Quantizing LLMs Step-by-Step: Changing FP16 Fashions to GGUF

On this article, you’ll learn the way quantization shrinks massive language fashions and the way to convert an FP16 checkpoint into an environment friendly GGUF file you possibly can share and run regionally.

Matters we’ll cowl embody:

• What precision sorts (FP32, FP16, 8-bit, 4-bit) imply for mannequin dimension and pace
• Tips on how to use huggingface_hub to fetch a mannequin and authenticate
• Tips on how to convert to GGUF with llama.cpp and add the outcome to Hugging Face

And away we go.

Quantizing LLMs Step-by-Step: Changing FP16 Fashions to GGUF
Picture by Writer

Introduction

Massive language fashions like LLaMA, Mistral, and Qwen have billions of parameters that demand a number of reminiscence and compute energy. For instance, operating LLaMA 7B in full precision can require over 12 GB of VRAM, making it impractical for a lot of customers. You’ll be able to verify the small print on this Hugging Face dialogue. Don’t fear about what “full precision” means but; we’ll break it down quickly. The principle concept is that this: these fashions are too large to run on commonplace {hardware} with out assist. Quantization is that assist.

Quantization permits unbiased researchers and hobbyists to run massive fashions on private computer systems by shrinking the dimensions of the mannequin with out severely impacting efficiency. On this information, we’ll discover how quantization works, what completely different precision codecs imply, after which stroll by way of quantizing a pattern FP16 mannequin right into a GGUF format and importing it to Hugging Face.

What Is Quantization?

At a really fundamental degree, quantization is about making a mannequin smaller with out breaking it. Massive language fashions are made up of billions of numerical values referred to as weights. These numbers management how strongly completely different elements of the community affect one another when producing an output. By default, these weights are saved utilizing high-precision codecs resembling FP32 or FP16, which implies each quantity takes up a number of reminiscence, and when you have got billions of them, issues get out of hand in a short time. Take a single quantity like 2.31384. In FP32, that one quantity alone makes use of 32 bits of reminiscence. Now think about storing billions of numbers like that. Because of this a 7B mannequin can simply take round 28 GB in FP32 and about 14 GB even in FP16. For many laptops and GPUs, that’s already an excessive amount of.

Quantization fixes this by saying: we don’t really need that a lot precision anymore. As a substitute of storing 2.31384 precisely, we retailer one thing near it utilizing fewer bits. Possibly it turns into 2.3 or a close-by integer worth below the hood. The quantity is barely much less correct, however the mannequin nonetheless behaves the identical in follow. Neural networks can tolerate these small errors as a result of the ultimate output depends upon billions of calculations, not a single quantity. Small variations common out, very similar to picture compression reduces file dimension with out ruining how the picture appears. However the payoff is large. A mannequin that wants 14 GB in FP16 can typically run in about 7 GB with 8-bit quantization, and even round 4 GB with 4-bit quantization. That is what makes it potential to run massive language fashions regionally as a substitute of counting on costly servers.

After quantizing, we frequently retailer the mannequin in a unified file format. One fashionable format is GGUF, created by Georgi Gerganov (creator of llama.cpp). GGUF is a single-file format that features each the quantized weights and helpful metadata. It’s optimized for fast loading and inference on CPUs or different light-weight runtimes. GGUF additionally helps a number of quantization sorts (like Q4_0, Q8_0) and works nicely on CPUs and low-end GPUs. Hopefully, this clarifies each the idea and the motivation behind quantization. Now let’s transfer on to writing some code.

Step-by-Step: Quantizing a Mannequin to GGUF

1. Putting in Dependencies and Logging to Hugging Face

Earlier than downloading or changing any mannequin, we have to set up the required Python packages and authenticate with Hugging Face. We’ll use huggingface_hub, Transformers, and SentencePiece. This ensures we are able to entry public or gated fashions with out errors:

2. Downloading a Pre-trained Mannequin

We’ll decide a small FP16 mannequin from Hugging Face. Right here we use TinyLlama 1.1B, which is sufficiently small to run in Colab however nonetheless provides an excellent demonstration. Utilizing Python, we are able to obtain it with huggingface_hub:

This command saves the mannequin recordsdata into the model_folder listing. You’ll be able to substitute model_id with any Hugging Face mannequin ID that you just wish to quantize. (If wanted, you too can use AutoModel.from_pretrained with torch.float16 to load it first, however snapshot_download is easy for grabbing the recordsdata.)

3. Setting Up the Conversion Instruments

Subsequent, we clone the llama.cpp repository, which accommodates the conversion scripts. In Colab:

This provides you entry to convert_hf_to_gguf.py. The Python necessities guarantee you have got all wanted libraries to run the script.

4. Changing the Mannequin to GGUF with Quantization

Now, run the conversion script, specifying the enter folder, output filename, and quantization sort. We’ll use q8_0 (8-bit quantization). This may roughly halve the reminiscence footprint of the mannequin:

Right here /content material/model_folder is the place we downloaded the mannequin, /content material/tinyllama-1.1b-chat.Q8_0.gguf is the output GGUF file, and the --outtype q8_0 flag means “quantize to 8-bit.” The script hundreds the FP16 weights, converts them into 8-bit values, and writes a single GGUF file. This file is now a lot smaller and prepared for inference with GGUF-compatible instruments.

You’ll be able to confirm the output:

You must see a file a couple of GB in dimension, decreased from the unique FP16 mannequin.

5. Importing the Quantized Mannequin to Hugging Face

Lastly, you possibly can publish the GGUF mannequin so others can simply obtain and use it utilizing the huggingface_hub Python library:

This creates a brand new repository (if it doesn’t exist) and uploads your quantized GGUF file. Anybody can now load it with llama.cpp, llama-cpp-python, or Ollama. You’ll be able to entry the quantized GGUF file that we created right here.

Wrapping Up

By following the steps above, you possibly can take any supported Hugging Face mannequin, quantize it (e.g. to 4-bit or 8-bit), and reserve it as GGUF. Then push it to Hugging Face to share or deploy. This makes it simpler than ever to compress and use massive language fashions on on a regular basis {hardware}.