a729ff14e9
Profiles select CPU vs Vulkan and Bonsai-27B vs Qwen3.5-9B Abliterated Q4_K_M. OpenHands is no longer optional; compose deploys every service. download-models.sh fetches both GGUF families as the profile requests.
26 lines
801 B
Bash
26 lines
801 B
Bash
# Profile: mini — Beelink / SER5-class mini PC
|
|
# Always CPU for inference (stable, no iGPU hassle). Full stack always deploys.
|
|
STACK_PROFILE=mini
|
|
|
|
# Ternary-Bonsai 27B fits in system RAM; keep CPU
|
|
BONSAI_BACKEND=cpu
|
|
BONSAI_NGL=0
|
|
BONSAI_CTX=16384
|
|
BONSAI_KV4=0
|
|
BONSAI_THREADS=12
|
|
BONSAI_TEMP=0.7
|
|
BONSAI_TOP_P=0.95
|
|
BONSAI_TOP_K=20
|
|
|
|
# Primary model (Prism Q2_0 — needs Prism llama-server image)
|
|
LLM_ID=bonsai27
|
|
BONSAI_HF_REPO=prism-ml/Ternary-Bonsai-27B-gguf
|
|
BONSAI_MAIN_FILE=Ternary-Bonsai-27B-Q2_0.gguf
|
|
BONSAI_MODEL_PATH=/models/Ternary-Bonsai-27B-Q2_0.gguf
|
|
OPENWEBUI_DEFAULT_MODEL=openai/Ternary-Bonsai-27B-Q2_0.gguf
|
|
OPENHANDS_MODEL=openai/Ternary-Bonsai-27B-Q2_0.gguf
|
|
|
|
# Also download mid-tier Qwen so you can switch without re-bootstrap
|
|
DOWNLOAD_MODEL=true
|
|
DOWNLOAD_QWEN=true
|
|
DOWNLOAD_VISION=true
|