| |
| """Cross-check data/fit-estimates.csv against the live https://localllmrig.com/api/recommend. |
| Expected result: identical fit labels; speeds within ±0.1 tok/s (KV rounding for sliding-window models).""" |
| import csv, json, urllib.request |
| rows = {(r["model_id"], r["hardware_id"], r["quant"]): r for r in csv.DictReader(open("data/fit-estimates.csv"))} |
| n = bad = 0 |
| for model in ["llama-3.3-70b", "qwen2.5-32b", "llama-3.1-8b", "gemma-2-27b", "qwen2.5-72b", "gemma-2-9b", "nemotron-3-5-bolt-74b"]: |
| for q in ["Q4", "Q8"]: |
| for budget in [800, 2000, 5000, 20000]: |
| url = f"https://localllmrig.com/api/recommend?model={model}&budget={budget}&quant={q}&context=8192" |
| for p in json.load(urllib.request.urlopen(url)).get("picks", []): |
| r = rows[(model, p["id"], q)]; n += 1 |
| if r["fit"] != p["fit"] or abs(float(r["est_decode_tok_s"]) - p["tokensPerSec"]) > 0.15: |
| bad += 1; print("MISMATCH", model, q, p["id"], r["fit"], r["est_decode_tok_s"], p["fit"], p["tokensPerSec"]) |
| print(f"checked {n} picks, {bad} outside tolerance") |
|
|
