Run QWEN3.8 27B on 16gb Nvidia GPUs

This kit provides a one-click installer and serving suite to run Qwen3.8-27B on a single consumer NVIDIA GPU (targeted at 16 GB but supports 12-32 GB cards) by selecting and deploying TURBODERP EXL3 quantized weights that fit detected VRAM. It creates an isolated Python environment, downloads resumable weight files, installs a prebuilt inference engine when available (with a compile fallback), and exposes an OpenAI-compatible /v1 API plus a browser chat UI. Required host specifics are NVIDIA GPUs with compute capability 7.5+, driver 570+, Python 3.11+, and Node 22.19+ for the UI; disk needs vary by quant (roughly 9.7-22.9 GB per quant). Everything is confined to the kit folder (.venv, models, logs, apps), needs no admin rights, and lets you switch quant sizes or reset the install later.

Windows and Linux use separate scripts with clear workflows: Windows provides windows/START-HERE.bat to install-and-start, start.bat to run installed models, tray integration, and VRAM preflight checks that warn about processes holding GPU memory; Linux uses linux/setup.sh and linux/start.sh (with -b for background) and serves the API on port 8888 and UI on 3080. The kit documents troubleshooting (insufficient VRAM, missing Python, CRLF issues, missing prebuilt wheels), configurable .env parameters like CONTEXT_SIZE and GPU_MEM_GB, and logs for diagnostics.