bitsandbytes/pr_2032/en/_app/immutable/nodes/7.D04ux_ZK.js

import"../chunks/DsnmJJEf.js";import{i as T,h as U,C as _,H as t,a,E as B,s as C}from"../chunks/CHailA3B.js";import{p as j,o as F,s as e,f as R,a as m,b as A,c as g,n as Z}from"../chunks/CMkz9frW.js";const Q='{"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}';var q=g('<meta name="hf:doc:metadata"/>'),v=g('<p></p> <!> <!> <p>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.</p> <p>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.</p> <blockquote class="tip"><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></blockquote> <!> <p>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_2032/en/reference/nn/linear4bit#bitsandbytes.nn.Linear4bit">Linear4bit</a> and <a href="/docs/bitsandbytes/pr_2032/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_2032/en/reference/nn/linear4bit#bitsandbytes.nn.Linear4bit">Linear4bit</a> with such as bfloat16, float16 or float32.</p> <p>You’ll typically access and configure this option from <a href="https://huggingface.co/docs/transformers/main/en/main_classes/quantization#transformers.BitsAndBytesConfig" rel="nofollow">transformers.BitsAndBytesConfig</a> 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.</p> <blockquote class="tip"><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></blockquote> <!> <p>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.</p> <!> <blockquote class="tip"><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></blockquote> <p>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>.</p> <p>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.</p> <p>Before you begin, make sure you have the latest libraries installed.</p> <!> <p>The important change that enables FSDP-QLoRA training is the <code>bnb_4bit_quant_storage</code> parameter in the <a href="https://huggingface.co/docs/transformers/main/en/main_classes/quantization#transformers.BitsAndBytesConfig" rel="nofollow">BitsAndBytesConfig</a> class. This allows you to set the storage data type of the quantized weights to a float data type.</p> <!> <p>Pass the <a href="https://huggingface.co/docs/transformers/main/en/main_classes/quantization#transformers.BitsAndBytesConfig" rel="nofollow">BitsAndBytesConfig</a> 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_2032/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_2032/en/reference/nn/linear4bit#bitsandbytes.nn.Linear4bit">Linear4bit</a> layer is wrapped individually.</p> <!> <p>Configure the <a href="https://huggingface.co/docs/peft/main/en/package_reference/lora#peft.LoraConfig" rel="nofollow">LoraConfig</a> class for QLoRA training by setting <code>target_modules="all-linear"</code>.</p> <!> <p>Now you can pass everything to the <a href="https://huggingface.co/docs/trl/main/en/sft_trainer#trl.SFTTrainer" rel="nofollow">SFTTrainer</a> for training.</p> <!> <!> <p>To learn more about FSDP and QLoRA, check out the following resources:</p> <ul><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></ul> <!> <p></p>',1);function k(M,w){j(w,!1),F(()=>{new URLSearchParams(window.location.search).get("fw")}),T();var o=v();U("r7h6z8",f=>{var u=q();C(u,"content",Q),m(f,u)});var n=e(R(o),2);_(n,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var s=e(n,2);t(s,{title:"FSDP-QLoRA",local:"fsdp-qlora",headingTag:"h1"});var l=e(s,8);t(l,{title:"Quantized data storage",local:"quantized-data-storage",headingTag:"h2"});var r=e(l,8);a(r,{code:"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",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});var i=e(r,4);t(i,{title:"Training",local:"training",headingTag:"h2"});var d=e(i,10);a(d,{code:"cGlwJTIwaW5zdGFsbCUyMC1VJTIwYml0c2FuZGJ5dGVzJTIwYWNjZWxlcmF0ZSUyMHRyYW5zZm9ybWVycyUyMHBlZnQlMjB0cmw=",highlighted:"pip install -U bitsandbytes accelerate transformers peft trl",lang:"bash",wrap:!1});var c=e(d,4);a(c,{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});var p=e(c,4);a(p,{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});var h=e(p,4);a(h,{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});var b=e(h,4);a(b,{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});var y=e(b,2);t(y,{title:"Resources",local:"resources",headingTag:"h2"});var J=e(y,6);B(J,{source:"https://github.com/bitsandbytes-foundation/bitsandbytes/blob/main/docs/source/fsdp_qlora.md"}),Z(2),m(M,o),A()}export{k as component};