vLLM Qwen 3.6-27B (AWQ-INT4) + DFlash speculative decoding on AMD Strix Halo (gfx1151 iGPU, 128 GB UMA, ROCm 7.13). 24.8 t/s single-stream, vision, tool calling, 256K context, OpenAI-compatible, Docker. Matches DGX Spark FP8+DFlash+MTP at a third of the cost. No CUDA.

Self-hosted, browser-based AI dungeon RPG, fully local on an AMD Strix Halo APU: narrator + per-character agents on one local LLM (Gemma 4 26B MoE via llama.cpp/Vulkan), local images (FLUX.2 klein/ComfyUI), expressive voice (Maya1 TTS). FastAPI brain, SQLite, vanilla JS, docker-compose with guided setup (CLI wizard or double-click HTML).

Model-agnostic NPU+GPU+CPU inference engine for AMD Strix Halo. FastFlowLM fully reverse-engineered and replaced with a native open-source XDNA 2 stack — zero proprietary code. GGUF/ONNX/1BP ternary (TQ2 2-bit). Zero Python at runtime. MIT.

Run Qwen 3.6-27B AWQ-INT4 models with DFlash speculative decoding on AMD Strix Halo hardware using vLLM for high-throughput inference.