VRAM Diary

StarCraft native engine (sc-native)

Wed, 28 Oct 2026 00:00:00 +0000

A from-scratch native Linux engine for StarCraft that reads the data from your own StarCraft (Remastered) install. Download an installer, and it compiles on your computer.

The Long Rot

Wed, 28 Oct 2026 00:00:00 +0000

A realistic first-person, open-world zombie survival game from my studio, Tensor Games. The world didn’t end in an hour. It rotted away over years.

Qwen: a self-hosted voice assistant

Wed, 28 Oct 2026 00:00:00 +0000

A voice-first assistant I run on my own hardware. It listens, thinks with a local language model, speaks back, and can use tools like search, notes, reminders and my homelab’s status.

Blender and 3D AI tools

Wed, 28 Oct 2026 00:00:00 +0000

A set of Open WebUI tools that let a local AI model drive Blender, turn images into textured 3D models, and slice and run 3D prints on my Bambu P2S.

3D-printed hardware

Wed, 28 Oct 2026 00:00:00 +0000

Custom printed parts for the homelab: fan shrouds for passive Tesla P100 GPUs, and a dragon-head PC case for the home server.

The homelab

Wed, 28 Oct 2026 00:00:00 +0000

The self-hosted server and GPU machines behind everything else, from file sync and media to local AI inference. Most of this blog documents it.

Silence Cutter

Wed, 28 Oct 2026 00:00:00 +0000

A single-file Python script that removes dead air from a video and gives you a tighter jump-cut edit.

Downloads

Wed, 28 Oct 2026 00:00:00 +0000

sc-native: native StarCraft engine for Linux

sc-native is my from-scratch native engine for StarCraft. Two installers are available for Ubuntu and Linux Mint (tested on Ubuntu 24.04 and 22.04, Debian 12, and Linux Mint 22). Both carry only the engine’s source code, which they compile on your own computer.

No game data included. The installers contain no StarCraft game data. You need an existing StarCraft (Remastered) install for the engine to read from. The installer looks for it automatically (Wine, Lutris and Bottles prefixes, ~/Games, mounted drives) or asks you for the folder.

Self-extracting installer .run

Per-user install, no root needed unless build packages are missing. Recommended.

Llama 3.2 Vision Doesn't Run Here (And Here's Why)

Tue, 07 Jul 2026 00:00:00 +0000

Not every homelab post ends with a working setup. This one’s about a model that, as far as I can tell, cannot currently run on my hardware through any tool I have access to — not because of a config mistake, but because of a real gap between two separate projects that neither one has closed yet.

I’m writing it up anyway, because “here’s the thing that didn’t work and exactly why” is worth just as much as another bake-off post, and because I burned real hours chasing this before the actual root cause became clear.

Five Models, One GTX 1060: A Laptop Local AI Bake-Off

Fri, 03 Jul 2026 00:00:00 +0000

Testing five small LLMs on an aging GTX 1060 laptop to find out which ones are actually usable — and which ones will make you question your life choices.

Getting this blog and its own email address live: a full night of self-inflicted problems

Fri, 03 Jul 2026 00:00:00 +0000

Registering a domain sounds like a five-minute task. Getting this actual blog live, plus a working self-hosted mailbox on that domain, took most of a night — almost none of it because anything was genuinely hard, and almost all of it because I kept making one small wrong assumption at a time. Writing the whole thing down, mistakes included, because that’s more useful than pretending it went smoothly.

Buying the domain

Landed on vramdiary.com — specific, a little funny, honest about what half these posts are actually about. Checked out on Namecheap, skipped the upsells (PremiumDNS, Stellar Web Hosting, all of it — none of it does anything a self-hosted setup needs).

About

Thu, 02 Jul 2026 00:00:00 +0000

I’m JR — an Operations Supervisor by day, running interplant shipments through SAP. Work only allows Copilot, so everything here happens on my own hardware, on my own time.

This blog is the build log for that: a small fleet of self-hosted machines running Linux, Docker, and increasingly, local AI models I run myself instead of renting from someone else’s API. Some of it is genuinely useful — automation, homelab infrastructure, model comparisons with real numbers instead of vibes. Some of it is just me chasing a VRAM budget at 11pm because I refuse to let a problem sit overnight.

Building out my gaming server on Linux Mint

Thu, 02 Jul 2026 00:00:00 +0000

Spent the last few weeks reworking the home lab, and this round was getting my gaming server fully set up — Steam library, storage layout, and streaming to the living room TV.

The hardware

Base machine is an ASRock motherboard with an Intel i5-12600KF, a Radeon RX 6750 XT (12GB), and 16GB of DDR4. Storage-wise: three 1TB NVMe drives and a 1.82TB Seagate 7200rpm hard drive. This box is slated for a bigger upgrade down the line as part of the August platform swap, but for now its job is straightforward: run my Steam library and stream it to the TV with zero hassle.

Can a local model actually design something? gpt-oss:20b vs Gemma 4 12B

Thu, 02 Jul 2026 00:00:00 +0000

Wanted to see if either of my two resident models could act as an actual design collaborator — not just answer questions, but iterate on real critique the way a person would. Gave both the identical starting brief I used when building this blog: design plan first, code later, avoid the generic AI-blog defaults.

Same four-stage escalation for both: initial brief, call out any cliché it lands on, push for one genuinely distinctive structural element, then (if it gets that far) flag the live-backend tradeoff that idea usually implies and see if it self-corrects.

Finding the Best Local LLM for a 16GB AMD GPU

Thu, 02 Jul 2026 00:00:00 +0000

If you’ve spent any time searching for “best local LLM for X GB VRAM,” you’ve probably noticed the same problem I ran into: almost every guide is a spec-sheet comparison. Parameter counts, benchmark scores, MMLU percentages — all useful in theory, but none of it tells you how a model actually behaves when you ask it something real.

So instead of trusting the spec sheets, I ran a real head-to-head test across 8 local models on my own hardware, using real prompts, and I’m writing up exactly what I found — including the dead ends, the wrong assumptions, and the actual numbers.

Now

Thu, 02 Jul 2026 00:00:00 +0000

What’s actively in motion, updated as things change.

Active

Late Orders Report — 10-sheet automated Excel report reading SAP exports, now with a password-protected commercial trial variant. Deploying update_report.py to the main LLM box next.
Model bake-offs — running the same sudo/GPU-nuance and hard bash/awk regex test suite across every machine to keep model picks honest instead of vibes-based.
Video editor pipeline — automated rough-cut assembler (ffmpeg silencedetect + PySceneDetect, MoviePy assembly) verified on a synthetic clip. Needs a real gameplay-footage test before I trust the default thresholds.

Planned — August hardware

Moving the RX 9060 XT, 32GB RAM, and boot drive from the main LLM box onto the gaming PC’s board, once a second 9060 XT is in hand, to build out a proper dual-GPU inference server.
A dedicated single-GPU ComfyUI box — still deciding between an RTX 5060 Ti 16GB and a used RX 7900 XT 20GB.

Someday

AM5 build with dual RX 7900 XTX for tensor-split inference.
A small 2D top-down RPG in Godot, once there’s spare bandwidth for something that isn’t infrastructure.

Open WebUI: why I stopped trusting Knowledge Bases and just typed the facts into Memory

Thu, 02 Jul 2026 00:00:00 +0000

Spent a chunk of a weekend debugging something that turned out not to be a bug I could fix — it was a known limitation I had to route around instead. Writing it down because “the AI can’t find anything in the knowledge base” is exactly the kind of vague symptom that’s easy to misdiagnose as something wrong with your setup.

The symptom

Kristen asked her assistant a simple question: what equipment is in her garage gym. She’d already uploaded a fitness profile into a Knowledge base and connected it to her assistant in Model Settings. The assistant came back with nothing. Not a wrong answer — no retrieval at all, like the knowledge base wasn’t there.

Three tests, one winner: what Gemma 4 got right that gpt-oss didn't

Thu, 02 Jul 2026 00:00:00 +0000

Didn’t set out to run a Google-vs-OpenAI bake-off. Just kept testing models on real tasks across two machines, and the pattern kept repeating itself until it was hard to ignore. Three separate tests, three different axes, same result each time.

Test 1: the 6GB laptop

Five models on my laptop’s GTX 1060, tested on live-measured VRAM plus a hard OOM-log regex challenge. batiai/gemma4-e2b:q6 used roughly half the VRAM of every Q4 competitor — while running at a higher precision quant — and was the only model whose hard-mode answer had valid syntax that would actually execute. Qwen3:8b used the most VRAM and took the longest to respond, for a regex answer with broken shell quoting that wouldn’t run at all. More thinking tokens, worse output.

Unresolved: intermittent response delays on the laptop, and everything I ruled out

Thu, 02 Jul 2026 00:00:00 +0000

Not every post here ends with an answer. This one’s a log of a bug I haven’t fixed yet, written down so the next session doesn’t start from zero.

The symptom

batiai/gemma4-e2b:q6 on my laptop (GTX 1060 6GB) — the model that won the earlier bake-off outright — has started showing inconsistent response latency. Not the model’s own thinking time, which is separate and expected. A silent gap before generation even starts, anywhere from nothing to 40-60+ seconds, with no obvious pattern:

Three Bugs Deep: Getting Open WebUI Running on My ASUS Laptop

Wed, 01 Jul 2026 00:00:00 +0000

I spent a night debugging what looked like a single broken Docker container on my ASUS laptop (i7-8750H, GTX 1060 6GB, Linux Mint). It turned out to be three separate, unrelated bugs stacked on top of each other. Here’s the full trail, in case it saves someone else the same night.

The Setup

The goal was simple: get Open WebUI v0.10.2 running in Docker on the laptop, talking to a local Ollama instance, so I’d have the same self-hosted AI chat interface on the go that I already run at home.

Finding the Right Local AI Model for a 6GB Laptop GPU

Tue, 30 Jun 2026 00:00:00 +0000

My daily-driver laptop is an ASUS laptop — an i7-8750H with a GTX 1060 6GB, running Linux Mint. Not a beefy AI rig by any means, but with the right model it’s more than capable of running a genuinely useful local assistant via Ollama and Open WebUI. The question was which model actually earns a permanent spot on 6GB of VRAM.

I tested five candidates head-to-head: three VRAM measurements per model, then two increasingly hard questions — first a simple factual one, then a real technical challenge that ended up separating “sounds right” from “actually works.”