Log In to continue learning
Quantization Fundamentals
Beginner
1h14m
Topics
Generative Models
Compression and Quantization
MultiModal
Transformers
In this lesson, we'll take a look at what you get when you apply state-of-the-art quantization to large language models. For example, is it possible for quantization to help you with fine-tuning an LLM? Spoiler alert, the answer is yes. Let's see how. As seen in the previous lessons, quantizationis about compressing model weights in a certain manner. Quantization applied to large language models brought a lot of interest in the open-source AI community as effectively quantizing those models with minimal performance degradation can open up a lot of cool opportunities for anyone. Many groundbreaking papers came out in a short period of time, and we're naming just a few of them here. Starting from the summer of 2022, LLM.INT8 proposed a no-performance degradation 8-bit quantization method by decomposing the underlying matrix multiplication in two stages. To mitigate emergent features from LLMs at scale, the authors proposed to decompose the MatMul in two stages, the outlier part in float16 and the non-outlier part in int8. QLoRA proposed to make LLMs much more accessible by quantizing them in 4-bit precision and being able to fine-tune what we call low-rank adapters on top of the model. Don't worry, we'll explain that a bit later. Therefore, making fine-tuning LLMs much more accessible to anyone. On the other hand, AWQ, GPTQ, and also SmoothQuant proposed to pre-calibrate the model so that the quantized model does not get affected by large activations caused by large models. Later on, came out more methods that proved promising results for 2-bit precision, such as "Q sharp" [QuIP#], HQQ, and more recently, AQLM. All this amazing work, which all aims at focusing on making LLMs smaller and faster, are open-sourced, meaning that you can directly get your hands on the official implementation and try them out on your own. We cited just a few papers, but you can easily find many other papers that work on this specific topic. At the same time, you may also be wondering are these methods generalizable for all models? Among these methods, some of them require a calibration procedure, meaning you need to first pre-calibrate the model by iterating over a dataset and by minimizing an error, the quantization error, to get the best quantization parameters. As the original work for these methods is adapted to LLMs, you might need to tweak these methods on your own projects and use cases. However, some other methods do not require this step, meaning they can be used out of the box regardless of the modality, usually by replacing all instances of linear layers with a new quantized module, as we've been seeing in our lessons for the linear quantization. Quantization makes also LLMs easier to distribute, since they are smaller. For example, a 70 billion parameter model would need 280 GB of storage in full precision, whereras this can be further reduced to 40 GB if stored in 4-bit precision, leading to a 7x reduction. This makes loading these models much more affordable and opens up opportunities for loading these LLMs in local computers, using for example the GGUF format with "llama.cpp" In the Hugging Face ecosystem, you can also find some powerful quantized model distributors, such as The Bloke that you can see here, that distributes to the community quantized weights that most of the time require some pre-calibration that might be quite costly for anyone to run it, such as AWQ or GPTQ. You may be also wondering how all these quantization methods affect the model performance with respect to its original version. One way to evaluate this is to check the performance of the LLM on different well-known benchmarks specifically crafted for large language models. For that, you could check the OpenLLM leaderboard from Hugging Face, and the leaderboard now supports running evaluations on state-of-the-art quantized models such as LLM.int8, Qlora, and GPTQ. So for the model you are interested in, you could check its performance on the leaderboard and check if you are happy with it. So to wrap up, I would like to quickly cover another topic, which is fine-tuning LLMs. You may be wondering if it's possible to fine-tune a quantized model. Well, there are two cases where you might be interested in this scenario. The first case being where you would like to fine-tune a model while being quantized to get the best-quantized model possible. And the second case would be useful for people who would like to adapt their model for specific use cases and applications, such as fine-tuning an LLM on their own dataset. The first scenario would be doable through Quantization Aware Training. In this case, we train the model to be more accurate once we quantize it. Note, this method is not compatible with all the methods that we shared before, which belong to the category of Post Training Quantization techniques. So for the second use case, we would leverage PEFT methods, or parameter-efficient fine-tuning methods. So PEFT methods aim at drastically reducing the number of trainable parameters of a model, while trying to keep the same performance as full fine-tuning. So we will specifically deep dive into PEFT plus QLoRA, which leverages both quantization and PEFT methods. And you can also check this example on how to train an Llama 7B model on a three-tier Google Colab instance. So this is an animated diagram that shows you how low-rank adapters, or LoRA, work. When doing LoRa, you simply have to attach extra trainable parameters, the blue weights that you can see on the left, to a frozen weight, that you can see on the right. Since the r parameter is usually extremely small compared to the input hidden state dimension, the final optimizer states end up being extremely small, thus making the training protocol much more accessible. QLoRA leverages this by quanitizing the base weights, so the blue weights that you can see on the left, in 4-bit precision, and making sure the data type of the activation of the quantized weight matches the data type of the LoRA weights. That way, we could perform the sum that you can see here, easily, and we also get the best from both worlds, quantization and parameter-efficient fine-tuning methods, and unlock many cool opportunities, such as the ability to fine-tune LLMs on a free-tier Google Colab instance. So that's it for this lesson. I hope this gave you some good insights on the current status of state-of-the-art quantization with large language models and gave you a good overview of what you can achieve with these methods and how you can apply them depending on your use case. So in the next lesson, we will wrap up this course together and review what you have done during this course.
