Will's CUDA Suite — a VRAM/stress bench for local CUDA inference on NVIDIA GPUs.
Find each model's safe context ceiling without exhausting the card.
wcx-suite is the CUDA sibling of wmx-suite. Same discipline, other side of the hardware fence: it finds the safe context ceiling for a model+context on your GPU by extrapolating from measurements taken below the hardware wall — never by probing into it.
On Apple Silicon the wall is the unified-memory working set; here it's the GPU's VRAM. A CUDA over-allocation usually raises an out-of-memory error and kills the workload rather than the machine — but on a GPU that also drives a display, exhausting VRAM can freeze the desktop or trip a driver reset. The rule holds: stay under the wall.
uv sync
uv run wcx-suite system # GPU, VRAM total/free, the wall, and the safe budget
uv run wcx-suite calibrate # measure this card's cold-start VRAM overhead (once per machine)
uv run wcx-suite characterize <hf_id> # safe probe → fitted safe context ceiling, storedcharacterize ramps context under the engine and refuses before it ever loads past the
VRAM wall, then fits and stores the safe ceiling — the same predict-don't-probe discipline
as wmx-suite. Footprint is read in binary GiB to match the wall, and weight/KV-cache
quantization (int4/int8/fp8 weights, hqq KV) is wired through the loader.
A functional measurement engine, not barebones anymore: system + calibrate +
characterize, plus governed generate and capability benchmark workers that
project-ara drives out-of-process
(python -m wcx_suite.generate / wcx_suite.benchmark). Both apply each model's own chat
template so template-strict instruct models (e.g. Gemma) behave correctly. 186 tests pass.
This is the engine project-ara installs on NVIDIA
hardware (ara install --engine wcx), the way wmx-suite is its Apple-Silicon engine.
Apache 2.0 © Will Sarg