nx_embed_bench.nx

nx_embed_bench.nx -- IN-PROCESS m-SWEEP: is prefill actually BATCHING? v1 measured cross-run and DISAGREED WITH ITSELF: 8tok=606ms/tok, 32tok=673, 125tok=403 (non-monotonic). Two causes, both avoidable: it raced continuous sibling build load, and it re-loaded a 491MB model every run (measured load variance alone 16554..20767ms, ~25%). nx_llm_forward_profile's own header already warned this host is too noisy for cross-run attribution below 20% -- I hit a limitation the ecosystem had already written down. LAW: WHEN A MEASUREMENT DISAGREES WITH ITSELF ACROSS RUNS, FIX THE INSTRUMENT BEFORE INTERPRETING IT. v2 loads the model ONCE and sweeps m INSIDE one process, so load variance and cross-run drift cancel and only the m-dependence remains. Two reps per m are printed so the reader SEES the noise instead of trusting one sample. FLAT ms_per_token across m => prefill is NOT batching (wire the m>1 packed path in lazy_matmul) FALLING ms_per_token => batching works; the cost lives elsewhere Compare against the m=1 DECODE cost of 350ms/token from nx_llm_forward_profile. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0