VRAM arithmetic you can do in your head
Fit is a function, not a badge: weights plus context-dependent cache plus runtime overhead. Change the context or batch size and the answer changes.
Budget weights, cache, and headroom
Estimate weight memory from parameters times bits per weight, add the runtime's KV-cache estimate, then reserve headroom for the OS, preprocessing, and prompt-ingestion spikes.
Make the memory budget executable
Finish a reusable estimator instead of calculating only one model. The configuration uses decimal gigabytes for a quick screening estimate; a real load test remains the authority.
Measure a fit matrix
For at least two models and three context lengths, record peak memory and generated tokens per second after one warm-up run. Keep every other setting fixed.
Fade the memory estimate
Study the first calculation, then finish a similar one without copying its final line.
Record measured peak memory
For the configuration you intend to use, save the highest allocation across prompt ingestion and generation. Describe model, format, and context in the note.
Save the fit table
Link the complete fit table and state which configuration you chose. The rejected rows are part of the evidence.