- Dockerfile.finetune: PyTorch 2.3/CUDA 12.1 base + unsloth + training stack
- finetune_local.py: auto-register model via Ollama HTTP API after GGUF
export; path-translate between finetune container mount and Ollama's view;
update config/llm.yaml automatically; DOCS_DIR env override for Docker
- prepare_training_data.py: DOCS_DIR env override so make prepare-training
works correctly inside the app container
- compose.yml: add finetune service (cpu/single-gpu/dual-gpu profiles);
DOCS_DIR=/docs injected into app + finetune containers
- compose.podman-gpu.yml: CDI device override for finetune service
- Makefile: make prepare-training + make finetune targets
scripts/preflight.py (stdlib-only, no psutil):
- Port probing: owned services auto-reassign to next free port; external
services (Ollama) show ✓ reachable / ⚠ not responding
- System resources: CPU cores, RAM (total + available), GPU VRAM via
nvidia-smi; works on Linux + macOS
- Profile recommendation: remote / cpu / single-gpu / dual-gpu
- vLLM KV cache offload: calculates CPU_OFFLOAD_GB when VRAM < 10 GB
free and RAM headroom > 4 GB (uses up to 25% of available headroom)
- Writes resolved values to .env for docker compose; single-service mode
(--service streamlit) for scripted port queries
- Exit 0 unless an owned port genuinely can't be resolved
scripts/manage-ui.sh:
- Calls preflight.py --service streamlit before bind; falls back to
pure-bash port scan if Python/yaml unavailable
compose.yml:
- vllm command: adds --cpu-offload-gb ${CPU_OFFLOAD_GB:-0}
Makefile:
- start / restart depend on preflight target
- PYTHON variable for env portability
- test target uses PYTHON variable
