I am a Member of Technical Staff at Black Forest Labs, specializing in optimizing the training and inference efficiency of Large Language Models (LLMs) and Large Vision Models (LVMs).

Previously, I worked at Databricks Mosaic AI, where I played a key role in developing the DBRX model by optimizing memory utilization, computational efficiency, and communication strategies during training to achieve state-of-the-art performance (Building DBRX-class Custom LLMs with Mosaic AI Training). I collaborated with NVIDIA to resolve FP8 training challenges in TransformerEngine, enabling FP8 training for Mosaic AI models. Additionally, I led the technical effort to optimize the inference of Llama and DBRX models.

Prior to Databricks, I was part of Microsoft DeepSpeed, where I enhanced the performance and usability of LLMs in production systems such as GitHub Copilot and DALL·E2. My work included developing cutting-edge AI system technologies and scaling Microsoft DeepSpeed into a leading AI framework.

I created llm-analysis, an open-source tool for analyzing latency and memory in transformer models, helping with resource planning and optimization. Check it out!