bitsandbytes/pr_1972/en/_app/immutable/nodes/7.84da6cc5.js

import{s as Ne,n as ze,o as We}from"../chunks/scheduler.8a2cc2fa.js";import{S as De,i as Ve,e as i,s as n,c as p,h as Oe,a as o,d as a,b as s,f as Ye,g as d,j as r,k as z,l as Ke,m as l,n as f,t as m,o as b,p as c}from"../chunks/index.7079e750.js";import{C as et,H as je,E as tt}from"../chunks/MermaidChart.svelte_svelte_type_style_lang.cb66a60e.js";import{C as X}from"../chunks/CodeBlock.7fce26df.js";function at(ve){let u,W,Y,D,g,V,w,O,T,Be='FSDP-QLoRA combines data parallelism (FSDP enables sharding model parameters, optimizer states, and gradients across GPUs), 4-bit quantization, and LoRA to train LLMs up to 70B parameters on a dual 24GB GPU system. This technique was released by <a href="https://www.answer.ai/posts/2024-03-06-fsdp-qlora" rel="nofollow">Answer.AI</a> in collaboration with bitsandbytes to make training LLMs more efficient and accessible for everyone.',K,J,Le="This guide provides a brief guide on how bitsandbytes supports storing quantized weights to enable FSDP-QLoRA, and how to run training with the Hugging Face libraries.",ee,y,Fe='<p>Other changes required for bitsandbytes to support FSDP-QLoRA, such as reconstructing the weights from the quantization metadata and preventing quantizing already quantized weights when they’re moved from a CPU to GPU, are documented in this <a href="https://github.com/bitsandbytes-foundation/bitsandbytes/pull/970" rel="nofollow">Pull Request</a> and described in the <a href="https://www.answer.ai/posts/2024-03-14-fsdp-qlora-deep-dive" rel="nofollow">Enabling 70B Finetuning on Consumer GPUs</a> blog post. We highly recommend reading these resources for a better understanding of FSDP-QLoRA!</p>',te,U,ae,_,Re='FSDP only supports sharding float data types which can be problematic because quantized weights are typically stored as integer data types (uint8). bitsandbytes doesn’t have this problem because it uses <code>StoreChar</code> to read and write quantized weights regardless of the data type storage. This makes it simple to add a <code>quant_storage</code> parameter to the <a href="/docs/bitsandbytes/pr_1972/en/reference/nn/linear4bit#bitsandbytes.nn.Linear4bit">Linear4bit</a> and <a href="/docs/bitsandbytes/pr_1972/en/reference/nn/linear4bit#bitsandbytes.nn.Params4bit">Params4bit</a> classes and set it to <code>torch.uint8</code> to maintain backward compatibility with the codebase. With the <code>quant_storage</code> parameter, you can select any of the FSDP supported data types to shard <a href="/docs/bitsandbytes/pr_1972/en/reference/nn/linear4bit#bitsandbytes.nn.Linear4bit">Linear4bit</a> with such as bfloat16, float16 or float32.',le,C,Ze="You’ll typically access and configure this option from <code>transformers.BitsAndBytesConfig</code> by setting the <code>bnb_4bit_quant_storage</code> parameter. It is very <strong>important</strong> the <code>quant_storage</code> data type matches the data types used throughout the model because FSDP can only wrap layers and modules that have the <em>same floating data type</em>. Making sure the data types are aligned will ensure the model is correctly sharded.",ne,h,Ae="<p>The <code>compute_dtype</code> is the data type used for computation inside the CUDA kernel, where the 4-bit quantized weights are unpacked from the data type in <code>quant_storage</code> and dequantized to <code>compute_dtype</code>. We recommend using torch.bfloat16 (if available on your hardware) for better numerical stability.</p>",se,$,ie,j,Qe='Check out this <a href="https://hf.co/docs/peft/main/en/accelerate/fsdp#use-peft-qlora-and-fsdp-for-finetuning-large-models-on-multiple-gpus" rel="nofollow">section</a> of the PEFT documentation for the config file and training code to run FSDP-QLoRA training.',oe,v,re,M,qe='<p>FSDP is a distributed training framework that needs to be launched as a distributed training job with a library like <a href="https://hf.co/docs/accelerate/index" rel="nofollow">Accelerate</a> or <a href="https://pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</a>. The launch command provided in this section uses Accelerate to launch the training script.</p>',pe,B,ke='bitsandbytes is deeply integrated with the Hugging Face ecosystem, making it easy to use with libraries like <a href="https://hf.co/docs/transformers" rel="nofollow">Transformers</a>, <a href="https://hf.co/docs/peft" rel="nofollow">PEFT</a>, and <a href="https://hf.co/docs/trl" rel="nofollow">TRL</a>.',de,L,Ee='PEFT provides a configuration file (<a href="https://github.com/huggingface/peft/blob/main/examples/sft/configs/fsdp_config_qlora.yaml" rel="nofollow">fsdp_config_qlora.yaml</a>), launch command (<a href="https://github.com/huggingface/peft/blob/main/examples/sft/run_peft_qlora_fsdp.sh" rel="nofollow">run_peft_qlora_fsdp.sh</a>), and training script (<a href="https://github.com/huggingface/peft/blob/main/examples/sft/train.py" rel="nofollow">train.py</a>) for running FSDP-QLoRA. To learn more, check out the <a href="https://huggingface.co/docs/peft/main/en/accelerate/fsdp#use-peft-qlora-and-fsdp-for-finetuning-large-models-on-multiple-gpus" rel="nofollow">Use PEFT QLoRA and FSDP for finetuning large models on multiple GPUs</a> documentation. This section briefly covers the steps to run FSDP-QLoRA training.',fe,F,xe="Before you begin, make sure you have the latest libraries installed.",me,R,be,Z,Ge="The important change that enables FSDP-QLoRA training is the <code>bnb_4bit_quant_storage</code> parameter in the <code>BitsAndBytesConfig</code> class. This allows you to set the storage data type of the quantized weights to a float data type.",ce,A,ue,Q,He='Pass the <code>BitsAndBytesConfig</code> to a model to set it up for FSDP-QLoRA. You should set the <code>torch_dtype</code> parameter to match <code>bnb_4bit_quant_storage</code> so that the <a href="/docs/bitsandbytes/pr_1972/en/reference/nn/linear4bit#bitsandbytes.nn.Linear4bit">Linear4bit</a> layers are wrapped identically to the <code>Linear</code> layers. If the storage types do not match, then each <a href="/docs/bitsandbytes/pr_1972/en/reference/nn/linear4bit#bitsandbytes.nn.Linear4bit">Linear4bit</a> layer is wrapped individually.',ye,q,he,k,Ie="Configure the <code>LoraConfig</code> class for QLoRA training by setting <code>target_modules=&quot;all-linear&quot;</code>.",Me,E,ge,x,Se="Now you can pass everything to the <code>SFTTrainer</code> for training.",we,G,Te,H,Je,I,Pe="To learn more about FSDP and QLoRA, check out the following resources:",Ue,S,Xe='<li>The <a href="https://github.com/AnswerDotAI/fsdp_qlora" rel="nofollow">AnswerDotAI/fsdp_qlora</a> repository.</li> <li>The introductory <a href="https://www.answer.ai/posts/2024-03-06-fsdp-qlora.html" rel="nofollow">You can now train a 70b language model at home</a> blog post by Answer.AI.</li> <li>For an introduction to FSDP, read the <a href="https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api" rel="nofollow">Introducing PyTorch Fully Sharded Data Parallel (FSDP) API</a> blog post.</li> <li>For more details about QLoRA, take a look at the <a href="https://huggingface.co/blog/4bit-transformers-bitsandbytes" rel="nofollow">Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA</a> blog post.</li>',_e,P,Ce,N,$e;return g=new et({props:{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"}}),w=new je({props:{title:"FSDP-QLoRA",local:"fsdp-qlora",headingTag:"h1"}}),U=new je({props:{title:"Quantized data storage",local:"quantized-data-storage",headingTag:"h2"}}),$=new X({props:{code:"ZnJvbSUyMHRyYW5zZm9ybWVycyUyMGltcG9ydCUyMEJpdHNBbmRCeXRlc0NvbmZpZyUyQyUyMEF1dG9Nb2RlbEZvckNhdXNhbExNJTBBJTBBYm5iX2NvbmZpZyUyMCUzRCUyMEJpdHNBbmRCeXRlc0NvbmZpZyglMEElMjAlMjAlMjAlMjBsb2FkX2luXzRiaXQlM0RUcnVlJTJDJTBBJTIwJTIwJTIwJTIwYm5iXzRiaXRfcXVhbnRfdHlwZSUzRCUyMm5mNCUyMiUyQyUwQSUyMCUyMCUyMCUyMGJuYl80Yml0X2NvbXB1dGVfZHR5cGUlM0R0b3JjaC5iZmxvYXQxNiUyQyUwQSUyMCUyMCUyMCUyMGJuYl80Yml0X3F1YW50X3N0b3JhZ2UlM0R0b3JjaC5iZmxvYXQxNiUyQyUwQSklMEElMEFtb2RlbCUyMCUzRCUyMEF1dG9Nb2RlbEZvckNhdXNhbExNLmZyb21fcHJldHJhaW5lZCglMEElMjAlMjAlMjAlMjAlMjJtZXRhLWxsYW1hJTJGTGxhbWEtMi03MGIlMjIlMkMlMEElMjAlMjAlMjAlMjBxdWFudGl6YXRpb25fY29uZmlnJTNEYm5iX2NvbmZpZyUyQyUwQSUyMCUyMCUyMCUyMHRvcmNoX2R0eXBlJTNEdG9yY2guYmZsb2F0MTYlMkMlMEEp",highlighted:`<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BitsAndBytesConfig, AutoModelForCausalLM

bnb_config = BitsAndBytesConfig(
    load_in_4bit=<span class="hljs-literal">True</span>,
    bnb_4bit_quant_type=<span class="hljs-string">&quot;nf4&quot;</span>,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_storage=torch.bfloat16,
)

model = AutoModelForCausalLM.from_pretrained(
    <span class="hljs-string">&quot;meta-llama/Llama-2-70b&quot;</span>,
    quantization_config=bnb_config,
    torch_dtype=torch.bfloat16,
)`,lang:"py",wrap:!1}}),v=new je({props:{title:"Training",local:"training",headingTag:"h2"}}),R=new X({props:{code:"cGlwJTIwaW5zdGFsbCUyMC1VJTIwYml0c2FuZGJ5dGVzJTIwYWNjZWxlcmF0ZSUyMHRyYW5zZm9ybWVycyUyMHBlZnQlMjB0cmw=",highlighted:"pip install -U bitsandbytes accelerate transformers peft trl",lang:"bash",wrap:!1}}),A=new X({props:{code:"ZnJvbSUyMHRyYW5zZm9ybWVycyUyMGltcG9ydCUyMEJpdHNBbmRCeXRlc0NvbmZpZyUwQSUwQWJuYl9jb25maWclMjAlM0QlMjBCaXRzQW5kQnl0ZXNDb25maWcoJTBBJTIwJTIwJTIwJTIwbG9hZF9pbl80Yml0JTNEVHJ1ZSUyQyUwQSUyMCUyMCUyMCUyMGJuYl80Yml0X3F1YW50X3R5cGUlM0QlMjJuZjQlMjIlMkMlMEElMjAlMjAlMjAlMjBibmJfNGJpdF9jb21wdXRlX2R0eXBlJTNEdG9yY2guYmZsb2F0MTYlMkMlMEElMjAlMjAlMjAlMjBibmJfNGJpdF91c2VfZG91YmxlX3F1YW50JTNEVHJ1ZSUyQyUwQSUyMCUyMCUyMCUyMGJuYl80Yml0X3F1YW50X3N0b3JhZ2UlM0R0b3JjaC5iZmxvYXQxNiUyQyUwQSk=",highlighted:`<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(
    load_in_4bit=<span class="hljs-literal">True</span>,
    bnb_4bit_quant_type=<span class="hljs-string">&quot;nf4&quot;</span>,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=<span class="hljs-literal">True</span>,
    bnb_4bit_quant_storage=torch.bfloat16,
)`,lang:"py",wrap:!1}}),q=new X({props:{code:"ZnJvbSUyMHRyYW5zZm9ybWVycyUyMGltcG9ydCUyMEF1dG9Nb2RlbEZvckNhdXNhbExNJTBBJTBBbW9kZWwlMjAlM0QlMjBBdXRvTW9kZWxGb3JDYXVzYWxMTS5mcm9tX3ByZXRyYWluZWQoJTBBJTIwJTIwJTIwJTIwJTIybWV0YS1sbGFtYSUyRkxsYW1hLTItNzBiJTIyJTJDJTBBJTIwJTIwJTIwJTIwcXVhbnRpemF0aW9uX2NvbmZpZyUzRGJuYl9jb25maWclMkMlMEElMjAlMjAlMjAlMjB0b3JjaF9kdHlwZSUzRHRvcmNoLmJmbG9hdDE2JTJDJTBBKQ==",highlighted:`<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    <span class="hljs-string">&quot;meta-llama/Llama-2-70b&quot;</span>,
    quantization_config=bnb_config,
    torch_dtype=torch.bfloat16,
)`,lang:"py",wrap:!1}}),E=new X({props:{code:"ZnJvbSUyMHBlZnQlMjBpbXBvcnQlMjBMb3JhQ29uZmlnJTBBJTBBcGVmdF9jb25maWclMjAlM0QlMjBMb3JhQ29uZmlnKCUwQSUyMCUyMCUyMCUyMGxvcmFfYWxwaGElM0QxNiUyQyUwQSUyMCUyMCUyMCUyMGxvcmFfZHJvcG91dCUzRDAuMSUyQyUwQSUyMCUyMCUyMCUyMHIlM0Q2NCUyQyUwQSUyMCUyMCUyMCUyMGJpYXMlM0QlMjJub25lJTIyJTJDJTBBJTIwJTIwJTIwJTIwdGFza190eXBlJTNEJTIyQ0FVU0FMX0xNJTIyJTJDJTBBJTIwJTIwJTIwJTIwdGFyZ2V0X21vZHVsZXMlM0QlMjJhbGwtbGluZWFyJTIyJTJDJTBBKQ==",highlighted:`<span class="hljs-keyword">from</span> peft <span class="hljs-keyword">import</span> LoraConfig

peft_config = LoraConfig(
    lora_alpha=<span class="hljs-number">16</span>,
    lora_dropout=<span class="hljs-number">0.1</span>,
    r=<span class="hljs-number">64</span>,
    bias=<span class="hljs-string">&quot;none&quot;</span>,
    task_type=<span class="hljs-string">&quot;CAUSAL_LM&quot;</span>,
    target_modules=<span class="hljs-string">&quot;all-linear&quot;</span>,
)`,lang:"py",wrap:!1}}),G=new X({props:{code:"ZnJvbSUyMHRybCUyMGltcG9ydCUyMFNGVFRyYWluZXIlMEElMEF0cmFpbmVyJTIwJTNEJTIwU0ZUVHJhaW5lciglMEElMjAlMjAlMjAlMjBtb2RlbCUzRG1vZGVsJTJDJTBBJTIwJTIwJTIwJTIwdHJhaW5fZGF0YXNldCUzRGRhdGFzZXQlMkMlMEElMjAlMjAlMjAlMjBwZWZ0X2NvbmZpZyUzRHBlZnRfY29uZmlnJTJDJTBBJTIwJTIwJTIwJTIwcHJvY2Vzc2luZ19jbGFzcyUzRHRva2VuaXplciUyQyUwQSUyMCUyMCUyMCUyMGFyZ3MlM0R0cmFpbmluZ19hcmd1bWVudHMlMkMlMEEpJTBBdHJhaW5lci50cmFpbigp",highlighted:`<span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> SFTTrainer

trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    peft_config=peft_config,
    processing_class=tokenizer,
    args=training_arguments,
)
trainer.train()`,lang:"py",wrap:!1}}),H=new je({props:{title:"Resources",local:"resources",headingTag:"h2"}}),P=new tt({props:{source:"https://github.com/bitsandbytes-foundation/bitsandbytes/blob/main/docs/source/fsdp_qlora.md"}}),{c(){u=i("meta"),W=n(),Y=i("p"),D=n(),p(g.$$.fragment),V=n(),p(w.$$.fragment),O=n(),T=i("p"),T.innerHTML=Be,K=n(),J=i("p"),J.textContent=Le,ee=n(),y=i("blockquote"),y.innerHTML=Fe,te=n(),p(U.$$.fragment),ae=n(),_=i("p"),_.innerHTML=Re,le=n(),C=i("p"),C.innerHTML=Ze,ne=n(),h=i("blockquote"),h.innerHTML=Ae,se=n(),p($.$$.fragment),ie=n(),j=i("p"),j.innerHTML=Qe,oe=n(),p(v.$$.fragment),re=n(),M=i("blockquote"),M.innerHTML=qe,pe=n(),B=i("p"),B.innerHTML=ke,de=n(),L=i("p"),L.innerHTML=Ee,fe=n(),F=i("p"),F.textContent=xe,me=n(),p(R.$$.fragment),be=n(),Z=i("p"),Z.innerHTML=Ge,ce=n(),p(A.$$.fragment),ue=n(),Q=i("p"),Q.innerHTML=He,ye=n(),p(q.$$.fragment),he=n(),k=i("p"),k.innerHTML=Ie,Me=n(),p(E.$$.fragment),ge=n(),x=i("p"),x.innerHTML=Se,we=n(),p(G.$$.fragment),Te=n(),p(H.$$.fragment),Je=n(),I=i("p"),I.textContent=Pe,Ue=n(),S=i("ul"),S.innerHTML=Xe,_e=n(),p(P.$$.fragment),Ce=n(),N=i("p"),this.h()},l(e){const t=Oe("svelte-u9bgzb",document.head);u=o(t,"META",{name:!0,content:!0}),t.forEach(a),W=s(e),Y=o(e,"P",{}),Ye(Y).forEach(a),D=s(e),d(g.$$.fragment,e),V=s(e),d(w.$$.fragment,e),O=s(e),T=o(e,"P",{"data-svelte-h":!0}),r(T)!=="svelte-5l5q4j"&&(T.innerHTML=Be),K=s(e),J=o(e,"P",{"data-svelte-h":!0}),r(J)!=="svelte-1naqdkg"&&(J.textContent=Le),ee=s(e),y=o(e,"BLOCKQUOTE",{class:!0,"data-svelte-h":!0}),r(y)!=="svelte-1bablpl"&&(y.innerHTML=Fe),te=s(e),d(U.$$.fragment,e),ae=s(e),_=o(e,"P",{"data-svelte-h":!0}),r(_)!=="svelte-1hnrguw"&&(_.innerHTML=Re),le=s(e),C=o(e,"P",{"data-svelte-h":!0}),r(C)!=="svelte-1chmkhe"&&(C.innerHTML=Ze),ne=s(e),h=o(e,"BLOCKQUOTE",{class:!0,"data-svelte-h":!0}),r(h)!=="svelte-5h53j4"&&(h.innerHTML=Ae),se=s(e),d($.$$.fragment,e),ie=s(e),j=o(e,"P",{"data-svelte-h":!0}),r(j)!=="svelte-1mcsrsf"&&(j.innerHTML=Qe),oe=s(e),d(v.$$.fragment,e),re=s(e),M=o(e,"BLOCKQUOTE",{class:!0,"data-svelte-h":!0}),r(M)!=="svelte-1nwuwp3"&&(M.innerHTML=qe),pe=s(e),B=o(e,"P",{"data-svelte-h":!0}),r(B)!=="svelte-11niy1x"&&(B.innerHTML=ke),de=s(e),L=o(e,"P",{"data-svelte-h":!0}),r(L)!=="svelte-eapz72"&&(L.innerHTML=Ee),fe=s(e),F=o(e,"P",{"data-svelte-h":!0}),r(F)!=="svelte-1x35k4g"&&(F.textContent=xe),me=s(e),d(R.$$.fragment,e),be=s(e),Z=o(e,"P",{"data-svelte-h":!0}),r(Z)!=="svelte-b8kirt"&&(Z.innerHTML=Ge),ce=s(e),d(A.$$.fragment,e),ue=s(e),Q=o(e,"P",{"data-svelte-h":!0}),r(Q)!=="svelte-obwwv2"&&(Q.innerHTML=He),ye=s(e),d(q.$$.fragment,e),he=s(e),k=o(e,"P",{"data-svelte-h":!0}),r(k)!=="svelte-1u3tuf"&&(k.innerHTML=Ie),Me=s(e),d(E.$$.fragment,e),ge=s(e),x=o(e,"P",{"data-svelte-h":!0}),r(x)!=="svelte-x2x2qz"&&(x.innerHTML=Se),we=s(e),d(G.$$.fragment,e),Te=s(e),d(H.$$.fragment,e),Je=s(e),I=o(e,"P",{"data-svelte-h":!0}),r(I)!=="svelte-ijbsql"&&(I.textContent=Pe),Ue=s(e),S=o(e,"UL",{"data-svelte-h":!0}),r(S)!=="svelte-47gauf"&&(S.innerHTML=Xe),_e=s(e),d(P.$$.fragment,e),Ce=s(e),N=o(e,"P",{}),Ye(N).forEach(a),this.h()},h(){z(u,"name","hf:doc:metadata"),z(u,"content",lt),z(y,"class","tip"),z(h,"class","tip"),z(M,"class","tip")},m(e,t){Ke(document.head,u),l(e,W,t),l(e,Y,t),l(e,D,t),f(g,e,t),l(e,V,t),f(w,e,t),l(e,O,t),l(e,T,t),l(e,K,t),l(e,J,t),l(e,ee,t),l(e,y,t),l(e,te,t),f(U,e,t),l(e,ae,t),l(e,_,t),l(e,le,t),l(e,C,t),l(e,ne,t),l(e,h,t),l(e,se,t),f($,e,t),l(e,ie,t),l(e,j,t),l(e,oe,t),f(v,e,t),l(e,re,t),l(e,M,t),l(e,pe,t),l(e,B,t),l(e,de,t),l(e,L,t),l(e,fe,t),l(e,F,t),l(e,me,t),f(R,e,t),l(e,be,t),l(e,Z,t),l(e,ce,t),f(A,e,t),l(e,ue,t),l(e,Q,t),l(e,ye,t),f(q,e,t),l(e,he,t),l(e,k,t),l(e,Me,t),f(E,e,t),l(e,ge,t),l(e,x,t),l(e,we,t),f(G,e,t),l(e,Te,t),f(H,e,t),l(e,Je,t),l(e,I,t),l(e,Ue,t),l(e,S,t),l(e,_e,t),f(P,e,t),l(e,Ce,t),l(e,N,t),$e=!0},p:ze,i(e){$e||(m(g.$$.fragment,e),m(w.$$.fragment,e),m(U.$$.fragment,e),m($.$$.fragment,e),m(v.$$.fragment,e),m(R.$$.fragment,e),m(A.$$.fragment,e),m(q.$$.fragment,e),m(E.$$.fragment,e),m(G.$$.fragment,e),m(H.$$.fragment,e),m(P.$$.fragment,e),$e=!0)},o(e){b(g.$$.fragment,e),b(w.$$.fragment,e),b(U.$$.fragment,e),b($.$$.fragment,e),b(v.$$.fragment,e),b(R.$$.fragment,e),b(A.$$.fragment,e),b(q.$$.fragment,e),b(E.$$.fragment,e),b(G.$$.fragment,e),b(H.$$.fragment,e),b(P.$$.fragment,e),$e=!1},d(e){e&&(a(W),a(Y),a(D),a(V),a(O),a(T),a(K),a(J),a(ee),a(y),a(te),a(ae),a(_),a(le),a(C),a(ne),a(h),a(se),a(ie),a(j),a(oe),a(re),a(M),a(pe),a(B),a(de),a(L),a(fe),a(F),a(me),a(be),a(Z),a(ce),a(ue),a(Q),a(ye),a(he),a(k),a(Me),a(ge),a(x),a(we),a(Te),a(Je),a(I),a(Ue),a(S),a(_e),a(Ce),a(N)),a(u),c(g,e),c(w,e),c(U,e),c($,e),c(v,e),c(R,e),c(A,e),c(q,e),c(E,e),c(G,e),c(H,e),c(P,e)}}}const lt='{"title":"FSDP-QLoRA","local":"fsdp-qlora","sections":[{"title":"Quantized data storage","local":"quantized-data-storage","sections":[],"depth":2},{"title":"Training","local":"training","sections":[],"depth":2},{"title":"Resources","local":"resources","sections":[],"depth":2}],"depth":1}';function nt(ve){return We(()=>{new URLSearchParams(window.location.search).get("fw")}),[]}class pt extends De{constructor(u){super(),Ve(this,u,nt,at,Ne,{})}}export{pt as component};