Records what the exact-KV reslice actually did, and corrects an instrument law that
would have misled the next measurement.
Boot ledger — measured, not projected
The pre-flight projection survives to the MiB: KV 16,384.00 MiB (= 32,768
B/token x 524,288, ledger-exact), RS 100.50 MiB (2 cells), compute 1,408.06 +
1,088.07 CUDA_Host, settled nvidia-smi21,176 used / 2,949 free, n_slots = 2, n_ctx_slot = 262144, DropInPaths empty, no clamp line. The ~2 GiB margin is
a desktop margin: the mmproj's 868 MiB worst case still lands after the engine's own
fit check.
Rates are floors, labelled as such
Another consumer drove the endpoint through the whole window (56 new prompts on the
other slot in three minutes, 96 % GPU-util), which is the condition the Benchmarking
Policy forbids. Reported honestly as a floor band:
Arm
q8_0 3-slot record
f16 2-slot, live co-tenancy
single-stream decode
168.7
134.1–141.5 (145.8 / 163.2 in a lighter minute)
concurrent aggregate
~300 across 3 (102.8 each)
256.6 / 262.4 / 265.1 across 2 (~130 each)
prefill at ~106k
2,877 at 106,418
3,159.5 at 106,282
decode at 106k depth
never measured
71.5
Per-stream concurrency improved. The true decode cost of the 1.882x KV byte bill is
still unmeasured; a quiet-window re-run is flagged in the runbook.
Instrument correction
On PrismML fork build 10777, a greedy ignore_eos run on the raw /completion path
answers 500 "The model produced output that does not match the expected Content-only
format". The per-request timings object for this endpoint has to be read off /v1/chat/completions. AGENTS.md named /completion as the instrument; corrected,
plus a new contract trap in the runbook.
Records what the exact-KV reslice actually did, and corrects an instrument law that
would have misled the next measurement.
## Boot ledger — measured, not projected
The pre-flight projection survives to the MiB: KV **16,384.00 MiB** (= 32,768
B/token x 524,288, ledger-exact), RS **100.50 MiB** (2 cells), compute 1,408.06 +
1,088.07 CUDA_Host, settled `nvidia-smi` **21,176 used / 2,949 free**, `n_slots =
2`, `n_ctx_slot = 262144`, `DropInPaths` empty, no clamp line. The ~2 GiB margin is
a desktop margin: the mmproj's 868 MiB worst case still lands after the engine's own
fit check.
## Rates are floors, labelled as such
Another consumer drove the endpoint through the whole window (56 new prompts on the
other slot in three minutes, 96 % GPU-util), which is the condition the Benchmarking
Policy forbids. Reported honestly as a floor band:
|Arm|q8_0 3-slot record|f16 2-slot, live co-tenancy|
|---|---|---|
|single-stream decode|168.7|134.1–141.5 (145.8 / 163.2 in a lighter minute)|
|concurrent aggregate|~300 across 3 (102.8 each)|256.6 / 262.4 / 265.1 across 2 (~130 each)|
|prefill at ~106k|2,877 at 106,418|**3,159.5 at 106,282**|
|decode at 106k depth|never measured|71.5|
Per-stream concurrency improved. The true decode cost of the 1.882x KV byte bill is
still unmeasured; a quiet-window re-run is flagged in the runbook.
## Instrument correction
On PrismML fork build 10777, a greedy `ignore_eos` run on the raw `/completion` path
answers **500 "The model produced output that does not match the expected Content-only
format"**. The per-request `timings` object for this endpoint has to be read off
`/v1/chat/completions`. `AGENTS.md` named `/completion` as the instrument; corrected,
plus a new contract trap in the runbook.
## Also
- `log.md` gets a 2026-10-18 entry (reslice + measurement + instrument law).
- Every remaining `26,112 MiB` mention in the repo is now either the corrected
explanation or frank-archon's own unrelated radiance arithmetic.
Tests: `./tests/run.sh` — 447 passed.
---
cwd: /home/blank/.omp/wt/wt-20261018-204501-6d1fc1f · session: 01a15243-3541-7380-9dae-4985fdc86843
Deployed shape verified live: KV 16,384.00 MiB (ledger-exact 32,768 B/token x
524,288), RS 100.50 MiB (2 cells), settled nvidia-smi 21,176 used / 2,949 free,
n_slots 2, n_ctx_slot 262144, DropInPaths empty.
Rates recorded as FLOORS: another consumer drove the endpoint through the whole
window (56 prompts on the other slot in three minutes, 96% GPU-util). Single
stream 134.1-141.5 (145.8/163.2 in a lighter minute) vs the 168.7 q8_0 record;
two-stream aggregate 256.6-265.1 vs ~300 across three; prefill 3,159.5 at
106,282 beats the 2,877 record while contended; decode at 106k depth 71.5 (new
baseline, no prior). Per-stream concurrency improved: ~130 each vs 102.8.
Instrument correction: the PrismML fork build 10777 answers 500 'The model
produced output that does not match the expected Content-only format' for a
greedy ignore_eos run on the raw /completion path, so the per-request timings
must be read off /v1/chat/completions. AGENTS.md Benchmarking Policy corrected.
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.
Records what the exact-KV reslice actually did, and corrects an instrument law that
would have misled the next measurement.
Boot ledger — measured, not projected
The pre-flight projection survives to the MiB: KV 16,384.00 MiB (= 32,768
B/token x 524,288, ledger-exact), RS 100.50 MiB (2 cells), compute 1,408.06 +
1,088.07 CUDA_Host, settled
nvidia-smi21,176 used / 2,949 free,n_slots = 2,n_ctx_slot = 262144,DropInPathsempty, no clamp line. The ~2 GiB margin isa desktop margin: the mmproj's 868 MiB worst case still lands after the engine's own
fit check.
Rates are floors, labelled as such
Another consumer drove the endpoint through the whole window (56 new prompts on the
other slot in three minutes, 96 % GPU-util), which is the condition the Benchmarking
Policy forbids. Reported honestly as a floor band:
Per-stream concurrency improved. The true decode cost of the 1.882x KV byte bill is
still unmeasured; a quiet-window re-run is flagged in the runbook.
Instrument correction
On PrismML fork build 10777, a greedy
ignore_eosrun on the raw/completionpathanswers 500 "The model produced output that does not match the expected Content-only
format". The per-request
timingsobject for this endpoint has to be read off/v1/chat/completions.AGENTS.mdnamed/completionas the instrument; corrected,plus a new contract trap in the runbook.
Also
log.mdgets a 2026-10-18 entry (reslice + measurement + instrument law).26,112 MiBmention in the repo is now either the correctedexplanation or frank-archon's own unrelated radiance arithmetic.
Tests:
./tests/run.sh— 447 passed.cwd: /home/blank/.omp/wt/wt-20261018-204501-6d1fc1f · session: 01a15243-3541-7380-9dae-4985fdc86843