adapter training
Generate a reproducible local config with npm run cortex:v9:lora:config then run the optional CUDA trainer using npm run v9:train. The trainer uses local model files and local JSONL data.
V9 MODEL OPERATIONS gpu inventory registered models LORA PIPELINE QUANTIZATION
GPU-aware model manager
inspect local hardware model artifacts quantization choices and serving placement without sending model metadata to a cloud AI API.
serving hardware
probing model manager…
placement planner
register GGUF or adapter artifacts to see serving plans.
GGUF serving export
With llama.cpp tools installed run npm run cortex:v9:quantize -- input.gguf output.gguf Q4_K_M cortex-local. The resulting artifact is added to the local model registry.