljs-meta">&gt;&gt;&gt; </span>predicted_token_class_ids = logits.argmax(-<span class="hljs-number">1</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># Note that tokens are classified rather then input words which means that</span>
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># there might be more predicted token classes than words.</span>
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># Multiple token classes might account for the same word</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>predicted_tokens_classes = [model.config.id2label[t.item()] <span class="hljs-keyword">for</span> t <span class="hljs-keyword">in</span> predicted_token_class_ids[<span class="hljs-number">0</span>]]

<span class="hljs-meta">&gt;&gt;&gt; </span>labels = predicted_token_class_ids
<span class="hljs-meta">&gt;&gt;&gt; </span>loss = model(**inputs, labels=labels).loss
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-built_in">round</span>(loss.item(), <span class="hljs-number">2</span>)
<span class="hljs-number">0.25</span>`,wrap:!1}}),{c(){e=p("p"),e.textContent=h,t=r(),_(s.$$.fragment)},l(o){e=u(o,"P",{"data-svelte-h":!0}),g(e)!=="svelte-11lpom8"&&(e.textContent=h),t=l(o),T(s.$$.fragment,o)},m(o,$){c(o,e,$),c(o,t,$),y(s,o,$),m=!0},p:re,i(o){m||(b(s.$$.fragment,o),m=!0)},o(o){M(s.$$.fragment,o),m=!1},d(o){o&&(d(e),d(t)),w(s,o)}}}function Jo(G){let e,h,t,s,m,o,$="The bare GPT Neo Model transformer outputting raw hidden-states without any specific head on top.",Ue,I,Ke=`This model inherits from <a href="/docs/transformers/pr_35674/en/main_classes/model#transformers.PreTrainedModel">PreTrainedModel</a>. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)`,ze,W,et=`This model is also a PyTorch <a href="https://pytorch.org/docs/stable/nn.html#torch.nn.Module" rel="nofollow">torch.nn.Module</a> subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.`,Ie,P,Me,We,Z,Xe='The <a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoModel">GPTNeoModel</a> forward method, overrides the <code>__call__</code> special method.',de,U,le,R,ie,Y,we,j,me,ee,E,ve=`The GPT Neo Model transformer with a language modeling head on top (linear layer with weights tied to the input
embeddings).`,ke,C,$e=`This model inherits from <a href="/docs/transformers/pr_35674/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.)`,Ze,B,Ee=`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.`,ce,z,O,Qe,D,fe='The <a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoForCausalLM">GPTNeoForCausalLM</a> forward method, overrides the <code>__call__</code> special method.',Se,L,ge,te,Q,xe,Ge,J,H,q,V,Ye=`The GPT-Neo Model transformer with a span classification head on top for extractive question-answering tasks like
SQuAD (a linear layer on top of the hidden-states output to compute <code>span start logits</code> and <code>span end logits</code>).`,_e,K,Ae=`This model inherits from <a href="/docs/transformers/pr_35674/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.)`,S,Te,k=`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.`,x,F,N,X,oe,lt='The <a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoForQuestionAnswering">GPTNeoForQuestionAnswering</a> forward method, overrides the <code>__call__</code> special method.',Oe,Je,n,f,pe,qe,kt,it,$t,A,dt,Ct,mt,Et="The GPTNeo Model transformer with a sequence classification head on top (linear layer).",Nt,ft,Qt=`<a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoForSequenceClassification">GPTNeoForSequenceClassification</a> uses the last token in order to do the classification, as other causal models
(e.g. GPT-1) do.`,jt,gt,St=`Since it does classification on the last token, it requires to know the position of the last token. If a
<code>pad_token_id</code> is defined in the configuration, it finds the last token that is not a padding token in each row. If
no <code>pad_token_id</code> is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
padding tokens when <code>inputs_embeds</code> are passed instead of <code>input_ids</code>, it does the same (take the last value in
each row of the batch).`,Ft,_t,At=`This model inherits from <a href="/docs/transformers/pr_35674/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.)`,Pt,Tt,Yt=`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.`,Ut,ye,ct,zt,yt,Ot='The <a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoForSequenceClassification">GPTNeoForSequenceClassification</a> forward method, overrides the <code>__call__</code> special method.',It,tt,Wt,ot,Zt,nt,xt,pt,Gt,ue,ut,Bt,bt,Dt=`GPT Neo model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
Named-Entity-Recognition (NER) tasks.`,qt,Mt,Kt=`This model inherits from <a href="/docs/transformers/pr_35674/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.)`,Vt,wt,eo=`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.`,Rt,Be,ht,Lt,vt,to='The <a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoForTokenClassification">GPTNeoForTokenClassification</a> forward method, overrides the <code>__call__</code> special method.',Ht,st,Xt,at,Jt;return e=new be({props:{title:"GPTNeoModel",local:"transformers.GPTNeoModel",headingTag:"h2"}}),s=new he({props:{name:"class transformers.GPTNeoModel",anchor:"transformers.GPTNeoModel",parameters:[{name:"config",val:""}],parametersDescription:[{anchor:"transformers.GPTNeoModel.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig">GPTNeoConfig</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_35674/en/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L618"}}),Me=new he({props:{name:"forward",anchor:"transformers.GPTNeoModel.forward",parameters:[{name:"input_ids",val:": typing.Optional[torch.Tensor] = None"},{name:"past_key_values",val:": typing.Union[transformers.cache_utils.Cache, typing.Tuple[torch.FloatTensor], NoneType] = None"},{name:"attention_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"token_type_ids",val:": typing.Optional[torch.Tensor] = None"},{name:"position_ids",val:": typing.Optional[torch.Tensor] = None"},{name:"head_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"inputs_embeds",val:": typing.Optional[torch.Tensor] = None"},{name:"use_cache",val:": typing.Optional[bool] = None"},{name:"output_attentions",val:": typing.Optional[bool] = None"},{name:"output_hidden_states",val:": typing.Optional[bool] = None"},{name:"return_dict",val:": typing.Optional[bool] = None"},{name:"cache_position",val:": typing.Optional[torch.LongTensor] = None"}],parametersDescription:[{anchor:"transformers.GPTNeoModel.forward.input_ids",description:`<strong>input_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>) &#x2014;
<code>input_ids_length</code> = <code>sequence_length</code> if <code>past_key_values</code> is <code>None</code> else
<code>past_key_values[0][0].shape[-2]</code> (<code>sequence_length</code> of input past key value states). Indices of input
sequence tokens in the vocabulary.</p>
<p>If <code>past_key_values</code> is used, only <code>input_ids</code> that do not have their past calculated should be passed as
<code>input_ids</code>.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_35674/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_35674/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_35674/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.GPTNeoModel.forward.past_key_values",description:`<strong>past_key_values</strong> (<code>Cache</code> or <code>tuple(tuple(torch.FloatTensor))</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>Two formats are allowed:</p>
<ul>
<li>a <a href="/docs/transformers/pr_35674/en/internal/generation_utils#transformers.Cache">Cache</a> instance, see our
<a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>;</li>
<li>Tuple of <code>tuple(torch.FloatTensor)</code> of length <code>config.n_layers</code>, with each tuple having 2 tensors of
shape <code>(batch_size, num_heads, sequence_length, embed_size_per_head)</code>). This is also known as the legacy
cache format.</li>
</ul>
<p>The model will output the same cache format that is fed as input. If no <code>past_key_values</code> are passed, the
legacy cache format will be returned.</p>
<p>If <code>past_key_values</code> are used, the user can optionally input only the last <code>input_ids</code> (those that don&#x2019;t
have their past key value states given to this model) of shape <code>(batch_size, 1)</code> instead of all <code>input_ids</code>
of shape <code>(batch_size, sequence_length)</code>.`,name:"past_key_values"},{anchor:"transformers.GPTNeoModel.forward.attention_mask",description:`<strong>attention_mask</strong> (<code>torch.FloatTensor</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.GPTNeoModel.forward.token_type_ids",description:`<strong>token_type_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>, <em>optional</em>) &#x2014;
Segment token indices to indicate first and second portions of the inputs. Indices are selected in <code>[0, 1]</code>:</p>
<ul>
<li>0 corresponds to a <em>sentence A</em> token,</li>
<li>1 corresponds to a <em>sentence B</em> token.</li>
</ul>
<p><a href="../glossary#token-type-ids">What are token type IDs?</a>`,name:"token_type_ids"},{anchor:"transformers.GPTNeoModel.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.max_position_embeddings - 1]</code>.</p>
<p><a href="../glossary#position-ids">What are position IDs?</a>`,name:"position_ids"},{anchor:"transformers.GPTNeoModel.forward.head_mask",description:`<strong>head_mask</strong> (<code>torch.FloatTensor</code> of shape <code>(num_heads,)</code> or <code>(num_layers, num_heads)</code>, <em>optional</em>) &#x2014;
Mask to nullify selected heads of the self-attention modules. Mask values selected in <code>[0, 1]</code>:</p>
<ul>
<li>1 indicates the head is <strong>not masked</strong>,</li>
<li>0 indicates the head is <strong>masked</strong>.</li>
</ul>`,name:"head_mask"},{anchor:"transformers.GPTNeoModel.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.</p>
<p>If <code>past_key_values</code> is used, optionally only the last <code>inputs_embeds</code> have to be input (see
<code>past_key_values</code>).`,name:"inputs_embeds"},{anchor:"transformers.GPTNeoModel.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.GPTNeoModel.forward.output_attentions",description:`<strong>output_attentions</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the attentions tensors of all attention layers. See <code>attentions</code> under returned
tensors for more detail.`,name:"output_attentions"},{anchor:"transformers.GPTNeoModel.forward.output_hidden_states",description:`<strong>output_hidden_states</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the hidden states of all layers. See <code>hidden_states</code> under returned tensors for
more detail.`,name:"output_hidden_states"},{anchor:"transformers.GPTNeoModel.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return a <a href="/docs/transformers/pr_35674/en/main_classes/output#transformers.utils.ModelOutput">ModelOutput</a> instead of a plain tuple.`,name:"return_dict"},{anchor:"transformers.GPTNeoModel.forward.cache_position",description:`<strong>cache_position</strong> (<code>torch.LongTensor</code> of shape <code>(sequence_length)</code>, <em>optional</em>) &#x2014;
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to <code>position_ids</code>,
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
the complete sequence length.`,name:"cache_position"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L643",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>A <a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_outputs.BaseModelOutputWithPastAndCrossAttentions"
>transformers.modeling_outputs.BaseModelOutputWithPastAndCrossAttentions</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_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig"
>GPTNeoConfig</a>) and inputs.</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>tuple(tuple(torch.FloatTensor))</code>, <em>optional</em>, returned when <code>use_cache=True</code> is passed or when <code>config.use_cache=True</code>) — Tuple of <code>tuple(torch.FloatTensor)</code> of length <code>config.n_layers</code>, with each tuple having 2 tensors of shape
<code>(batch_size, num_heads, sequence_length, embed_size_per_head)</code>) and optionally if
<code>config.is_encoder_decoder=True</code> 2 additional tensors of shape <code>(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)</code>.</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>
<li>
<p><strong>cross_attentions</strong> (<code>tuple(torch.FloatTensor)</code>, <em>optional</em>, returned when <code>output_attentions=True</code> and <code>config.add_cross_attention=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 of the decoder’s cross-attention layer, after the attention softmax, used to compute the
weighted average in the cross-attention heads.</p>
</li>
</ul>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_outputs.BaseModelOutputWithPastAndCrossAttentions"
>transformers.modeling_outputs.BaseModelOutputWithPastAndCrossAttentions</a> or <code>tuple(torch.FloatTensor)</code></p>
`}}),U=new rt({props:{$$slots:{default:[go]},$$scope:{ctx:G}}}),R=new De({props:{anchor:"transformers.GPTNeoModel.forward.example",$$slots:{default:[_o]},$$scope:{ctx:G}}}),Y=new be({props:{title:"GPTNeoForCausalLM",local:"transformers.GPTNeoForCausalLM",headingTag:"h2"}}),me=new he({props:{name:"class transformers.GPTNeoForCausalLM",anchor:"transformers.GPTNeoForCausalLM",parameters:[{name:"config",val:""}],parametersDescription:[{anchor:"transformers.GPTNeoForCausalLM.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig">GPTNeoConfig</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_35674/en/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L909"}}),O=new he({props:{name:"forward",anchor:"transformers.GPTNeoForCausalLM.forward",parameters:[{name:"input_ids",val:": typing.Optional[torch.Tensor] = None"},{name:"past_key_values",val:": typing.Union[transformers.cache_utils.Cache, typing.Tuple[torch.FloatTensor], NoneType] = None"},{name:"attention_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"token_type_ids",val:": typing.Optional[torch.Tensor] = None"},{name:"position_ids",val:": typing.Optional[torch.Tensor] = None"},{name:"head_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"inputs_embeds",val:": typing.Optional[torch.Tensor] = None"},{name:"labels",val:": typing.Optional[torch.Tensor] = None"},{name:"use_cache",val:": typing.Optional[bool] = None"},{name:"output_attentions",val:": typing.Optional[bool] = None"},{name:"output_hidden_states",val:": typing.Optional[bool] = None"},{name:"return_dict",val:": typing.Optional[bool] = None"},{name:"cache_position",val:": typing.Optional[torch.LongTensor] = None"}],parametersDescription:[{anchor:"transformers.GPTNeoForCausalLM.forward.input_ids",description:`<strong>input_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>) &#x2014;
<code>input_ids_length</code> = <code>sequence_length</code> if <code>past_key_values</code> is <code>None</code> else
<code>past_key_values[0][0].shape[-2]</code> (<code>sequence_length</code> of input past key value states). Indices of input
sequence tokens in the vocabulary.</p>
<p>If <code>past_key_values</code> is used, only <code>input_ids</code> that do not have their past calculated should be passed as
<code>input_ids</code>.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_35674/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_35674/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_35674/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.GPTNeoForCausalLM.forward.past_key_values",description:`<strong>past_key_values</strong> (<code>Cache</code> or <code>tuple(tuple(torch.FloatTensor))</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>Two formats are allowed:</p>
<ul>
<li>a <a href="/docs/transformers/pr_35674/en/internal/generation_utils#transformers.Cache">Cache</a> instance, see our
<a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>;</li>
<li>Tuple of <code>tuple(torch.FloatTensor)</code> of length <code>config.n_layers</code>, with each tuple having 2 tensors of
shape <code>(batch_size, num_heads, sequence_length, embed_size_per_head)</code>). This is also known as the legacy
cache format.</li>
</ul>
<p>The model will output the same cache format that is fed as input. If no <code>past_key_values</code> are passed, the
legacy cache format will be returned.</p>
<p>If <code>past_key_values</code> are used, the user can optionally input only the last <code>input_ids</code> (those that don&#x2019;t
have their past key value states given to this model) of shape <code>(batch_size, 1)</code> instead of all <code>input_ids</code>
of shape <code>(batch_size, sequence_length)</code>.`,name:"past_key_values"},{anchor:"transformers.GPTNeoForCausalLM.forward.attention_mask",description:`<strong>attention_mask</strong> (<code>torch.FloatTensor</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.GPTNeoForCausalLM.forward.token_type_ids",description:`<strong>token_type_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>, <em>optional</em>) &#x2014;
Segment token indices to indicate first and second portions of the inputs. Indices are selected in <code>[0, 1]</code>:</p>
<ul>
<li>0 corresponds to a <em>sentence A</em> token,</li>
<li>1 corresponds to a <em>sentence B</em> token.</li>
</ul>
<p><a href="../glossary#token-type-ids">What are token type IDs?</a>`,name:"token_type_ids"},{anchor:"transformers.GPTNeoForCausalLM.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.max_position_embeddings - 1]</code>.</p>
<p><a href="../glossary#position-ids">What are position IDs?</a>`,name:"position_ids"},{anchor:"transformers.GPTNeoForCausalLM.forward.head_mask",description:`<strong>head_mask</strong> (<code>torch.FloatTensor</code> of shape <code>(num_heads,)</code> or <code>(num_layers, num_heads)</code>, <em>optional</em>) &#x2014;
Mask to nullify selected heads of the self-attention modules. Mask values selected in <code>[0, 1]</code>:</p>
<ul>
<li>1 indicates the head is <strong>not masked</strong>,</li>
<li>0 indicates the head is <strong>masked</strong>.</li>
</ul>`,name:"head_mask"},{anchor:"transformers.GPTNeoForCausalLM.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.</p>
<p>If <code>past_key_values</code> is used, optionally only the last <code>inputs_embeds</code> have to be input (see
<code>past_key_values</code>).`,name:"inputs_embeds"},{anchor:"transformers.GPTNeoForCausalLM.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.GPTNeoForCausalLM.forward.output_attentions",description:`<strong>output_attentions</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the attentions tensors of all attention layers. See <code>attentions</code> under returned
tensors for more detail.`,name:"output_attentions"},{anchor:"transformers.GPTNeoForCausalLM.forward.output_hidden_states",description:`<strong>output_hidden_states</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the hidden states of all layers. See <code>hidden_states</code> under returned tensors for
more detail.`,name:"output_hidden_states"},{anchor:"transformers.GPTNeoForCausalLM.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return a <a href="/docs/transformers/pr_35674/en/main_classes/output#transformers.utils.ModelOutput">ModelOutput</a> instead of a plain tuple.`,name:"return_dict"},{anchor:"transformers.GPTNeoForCausalLM.forward.cache_position",description:`<strong>cache_position</strong> (<code>torch.LongTensor</code> of shape <code>(sequence_length)</code>, <em>optional</em>) &#x2014;
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to <code>position_ids</code>,
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
the complete sequence length.`,name:"cache_position"},{anchor:"transformers.GPTNeoForCausalLM.forward.labels",description:`<strong>labels</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, sequence_length)</code>, <em>optional</em>) &#x2014;
Labels for language modeling. Note that the labels <strong>are shifted</strong> inside the model, i.e. you can set
<code>labels = input_ids</code> Indices are selected in <code>[-100, 0, ..., config.vocab_size]</code> All labels set to <code>-100</code>
are ignored (masked), the loss is only computed for labels in <code>[0, ..., config.vocab_size]</code>`,name:"labels"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L933",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>A <a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_outputs.CausalLMOutputWithCrossAttentions"
>transformers.modeling_outputs.CausalLMOutputWithCrossAttentions</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_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig"
>GPTNeoConfig</a>) and inputs.</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>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>
<li>
<p><strong>cross_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>Cross attentions weights after the attention softmax, used to compute the weighted average in the
cross-attention heads.</p>
</li>
<li>
<p><strong>past_key_values</strong> (<code>tuple(tuple(torch.FloatTensor))</code>, <em>optional</em>, returned when <code>use_cache=True</code> is passed or when <code>config.use_cache=True</code>) — Tuple of <code>torch.FloatTensor</code> tuples of length <code>config.n_layers</code>, with each tuple containing the cached key,
value states of the self-attention and the cross-attention layers if model is used in encoder-decoder
setting. Only relevant if <code>config.is_decoder = True</code>.</p>
<p>Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
<code>past_key_values</code> input) to speed up sequential decoding.</p>
</li>
</ul>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_outputs.CausalLMOutputWithCrossAttentions"
>transformers.modeling_outputs.CausalLMOutputWithCrossAttentions</a> or <code>tuple(torch.FloatTensor)</code></p>
`}}),L=new rt({props:{$$slots:{default:[To]},$$scope:{ctx:G}}}),te=new De({props:{anchor:"transformers.GPTNeoForCausalLM.forward.example",$$slots:{default:[yo]},$$scope:{ctx:G}}}),xe=new be({props:{title:"GPTNeoForQuestionAnswering",local:"transformers.GPTNeoForQuestionAnswering",headingTag:"h2"}}),H=new he({props:{name:"class transformers.GPTNeoForQuestionAnswering",anchor:"transformers.GPTNeoForQuestionAnswering",parameters:[{name:"config",val:""}],parametersDescription:[{anchor:"transformers.GPTNeoForQuestionAnswering.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig">GPTNeoConfig</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_35674/en/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L1240"}}),N=new he({props:{name:"forward",anchor:"transformers.GPTNeoForQuestionAnswering.forward",parameters:[{name:"input_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"attention_mask",val:": typing.Optional[torch.FloatTensor] = None"},{name:"token_type_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"position_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"head_mask",val:": typing.Optional[torch.FloatTensor] = None"},{name:"inputs_embeds",val:": typing.Optional[torch.FloatTensor] = None"},{name:"start_positions",val:": typing.Optional[torch.LongTensor] = None"},{name:"end_positions",val:": typing.Optional[torch.LongTensor] = None"},{name:"output_attentions",val:": typing.Optional[bool] = None"},{name:"output_hidden_states",val:": typing.Optional[bool] = None"},{name:"return_dict",val:": typing.Optional[bool] = None"}],parametersDescription:[{anchor:"transformers.GPTNeoForQuestionAnswering.forward.input_ids",description:`<strong>input_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>) &#x2014;
<code>input_ids_length</code> = <code>sequence_length</code> if <code>past_key_values</code> is <code>None</code> else
<code>past_key_values[0][0].shape[-2]</code> (<code>sequence_length</code> of input past key value states). Indices of input
sequence tokens in the vocabulary.</p>
<p>If <code>past_key_values</code> is used, only <code>input_ids</code> that do not have their past calculated should be passed as
<code>input_ids</code>.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_35674/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_35674/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_35674/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.GPTNeoForQuestionAnswering.forward.past_key_values",description:`<strong>past_key_values</strong> (<code>Cache</code> or <code>tuple(tuple(torch.FloatTensor))</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>Two formats are allowed:</p>
<ul>
<li>a <a href="/docs/transformers/pr_35674/en/internal/generation_utils#transformers.Cache">Cache</a> instance, see our
<a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>;</li>
<li>Tuple of <code>tuple(torch.FloatTensor)</code> of length <code>config.n_layers</code>, with each tuple having 2 tensors of
shape <code>(batch_size, num_heads, sequence_length, embed_size_per_head)</code>). This is also known as the legacy
cache format.</li>
</ul>
<p>The model will output the same cache format that is fed as input. If no <code>past_key_values</code> are passed, the
legacy cache format will be returned.</p>
<p>If <code>past_key_values</code> are used, the user can optionally input only the last <code>input_ids</code> (those that don&#x2019;t
have their past key value states given to this model) of shape <code>(batch_size, 1)</code> instead of all <code>input_ids</code>
of shape <code>(batch_size, sequence_length)</code>.`,name:"past_key_values"},{anchor:"transformers.GPTNeoForQuestionAnswering.forward.attention_mask",description:`<strong>attention_mask</strong> (<code>torch.FloatTensor</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.GPTNeoForQuestionAnswering.forward.token_type_ids",description:`<strong>token_type_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>, <em>optional</em>) &#x2014;
Segment token indices to indicate first and second portions of the inputs. Indices are selected in <code>[0, 1]</code>:</p>
<ul>
<li>0 corresponds to a <em>sentence A</em> token,</li>
<li>1 corresponds to a <em>sentence B</em> token.</li>
</ul>
<p><a href="../glossary#token-type-ids">What are token type IDs?</a>`,name:"token_type_ids"},{anchor:"transformers.GPTNeoForQuestionAnswering.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.max_position_embeddings - 1]</code>.</p>
<p><a href="../glossary#position-ids">What are position IDs?</a>`,name:"position_ids"},{anchor:"transformers.GPTNeoForQuestionAnswering.forward.head_mask",description:`<strong>head_mask</strong> (<code>torch.FloatTensor</code> of shape <code>(num_heads,)</code> or <code>(num_layers, num_heads)</code>, <em>optional</em>) &#x2014;
Mask to nullify selected heads of the self-attention modules. Mask values selected in <code>[0, 1]</code>:</p>
<ul>
<li>1 indicates the head is <strong>not masked</strong>,</li>
<li>0 indicates the head is <strong>masked</strong>.</li>
</ul>`,name:"head_mask"},{anchor:"transformers.GPTNeoForQuestionAnswering.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.</p>
<p>If <code>past_key_values</code> is used, optionally only the last <code>inputs_embeds</code> have to be input (see
<code>past_key_values</code>).`,name:"inputs_embeds"},{anchor:"transformers.GPTNeoForQuestionAnswering.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.GPTNeoForQuestionAnswering.forward.output_attentions",description:`<strong>output_attentions</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the attentions tensors of all attention layers. See <code>attentions</code> under returned
tensors for more detail.`,name:"output_attentions"},{anchor:"transformers.GPTNeoForQuestionAnswering.forward.output_hidden_states",description:`<strong>output_hidden_states</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the hidden states of all layers. See <code>hidden_states</code> under returned tensors for
more detail.`,name:"output_hidden_states"},{anchor:"transformers.GPTNeoForQuestionAnswering.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return a <a href="/docs/transformers/pr_35674/en/main_classes/output#transformers.utils.ModelOutput">ModelOutput</a> instead of a plain tuple.`,name:"return_dict"},{anchor:"transformers.GPTNeoForQuestionAnswering.forward.cache_position",description:`<strong>cache_position</strong> (<code>torch.LongTensor</code> of shape <code>(sequence_length)</code>, <em>optional</em>) &#x2014;
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to <code>position_ids</code>,
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
the complete sequence length.`,name:"cache_position"},{anchor:"transformers.GPTNeoForQuestionAnswering.forward.start_positions",description:`<strong>start_positions</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size,)</code>, <em>optional</em>) &#x2014;
Labels for position (index) of the start of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (<code>sequence_length</code>). Position outside of the sequence
are not taken into account for computing the loss.`,name:"start_positions"},{anchor:"transformers.GPTNeoForQuestionAnswering.forward.end_positions",description:`<strong>end_positions</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size,)</code>, <em>optional</em>) &#x2014;
Labels for position (index) of the end of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (<code>sequence_length</code>). Position outside of the sequence
are not taken into account for computing the loss.`,name:"end_positions"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L1257",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>A <a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_outputs.QuestionAnsweringModelOutput"
>transformers.modeling_outputs.QuestionAnsweringModelOutput</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_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig"
>GPTNeoConfig</a>) and inputs.</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) — Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.</p>
</li>
<li>
<p><strong>start_logits</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, sequence_length)</code>) — Span-start scores (before SoftMax).</p>
</li>
<li>
<p><strong>end_logits</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, sequence_length)</code>) — Span-end scores (before SoftMax).</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>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_outputs.QuestionAnsweringModelOutput"
>transformers.modeling_outputs.QuestionAnsweringModelOutput</a> or <code>tuple(torch.FloatTensor)</code></p>
`}}),Je=new rt({props:{$$slots:{default:[bo]},$$scope:{ctx:G}}}),f=new rt({props:{warning:!0,$$slots:{default:[Mo]},$$scope:{ctx:G}}}),qe=new De({props:{anchor:"transformers.GPTNeoForQuestionAnswering.forward.example",$$slots:{default:[wo]},$$scope:{ctx:G}}}),it=new be({props:{title:"GPTNeoForSequenceClassification",local:"transformers.GPTNeoForSequenceClassification",headingTag:"h2"}}),dt=new he({props:{name:"class transformers.GPTNeoForSequenceClassification",anchor:"transformers.GPTNeoForSequenceClassification",parameters:[{name:"config",val:""}],parametersDescription:[{anchor:"transformers.GPTNeoForSequenceClassification.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig">GPTNeoConfig</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_35674/en/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L1026"}}),ct=new he({props:{name:"forward",anchor:"transformers.GPTNeoForSequenceClassification.forward",parameters:[{name:"input_ids",val:": typing.Optional[torch.Tensor] = None"},{name:"past_key_values",val:": typing.Union[transformers.cache_utils.Cache, typing.Tuple[torch.FloatTensor], NoneType] = None"},{name:"attention_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"token_type_ids",val:": typing.Optional[torch.Tensor] = None"},{name:"position_ids",val:": typing.Optional[torch.Tensor] = None"},{name:"head_mask",val:": typing.Optional[torch.Tensor] = None"},{name:"inputs_embeds",val:": typing.Optional[torch.Tensor] = None"},{name:"labels",val:": typing.Optional[torch.Tensor] = None"},{name:"use_cache",val:": typing.Optional[bool] = None"},{name:"output_attentions",val:": typing.Optional[bool] = None"},{name:"output_hidden_states",val:": typing.Optional[bool] = None"},{name:"return_dict",val:": typing.Optional[bool] = None"}],parametersDescription:[{anchor:"transformers.GPTNeoForSequenceClassification.forward.input_ids",description:`<strong>input_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>) &#x2014;
<code>input_ids_length</code> = <code>sequence_length</code> if <code>past_key_values</code> is <code>None</code> else
<code>past_key_values[0][0].shape[-2]</code> (<code>sequence_length</code> of input past key value states). Indices of input
sequence tokens in the vocabulary.</p>
<p>If <code>past_key_values</code> is used, only <code>input_ids</code> that do not have their past calculated should be passed as
<code>input_ids</code>.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_35674/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_35674/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_35674/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.GPTNeoForSequenceClassification.forward.past_key_values",description:`<strong>past_key_values</strong> (<code>Cache</code> or <code>tuple(tuple(torch.FloatTensor))</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>Two formats are allowed:</p>
<ul>
<li>a <a href="/docs/transformers/pr_35674/en/internal/generation_utils#transformers.Cache">Cache</a> instance, see our
<a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>;</li>
<li>Tuple of <code>tuple(torch.FloatTensor)</code> of length <code>config.n_layers</code>, with each tuple having 2 tensors of
shape <code>(batch_size, num_heads, sequence_length, embed_size_per_head)</code>). This is also known as the legacy
cache format.</li>
</ul>
<p>The model will output the same cache format that is fed as input. If no <code>past_key_values</code> are passed, the
legacy cache format will be returned.</p>
<p>If <code>past_key_values</code> are used, the user can optionally input only the last <code>input_ids</code> (those that don&#x2019;t
have their past key value states given to this model) of shape <code>(batch_size, 1)</code> instead of all <code>input_ids</code>
of shape <code>(batch_size, sequence_length)</code>.`,name:"past_key_values"},{anchor:"transformers.GPTNeoForSequenceClassification.forward.attention_mask",description:`<strong>attention_mask</strong> (<code>torch.FloatTensor</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.GPTNeoForSequenceClassification.forward.token_type_ids",description:`<strong>token_type_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>, <em>optional</em>) &#x2014;
Segment token indices to indicate first and second portions of the inputs. Indices are selected in <code>[0, 1]</code>:</p>
<ul>
<li>0 corresponds to a <em>sentence A</em> token,</li>
<li>1 corresponds to a <em>sentence B</em> token.</li>
</ul>
<p><a href="../glossary#token-type-ids">What are token type IDs?</a>`,name:"token_type_ids"},{anchor:"transformers.GPTNeoForSequenceClassification.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.max_position_embeddings - 1]</code>.</p>
<p><a href="../glossary#position-ids">What are position IDs?</a>`,name:"position_ids"},{anchor:"transformers.GPTNeoForSequenceClassification.forward.head_mask",description:`<strong>head_mask</strong> (<code>torch.FloatTensor</code> of shape <code>(num_heads,)</code> or <code>(num_layers, num_heads)</code>, <em>optional</em>) &#x2014;
Mask to nullify selected heads of the self-attention modules. Mask values selected in <code>[0, 1]</code>:</p>
<ul>
<li>1 indicates the head is <strong>not masked</strong>,</li>
<li>0 indicates the head is <strong>masked</strong>.</li>
</ul>`,name:"head_mask"},{anchor:"transformers.GPTNeoForSequenceClassification.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.</p>
<p>If <code>past_key_values</code> is used, optionally only the last <code>inputs_embeds</code> have to be input (see
<code>past_key_values</code>).`,name:"inputs_embeds"},{anchor:"transformers.GPTNeoForSequenceClassification.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.GPTNeoForSequenceClassification.forward.output_attentions",description:`<strong>output_attentions</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the attentions tensors of all attention layers. See <code>attentions</code> under returned
tensors for more detail.`,name:"output_attentions"},{anchor:"transformers.GPTNeoForSequenceClassification.forward.output_hidden_states",description:`<strong>output_hidden_states</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the hidden states of all layers. See <code>hidden_states</code> under returned tensors for
more detail.`,name:"output_hidden_states"},{anchor:"transformers.GPTNeoForSequenceClassification.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return a <a href="/docs/transformers/pr_35674/en/main_classes/output#transformers.utils.ModelOutput">ModelOutput</a> instead of a plain tuple.`,name:"return_dict"},{anchor:"transformers.GPTNeoForSequenceClassification.forward.cache_position",description:`<strong>cache_position</strong> (<code>torch.LongTensor</code> of shape <code>(sequence_length)</code>, <em>optional</em>) &#x2014;
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to <code>position_ids</code>,
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
the complete sequence length.`,name:"cache_position"},{anchor:"transformers.GPTNeoForSequenceClassification.forward.labels",description:`<strong>labels</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size,)</code>, <em>optional</em>) &#x2014;
Labels for computing the sequence classification/regression loss. Indices should be in <code>[0, ..., config.num_labels - 1]</code>. If <code>config.num_labels == 1</code> a regression loss is computed (Mean-Square loss), If
<code>config.num_labels &gt; 1</code> a classification loss is computed (Cross-Entropy).`,name:"labels"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L1051",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>A <code>transformers.modeling_outputs.SequenceClassifierOutputWithPast</code> 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_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig"
>GPTNeoConfig</a>) and inputs.</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) — Classification (or regression if config.num_labels==1) loss.</p>
</li>
<li>
<p><strong>logits</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, config.num_labels)</code>) — Classification (or regression if config.num_labels==1) scores (before SoftMax).</p>
</li>
<li>
<p><strong>past_key_values</strong> (<code>tuple(tuple(torch.FloatTensor))</code>, <em>optional</em>, returned when <code>use_cache=True</code> is passed or when <code>config.use_cache=True</code>) — Tuple of <code>tuple(torch.FloatTensor)</code> of length <code>config.n_layers</code>, with each tuple having 2 tensors of shape
<code>(batch_size, num_heads, sequence_length, embed_size_per_head)</code>)</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>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><code>transformers.modeling_outputs.SequenceClassifierOutputWithPast</code> or <code>tuple(torch.FloatTensor)</code></p>
`}}),tt=new rt({props:{$$slots:{default:[vo]},$$scope:{ctx:G}}}),ot=new De({props:{anchor:"transformers.GPTNeoForSequenceClassification.forward.example",$$slots:{default:[ko]},$$scope:{ctx:G}}}),nt=new De({props:{anchor:"transformers.GPTNeoForSequenceClassification.forward.example-2",$$slots:{default:[$o]},$$scope:{ctx:G}}}),pt=new be({props:{title:"GPTNeoForTokenClassification",local:"transformers.GPTNeoForTokenClassification",headingTag:"h2"}}),ut=new he({props:{name:"class transformers.GPTNeoForTokenClassification",anchor:"transformers.GPTNeoForTokenClassification",parameters:[{name:"config",val:""}],parametersDescription:[{anchor:"transformers.GPTNeoForTokenClassification.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig">GPTNeoConfig</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_35674/en/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L1155"}}),ht=new he({props:{name:"forward",anchor:"transformers.GPTNeoForTokenClassification.forward",parameters:[{name:"input_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"past_key_values",val:": typing.Union[transformers.cache_utils.Cache, typing.Tuple[typing.Tuple[torch.Tensor]], NoneType] = None"},{name:"attention_mask",val:": typing.Optional[torch.FloatTensor] = None"},{name:"token_type_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"position_ids",val:": typing.Optional[torch.LongTensor] = None"},{name:"head_mask",val:": typing.Optional[torch.FloatTensor] = None"},{name:"inputs_embeds",val:": typing.Optional[torch.FloatTensor] = None"},{name:"labels",val:": typing.Optional[torch.LongTensor] = None"},{name:"use_cache",val:": typing.Optional[bool] = None"},{name:"output_attentions",val:": typing.Optional[bool] = None"},{name:"output_hidden_states",val:": typing.Optional[bool] = None"},{name:"return_dict",val:": typing.Optional[bool] = None"}],parametersDescription:[{anchor:"transformers.GPTNeoForTokenClassification.forward.input_ids",description:`<strong>input_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>) &#x2014;
<code>input_ids_length</code> = <code>sequence_length</code> if <code>past_key_values</code> is <code>None</code> else
<code>past_key_values[0][0].shape[-2]</code> (<code>sequence_length</code> of input past key value states). Indices of input
sequence tokens in the vocabulary.</p>
<p>If <code>past_key_values</code> is used, only <code>input_ids</code> that do not have their past calculated should be passed as
<code>input_ids</code>.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_35674/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_35674/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_35674/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.GPTNeoForTokenClassification.forward.past_key_values",description:`<strong>past_key_values</strong> (<code>Cache</code> or <code>tuple(tuple(torch.FloatTensor))</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>Two formats are allowed:</p>
<ul>
<li>a <a href="/docs/transformers/pr_35674/en/internal/generation_utils#transformers.Cache">Cache</a> instance, see our
<a href="https://huggingface.co/docs/transformers/en/kv_cache" rel="nofollow">kv cache guide</a>;</li>
<li>Tuple of <code>tuple(torch.FloatTensor)</code> of length <code>config.n_layers</code>, with each tuple having 2 tensors of
shape <code>(batch_size, num_heads, sequence_length, embed_size_per_head)</code>). This is also known as the legacy
cache format.</li>
</ul>
<p>The model will output the same cache format that is fed as input. If no <code>past_key_values</code> are passed, the
legacy cache format will be returned.</p>
<p>If <code>past_key_values</code> are used, the user can optionally input only the last <code>input_ids</code> (those that don&#x2019;t
have their past key value states given to this model) of shape <code>(batch_size, 1)</code> instead of all <code>input_ids</code>
of shape <code>(batch_size, sequence_length)</code>.`,name:"past_key_values"},{anchor:"transformers.GPTNeoForTokenClassification.forward.attention_mask",description:`<strong>attention_mask</strong> (<code>torch.FloatTensor</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.GPTNeoForTokenClassification.forward.token_type_ids",description:`<strong>token_type_ids</strong> (<code>torch.LongTensor</code> of shape <code>(batch_size, input_ids_length)</code>, <em>optional</em>) &#x2014;
Segment token indices to indicate first and second portions of the inputs. Indices are selected in <code>[0, 1]</code>:</p>
<ul>
<li>0 corresponds to a <em>sentence A</em> token,</li>
<li>1 corresponds to a <em>sentence B</em> token.</li>
</ul>
<p><a href="../glossary#token-type-ids">What are token type IDs?</a>`,name:"token_type_ids"},{anchor:"transformers.GPTNeoForTokenClassification.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.max_position_embeddings - 1]</code>.</p>
<p><a href="../glossary#position-ids">What are position IDs?</a>`,name:"position_ids"},{anchor:"transformers.GPTNeoForTokenClassification.forward.head_mask",description:`<strong>head_mask</strong> (<code>torch.FloatTensor</code> of shape <code>(num_heads,)</code> or <code>(num_layers, num_heads)</code>, <em>optional</em>) &#x2014;
Mask to nullify selected heads of the self-attention modules. Mask values selected in <code>[0, 1]</code>:</p>
<ul>
<li>1 indicates the head is <strong>not masked</strong>,</li>
<li>0 indicates the head is <strong>masked</strong>.</li>
</ul>`,name:"head_mask"},{anchor:"transformers.GPTNeoForTokenClassification.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.</p>
<p>If <code>past_key_values</code> is used, optionally only the last <code>inputs_embeds</code> have to be input (see
<code>past_key_values</code>).`,name:"inputs_embeds"},{anchor:"transformers.GPTNeoForTokenClassification.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.GPTNeoForTokenClassification.forward.output_attentions",description:`<strong>output_attentions</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the attentions tensors of all attention layers. See <code>attentions</code> under returned
tensors for more detail.`,name:"output_attentions"},{anchor:"transformers.GPTNeoForTokenClassification.forward.output_hidden_states",description:`<strong>output_hidden_states</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the hidden states of all layers. See <code>hidden_states</code> under returned tensors for
more detail.`,name:"output_hidden_states"},{anchor:"transformers.GPTNeoForTokenClassification.forward.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return a <a href="/docs/transformers/pr_35674/en/main_classes/output#transformers.utils.ModelOutput">ModelOutput</a> instead of a plain tuple.`,name:"return_dict"},{anchor:"transformers.GPTNeoForTokenClassification.forward.cache_position",description:`<strong>cache_position</strong> (<code>torch.LongTensor</code> of shape <code>(sequence_length)</code>, <em>optional</em>) &#x2014;
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to <code>position_ids</code>,
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
the complete sequence length.`,name:"cache_position"},{anchor:"transformers.GPTNeoForTokenClassification.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 sequence classification/regression loss. Indices should be in <code>[0, ..., config.num_labels - 1]</code>. If <code>config.num_labels == 1</code> a regression loss is computed (Mean-Square loss), If
<code>config.num_labels &gt; 1</code> a classification loss is computed (Cross-Entropy).`,name:"labels"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_gpt_neo.py#L1174",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>A <a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_outputs.TokenClassifierOutput"
>transformers.modeling_outputs.TokenClassifierOutput</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_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig"
>GPTNeoConfig</a>) and inputs.</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)  — Classification loss.</p>
</li>
<li>
<p><strong>logits</strong> (<code>torch.FloatTensor</code> of shape <code>(batch_size, sequence_length, config.num_labels)</code>) — Classification scores (before SoftMax).</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>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_outputs.TokenClassifierOutput"
>transformers.modeling_outputs.TokenClassifierOutput</a> or <code>tuple(torch.FloatTensor)</code></p>
`}}),st=new rt({props:{$$slots:{default:[xo]},$$scope:{ctx:G}}}),at=new De({props:{anchor:"transformers.GPTNeoForTokenClassification.forward.example",$$slots:{default:[Go]},$$scope:{ctx:G}}}),{c(){_(e.$$.fragment),h=r(),t=p("div"),_(s.$$.fragment),m=r(),o=p("p"),o.textContent=$,Ue=r(),I=p("p"),I.innerHTML=Ke,ze=r(),W=p("p"),W.innerHTML=et,Ie=r(),P=p("div"),_(Me.$$.fragment),We=r(),Z=p("p"),Z.innerHTML=Xe,de=r(),_(U.$$.fragment),le=r(),_(R.$$.fragment),ie=r(),_(Y.$$.fragment),we=r(),j=p("div"),_(me.$$.fragment),ee=r(),E=p("p"),E.textContent=ve,ke=r(),C=p("p"),C.innerHTML=$e,Ze=r(),B=p("p"),B.innerHTML=Ee,ce=r(),z=p("div"),_(O.$$.fragment),Qe=r(),D=p("p"),D.innerHTML=fe,Se=r(),_(L.$$.fragment),ge=r(),_(te.$$.fragment),Q=r(),_(xe.$$.fragment),Ge=r(),J=p("div"),_(H.$$.fragment),q=r(),V=p("p"),V.innerHTML=Ye,_e=r(),K=p("p"),K.innerHTML=Ae,S=r(),Te=p("p"),Te.innerHTML=k,x=r(),F=p("div"),_(N.$$.fragment),X=r(),oe=p("p"),oe.innerHTML=lt,Oe=r(),_(Je.$$.fragment),n=r(),_(f.$$.fragment),pe=r(),_(qe.$$.fragment),kt=r(),_(it.$$.fragment),$t=r(),A=p("div"),_(dt.$$.fragment),Ct=r(),mt=p("p"),mt.textContent=Et,Nt=r(),ft=p("p"),ft.innerHTML=Qt,jt=r(),gt=p("p"),gt.innerHTML=St,Ft=r(),_t=p("p"),_t.innerHTML=At,Pt=r(),Tt=p("p"),Tt.innerHTML=Yt,Ut=r(),ye=p("div"),_(ct.$$.fragment),zt=r(),yt=p("p"),yt.innerHTML=Ot,It=r(),_(tt.$$.fragment),Wt=r(),_(ot.$$.fragment),Zt=r(),_(nt.$$.fragment),xt=r(),_(pt.$$.fragment),Gt=r(),ue=p("div"),_(ut.$$.fragment),Bt=r(),bt=p("p"),bt.textContent=Dt,qt=r(),Mt=p("p"),Mt.innerHTML=Kt,Vt=r(),wt=p("p"),wt.innerHTML=eo,Rt=r(),Be=p("div"),_(ht.$$.fragment),Lt=r(),vt=p("p"),vt.innerHTML=to,Ht=r(),_(st.$$.fragment),Xt=r(),_(at.$$.fragment),this.h()},l(a){T(e.$$.fragment,a),h=l(a),t=u(a,"DIV",{class:!0});var v=se(t);T(s.$$.fragment,v),m=l(v),o=u(v,"P",{"data-svelte-h":!0}),g(o)!=="svelte-1vkbqia"&&(o.textContent=$),Ue=l(v),I=u(v,"P",{"data-svelte-h":!0}),g(I)!=="svelte-az7ywp"&&(I.innerHTML=Ke),ze=l(v),W=u(v,"P",{"data-svelte-h":!0}),g(W)!=="svelte-hswkmf"&&(W.innerHTML=et),Ie=l(v),P=u(v,"DIV",{class:!0});var Ve=se(P);T(Me.$$.fragment,Ve),We=l(Ve),Z=u(Ve,"P",{"data-svelte-h":!0}),g(Z)!=="svelte-10gczlo"&&(Z.innerHTML=Xe),de=l(Ve),T(U.$$.fragment,Ve),le=l(Ve),T(R.$$.fragment,Ve),Ve.forEach(d),v.forEach(d),ie=l(a),T(Y.$$.fragment,a),we=l(a),j=u(a,"DIV",{class:!0});var Ce=se(j);T(me.$$.fragment,Ce),ee=l(Ce),E=u(Ce,"P",{"data-svelte-h":!0}),g(E)!=="svelte-ygetm0"&&(E.textContent=ve),ke=l(Ce),C=u(Ce,"P",{"data-svelte-h":!0}),g(C)!=="svelte-az7ywp"&&(C.innerHTML=$e),Ze=l(Ce),B=u(Ce,"P",{"data-svelte-h":!0}),g(B)!=="svelte-hswkmf"&&(B.innerHTML=Ee),ce=l(Ce),z=u(Ce,"DIV",{class:!0});var Re=se(z);T(O.$$.fragment,Re),Qe=l(Re),D=u(Re,"P",{"data-svelte-h":!0}),g(D)!=="svelte-1627nfw"&&(D.innerHTML=fe),Se=l(Re),T(L.$$.fragment,Re),ge=l(Re),T(te.$$.fragment,Re),Re.forEach(d),Ce.forEach(d),Q=l(a),T(xe.$$.fragment,a),Ge=l(a),J=u(a,"DIV",{class:!0});var Ne=se(J);T(H.$$.fragment,Ne),q=l(Ne),V=u(Ne,"P",{"data-svelte-h":!0}),g(V)!=="svelte-gjginq"&&(V.innerHTML=Ye),_e=l(Ne),K=u(Ne,"P",{"data-svelte-h":!0}),g(K)!=="svelte-az7ywp"&&(K.innerHTML=Ae),S=l(Ne),Te=u(Ne,"P",{"data-svelte-h":!0}),g(Te)!=="svelte-hswkmf"&&(Te.innerHTML=k),x=l(Ne),F=u(Ne,"DIV",{class:!0});var je=se(F);T(N.$$.fragment,je),X=l(je),oe=u(je,"P",{"data-svelte-h":!0}),g(oe)!=="svelte-1rbsiro"&&(oe.innerHTML=lt),Oe=l(je),T(Je.$$.fragment,je),n=l(je),T(f.$$.fragment,je),pe=l(je),T(qe.$$.fragment,je),je.forEach(d),Ne.forEach(d),kt=l(a),T(it.$$.fragment,a),$t=l(a),A=u(a,"DIV",{class:!0});var ne=se(A);T(dt.$$.fragment,ne),Ct=l(ne),mt=u(ne,"P",{"data-svelte-h":!0}),g(mt)!=="svelte-17hgce5"&&(mt.textContent=Et),Nt=l(ne),ft=u(ne,"P",{"data-svelte-h":!0}),g(ft)!=="svelte-1wnbuuy"&&(ft.innerHTML=Qt),jt=l(ne),gt=u(ne,"P",{"data-svelte-h":!0}),g(gt)!=="svelte-10ugs3m"&&(gt.innerHTML=St),Ft=l(ne),_t=u(ne,"P",{"data-svelte-h":!0}),g(_t)!=="svelte-az7ywp"&&(_t.innerHTML=At),Pt=l(ne),Tt=u(ne,"P",{"data-svelte-h":!0}),g(Tt)!=="svelte-hswkmf"&&(Tt.innerHTML=Yt),Ut=l(ne),ye=u(ne,"DIV",{class:!0});var Fe=se(ye);T(ct.$$.fragment,Fe),zt=l(Fe),yt=u(Fe,"P",{"data-svelte-h":!0}),g(yt)!=="svelte-11lwf1s"&&(yt.innerHTML=Ot),It=l(Fe),T(tt.$$.fragment,Fe),Wt=l(Fe),T(ot.$$.fragment,Fe),Zt=l(Fe),T(nt.$$.fragment,Fe),Fe.forEach(d),ne.forEach(d),xt=l(a),T(pt.$$.fragment,a),Gt=l(a),ue=u(a,"DIV",{class:!0});var Pe=se(ue);T(ut.$$.fragment,Pe),Bt=l(Pe),bt=u(Pe,"P",{"data-svelte-h":!0}),g(bt)!=="svelte-1pit5xw"&&(bt.textContent=Dt),qt=l(Pe),Mt=u(Pe,"P",{"data-svelte-h":!0}),g(Mt)!=="svelte-az7ywp"&&(Mt.innerHTML=Kt),Vt=l(Pe),wt=u(Pe,"P",{"data-svelte-h":!0}),g(wt)!=="svelte-hswkmf"&&(wt.innerHTML=eo),Rt=l(Pe),Be=u(Pe,"DIV",{class:!0});var Le=se(Be);T(ht.$$.fragment,Le),Lt=l(Le),vt=u(Le,"P",{"data-svelte-h":!0}),g(vt)!=="svelte-10krurm"&&(vt.innerHTML=to),Ht=l(Le),T(st.$$.fragment,Le),Xt=l(Le),T(at.$$.fragment,Le),Le.forEach(d),Pe.forEach(d),this.h()},h(){ae(P,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(t,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(z,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(j,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(F,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(J,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(ye,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(A,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(Be,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(ue,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8")},m(a,v){y(e,a,v),c(a,h,v),c(a,t,v),y(s,t,null),i(t,m),i(t,o),i(t,Ue),i(t,I),i(t,ze),i(t,W),i(t,Ie),i(t,P),y(Me,P,null),i(P,We),i(P,Z),i(P,de),y(U,P,null),i(P,le),y(R,P,null),c(a,ie,v),y(Y,a,v),c(a,we,v),c(a,j,v),y(me,j,null),i(j,ee),i(j,E),i(j,ke),i(j,C),i(j,Ze),i(j,B),i(j,ce),i(j,z),y(O,z,null),i(z,Qe),i(z,D),i(z,Se),y(L,z,null),i(z,ge),y(te,z,null),c(a,Q,v),y(xe,a,v),c(a,Ge,v),c(a,J,v),y(H,J,null),i(J,q),i(J,V),i(J,_e),i(J,K),i(J,S),i(J,Te),i(J,x),i(J,F),y(N,F,null),i(F,X),i(F,oe),i(F,Oe),y(Je,F,null),i(F,n),y(f,F,null),i(F,pe),y(qe,F,null),c(a,kt,v),y(it,a,v),c(a,$t,v),c(a,A,v),y(dt,A,null),i(A,Ct),i(A,mt),i(A,Nt),i(A,ft),i(A,jt),i(A,gt),i(A,Ft),i(A,_t),i(A,Pt),i(A,Tt),i(A,Ut),i(A,ye),y(ct,ye,null),i(ye,zt),i(ye,yt),i(ye,It),y(tt,ye,null),i(ye,Wt),y(ot,ye,null),i(ye,Zt),y(nt,ye,null),c(a,xt,v),y(pt,a,v),c(a,Gt,v),c(a,ue,v),y(ut,ue,null),i(ue,Bt),i(ue,bt),i(ue,qt),i(ue,Mt),i(ue,Vt),i(ue,wt),i(ue,Rt),i(ue,Be),y(ht,Be,null),i(Be,Lt),i(Be,vt),i(Be,Ht),y(st,Be,null),i(Be,Xt),y(at,Be,null),Jt=!0},p(a,v){const Ve={};v&2&&(Ve.$$scope={dirty:v,ctx:a}),U.$set(Ve);const Ce={};v&2&&(Ce.$$scope={dirty:v,ctx:a}),R.$set(Ce);const Re={};v&2&&(Re.$$scope={dirty:v,ctx:a}),L.$set(Re);const Ne={};v&2&&(Ne.$$scope={dirty:v,ctx:a}),te.$set(Ne);const je={};v&2&&(je.$$scope={dirty:v,ctx:a}),Je.$set(je);const ne={};v&2&&(ne.$$scope={dirty:v,ctx:a}),f.$set(ne);const Fe={};v&2&&(Fe.$$scope={dirty:v,ctx:a}),qe.$set(Fe);const Pe={};v&2&&(Pe.$$scope={dirty:v,ctx:a}),tt.$set(Pe);const Le={};v&2&&(Le.$$scope={dirty:v,ctx:a}),ot.$set(Le);const oo={};v&2&&(oo.$$scope={dirty:v,ctx:a}),nt.$set(oo);const no={};v&2&&(no.$$scope={dirty:v,ctx:a}),st.$set(no);const so={};v&2&&(so.$$scope={dirty:v,ctx:a}),at.$set(so)},i(a){Jt||(b(e.$$.fragment,a),b(s.$$.fragment,a),b(Me.$$.fragment,a),b(U.$$.fragment,a),b(R.$$.fragment,a),b(Y.$$.fragment,a),b(me.$$.fragment,a),b(O.$$.fragment,a),b(L.$$.fragment,a),b(te.$$.fragment,a),b(xe.$$.fragment,a),b(H.$$.fragment,a),b(N.$$.fragment,a),b(Je.$$.fragment,a),b(f.$$.fragment,a),b(qe.$$.fragment,a),b(it.$$.fragment,a),b(dt.$$.fragment,a),b(ct.$$.fragment,a),b(tt.$$.fragment,a),b(ot.$$.fragment,a),b(nt.$$.fragment,a),b(pt.$$.fragment,a),b(ut.$$.fragment,a),b(ht.$$.fragment,a),b(st.$$.fragment,a),b(at.$$.fragment,a),Jt=!0)},o(a){M(e.$$.fragment,a),M(s.$$.fragment,a),M(Me.$$.fragment,a),M(U.$$.fragment,a),M(R.$$.fragment,a),M(Y.$$.fragment,a),M(me.$$.fragment,a),M(O.$$.fragment,a),M(L.$$.fragment,a),M(te.$$.fragment,a),M(xe.$$.fragment,a),M(H.$$.fragment,a),M(N.$$.fragment,a),M(Je.$$.fragment,a),M(f.$$.fragment,a),M(qe.$$.fragment,a),M(it.$$.fragment,a),M(dt.$$.fragment,a),M(ct.$$.fragment,a),M(tt.$$.fragment,a),M(ot.$$.fragment,a),M(nt.$$.fragment,a),M(pt.$$.fragment,a),M(ut.$$.fragment,a),M(ht.$$.fragment,a),M(st.$$.fragment,a),M(at.$$.fragment,a),Jt=!1},d(a){a&&(d(h),d(t),d(ie),d(we),d(j),d(Q),d(Ge),d(J),d(kt),d($t),d(A),d(xt),d(Gt),d(ue)),w(e,a),w(s),w(Me),w(U),w(R),w(Y,a),w(me),w(O),w(L),w(te),w(xe,a),w(H),w(N),w(Je),w(f),w(qe),w(it,a),w(dt),w(ct),w(tt),w(ot),w(nt),w(pt,a),w(ut),w(ht),w(st),w(at)}}}function Co(G){let e,h;return e=new ao({props:{$$slots:{default:[Jo]},$$scope:{ctx:G}}}),{c(){_(e.$$.fragment)},l(t){T(e.$$.fragment,t)},m(t,s){y(e,t,s),h=!0},p(t,s){const m={};s&2&&(m.$$scope={dirty:s,ctx:t}),e.$set(m)},i(t){h||(b(e.$$.fragment,t),h=!0)},o(t){M(e.$$.fragment,t),h=!1},d(t){w(e,t)}}}function No(G){let e,h=`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.`;return{c(){e=p("p"),e.innerHTML=h},l(t){e=u(t,"P",{"data-svelte-h":!0}),g(e)!=="svelte-fincs2"&&(e.innerHTML=h)},m(t,s){c(t,e,s)},p:re,d(t){t&&d(e)}}}function jo(G){let e,h="Example:",t,s,m;return s=new He({props:{code:"ZnJvbSUyMHRyYW5zZm9ybWVycyUyMGltcG9ydCUyMEF1dG9Ub2tlbml6ZXIlMkMlMjBGbGF4R1BUTmVvTW9kZWwlMEElMEF0b2tlbml6ZXIlMjAlM0QlMjBBdXRvVG9rZW5pemVyLmZyb21fcHJldHJhaW5lZCglMjJFbGV1dGhlckFJJTJGZ3B0LW5lby0xLjNCJTIyKSUwQW1vZGVsJTIwJTNEJTIwRmxheEdQVE5lb01vZGVsLmZyb21fcHJldHJhaW5lZCglMjJFbGV1dGhlckFJJTJGZ3B0LW5lby0xLjNCJTIyKSUwQSUwQWlucHV0cyUyMCUzRCUyMHRva2VuaXplciglMjJIZWxsbyUyQyUyMG15JTIwZG9nJTIwaXMlMjBjdXRlJTIyJTJDJTIwcmV0dXJuX3RlbnNvcnMlM0QlMjJqYXglMjIpJTBBb3V0cHV0cyUyMCUzRCUyMG1vZGVsKCoqaW5wdXRzKSUwQSUwQWxhc3RfaGlkZGVuX3N0YXRlcyUyMCUzRCUyMG91dHB1dHMubGFzdF9oaWRkZW5fc3RhdGU=",highlighted:`<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer, FlaxGPTNeoModel

<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = FlaxGPTNeoModel.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>inputs = tokenizer(<span class="hljs-string">&quot;Hello, my dog is cute&quot;</span>, return_tensors=<span class="hljs-string">&quot;jax&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>outputs = model(**inputs)

<span class="hljs-meta">&gt;&gt;&gt; </span>last_hidden_states = outputs.last_hidden_state`,wrap:!1}}),{c(){e=p("p"),e.textContent=h,t=r(),_(s.$$.fragment)},l(o){e=u(o,"P",{"data-svelte-h":!0}),g(e)!=="svelte-11lpom8"&&(e.textContent=h),t=l(o),T(s.$$.fragment,o)},m(o,$){c(o,e,$),c(o,t,$),y(s,o,$),m=!0},p:re,i(o){m||(b(s.$$.fragment,o),m=!0)},o(o){M(s.$$.fragment,o),m=!1},d(o){o&&(d(e),d(t)),w(s,o)}}}function Fo(G){let e,h=`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.`;return{c(){e=p("p"),e.innerHTML=h},l(t){e=u(t,"P",{"data-svelte-h":!0}),g(e)!=="svelte-fincs2"&&(e.innerHTML=h)},m(t,s){c(t,e,s)},p:re,d(t){t&&d(e)}}}function Po(G){let e,h="Example:",t,s,m;return s=new He({props:{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, FlaxGPTNeoForCausalLM

<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>model = FlaxGPTNeoForCausalLM.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>inputs = tokenizer(<span class="hljs-string">&quot;Hello, my dog is cute&quot;</span>, return_tensors=<span class="hljs-string">&quot;np&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>outputs = model(**inputs)

<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-comment"># retrieve logts for next token</span>
<span class="hljs-meta">&gt;&gt;&gt; </span>next_token_logits = outputs.logits[:, -<span class="hljs-number">1</span>]`,wrap:!1}}),{c(){e=p("p"),e.textContent=h,t=r(),_(s.$$.fragment)},l(o){e=u(o,"P",{"data-svelte-h":!0}),g(e)!=="svelte-11lpom8"&&(e.textContent=h),t=l(o),T(s.$$.fragment,o)},m(o,$){c(o,e,$),c(o,t,$),y(s,o,$),m=!0},p:re,i(o){m||(b(s.$$.fragment,o),m=!0)},o(o){M(s.$$.fragment,o),m=!1},d(o){o&&(d(e),d(t)),w(s,o)}}}function Uo(G){let e,h,t,s,m,o,$="The bare GPTNeo Model transformer outputting raw hidden-states without any specific head on top.",Ue,I,Ke=`This model inherits from <a href="/docs/transformers/pr_35674/en/main_classes/model#transformers.FlaxPreTrainedModel">FlaxPreTrainedModel</a>. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)`,ze,W,et=`This model is also a Flax Linen
<a href="https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html" rel="nofollow">flax.nn.Module</a> subclass. Use it as a
regular Flax Module and refer to the Flax documentation for all matter related to general usage and behavior.`,Ie,P,Me="Finally, this model supports inherent JAX features such as:",We,Z,Xe='<li><a href="https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit" rel="nofollow">Just-In-Time (JIT) compilation</a></li> <li><a href="https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation" rel="nofollow">Automatic Differentiation</a></li> <li><a href="https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap" rel="nofollow">Vectorization</a></li> <li><a href="https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap" rel="nofollow">Parallelization</a></li>',de,U,le,R,ie,Y="The <code>FlaxGPTNeoPreTrainedModel</code> forward method, overrides the <code>__call__</code> special method.",we,j,me,ee,E,ve,ke,C,$e,Ze,B,Ee=`The GPTNeo Model transformer with a language modeling head on top (linear layer with weights tied to the input
embeddings).`,ce,z,O=`This model inherits from <a href="/docs/transformers/pr_35674/en/main_classes/model#transformers.FlaxPreTrainedModel">FlaxPreTrainedModel</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.)`,Qe,D,fe=`This model is also a Flax Linen
<a href="https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html" rel="nofollow">flax.nn.Module</a> subclass. Use it as a
regular Flax Module and refer to the Flax documentation for all matter related to general usage and behavior.`,Se,L,ge="Finally, this model supports inherent JAX features such as:",te,Q,xe='<li><a href="https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit" rel="nofollow">Just-In-Time (JIT) compilation</a></li> <li><a href="https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation" rel="nofollow">Automatic Differentiation</a></li> <li><a href="https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap" rel="nofollow">Vectorization</a></li> <li><a href="https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap" rel="nofollow">Parallelization</a></li>',Ge,J,H,q,V,Ye="The <code>FlaxGPTNeoPreTrainedModel</code> forward method, overrides the <code>__call__</code> special method.",_e,K,Ae,S,Te;return e=new be({props:{title:"FlaxGPTNeoModel",local:"transformers.FlaxGPTNeoModel",headingTag:"h2"}}),s=new he({props:{name:"class transformers.FlaxGPTNeoModel",anchor:"transformers.FlaxGPTNeoModel",parameters:[{name:"config",val:": GPTNeoConfig"},{name:"input_shape",val:": typing.Tuple = (1, 1)"},{name:"seed",val:": int = 0"},{name:"dtype",val:": dtype = <class 'jax.numpy.float32'>"},{name:"_do_init",val:": bool = True"},{name:"**kwargs",val:""}],parametersDescription:[{anchor:"transformers.FlaxGPTNeoModel.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig">GPTNeoConfig</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_35674/en/main_classes/model#transformers.FlaxPreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"},{anchor:"transformers.FlaxGPTNeoModel.dtype",description:`<strong>dtype</strong> (<code>jax.numpy.dtype</code>, <em>optional</em>, defaults to <code>jax.numpy.float32</code>) &#x2014;
The data type of the computation. Can be one of <code>jax.numpy.float32</code>, <code>jax.numpy.float16</code> (on GPUs) and
<code>jax.numpy.bfloat16</code> (on TPUs).</p>
<p>This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
specified all the computation will be performed with the given <code>dtype</code>.</p>
<p><strong>Note that this only specifies the dtype of the computation and does not influence the dtype of model
parameters.</strong></p>
<p>If you wish to change the dtype of the model parameters, see <a href="/docs/transformers/pr_35674/en/main_classes/model#transformers.FlaxPreTrainedModel.to_fp16">to_fp16()</a> and
<a href="/docs/transformers/pr_35674/en/main_classes/model#transformers.FlaxPreTrainedModel.to_bf16">to_bf16()</a>.`,name:"dtype"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_flax_gpt_neo.py#L587"}}),le=new he({props:{name:"__call__",anchor:"transformers.FlaxGPTNeoModel.__call__",parameters:[{name:"input_ids",val:""},{name:"attention_mask",val:" = None"},{name:"position_ids",val:" = None"},{name:"params",val:": dict = None"},{name:"past_key_values",val:": dict = None"},{name:"dropout_rng",val:": <function PRNGKey at 0x7f65f3f1eb90> = None"},{name:"train",val:": bool = False"},{name:"output_attentions",val:": typing.Optional[bool] = None"},{name:"output_hidden_states",val:": typing.Optional[bool] = None"},{name:"return_dict",val:": typing.Optional[bool] = None"}],parametersDescription:[{anchor:"transformers.FlaxGPTNeoModel.__call__.input_ids",description:`<strong>input_ids</strong> (<code>numpy.ndarray</code> of shape <code>(batch_size, input_ids_length)</code>) &#x2014;
<code>input_ids_length</code> = <code>sequence_length</code>. Indices of input sequence tokens in the vocabulary.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_35674/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_35674/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_35674/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.FlaxGPTNeoModel.__call__.attention_mask",description:`<strong>attention_mask</strong> (<code>numpy.ndarray</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.FlaxGPTNeoModel.__call__.position_ids",description:`<strong>position_ids</strong> (<code>numpy.ndarray</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.max_position_embeddings - 1]</code>.`,name:"position_ids"},{anchor:"transformers.FlaxGPTNeoModel.__call__.past_key_values",description:`<strong>past_key_values</strong> (<code>Dict[str, np.ndarray]</code>, <em>optional</em>, returned by <code>init_cache</code> or when passing previous <code>past_key_values</code>) &#x2014;
Dictionary of pre-computed hidden-states (key and values in the attention blocks) that can be used for fast
auto-regressive decoding. Pre-computed key and value hidden-states are of shape <em>[batch_size, max_length]</em>.`,name:"past_key_values"},{anchor:"transformers.FlaxGPTNeoModel.__call__.output_attentions",description:`<strong>output_attentions</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the attentions tensors of all attention layers. See <code>attentions</code> under returned
tensors for more detail.`,name:"output_attentions"},{anchor:"transformers.FlaxGPTNeoModel.__call__.output_hidden_states",description:`<strong>output_hidden_states</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the hidden states of all layers. See <code>hidden_states</code> under returned tensors for
more detail.`,name:"output_hidden_states"},{anchor:"transformers.FlaxGPTNeoModel.__call__.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return a <a href="/docs/transformers/pr_35674/en/main_classes/output#transformers.utils.ModelOutput">ModelOutput</a> instead of a plain tuple.`,name:"return_dict"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_flax_gpt_neo.py#L401",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>A <a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_flax_outputs.FlaxBaseModelOutput"
>transformers.modeling_flax_outputs.FlaxBaseModelOutput</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_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig"
>GPTNeoConfig</a>) and inputs.</p>
<ul>
<li>
<p><strong>last_hidden_state</strong> (<code>jnp.ndarray</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>
</li>
<li>
<p><strong>hidden_states</strong> (<code>tuple(jnp.ndarray)</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>jnp.ndarray</code> (one for the output of the embeddings + 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 initial embedding outputs.</p>
</li>
<li>
<p><strong>attentions</strong> (<code>tuple(jnp.ndarray)</code>, <em>optional</em>, returned when <code>output_attentions=True</code> is passed or when <code>config.output_attentions=True</code>) — Tuple of <code>jnp.ndarray</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>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_flax_outputs.FlaxBaseModelOutput"
>transformers.modeling_flax_outputs.FlaxBaseModelOutput</a> or <code>tuple(torch.FloatTensor)</code></p>
`}}),j=new rt({props:{$$slots:{default:[No]},$$scope:{ctx:G}}}),ee=new De({props:{anchor:"transformers.FlaxGPTNeoModel.__call__.example",$$slots:{default:[jo]},$$scope:{ctx:G}}}),ve=new be({props:{title:"FlaxGPTNeoForCausalLM",local:"transformers.FlaxGPTNeoForCausalLM",headingTag:"h2"}}),$e=new he({props:{name:"class transformers.FlaxGPTNeoForCausalLM",anchor:"transformers.FlaxGPTNeoForCausalLM",parameters:[{name:"config",val:": GPTNeoConfig"},{name:"input_shape",val:": typing.Tuple = (1, 1)"},{name:"seed",val:": int = 0"},{name:"dtype",val:": dtype = <class 'jax.numpy.float32'>"},{name:"_do_init",val:": bool = True"},{name:"**kwargs",val:""}],parametersDescription:[{anchor:"transformers.FlaxGPTNeoForCausalLM.config",description:`<strong>config</strong> (<a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig">GPTNeoConfig</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_35674/en/main_classes/model#transformers.FlaxPreTrainedModel.from_pretrained">from_pretrained()</a> method to load the model weights.`,name:"config"},{anchor:"transformers.FlaxGPTNeoForCausalLM.dtype",description:`<strong>dtype</strong> (<code>jax.numpy.dtype</code>, <em>optional</em>, defaults to <code>jax.numpy.float32</code>) &#x2014;
The data type of the computation. Can be one of <code>jax.numpy.float32</code>, <code>jax.numpy.float16</code> (on GPUs) and
<code>jax.numpy.bfloat16</code> (on TPUs).</p>
<p>This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
specified all the computation will be performed with the given <code>dtype</code>.</p>
<p><strong>Note that this only specifies the dtype of the computation and does not influence the dtype of model
parameters.</strong></p>
<p>If you wish to change the dtype of the model parameters, see <a href="/docs/transformers/pr_35674/en/main_classes/model#transformers.FlaxPreTrainedModel.to_fp16">to_fp16()</a> and
<a href="/docs/transformers/pr_35674/en/main_classes/model#transformers.FlaxPreTrainedModel.to_bf16">to_bf16()</a>.`,name:"dtype"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_flax_gpt_neo.py#L647"}}),H=new he({props:{name:"__call__",anchor:"transformers.FlaxGPTNeoForCausalLM.__call__",parameters:[{name:"input_ids",val:""},{name:"attention_mask",val:" = None"},{name:"position_ids",val:" = None"},{name:"params",val:": dict = None"},{name:"past_key_values",val:": dict = None"},{name:"dropout_rng",val:": <function PRNGKey at 0x7f65f3f1eb90> = None"},{name:"train",val:": bool = False"},{name:"output_attentions",val:": typing.Optional[bool] = None"},{name:"output_hidden_states",val:": typing.Optional[bool] = None"},{name:"return_dict",val:": typing.Optional[bool] = None"}],parametersDescription:[{anchor:"transformers.FlaxGPTNeoForCausalLM.__call__.input_ids",description:`<strong>input_ids</strong> (<code>numpy.ndarray</code> of shape <code>(batch_size, input_ids_length)</code>) &#x2014;
<code>input_ids_length</code> = <code>sequence_length</code>. Indices of input sequence tokens in the vocabulary.</p>
<p>Indices can be obtained using <a href="/docs/transformers/pr_35674/en/model_doc/auto#transformers.AutoTokenizer">AutoTokenizer</a>. See <a href="/docs/transformers/pr_35674/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.encode">PreTrainedTokenizer.encode()</a> and
<a href="/docs/transformers/pr_35674/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.FlaxGPTNeoForCausalLM.__call__.attention_mask",description:`<strong>attention_mask</strong> (<code>numpy.ndarray</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.FlaxGPTNeoForCausalLM.__call__.position_ids",description:`<strong>position_ids</strong> (<code>numpy.ndarray</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.max_position_embeddings - 1]</code>.`,name:"position_ids"},{anchor:"transformers.FlaxGPTNeoForCausalLM.__call__.past_key_values",description:`<strong>past_key_values</strong> (<code>Dict[str, np.ndarray]</code>, <em>optional</em>, returned by <code>init_cache</code> or when passing previous <code>past_key_values</code>) &#x2014;
Dictionary of pre-computed hidden-states (key and values in the attention blocks) that can be used for fast
auto-regressive decoding. Pre-computed key and value hidden-states are of shape <em>[batch_size, max_length]</em>.`,name:"past_key_values"},{anchor:"transformers.FlaxGPTNeoForCausalLM.__call__.output_attentions",description:`<strong>output_attentions</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the attentions tensors of all attention layers. See <code>attentions</code> under returned
tensors for more detail.`,name:"output_attentions"},{anchor:"transformers.FlaxGPTNeoForCausalLM.__call__.output_hidden_states",description:`<strong>output_hidden_states</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return the hidden states of all layers. See <code>hidden_states</code> under returned tensors for
more detail.`,name:"output_hidden_states"},{anchor:"transformers.FlaxGPTNeoForCausalLM.__call__.return_dict",description:`<strong>return_dict</strong> (<code>bool</code>, <em>optional</em>) &#x2014;
Whether or not to return a <a href="/docs/transformers/pr_35674/en/main_classes/output#transformers.utils.ModelOutput">ModelOutput</a> instead of a plain tuple.`,name:"return_dict"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/modeling_flax_gpt_neo.py#L401",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>A <a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_flax_outputs.FlaxMaskedLMOutput"
>transformers.modeling_flax_outputs.FlaxMaskedLMOutput</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_35674/en/model_doc/gpt_neo#transformers.GPTNeoConfig"
>GPTNeoConfig</a>) and inputs.</p>
<ul>
<li>
<p><strong>logits</strong> (<code>jnp.ndarray</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>hidden_states</strong> (<code>tuple(jnp.ndarray)</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>jnp.ndarray</code> (one for the output of the embeddings + 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 initial embedding outputs.</p>
</li>
<li>
<p><strong>attentions</strong> (<code>tuple(jnp.ndarray)</code>, <em>optional</em>, returned when <code>output_attentions=True</code> is passed or when <code>config.output_attentions=True</code>) — Tuple of <code>jnp.ndarray</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>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><a
  href="/docs/transformers/pr_35674/en/main_classes/output#transformers.modeling_flax_outputs.FlaxMaskedLMOutput"
>transformers.modeling_flax_outputs.FlaxMaskedLMOutput</a> or <code>tuple(torch.FloatTensor)</code></p>
`}}),K=new rt({props:{$$slots:{default:[Fo]},$$scope:{ctx:G}}}),S=new De({props:{anchor:"transformers.FlaxGPTNeoForCausalLM.__call__.example",$$slots:{default:[Po]},$$scope:{ctx:G}}}),{c(){_(e.$$.fragment),h=r(),t=p("div"),_(s.$$.fragment),m=r(),o=p("p"),o.textContent=$,Ue=r(),I=p("p"),I.innerHTML=Ke,ze=r(),W=p("p"),W.innerHTML=et,Ie=r(),P=p("p"),P.textContent=Me,We=r(),Z=p("ul"),Z.innerHTML=Xe,de=r(),U=p("div"),_(le.$$.fragment),R=r(),ie=p("p"),ie.innerHTML=Y,we=r(),_(j.$$.fragment),me=r(),_(ee.$$.fragment),E=r(),_(ve.$$.fragment),ke=r(),C=p("div"),_($e.$$.fragment),Ze=r(),B=p("p"),B.textContent=Ee,ce=r(),z=p("p"),z.innerHTML=O,Qe=r(),D=p("p"),D.innerHTML=fe,Se=r(),L=p("p"),L.textContent=ge,te=r(),Q=p("ul"),Q.innerHTML=xe,Ge=r(),J=p("div"),_(H.$$.fragment),q=r(),V=p("p"),V.innerHTML=Ye,_e=r(),_(K.$$.fragment),Ae=r(),_(S.$$.fragment),this.h()},l(k){T(e.$$.fragment,k),h=l(k),t=u(k,"DIV",{class:!0});var x=se(t);T(s.$$.fragment,x),m=l(x),o=u(x,"P",{"data-svelte-h":!0}),g(o)!=="svelte-1k424oq"&&(o.textContent=$),Ue=l(x),I=u(x,"P",{"data-svelte-h":!0}),g(I)!=="svelte-19brmut"&&(I.innerHTML=Ke),ze=l(x),W=u(x,"P",{"data-svelte-h":!0}),g(W)!=="svelte-idybz1"&&(W.innerHTML=et),Ie=l(x),P=u(x,"P",{"data-svelte-h":!0}),g(P)!=="svelte-1pplc4a"&&(P.textContent=Me),We=l(x),Z=u(x,"UL",{"data-svelte-h":!0}),g(Z)!=="svelte-1w7z84m"&&(Z.innerHTML=Xe),de=l(x),U=u(x,"DIV",{class:!0});var F=se(U);T(le.$$.fragment,F),R=l(F),ie=u(F,"P",{"data-svelte-h":!0}),g(ie)!=="svelte-f61feg"&&(ie.innerHTML=Y),we=l(F),T(j.$$.fragment,F),me=l(F),T(ee.$$.fragment,F),F.forEach(d),x.forEach(d),E=l(k),T(ve.$$.fragment,k),ke=l(k),C=u(k,"DIV",{class:!0});var N=se(C);T($e.$$.fragment,N),Ze=l(N),B=u(N,"P",{"data-svelte-h":!0}),g(B)!=="svelte-1889ma8"&&(B.textContent=Ee),ce=l(N),z=u(N,"P",{"data-svelte-h":!0}),g(z)!=="svelte-19brmut"&&(z.innerHTML=O),Qe=l(N),D=u(N,"P",{"data-svelte-h":!0}),g(D)!=="svelte-idybz1"&&(D.innerHTML=fe),Se=l(N),L=u(N,"P",{"data-svelte-h":!0}),g(L)!=="svelte-1pplc4a"&&(L.textContent=ge),te=l(N),Q=u(N,"UL",{"data-svelte-h":!0}),g(Q)!=="svelte-1w7z84m"&&(Q.innerHTML=xe),Ge=l(N),J=u(N,"DIV",{class:!0});var X=se(J);T(H.$$.fragment,X),q=l(X),V=u(X,"P",{"data-svelte-h":!0}),g(V)!=="svelte-f61feg"&&(V.innerHTML=Ye),_e=l(X),T(K.$$.fragment,X),Ae=l(X),T(S.$$.fragment,X),X.forEach(d),N.forEach(d),this.h()},h(){ae(U,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(t,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(J,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),ae(C,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8")},m(k,x){y(e,k,x),c(k,h,x),c(k,t,x),y(s,t,null),i(t,m),i(t,o),i(t,Ue),i(t,I),i(t,ze),i(t,W),i(t,Ie),i(t,P),i(t,We),i(t,Z),i(t,de),i(t,U),y(le,U,null),i(U,R),i(U,ie),i(U,we),y(j,U,null),i(U,me),y(ee,U,null),c(k,E,x),y(ve,k,x),c(k,ke,x),c(k,C,x),y($e,C,null),i(C,Ze),i(C,B),i(C,ce),i(C,z),i(C,Qe),i(C,D),i(C,Se),i(C,L),i(C,te),i(C,Q),i(C,Ge),i(C,J),y(H,J,null),i(J,q),i(J,V),i(J,_e),y(K,J,null),i(J,Ae),y(S,J,null),Te=!0},p(k,x){const F={};x&2&&(F.$$scope={dirty:x,ctx:k}),j.$set(F);const N={};x&2&&(N.$$scope={dirty:x,ctx:k}),ee.$set(N);const X={};x&2&&(X.$$scope={dirty:x,ctx:k}),K.$set(X);const oe={};x&2&&(oe.$$scope={dirty:x,ctx:k}),S.$set(oe)},i(k){Te||(b(e.$$.fragment,k),b(s.$$.fragment,k),b(le.$$.fragment,k),b(j.$$.fragment,k),b(ee.$$.fragment,k),b(ve.$$.fragment,k),b($e.$$.fragment,k),b(H.$$.fragment,k),b(K.$$.fragment,k),b(S.$$.fragment,k),Te=!0)},o(k){M(e.$$.fragment,k),M(s.$$.fragment,k),M(le.$$.fragment,k),M(j.$$.fragment,k),M(ee.$$.fragment,k),M(ve.$$.fragment,k),M($e.$$.fragment,k),M(H.$$.fragment,k),M(K.$$.fragment,k),M(S.$$.fragment,k),Te=!1},d(k){k&&(d(h),d(t),d(E),d(ke),d(C)),w(e,k),w(s),w(le),w(j),w(ee),w(ve,k),w($e),w(H),w(K),w(S)}}}function zo(G){let e,h;return e=new ao({props:{$$slots:{default:[Uo]},$$scope:{ctx:G}}}),{c(){_(e.$$.fragment)},l(t){T(e.$$.fragment,t)},m(t,s){y(e,t,s),h=!0},p(t,s){const m={};s&2&&(m.$$scope={dirty:s,ctx:t}),e.$set(m)},i(t){h||(b(e.$$.fragment,t),h=!0)},o(t){M(e.$$.fragment,t),h=!1},d(t){w(e,t)}}}function Io(G){let e,h,t,s,m,o,$,Ue,I,Ke=`The GPTNeo model was released in the <a href="https://github.com/EleutherAI/gpt-neo" rel="nofollow">EleutherAI/gpt-neo</a> repository by Sid
Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy. It is a GPT2 like causal language model trained on the
<a href="https://pile.eleuther.ai/" rel="nofollow">Pile</a> dataset.`,ze,W,et=`The architecture is similar to GPT2 except that GPT Neo uses local attention in every other layer with a window size of
256 tokens.`,Ie,P,Me='This model was contributed by <a href="https://huggingface.co/valhalla" rel="nofollow">valhalla</a>.',We,Z,Xe,de,U="The <code>generate()</code> method can be used to generate text using GPT Neo model.",le,R,ie,Y,we,j,me='First, make sure to install the latest version of Flash Attention 2 to include the sliding window attention feature, and make sure your hardware is compatible with Flash-Attention 2. More details are available <a href="https://huggingface.co/docs/transformers/perf_infer_gpu_one#flashattention-2" rel="nofollow">here</a> concerning the installation.',ee,E,ve="Make sure as well to load your model in half-precision (e.g. <code>torch.float16</code>).",ke,C,$e="To load and run a model using Flash Attention 2, refer to the snippet below:",Ze,B,Ee,ce,z,O,Qe=`Below is an expected speedup diagram that compares pure inference time between the native implementation in transformers using <code>EleutherAI/gpt-neo-2.7B</code> checkpoint and the Flash Attention 2 version of the model.
Note that for GPT-Neo it is not possible to train / run on very long context as the max <a href="https://huggingface.co/EleutherAI/gpt-neo-2.7B/blob/main/config.json#L58" rel="nofollow">position embeddings</a> is limited to 2048 - but this is applicable to all gpt-neo models and not specific to FA-2`,D,fe,Se='<img src="https://user-images.githubusercontent.com/49240599/272241893-b1c66b75-3a48-4265-bc47-688448568b3d.png"/>',L,ge,te,Q,xe='<li><a href="../tasks/sequence_classification">Text classification task guide</a></li> <li><a href="../tasks/language_modeling">Causal language modeling task guide</a></li>',Ge,J,H,q,V,Ye,_e,K=`This is the configuration class to store the configuration of a <a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoModel">GPTNeoModel</a>. It is used to instantiate a GPT
Neo 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 GPTNeo
<a href="https://huggingface.co/EleutherAI/gpt-neo-1.3B" rel="nofollow">EleutherAI/gpt-neo-1.3B</a> architecture.`,Ae,S,Te=`Configuration objects inherit from <a href="/docs/transformers/pr_35674/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_35674/en/main_classes/configuration#transformers.PretrainedConfig">PretrainedConfig</a> for more information.`,k,x,F,N,X,oe,lt,Oe,Je;return m=new be({props:{title:"GPT Neo",local:"gpt-neo",headingTag:"h1"}}),$=new be({props:{title:"Overview",local:"overview",headingTag:"h2"}}),Z=new be({props:{title:"Usage example",local:"usage-example",headingTag:"h2"}}),R=new He({props:{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> GPTNeoForCausalLM, GPT2Tokenizer

<span class="hljs-meta">&gt;&gt;&gt; </span>model = GPTNeoForCausalLM.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = GPT2Tokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-1.3B&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>prompt = (
<span class="hljs-meta">... </span>    <span class="hljs-string">&quot;In a shocking finding, scientists discovered a herd of unicorns living in a remote, &quot;</span>
<span class="hljs-meta">... </span>    <span class="hljs-string">&quot;previously unexplored valley, in the Andes Mountains. Even more surprising to the &quot;</span>
<span class="hljs-meta">... </span>    <span class="hljs-string">&quot;researchers was the fact that the unicorns spoke perfect English.&quot;</span>
<span class="hljs-meta">... </span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>input_ids = tokenizer(prompt, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>).input_ids

<span class="hljs-meta">&gt;&gt;&gt; </span>gen_tokens = model.generate(
<span class="hljs-meta">... </span>    input_ids,
<span class="hljs-meta">... </span>    do_sample=<span class="hljs-literal">True</span>,
<span class="hljs-meta">... </span>    temperature=<span class="hljs-number">0.9</span>,
<span class="hljs-meta">... </span>    max_length=<span class="hljs-number">100</span>,
<span class="hljs-meta">... </span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>gen_text = tokenizer.batch_decode(gen_tokens)[<span class="hljs-number">0</span>]`,wrap:!1}}),Y=new be({props:{title:"Combining GPT-Neo and Flash Attention 2",local:"combining-gpt-neo-and-flash-attention-2",headingTag:"h2"}}),B=new He({props:{code:"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",highlighted:`<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">import</span> torch
<span class="hljs-meta">&gt;&gt;&gt; </span><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM, AutoTokenizer
<span class="hljs-meta">&gt;&gt;&gt; </span>device = <span class="hljs-string">&quot;cuda&quot;</span> <span class="hljs-comment"># the device to load the model onto</span>

<span class="hljs-meta">&gt;&gt;&gt; </span>model = AutoModelForCausalLM.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-2.7B&quot;</span>, torch_dtype=torch.float16, attn_implementation=<span class="hljs-string">&quot;flash_attention_2&quot;</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">&quot;EleutherAI/gpt-neo-2.7B&quot;</span>)

<span class="hljs-meta">&gt;&gt;&gt; </span>prompt = <span class="hljs-string">&quot;def hello_world():&quot;</span>

<span class="hljs-meta">&gt;&gt;&gt; </span>model_inputs = tokenizer([prompt], return_tensors=<span class="hljs-string">&quot;pt&quot;</span>).to(device)
<span class="hljs-meta">&gt;&gt;&gt; </span>model.to(device)

<span class="hljs-meta">&gt;&gt;&gt; </span>generated_ids = model.generate(**model_inputs, max_new_tokens=<span class="hljs-number">100</span>, do_sample=<span class="hljs-literal">True</span>)
<span class="hljs-meta">&gt;&gt;&gt; </span>tokenizer.batch_decode(generated_ids)[<span class="hljs-number">0</span>]
<span class="hljs-string">&quot;def hello_world():\\n    &gt;&gt;&gt; run_script(&quot;</span>hello.py<span class="hljs-string">&quot;)\\n    &gt;&gt;&gt; exit(0)\\n&lt;|endoftext|&gt;&quot;</span>`,wrap:!1}}),ce=new be({props:{title:"Expected speedups",local:"expected-speedups",headingTag:"h3"}}),ge=new be({props:{title:"Resources",local:"resources",headingTag:"h2"}}),J=new be({props:{title:"GPTNeoConfig",local:"transformers.GPTNeoConfig",headingTag:"h2"}}),V=new he({props:{name:"class transformers.GPTNeoConfig",anchor:"transformers.GPTNeoConfig",parameters:[{name:"vocab_size",val:" = 50257"},{name:"max_position_embeddings",val:" = 2048"},{name:"hidden_size",val:" = 2048"},{name:"num_layers",val:" = 24"},{name:"attention_types",val:" = [[['global', 'local'], 12]]"},{name:"num_heads",val:" = 16"},{name:"intermediate_size",val:" = None"},{name:"window_size",val:" = 256"},{name:"activation_function",val:" = 'gelu_new'"},{name:"resid_dropout",val:" = 0.0"},{name:"embed_dropout",val:" = 0.0"},{name:"attention_dropout",val:" = 0.0"},{name:"classifier_dropout",val:" = 0.1"},{name:"layer_norm_epsilon",val:" = 1e-05"},{name:"initializer_range",val:" = 0.02"},{name:"use_cache",val:" = True"},{name:"bos_token_id",val:" = 50256"},{name:"eos_token_id",val:" = 50256"},{name:"**kwargs",val:""}],parametersDescription:[{anchor:"transformers.GPTNeoConfig.vocab_size",description:`<strong>vocab_size</strong> (<code>int</code>, <em>optional</em>, defaults to 50257) &#x2014;
Vocabulary size of the GPT Neo model. Defines the number of different tokens that can be represented by the
<code>inputs_ids</code> passed when calling <a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoModel">GPTNeoModel</a>. Vocabulary size of the model. Defines the different
tokens that can be represented by the <em>inputs_ids</em> passed to the forward method of <a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoModel">GPTNeoModel</a>.`,name:"vocab_size"},{anchor:"transformers.GPTNeoConfig.max_position_embeddings",description:`<strong>max_position_embeddings</strong> (<code>int</code>, <em>optional</em>, defaults to 2048) &#x2014;
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).`,name:"max_position_embeddings"},{anchor:"transformers.GPTNeoConfig.hidden_size",description:`<strong>hidden_size</strong> (<code>int</code>, <em>optional</em>, defaults to 2048) &#x2014;
Dimensionality of the encoder layers and the pooler layer.`,name:"hidden_size"},{anchor:"transformers.GPTNeoConfig.num_layers",description:`<strong>num_layers</strong> (<code>int</code>, <em>optional</em>, defaults to 24) &#x2014;
Number of hidden layers in the Transformer encoder.`,name:"num_layers"},{anchor:"transformers.GPTNeoConfig.attention_types",description:`<strong>attention_types</strong> (<code>List</code>, <em>optional</em>, defaults to <code>[[[&apos;global&apos;, &apos;local&apos;], 12]]</code>) &#x2014;
The type of attention for each layer in a <code>List</code> of the following format <code>[[[&quot;attention_type&quot;], num_layerss]]</code> e.g. for a 24 layer model <code>[[[&quot;global&quot;], 24]]</code> or <code>[[[&quot;global&quot;, &quot;local&quot;], 12]]</code> Choose the
value of <code>attention_type</code> from <code>[&quot;global&quot;, &quot;local&quot;]</code>`,name:"attention_types"},{anchor:"transformers.GPTNeoConfig.num_heads",description:`<strong>num_heads</strong> (<code>int</code>, <em>optional</em>, defaults to 16) &#x2014;
Number of attention heads for each attention layer in the Transformer encoder.`,name:"num_heads"},{anchor:"transformers.GPTNeoConfig.intermediate_size",description:`<strong>intermediate_size</strong> (<code>int</code>, <em>optional</em>, defaults to 8192) &#x2014;
Dimensionality of the &#x201C;intermediate&#x201D; (i.e., feed-forward) layer in the Transformer encoder.`,name:"intermediate_size"},{anchor:"transformers.GPTNeoConfig.window_size",description:`<strong>window_size</strong> (<code>int</code>, <em>optional</em>, defaults to 256) &#x2014;
The size of the sliding window for local attention.`,name:"window_size"},{anchor:"transformers.GPTNeoConfig.activation_function",description:`<strong>activation_function</strong> (<code>str</code> or <code>function</code>, <em>optional</em>, defaults to <code>&quot;gelu_new&quot;</code>) &#x2014;
The non-linear activation function (function or string) in the encoder and pooler. If string, <code>&quot;gelu&quot;</code>,
<code>&quot;relu&quot;</code>, <code>&quot;selu&quot;</code> and <code>&quot;gelu_new&quot;</code> are supported.`,name:"activation_function"},{anchor:"transformers.GPTNeoConfig.resid_dropout",description:`<strong>resid_dropout</strong> (<code>float</code>, <em>optional</em>, defaults to 0.0) &#x2014;
Residual dropout used in the attention pattern.`,name:"resid_dropout"},{anchor:"transformers.GPTNeoConfig.embed_dropout",description:`<strong>embed_dropout</strong> (<code>float</code>, <em>optional</em>, defaults to 0.0) &#x2014;
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.`,name:"embed_dropout"},{anchor:"transformers.GPTNeoConfig.attention_dropout",description:`<strong>attention_dropout</strong> (<code>float</code>, <em>optional</em>, defaults to 0.0) &#x2014;
The dropout ratio for the attention probabilities.`,name:"attention_dropout"},{anchor:"transformers.GPTNeoConfig.classifier_dropout",description:`<strong>classifier_dropout</strong> (<code>float</code>, <em>optional</em>, defaults to 0.1) &#x2014;
Argument used when doing token classification, used in the model <a href="/docs/transformers/pr_35674/en/model_doc/gpt_neo#transformers.GPTNeoForTokenClassification">GPTNeoForTokenClassification</a>. The
dropout ratio for the hidden layer.`,name:"classifier_dropout"},{anchor:"transformers.GPTNeoConfig.layer_norm_epsilon",description:`<strong>layer_norm_epsilon</strong> (<code>float</code>, <em>optional</em>, defaults to 1e-05) &#x2014;
The epsilon used by the layer normalization layers.`,name:"layer_norm_epsilon"},{anchor:"transformers.GPTNeoConfig.initializer_range",description:`<strong>initializer_range</strong> (<code>float</code>, <em>optional</em>, defaults to 0.02) &#x2014;
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.`,name:"initializer_range"},{anchor:"transformers.GPTNeoConfig.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>.`,name:"use_cache"},{anchor:"transformers.GPTNeoConfig.bos_token_id",description:`<strong>bos_token_id</strong> (<code>int</code>, <em>optional</em>, defaults to 50256) &#x2014;
The id of the beginning of sentence token in the vocabulary.`,name:"bos_token_id"},{anchor:"transformers.GPTNeoConfig.eos_token_id",description:`<strong>eos_token_id</strong> (<code>int</code>, <em>optional</em>, defaults to 50256) &#x2014;
The id of the end of sentence token in the vocabulary.`,name:"eos_token_id"}],source:"https://github.com/huggingface/transformers/blob/vr_35674/src/transformers/models/gpt_neo/configuration_gpt_neo.py#L29"}}),x=new De({props:{anchor:"transformers.GPTNeoConfig.example",$$slots:{default:[fo]},$$scope:{ctx:G}}}),N=new ho({props:{pytorch:!0,tensorflow:!1,jax:!0,$$slots:{jax:[zo],pytorch:[Co]},$$scope:{ctx:G}}}),oe=new mo({props:{source:"https://github.com/huggingface/transformers/blob/main/docs/source/en/model_doc/gpt_neo.md"}}),{c(){e=p("meta"),h=r(),t=p("p"),s=r(),_(m.$$.fragment),o=r(),_($.$$.fragment),Ue=r(),I=p("p"),I.innerHTML=Ke,ze=r(),W=p("p"),W.textContent=et,Ie=r(),P=p("p"),P.innerHTML=Me,We=r(),_(Z.$$.fragment),Xe=r(),de=p("p"),de.innerHTML=U,le=r(),_(R.$$.fragment),ie=r(),_(Y.$$.fragment),we=r(),j=p("p"),j.innerHTML=me,ee=r(),E=p("p"),E.innerHTML=ve,ke=r(),C=p("p"),C.textContent=$e,Ze=r(),_(B.$$.fragment),Ee=r(),_(ce.$$.fragment),z=r(),O=p("p"),O.innerHTML=Qe,D=r(),fe=p("div"),fe.innerHTML=Se,L=r(),_(ge.$$.fragment),te=r(),Q=p("ul"),Q.innerHTML=xe,Ge=r(),_(J.$$.fragment),H=r(),q=p("div"),_(V.$$.fragment),Ye=r(),_e=p("p"),_e.innerHTML=K,Ae=r(),S=p("p"),S.innerHTML=Te,k=r(),_(x.$$.fragment),F=r(),_(N.$$.fragment),X=r(),_(oe.$$.fragment),lt=r(),Oe=p("p"),this.h()},l(n){const f=po("svelte-u9bgzb",document.head);e=u(f,"META",{name:!0,content:!0}),f.forEach(d),h=l(n),t=u(n,"P",{}),se(t).forEach(d),s=l(n),T(m.$$.fragment,n),o=l(n),T($.$$.fragment,n),Ue=l(n),I=u(n,"P",{"data-svelte-h":!0}),g(I)!=="svelte-1bj8qi3"&&(I.innerHTML=Ke),ze=l(n),W=u(n,"P",{"data-svelte-h":!0}),g(W)!=="svelte-1wf6p9c"&&(W.textContent=et),Ie=l(n),P=u(n,"P",{"data-svelte-h":!0}),g(P)!=="svelte-z2se1l"&&(P.innerHTML=Me),We=l(n),T(Z.$$.fragment,n),Xe=l(n),de=u(n,"P",{"data-svelte-h":!0}),g(de)!=="svelte-mhi95y"&&(de.innerHTML=U),le=l(n),T(R.$$.fragment,n),ie=l(n),T(Y.$$.fragment,n),we=l(n),j=u(n,"P",{"data-svelte-h":!0}),g(j)!=="svelte-i9dlab"&&(j.innerHTML=me),ee=l(n),E=u(n,"P",{"data-svelte-h":!0}),g(E)!=="svelte-1g1dfjg"&&(E.innerHTML=ve),ke=l(n),C=u(n,"P",{"data-svelte-h":!0}),g(C)!=="svelte-14hchid"&&(C.textContent=$e),Ze=l(n),T(B.$$.fragment,n),Ee=l(n),T(ce.$$.fragment,n),z=l(n),O=u(n,"P",{"data-svelte-h":!0}),g(O)!=="svelte-1r7180s"&&(O.innerHTML=Qe),D=l(n),fe=u(n,"DIV",{style:!0,"data-svelte-h":!0}),g(fe)!=="svelte-duh08u"&&(fe.innerHTML=Se),L=l(n),T(ge.$$.fragment,n),te=l(n),Q=u(n,"UL",{"data-svelte-h":!0}),g(Q)!=="svelte-17u5l9r"&&(Q.innerHTML=xe),Ge=l(n),T(J.$$.fragment,n),H=l(n),q=u(n,"DIV",{class:!0});var pe=se(q);T(V.$$.fragment,pe),Ye=l(pe),_e=u(pe,"P",{"data-svelte-h":!0}),g(_e)!=="svelte-jn0idi"&&(_e.innerHTML=K),Ae=l(pe),S=u(pe,"P",{"data-svelte-h":!0}),g(S)!=="svelte-f39fgb"&&(S.innerHTML=Te),k=l(pe),T(x.$$.fragment,pe),pe.forEach(d),F=l(n),T(N.$$.fragment,n),X=l(n),T(oe.$$.fragment,n),lt=l(n),Oe=u(n,"P",{}),se(Oe).forEach(d),this.h()},h(){ae(e,"name","hf:doc:metadata"),ae(e,"content",Wo),uo(fe,"text-align","center"),ae(q,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8")},m(n,f){i(document.head,e),c(n,h,f),c(n,t,f),c(n,s,f),y(m,n,f),c(n,o,f),y($,n,f),c(n,Ue,f),c(n,I,f),c(n,ze,f),c(n,W,f),c(n,Ie,f),c(n,P,f),c(n,We,f),y(Z,n,f),c(n,Xe,f),c(n,de,f),c(n,le,f),y(R,n,f),c(n,ie,f),y(Y,n,f),c(n,we,f),c(n,j,f),c(n,ee,f),c(n,E,f),c(n,ke,f),c(n,C,f),c(n,Ze,f),y(B,n,f),c(n,Ee,f),y(ce,n,f),c(n,z,f),c(n,O,f),c(n,D,f),c(n,fe,f),c(n,L,f),y(ge,n,f),c(n,te,f),c(n,Q,f),c(n,Ge,f),y(J,n,f),c(n,H,f),c(n,q,f),y(V,q,null),i(q,Ye),i(q,_e),i(q,Ae),i(q,S),i(q,k),y(x,q,null),c(n,F,f),y(N,n,f),c(n,X,f),y(oe,n,f),c(n,lt,f),c(n,Oe,f),Je=!0},p(n,[f]){const pe={};f&2&&(pe.$$scope={dirty:f,ctx:n}),x.$set(pe);const qe={};f&2&&(qe.$$scope={dirty:f,ctx:n}),N.$set(qe)},i(n){Je||(b(m.$$.fragment,n),b($.$$.fragment,n),b(Z.$$.fragment,n),b(R.$$.fragment,n),b(Y.$$.fragment,n),b(B.$$.fragment,n),b(ce.$$.fragment,n),b(ge.$$.fragment,n),b(J.$$.fragment,n),b(V.$$.fragment,n),b(x.$$.fragment,n),b(N.$$.fragment,n),b(oe.$$.fragment,n),Je=!0)},o(n){M(m.$$.fragment,n),M($.$$.fragment,n),M(Z.$$.fragment,n),M(R.$$.fragment,n),M(Y.$$.fragment,n),M(B.$$.fragment,n),M(ce.$$.fragment,n),M(ge.$$.fragment,n),M(J.$$.fragment,n),M(V.$$.fragment,n),M(x.$$.fragment,n),M(N.$$.fragment,n),M(oe.$$.fragment,n),Je=!1},d(n){n&&(d(h),d(t),d(s),d(o),d(Ue),d(I),d(ze),d(W),d(Ie),d(P),d(We),d(Xe),d(de),d(le),d(ie),d(we),d(j),d(ee),d(E),d(ke),d(C),d(Ze),d(Ee),d(z),d(O),d(D),d(fe),d(L),d(te),d(Q),d(Ge),d(H),d(q),d(F),d(X),d(lt),d(Oe)),d(e),w(m,n),w($,n),w(Z,n),w(R,n),w(Y,n),w(B,n),w(ce,n),w(ge,n),w(J,n),w(V),w(x),w(N,n),w(oe,n)}}}const Wo='{"title":"GPT Neo","local":"gpt-neo","sections":[{"title":"Overview","local":"overview","sections":[],"depth":2},{"title":"Usage example","local":"usage-example","sections":[],"depth":2},{"title":"Combining GPT-Neo and Flash Attention 2","local":"combining-gpt-neo-and-flash-attention-2","sections":[{"title":"Expected speedups","local":"expected-speedups","sections":[],"depth":3}],"depth":2},{"title":"Resources","local":"resources","sections":[],"depth":2},{"title":"GPTNeoConfig","local":"transformers.GPTNeoConfig","sections":[],"depth":2},{"title":"GPTNeoModel","local":"transformers.GPTNeoModel","sections":[],"depth":2},{"title":"GPTNeoForCausalLM","local":"transformers.GPTNeoForCausalLM","sections":[],"depth":2},{"title":"GPTNeoForQuestionAnswering","local":"transformers.GPTNeoForQuestionAnswering","sections":[],"depth":2},{"title":"GPTNeoForSequenceClassification","local":"transformers.GPTNeoForSequenceClassification","sections":[],"depth":2},{"title":"GPTNeoForTokenClassification","local":"transformers.GPTNeoForTokenClassification","sections":[],"depth":2},{"title":"FlaxGPTNeoModel","local":"transformers.FlaxGPTNeoModel","sections":[],"depth":2},{"title":"FlaxGPTNeoForCausalLM","local":"transformers.FlaxGPTNeoForCausalLM","sections":[],"depth":2}],"depth":1}';function Zo(G){return lo(()=>{new URLSearchParams(window.location.search).get("fw")}),[]}class Qo extends io{constructor(e){super(),co(this,e,Zo,Io,ro,{})}}export{Qo as component};