Large Language Models (LLMs) have become a popular topic, but how can individuals deploy their own LLM inference services? In this session, I will share my experience deploying an LLM inference service (Ollama) on k0s with NVIDIA GPUs, including how to manage GPU resources within k0s. After deployment, one common challenge is that the inference service may run slowly or consume significant energy. To address this, I propose a testbed called Inference Testbed for Performance & Energy (ITPE) for evaluating both performance (via GenAI-Perf) and energy efficiency (via Kepler) across heterogeneous hardware platforms. I will also introduce the architecture of ITPE and explain how its components are managed using FluxCD.

網路一直都是生活必備品，而在一場會議提供一個穩定的網路則是一大難題。本場議程將分享講者過去在研討會中，是如何自建超高速網路 (10G+ 頻寬) 到會場，並調整骨幹路由的經驗。

LLM are rapidly growing in size and complexity, often exceeding the capacity of a single GPU. This talk explores practical strategies for running LLMs efficiently across multiple GPUs.

In a microservice email system, the emails can go through multiple services, which makes email tracing more difficult and troublesome. Mailtrace is a tool that makes tracing email flows in a microservice email system easier and faster.

In this talk, I would like to introduce DPI mechanism under 5G core network background. I will focus on DPI packet block-rate and the impact of throughput. I will introduce 3 different method to achieve DPI and describe the pros and cons. It will be fit if you are interest in 5G or networking.

本研究聚焦於在缺乏真實金融交易資料的情況下，如何基於程式生成的合成交易資料（synthetic transaction data）建立有效的風險評估基準（benchmark）。由於研究團隊所採用的資料完全依循關聯式資料庫結構與預設規則進行模擬生成，本報告提出了一個以「架構驅動」（schema-driven）為核心的合成金融資料驗證流程。該流程結合了金融風險理論（如信用風險、流動性風險和異常交易偵測）與資料生成機制中的可控參數，設計出可量化的風險指標並進行對照實驗。 透過模擬多種風險場景（例如高違約率、交易集中度增加、以及涉及洗錢的模式等），我們構建了一套多層次基準測試集合，以檢驗交易風險模型在敏感度與穩健性方面的表現。這將為後續的交易預測與風險分析提供牢固且可信的評估依據。

Day1 Sprint

你是否曾想過：「如果我能清楚描述想法，AI 是否就能做出產品？」 讓我們將模糊想法轉換成清晰可執行的規格，使用 AI 開發出你預期的功能！