Two Open-Source Tools Simplify AMD GPU Setup and Benchmarking for Local LLMs

·11 views

Running large language models locally on AMD GPUs has been complicated by missing drivers and unrecognised GPU architectures requiring manual environment variable tweaks. A tool called ROCmFix addresses this by automatically detecting GPU PCI IDs and setting the HSA_OVERRIDE_GFX_VERSION variable across multiple shells on both Windows and Linux. A companion tool, InferBench, helps users determine whether Vulkan or ROCm/HIP backends deliver better performance by running structured benchmarks with warm-up queries and forced VRAM unloads between runs. InferBench measures key metrics including median tokens per second and Time-to-First-Token to reduce caching bias in results. Both tools are available as open-source repositories aimed at AMD GPU users seeking a more reliable local AI inference experience.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Architect builds open-source AI agent that draws in CAD apps and exports IFC models

An architect frustrated with repetitive drafting work built Axio Coder, an open-source AI agent designed to operate CAD and design software from within a user's project environment. Unlike typical coding agents that only read and write files, Axio Coder interacts with programs that have no API — such as CAD windows — by reading the operating system's UI Automation tree rather than guessing pixel coordinates from screenshots. The agent can visually inspect its own output, comparing generated geometry against a reference image or drawing, then iterating corrections before exporting files in industry-standard formats including IFC, DXF, and STEP. It retains memory across sessions, logs every edit as a before-and-after diff, and includes a self-correction mechanism to break out of repetitive loops. Currently a single-user desktop app built on Electron and Python, the tool aims to automate the repetitive 80% of architectural modelling so designers can focus on work requiring human judgement.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Mapping CISA's September 2026 KEV Additions to Real-World Internet Exposure

CISA's Known Exploited Vulnerabilities catalog identifies flaws actively exploited in the wild, but does not indicate which affected systems are actually reachable in a given environment. September 2026 KEV additions notably include services such as GitLab, LiteLLM, and Cisco Secure Firewall Management Center — tools frequently internet-facing and holding credentials or controlling other systems. Internet measurement techniques, including fingerprinting via tools like ZoomEye, can partially bridge the gap by identifying reachable instances of these services. However, a fingerprint match confirms only that a service was observed at an address, not that it runs a vulnerable version or has been compromised. For services that cannot be reliably fingerprinted, internal inventory and continuous discovery remain the more dependable methods for assessing true exposure.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Engineering Mentor Teaches OOP Concepts Using Sandwich-Making Before Any Coding

A software engineering mentor ran an experiment with an apprentice named John, aiming to build the conceptual foundation of Object-Oriented Programming without writing any code first. Using a sandwich-making exercise, the apprentice learned to distinguish entities, tools, containers, and operations, and discovered how actions trigger state transitions rather than creating entirely new objects. The exercise was then extended to a bank account model, through which the apprentice independently arrived at the principle of Command-Query Separation, distinguishing commands that mutate state from queries that only read it. A final challenge involving 100 cars in a racing simulation led the apprentice to conceptually discover the idea of a class as a reusable template. The mentor argues this design-first approach helps beginners move beyond memorizing syntax and develop genuine problem-decomposition skills.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Detect and Fix Memory Leaks in Node.js Production Apps

Node.js applications running in production are susceptible to memory leaks, where objects no longer in use are still referenced and cannot be freed by the V8 Garbage Collector. Over time, this causes the application's memory footprint to grow, leading to performance degradation, high latency, and eventual heap-out-of-memory crashes. Memory leaks in Node.js most commonly originate from undeclared global variables, unremoved event listeners on long-lived objects, and closures that inadvertently retain references to outer-scope variables. Developers can address global variable leaks by enforcing strict mode or using linters like ESLint, while event listener leaks can be resolved by explicitly removing listeners once a request or task completes. Understanding how Node.js organizes memory — particularly the V8 heap's New Space and Old Space generations — is essential for diagnosing and preventing these issues in production environments.

0 comments Read more at DEV Community