Lauritz Rasbach
| lauritz.rasbach@gwdg.de |
Biography
Lauritz Rasbach is a PhD Candidate working on the analysis and optimization of generative AI. Using a whitebox approach, he researches the internal mechanisms of transformer models instead of treating them as a black box, with attention as his central subject. He aims to turn insight into a model's internals into more efficient systems, cutting computational demand, memory footprint and latency while preserving output quality.
Research Interests
- Analysis and optimization of generative AI
- Whitebox optimization of transformer models
- Attention mechanisms and sparse attention
- Efficient inference
Teaching
Open Thesis Topics
Speculative KV Caches for Efficient LLM InferenceApply
Key-Value caches are a critical component of Large Language Model inference, but they introduce significant memory and latency overheads. Speculative KV caches are techniques that predict and manage cache entries ahead of time, making them available when needed and improving inference efficiency. This thesis explores the design, implementation, and evaluation of such techniques for modern LLMs.
Adaptive Attention Budgeting for Efficient Large Language Model InferenceApply
This thesis investigates dynamic allocation of attention computation across transformer layers and attention heads based on input characteristics. The goal is to reduce inference cost while preserving model quality by learning or estimating per-input attention budgets.
Automatic Prompt Classification for Intelligent Language Model RoutingApply
On the ChatAI platform, the default model accounts for one of the highest proportions of user interactions, suggesting that many users rely on the preselected option rather than manually choosing a model. This thesis focuses on classifying user prompts into task categories (such as reasoning, coding, creative writing, summarization, and information retrieval). Based on the predicted task, requests are dynamically assigned to specialized models, with the goal of outperforming a single default-model approach in terms of accuracy, latency, and cost.