Home / Method notes
Method notes
How we label tok/s on named hardware. Method lives here; tables live on Tok Lanes.
Hardware named
- House desk: Rumble — RTX 5080 16GB (Ollama / FreeToken WSL2 smokes).
- Speed lane ≤32GB: B70-class ballpark from the lanes map research pass.
- Capacity ~128GB: Spark / Halo / Mac Max — community ballparks called out as such on the lanes map.
A number without a box and a stack is not a number we lead with.
Engine tok/s vs wall-clock
- Engine / decode tok/s — what the runtime logs while generating (example: FreeToken MoE engine ~130–150 decode).
- Wall — what you feel end-to-end on short prompts (same MoE smoke: wall ~60–90).
When they differ, both get labeled. We do not paste engine rates as if they were wall.
Fit vs spill
If weights + KV stay on the GPU, decode holds. Spill to system RAM and speed can fall off a cliff — house 5080 notes show ~12–15 tok/s when a 27B IQ4 fits, and under 1 tok/s after spill. Quant and ctx decide fit; copying a YouTube layer count is not a plan on 16GB.
What gets a Build card
- A real session on named hardware with a dated result (pass, fail, or tie).
- One-line result on the index card; full notes on the Build page.
- Negative results included (Flash-Next no; spilled quants stay spilled).
- Empty video slot = not run yet / placeholder — never a fake video ID.