The measured tap cost, and the one lever that is honest to pull.
Measurement (engine counters, 2026-10-29)
sample
prefill_tps
steps_prefill
preemptions
t0 (uptime 147 s)
71.0
4
0
t1 (+180 s)
70.7
11
3
7 steps per 180 s at the capped 2,048 tokens → ~71 tok/s. One 7,168-token document ≈ 100 s; round 1
(2,367 docs / 10.03M tokens) ≈ 39 h; the 12M floor ≈ 47 h of drained production.
Why a wider calib step is legal
kldref and score inherit production's step and slots to the letter because their KL numbers are
only comparable at an identical shape. Calib produces layer Hessians — no number any report
compares — and its cost is read volume per document: at 2,048-token steps a document is four steps,
each re-streaming every layer's routed expert union from the container (43.4 GiB a rank live on
disk). Q38_CALIB_MNBT / Q38_CALIB_SEQS are now calib's own, defaulting to the inherited values.
Test
./tests/run.sh → 589 ok / 0 fail; the new case asserts the knob reaches calib's argv, does
not reach kldref's, and is refused by name when non-numeric.
Next: probe at Q38_CALIB_MNBT=8192 (one 7,168-token document in a single step). If ≥1.5×, keep
the full 12M (~1.1 days); if not, cut the target to 6M (~1 day).
The measured tap cost, and the one lever that is honest to pull.
## Measurement (engine counters, 2026-10-29)
| sample | prefill_tps | steps_prefill | preemptions |
|---|---|---|---|
| t0 (uptime 147 s) | 71.0 | 4 | 0 |
| t1 (+180 s) | 70.7 | 11 | 3 |
7 steps per 180 s at the capped 2,048 tokens → ~71 tok/s. One 7,168-token document ≈ 100 s; round 1
(2,367 docs / 10.03M tokens) ≈ 39 h; the 12M floor ≈ **47 h** of drained production.
## Why a wider calib step is legal
`kldref` and `score` inherit production's step and slots to the letter because their KL numbers are
only comparable at an identical shape. Calib produces **layer Hessians** — no number any report
compares — and its cost is read volume per document: at 2,048-token steps a document is four steps,
each re-streaming every layer's routed expert union from the container (43.4 GiB a rank live on
disk). `Q38_CALIB_MNBT` / `Q38_CALIB_SEQS` are now calib's own, defaulting to the inherited values.
## Test
`./tests/run.sh` → **589 ok / 0 fail**; the new case asserts the knob reaches calib's argv, does
not reach kldref's, and is refused by name when non-numeric.
Next: probe at `Q38_CALIB_MNBT=8192` (one 7,168-token document in a single step). If ≥1.5×, keep
the full 12M (~1.1 days); if not, cut the target to 6M (~1 day).
---
cwd: /home/blank/.omp/wt/wt-20261027-142335-6d1fc1f · session: 01a17f3f-3e2e-7509-aef8-be516c27cf98
Measured from the engine's own counters on the 2026-10-29 drain: prefill 71 tok/s steady state, a
180 s window showing 7 prefill steps of the capped 2,048 tokens and 3 preemptions. One 7,168-token
document is therefore ~100 s, round 1 (2,367 docs, 10.03M tokens) ~39 h, and the 12M-position
floor ~47 h of drained production — not the ~18 h the brief assumed.
The tap is the only phase whose cost is read volume per document: at the inherited 2,048-token step
a document is four steps, and each step re-streams every layer's routed expert union from the
container (43.4 GiB a rank live on disk). One wide step reads each layer's union once.
The 'one step, one slot count' audit law binds kldref and score because their KL numbers are only
comparable at an identical shape. Calib produces layer Hessians and no number any report compares,
so Q38_CALIB_MNBT / Q38_CALIB_SEQS are its own, defaulting to the inherited value — nothing moves
unless a probe asks. The KL modes ignore both.
Tests: 589/0 — the knob reaches calib's argv, does not reach kldref's, and a non-numeric value is
refused by name.
dank
merged commit 9f997f0841 into main2026-10-29 16:50:51 -07:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
The measured tap cost, and the one lever that is honest to pull.
Measurement (engine counters, 2026-10-29)
7 steps per 180 s at the capped 2,048 tokens → ~71 tok/s. One 7,168-token document ≈ 100 s; round 1
(2,367 docs / 10.03M tokens) ≈ 39 h; the 12M floor ≈ 47 h of drained production.
Why a wider calib step is legal
kldrefandscoreinherit production's step and slots to the letter because their KL numbers areonly comparable at an identical shape. Calib produces layer Hessians — no number any report
compares — and its cost is read volume per document: at 2,048-token steps a document is four steps,
each re-streaming every layer's routed expert union from the container (43.4 GiB a rank live on
disk).
Q38_CALIB_MNBT/Q38_CALIB_SEQSare now calib's own, defaulting to the inherited values.Test
./tests/run.sh→ 589 ok / 0 fail; the new case asserts the knob reaches calib's argv, doesnot reach kldref's, and is refused by name when non-numeric.
Next: probe at
Q38_CALIB_MNBT=8192(one 7,168-token document in a single step). If ≥1.5×, keepthe full 12M (~1.1 days); if not, cut the target to 6M (~1 day).
cwd: /home/blank/.omp/wt/wt-20261027-142335-6d1fc1f · session: 01a17f3f-3e2e-7509-aef8-be516c27cf98