ABLITAI LAB — DRAFT PROTOCOL
PROPOSED STUDY — NO RESULTS

What fits on a consumer GPU?

Question
Where does a usable setup run out of memory?

Setup
Record the exact GPU and VRAM, driver, model revision, weight format, serving runtime and cache format. Keep the host and runtime settings fixed within each comparison.

Proposed method
1. Measure idle memory after loading the model and record the available cache capacity.
2. Increase input length in bounded steps using the same prompts and output allowance.
3. Measure first-token delay, generation speed and peak memory for each completed request.
4. Report any reduced precision, CPU offload or concurrency limit alongside the result. Keep stability testing separate from a short fit check.

Measures
- Successful input and output allowance
- Peak VRAM and host memory
- First-token delay and generation speed
- Concurrency and observed failure modes

Limitations
A successful single request does not establish multi-user capacity or long-term reliability. Different quantizations can change answer quality. Rental pricing and hardware availability also need a dated source.

Next step
Choose the exact model, runtime and hardware comparison before taking measurements.