Systems and local-AI engineering on 8 GB VRAM + AMD 3D V-Cache (96 MiB L3) under Omarchy Linux (Hyprland / Wayland). Everything here focuses on squeezing maximum throughput out of consumer hardware—pushing 35B MoE models to 76.3 tok/s, running 1M-vector RAG inside CPU L3 in 163 µs, and building native Wayland/PipeWire/DKMS tooling with zero cloud lock-in.
Important
Local-first by default. Consumer 8 GB GPUs hit a VRAM wall fast; pairing llama.cpp hot-expert caching with Ryzen 3D V-Cache core pinning (vcache-run -c 2-7) and non-temporal DDR5 prefetching (GGML_MOE_NTA=1) turns bandwidth-starved spillovers into smooth, interactive inference.
Tip
Strict security & hygiene baseline. All public repositories follow 0700/0600 permission isolation, shellcheck/clippy/unit-test gates, and pre-push secret/path auditing.
Hierarchical multi-tier routing topology balancing sub-millisecond L3 V-Cache retrieval, 8 GB CUDA inference, and edge/cloud overflow:
| Repository | Domain | Highlights & Measured Results |
|---|---|---|
zen-vcache-ai |
Local AI / C / AVX-512 | 35B MoE @ 76.3 tok/s (+25.5% gen, +22.2% pp128 via GGML_MOE_NTA=1); 1M-vector RAG in 162.8 µs (6,144 QPS, 366 GB/s); 1.2x–10.8x code-editing speedup with 0 MiB extra VRAM |
omarchy-qwenimage |
Generative AI / Python | QIS — local-first generative studio (Text-to-Image, Image-to-Image, Turbo diffusion, WD14 Danbooru interrogator, Qwen3-VL vision reverse-engineering, SIMD Lanczos resizer, ComfyUI backend) (Omarchy Marketplace) |
omarchy-bt-audio-eq |
PipeWire / QML / Bash | Bluetooth codec & quality switcher (LDAC, AAC, SBC-XQ, SBC), PipeWire sample-rate selector, live telemetry, and 2-stage 10-band EQ with AutoEq (Omarchy Marketplace) |
omarchy-vkeyboard |
Wayland / Quickshell | Non-focus-stealing OSK (WlrLayer.Overlay) with wtype injection, multi-language XKB switcher, and CRT Monolithic Grid styling (Omarchy Marketplace) |
btusb-mt7902-dkms |
Linux Kernel / DKMS | Standalone DKMS driver patch & power fix for MediaTek MT7902 Bluetooth (0489:e156) on Arch / Omarchy kernels |
Workstation Spec, Stack & Quickstart Commands
| Component | Specification |
|---|---|
| CPU | AMD Ryzen 7 7800X3D (8C/16T, 96 MiB 3D V-Cache, AVX-512) |
| GPU | NVIDIA GeForce RTX 3070 8 GB |
| Memory | 32 GB Dual-Channel DDR5-6000 |
| OS & WM | Omarchy Linux (Arch-based) · Hyprland (Wayland) · Quickshell / Qt6 |
| Languages | C / C++, Python, Rust, QML / Qt6, Bash |
# Run 3D V-Cache MoE inference pinned to 6 cores (2-7)
GGML_MOE_NTA=1 ./bin/vcache-run -c 2-7 -- llama-server -m model.gguf -ngl 99 -cmoe --moe-cache-mib 4096 -t 6
# Install Omarchy plugins directly
omarchy plugin add https://github.com/simonezpx3/omarchy-qwenimage.git --enable
omarchy plugin add https://github.com/simonezpx3/omarchy-bt-audio-eq.git --enable
omarchy plugin add https://github.com/simonezpx3/omarchy-vkeyboard.git --enableSupport / Hardware Fund (RTX 3090 24 GB & Zen 5 X3D benchmarks): Ko-fi · Lightning (arcicml@coinos.io) · Bitcoin