Maple Claims Gemma 4 26B Runs at 45 Tokens/Sec on a 6GB GPU via MoE Pruning

·1 views

A Hacker News post introduces Maple, a project that applies sub-2-bits-per-weight mixture-of-experts pruning. The developer says the technique lets Google's Gemma 4 26B model run at roughly 45 tokens per second on a GPU with 6GB of memory. The project is shared as a GitHub repository, and the post received 3 points and one comment at the time of listing. The source material is limited to the repository link and discussion metadata, so the performance claims have not been independently verified.

Read the full story at Hacker News

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

Discussion (0)

Why an Empty Test Result Should Not Count as a Pass

The article argues that a test run producing no output means no evidence was collected, so the outcome should be labeled inconclusive rather than passed. A pass is a positive claim requiring that a command executed and showed the behavior holds, which an empty result does not provide. Missing scripts, unrun branches, or unavailable dependencies are common causes that point to process gaps instead of clean code. Treating silence as success can hide defects, especially with AI-generated code where confident summaries may imply verification that never occurred. The recommended approach is to keep gaps visible, mark items as unverified, and close them by adding runnable checks that produce captured output.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CompTIA Network+ N10-009 Rewards Troubleshooting Over Memorization

The CompTIA Network+ N10-009 exam is described as a troubleshooting test that uses networking vocabulary rather than a vocabulary quiz. It has 90 questions with a passing score of 720 on a 100-900 scale, split across five domains. Troubleshooting is the largest at 24 percent, and the article argues it also appears within other domains. The author recommends practicing layer-by-layer reasoning with symptom scenarios instead of relying on flashcards. The piece also advises short daily subnetting practice over cramming, since subnetting is a common source of lost points.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Rebuilding a Tailwind v4 Theme from an Existing Website

A developer tasked with matching a new app to a client's existing site, which lacks Tailwind and a style guide, describes extracting a usable theme from the live page. The approach favors naming colors by role, such as primary, background, text, muted and border, over listing every hex code. Fonts should be checked via the browser's computed styles, with free substitutes chosen for commercial typefaces, and type sizes, weights, line heights and tracking recorded. The resulting values are placed in a Tailwind CSS v4 @theme block, which replaces Tailwind's defaults. The author also built a browser extension, Stylasso, to automate these steps, and notes that it is their own product.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How ATS Parsers Read Location, Remote and Relocation Details on Resumes

Applicant tracking systems usually convert a resume's location line into structured fields such as city, state or region, and country, rather than treating it as free text. Remote and relocation preferences are often stored separately as flags or notes, so they may not be captured from the resume alone. Vague entries like a bare 'Remote' with no city, or multiple cities crowded onto one line, can confuse field extraction and cause a candidate to be filtered out. The article recommends a plain 'City, State, Country' line with clear wording about remote or relocation next to it, and avoiding icons, tables, and graphics. It also notes that configurations vary across vendors such as Greenhouse, Workday, Lever, iCIMS, and Taleo, so candidates should also complete the application's own preference questions consistently with their resume.

0 comments Read more at DEV Community