Resource budgets and presets

hearthwork budget shows three separate things for each resource:

Column	Meaning
Configured	The budget you or a preset set.
Estimated	The need of the running model: weights + KV cache for its context + about 1.5 GB runtime.
Measured	What is used now: llama-server’s RAM from the OS, GPU memory from nvidia-smi when available, else “unavailable”.

hearthwork status ends with the same table.

Presets

Preset	Command	Effect
dedicated	budget use dedicated	Keeps 0.5 GB of VRAM free for other programs; gives the AI the most.
balanced (default)	budget use balanced	The setup’s own settings.
work	budget use work	For gaming or work: keeps about 5 GB of VRAM free, 1 slot, half the CPU threads, unloads after 15 idle minutes.
custom	budget set vram=10GB ram=12GB slots=1 threads=8 idle=20m	Your own numbers on top of the current preset. budget reset returns to balanced.

What is enforced

• Slots (-np), threads (-t) and idle unload are llama.cpp flags or host settings.
• The VRAM budget becomes llama.cpp’s --fit-target (best effort).
• The OS cannot cap RAM, so Hearthwork estimates the need before every start or switch and compares it with the budget and the RAM that is free now (the model being replaced counts as free). The default budget only warns about RAM. Presets and an explicit ram= are binding.

--over-budget and --force

An impossible start is refused with the numbers and options (a smaller quantization, an explicit smaller context, another preset). The context is never shrunk silently. hearthwork start --over-budget starts anyway.

--force is separate: it interrupts running requests.

Changing the budget

Changing the budget while a model runs restarts it through the host’s switch path. status shows pending: ... when the saved settings differ from the running server, and hearthwork start applies them when nothing is running. On unified memory (Apple silicon), memory is counted once.