Coursework on Ivris Raymond

Applying Graph Algorithms for MoE Token Batching

Sun, 01 Sep 2024 00:00:00 +0000

Overview

Current LLMs often leverage Expert Parallelism to improve inference throughput. This poses challenges for routing tokens between physical devices while also batching effectively for throughput optimization, particularly for long context inputs. We applied a graph-based algorithm in this project and measured its effectiveness for routing tokens among physical devices assuming a model was deployed using expert parallelism.

Efficient Memory Access Monitoring and Bug Detection

Fri, 01 Sep 2023 00:00:00 +0000

Overview

Many existing tools for memory tracing and bug detection rely on dynamic program analysis. For this project, we implemented ideas presented in Spindle (Wang, et al.) to perform memory tracing and bug detection using static analysis, achieving constant execution time and trace size scaling with array element count compared to the increasing time and size scaling of tools like Intel Pin and Valgrind. Our implementation is available on GitHub at github.com/EECS-583-Group-4/final-project.git.

RTOS Driven Autonomous Vehicle

Mon, 22 Aug 2022 00:00:00 +0000

Overview

Developed an autonomous vehicle capable of navigating around a track of reflective tape using two simple light sensors using an FPGA board. The system implemented a MicroBlaze SoC on the FPGA and utilized the FreeRTOS Kernel for scheduling.

Exploration of Rowhammering Techniques

Tue, 18 Jan 2022 00:00:00 +0000

Overview

Conducted a survey of various techniques for rowhammering and their implementation schemes, then produced a report on how a synthesis of new techniques may be capable of breaking some of the finalists in the NIST PQC Standardization Process by serving as the fault-injection mechanism.

Power Usage of AES Power Analysis Attack Mitigation Strategies

Mon, 17 Jan 2022 00:00:00 +0000

Overview

Analyzed the power usage characteristics of a microcontroller performing AES encryptions across a variety of AES implementations. These implementations used many variations and combinations of masking and hiding techniques to improve their resiliency against power analysis attacks.