AI

I've always used rough formulas to estimate the relationship between the scale of my models and the GPU VRAM consumption; after all, there are too many variables involved—model architecture, number of layers, attention mechanism implementation, sequence length, batch size, data precision used in training or inference... all of these affect our final calculation results.

I've been intermittently reading about a fine-tuning method called Kahneman-Tversky Optimization (KTO) from various sources like HuggingFace's official documents and other online materials. It's similar to DPO as a way to align models with human values, but KTO's data preparation format is much more convenient, so I'm quickly applying it to my current tasks before making time to study the detailed content in the related papers.

In the process of training and fine-tuning deep neural networks, the most important and scarce resource is undoubtedly the GPU's VRAM. Therefore, making every bit perform at its best is a critical task.

Problem Description

Recently, I've achieved some good application results by fine-tuning Gemma-2. However, I encountered various errors when deploying it on the client's equipment, which was quite frustrating. Currently, there isn't a systematic troubleshooting guide online, so I'm documenting it here.

When applying Large Language Models (LLMs) in real-world scenarios, it's often not just about letting the model generate text freely. We might want the model to return specific structures, such as multiple-choice questions or providing a rating. In such cases, transformers-based models can directly use the outlines tool.

Currently, LLM services cover a wide range of fields, and Prompt Injection and Jailbreak threats to LLMs are growing by the day. A few months ago, a customer service LLM even provided incorrect information, leading to a loss of customer rights (although that wasn't caused by a prompt attack).

Microsoft's open-source BIPIA (Benchmarking and Defending Against Indirect Prompt Injection Attacks on Large Language Models) evaluation method, although tested six months ago without significant updates since, remains a simple and convenient testing method for the tasks I have at hand.