⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.
clm-0087
measured-herehigh ●●●
citable URL: https://halobench.com/records/clm-0087/ — this address never moves; the anchor /records/#clm-0087 keeps resolving
NVIDIA Nemotron 3.5 Lightning 30B-A3B's native NextN head is genuinely active through llama.cpp 7077abb's draft-MTP path on ROCm. All n_max=1, 2 and 3 arms created an explicit MTP draft context, produced exact nonzero draft counters, kept all 15 varied sanity prompts clean and passed the house guard 4/4. Against the same-session 65.3310 tok/s plain fixed-workload mean, n1 reached 75.5813 tok/s (1.1569x), n2 69.9942 (1.0714x), and n3 61.6298 (0.9433x), with population CVs of 1.0331%, 0.0732%, and 1.0110%. Fixed-run draft acceptance declined from 118/164 (71.95%) to 147/274 (53.65%) and 155/381 (40.68%) as depth increased. N1 is the best bounded configuration and earns only the card's conditional five-task smoke; n3 is rejected. Do not promote the model or run full tau2 from this performance evidence alone. Gross whole-window energy was 1.2437 Wh plain, 1.3166 Wh n1, 1.2939 Wh n2, and 1.5277 Wh n3; these startup/probe/guard windows establish neither Wh/token nor Wh/correct.
Successful scientific root: aihydra /home/aihydra/bench-results/nemotron35-lightning-mtp-ab-20260819T112819Z. Exact commands are preserved there as command-plain.sh and command-mtp-n1..3.sh; runner sha256 004b50885c6fdc258130a444a087253bb5959b6f89230c7b273988a5cd54e51b; probe sha256 27751e8c83918417d331221bd5b30222ea550e681598eab77e3da64a6ad8f284; model sha256 6110e2e2e6cd324e6ee69ddced5a6b34fad6c94ca9827222a1e420fb92e3c90b. The earlier HG-002 attempts failed on guard integration, exact-marker matching, and supervisor source-grep logic. They are infrastructure incidents, not scientific evidence, and are not included here.