The goal of this blog post is to explain flash attention in such a way that hopefully anyone who already understands attention will ask themselves:

“Why didn’t I think of this before?” followed by “It’s so easy”.

https://gordicaleksa.medium.com/eli5-flash-attention-5c44017022ad

This repo shows you how to fine-tune an embedding model to improve RAG performance even if you don’t have labelled data (i.e. positive pairs of query/relevant documents).

We walkthrough step-by-step the process of generating a synthetic dataset with LLM, finetuning an opensource embedding model, and finally evaluating the finetuned model.

We experiment with a small scale dataset of financial PDF documents, and show that finetuning the embedding model can substantially improve retrieval performance.

https://github.com/run-llama/finetune-embedding

• Fastest Inference: mlc stands out as the fastest, prompting a need to assess its quality despite its impressive speed.
• Favorite Tool: CTranslate2 is the preferred choice due to its speed and user-friendliness, supported by excellent documentation. It lacks distributed inference unlike vLLM.
• vLLM Performance: vLLM is also fast but CTranslate outperforms it in speed. However, vLLM supports distributed inference, making it suitable for larger models.
• Text Generation Inference (TGI): An acceptable choice for deploying HuggingFace LLMs traditionally, but not as swift as vLLM. Offers features like telemetry and HF ecosystem integration. Note that TGI’s licensing has become more restrictive as of 7/28/2023, potentially limiting certain commercial uses.

https://cameronrwolfe.substack.com/p/advanced-prompt-engineering

The emergence of large language models (LLMs) has revolutionized problem-solving approaches. In the past, tasks like document reformatting or sentence classification necessitated creating specific computer programs. LLMs have transformed this process, enabling tasks to be accomplished through textual prompts. For instance, reformatting documents can be achieved by instructing an LLM. This shift was exemplified by GPT-3’s ability to achieve accurate results with minimal guidance.

As LLM research progressed, more sophisticated techniques emerged beyond basic prompting methods like zero/few-shot learning. Instruction-following LLMs (e.g., InstructGPT, ChatGPT) prompted investigations into tackling complex tasks. The goal was to extend LLMs beyond simple problems, requiring them to comprehend intricate instructions and execute multi-step reasoning. However, such challenges demand advanced prompting strategies due to their complexity.

https://lilianweng.github.io/posts/2023-01-10-inference-optimization/#quantization

Large transformer models are mainstream nowadays, creating SoTA results for a variety of tasks. They are powerful but very expensive to train and use. The extremely high inference cost, in both time and memory, is a big bottleneck for adopting a powerful transformer for solving real-world tasks at scale.

Why is it hard to run inference for large transformer models? Besides the increasing size of SoTA models, there are two main factors contributing to the inference challenge (Pope et al. 2022):

1. Large memory footprint. Both model parameters and intermediate states are needed in memory at inference time. For example,
   • The KV cache should be stored in memory during decoding time; E.g. For a batch size of 512 and context length of 2048, the KV cache totals 3TB, that is 3x the model size (!).
   • Inference cost from the attention mechanism scales quadratically with input sequence length.
2. Low parallelizability. Inference generation is executed in an autoregressive fashion, making the decoding process hard to parallel.

In this post, we will look into several approaches for making transformer inference more efficient. Some are general network compression methods, while others are specific to transformer architecture.

https://llm-attacks.org/

This research examines the safety of large language models (LLMs) such as ChatGPT, Bard, and Claude. It demonstrates the potential for automated creation of adversarial attacks, using character sequences added to user queries that manipulate the LLM into following harmful commands. Unlike traditional « jailbreaks, » these attacks are automated and can affect both open-source and closed-source chatbots. The study raises concerns about the effectiveness of mitigation measures and suggests that the challenges posed by adversarial behavior might persist due to the nature of deep learning models. The findings highlight the need for careful consideration of the safety implications as LLMs become more integrated into various applications.