Phase 1 of the voice-over goal (Gitea issue #2). One stdlib-only HTTP facade in
front of whichever TTS engine wins the engine decision: POST /v1/audio/speech
with voice=<name> answers audio in THAT voice, where a name is real exactly when
/voices/<name>/reference.wav exists — an unknown name is a 400 listing the
voices that are, never a silent fallback (a wrong-voice endpoint stays wrong for
weeks). voice_id works as the alias the omp harness sends; response_format=mp3
is really mp3 (ffmpeg re-encode); oversized input is refused with 414 instead of
half-narrated; /health reports the backend and the resolved voice list.
Two backends behind the one surface, chosen by TTS_BACKEND: omni-http forwards to
an OpenAI-shaped upstream carrying the voice as base64 reference_audio, and
llama-tts-cli runs a local binary per request with --tts-speaker-file at nice 19
so narration never starves the text endpoints. Both are driven in the seam
through stubs — no GPU, no containers.
Named frank-tts, a first-class fleet model endpoint: port 7772 in the 7770/7771
family, alias naming per the StarCraft law, rootless boot-persistent self-healing
quadlet with a read-only /home/dank/voices mount. Voice-pack law recorded in the
env header: reference.wav PLUS the engine's own speaker artifact (required, so a
voice is reusable in other runtimes), plus manifest.json provenance with offsets
always past the first 10 s of the source file.
Not deployed: the engine decision, the pack builder, deployer/homepage/Prometheus
registration and the omp models.yml runner row are phase 2.
Frank's 02:36 instruction (archive 01a14c93) ordered the 27B PTQ1_0 arm at
2 x 200,704 with the Q8 mmproj and batch/ubatch 3072/1536; it served
production 03:14:39-03:38:57 (pid 191358, journal type ptq1_0: 402 tensors).
Recovered its numbers from that process's own print_timing lines: ~48 t/s
single-stream decode, 25-27 t/s per stream at two streams, prefill 1,295 t/s
at 30,637 -> 943 at 97,856 -> 801 at 141,365. It bought the promised 1,201 MiB
and lost 37% of decode against the PQ2_0 record, so the publisher's Ampere
warning generalizes to GA102; Frank scratched the 27B plan at 03:33 for the 9B.
Every 'unfinished experiment' claim is re-keyed, and the AGENTS.md law gains
the durable lesson: a bench verdict that never reaches the repo is
indistinguishable from a rogue arm.
Deployed 2026-10-18 (e2e 33/33). The boot ledger prints
`reserving the KV pool, 401408 positions (about 11.0 GiB)` and
`62.9 GiB of weights locked … 11.0 GiB of KV pool, 5.7 GiB of working memory,
79.6 GiB in all`, so the ~10.8 GiB that every page quoted from the
~28 KiB/position arithmetic is replaced by the reservation the engine actually
makes (11.0) and its preflight bill (11.3). Host MemAvailable fell 26 -> 22 GiB
across the raise: the FLAGS.md collision (pool RAM is n-gram page cache) paid in
full, so the warm prompt-cache ratio is flagged as the number to watch.
All "committed but not yet deployed" clauses retired: AGENTS.md topology row,
KV-pool law and memory ledger, frank-dragoon.env lore, both index pages, the
iommu page, and the runbook chronology (third production boot recorded).
The sweep also caught the LLM index still describing frank-zealot as
Bonsai-2-27B at 2 x 200,704 in its intro and topology diagram - re-keyed to the
live 9B: 8 full-attention + 24 Gated DeltaNet + blk.32 MTP, 3 x 262,144,
tensor type 141.
The cutover deploy's FIRST boot was not the shipped payload: a stale
frank-zealot.service.d/ptq1.conf from the bench window still pointed
FRANK_ZEALOT_ENV_FILE at /tmp/zealot-arm-current.env, so the unit booted the
unmeasured 27B PTQ1_0 arm (3 slots of 133,888, 21,188 MiB) while /health
answered 200, /v1/models echoed the alias and the e2e gate passed 33/33.
Drop-in retired to ~/.AGENTTRASH/20261018-zealot-ptq1-arm/; the 13:19 restart
is the shape the docs now quote (n_ctx_slot 262144, KV 13,056.00 MiB,
settled 17,922/6,203).
Corrections the production ledger forced:
- RS buffer is 150.75 MiB (3 cells) on the shipped spec-off shape; 452.25 MiB
belongs to the SPEC-ARMED arms, where MTP depth 2 bills three seqs per slot.
The drafter triples the recurrent state on top of its 3,072 MiB of f16 KV.
- Decode measured 168.4-168.7 t/s single-stream and ~300 t/s aggregate over
three streams (102.8 each). Wall-clock timing of a generation that stops on
its own reads 4x low (72 t/s for a 24-token countdown) - measure with
ignore_eos and read the engine's timings.
- OpenWebUI stores both decoration values inside JSON columns (params/meta),
and frank-dragoon's context.length had drifted to 262,144 against its own
200,704 ceiling; both re-keyed in the live DB.
New law in AGENTS.md: after deploying a recently-benched LLM service, check
DropInPaths, then /proc/<pid>/cmdline, then the ledger. Health checks cannot
see the payload.
Swap the 3090 endpoint from Ternary-Bonsai-2-27B PQ2_0 (2 x 200,704,
fork build 10754) to Ternary-Bonsai-2-9B PQ2_0 with the MTP head IN the
file, three conversations at the trained 262,144 window
(--ctx-size 786432 --parallel 3), on vendor fork build 10777 -- the
minimum build that pairs MTP with a quantized V cache. q8_0 K + q8_0 V
is the only KV that fits: 13,056 MiB against a 26,112 MiB f16 pair.
Settled boot 17,876 used / 6,249 free.
MTP is armed and then DECLINED by measurement: the draft context bills
its own f16 KV over the whole pool (3,072 MiB) and at UBATCH_SIZE=2048
that halves decode (167.2 -> 109.2 t/s). The documented MTP arm
(SPEC_TYPE=draft-mtp + UBATCH_SIZE=1024, 213.5 t/s) is recorded in the
env header rather than left to rediscovery.
Two bugs the trial caught: ${SPEC_PRESENCE_PENALTY:=0.0} was a
default-assignment, so the env value always won and the no-presence-penalty
MTP arm was unreachable (now ${VAR=0.0}); and the Q8_0 mmproj is the
publisher's requirement for concurrent image requests, which matters at
--parallel 3.
contextWindow is now 262144 -- the widest in the fleet -- so :7771 is the
compaction hub. 27B runbook re-banned RETIRED with its bench-window
section closed; indexes, homepage card, OpenWebUI doc and the worklog
re-keyed.
frank-dragoon's total in-flight context was HALF frank-zealot's: zealot runs
one 401,408-position pool split into two 200,704 conversations, while dragoon
ran ONE 200,704-position pool behind its two slots, so a single maximum-length
request filled it and two long conversations had to stay under ~100k each.
HALOGEN_KV_POOL_POSITIONS 200704 -> 401408 = slots x ctx, the top of the seam's
legal band (ctx <= pool <= slots * ctx). The PER-REQUEST ceiling does not move:
HALOGEN_CTX stays 200704, because the checkpoint's native window is 262,144 and
the engine's own 400 text refuses to exceed it without HALOGEN_ROPE_YARN. What
this buys is concurrency, not a longer conversation, which is why the harness
contextWindow (155648 = ceiling - 45056) is untouched.
Cost: ~10.8 GiB of KV where 200,704 cost 5.5 (262,144 cost 7.2; the
2026-10-06 shape's 524,288 measured 14.6), taken from the page cache the 51.2 GB
n-gram table reads through — the warm prompt-cache ratio is the number to
re-measure after the boot. Engine-side total goes 74.1 -> ~79.4 GiB of 121.2.
Also carries scripts/download-frank-zealot-bonsai2-ptq1.sh: the PTQ1_0 candidate
payload fetcher (5,946,648,928 B, LFS sha256-verified, mmproj hardlinked), with
the publisher's own throughput table in its header — the packing is smaller but
the card measures it SLOWER on Ampere, so the swap is decided on our card, not
on the card's A100 row.
The scrape has never worked: lastError is "server returned HTTP status 501 Not
Implemented" on http://host.containers.internal:7771/metrics, and the oldest
retained sample of up{job="frank-zealot"} (2026-10-11) is already 0 -- the target
has been red across four payload generations while /health answered 200 the whole
time, keeping the critical LlmEndpointDown alert permanently active. llama.cpp
gates /metrics behind --metrics; halogen serves it natively, which is why the
asymmetry looked like a 3090-only problem.
METRICS=1 in the shipped env, --metrics reaches the argv through EXTRA_ARGS, and
the seam asserts both directions (default ships the flag, METRICS=0 strips it).
Diagnosed via /api/v1/targets lastError rather than the alert list, which is also
the lesson: Alertmanager here has no persistence, so a reboot re-stamps startsAt
and a weeks-old fault looks fresh.
/tmp is tmpfs: the 2026-10-18 IOMMU reboot deleted /tmp/halogen-eval mid-A/B and
the off-arm numbers nearly went unreproducible. One command restores the corpus
(wikitext-2 raw test, blanks dropped, 2891 lines / 1287656 B -- the byte count is
the identity check because battery.py sizes haystacks in characters), the 512px
probe PNG, and battery.py itself; WARM=1 also pre-reads the 51.2 GB n-gram table
so a cold page cache cannot masquerade as a measured regression. Must run through
a login shell: uv lives in ~/.local/bin.
The pending-entry grep anchored on "menuentry 'Ubuntu'" with -A4 and matched no
linux line, so the script printed nothing where it should have shown the next
boot's command line; the PENDING line then said "is in grub.cfg" when the loop
only ever compared the tracked file against /proc/cmdline. Now the default
entry's linux line is parsed with awk (first linux line of the first Ubuntu
menuentry, which is what GRUB_DEFAULT=0 selects), its iommu token is printed,
and a mismatch against the tracked file warns by name.
/etc/default/grub was host-only state while the IOMMU boot line was the most
consequential kernel decision on the box (the 2026-09-11 amd_iommu=off adoption
lived only in the decision record). Fetched byte-exact (1581 B) and paired with
scripts/set-dankclaw-bootline.sh: refuses any host but dankclaw, shows the diff,
installs 0644 root:root, runs update-grub, names the pending token, and reboots
only on --reboot. The deployer plans nothing under dankclaw/host/ by design —
kernel state is not a container restart.
Frank's call the same evening as the native-payload cutover: HALOGEN_CTX and
HALOGEN_KV_POOL_POSITIONS both 262144 -> 200704 (a multiple of 4096, the rung
frank-zealot already runs). The checkpoint's native window is 262,144 and the
engine serves anything below it, so this is a RAM decision, not a capability
limit: the KV reservation drops 7.2 GiB -> ~5.4 (the 2026-10-06 pool measured
14.6), and by the FLAGS.md law -- RAM the pool takes is RAM that n-gram table
loses -- the freed positions become page cache for the 47.68 GiB table that
long prompts read. Engine-side total ~74 GiB instead of 75.8.
The 400-zone law is invariant to the value, so the harness moves WITH the
ceiling: models.yml contextWindow 217088 -> 155648 = ceiling - 45056, keeping
the gap that absorbs the chars/4 under-count (worst-case request ~156.2k vs
the 200,704 ceiling, ~44k margin). Compaction now fires at 85% = 132,300 est
(~136k measured); the 180k target of the 262,144 era is unreachable because
the window itself is 155,648. maxTokens 45056 and the coupled
cap/timeout 65536/7200 are untouched.
Cost stated plainly: with pool == ctx, ONE maximum-length request fills the
pool, so two long conversations never coexist (two of <=100k do). Acceptable
because frank-dragoon carries only the slow role and the security-reviewer
subagent, not the interactive workhorse.
Seam re-keyed: the fit-law test now proves the invariant at the new pair, and
the pool law (ctx <= pool <= slots * ctx) still holds with pool == ctx.
The three artifacts that produced the native-vs-swift15 verdict, so the next
payload swap starts from instruments instead of ad-hoc scripts:
census-frank-dragoon.sh — the engine's own checkpoint tools (verify/ppl/niah)
over BOTH payloads under one knob set; the pool stays 524288 deliberately
because the ppl pass scores 297,053 ids as two 262,144-token sequences.
battery-frank-dragoon.py — behavior battery against the live service: refusal
census, thinking-token census, tool-call turns, the red-PNG vision probe,
needle-in-haystack, decode/prefill/prompt-cache rates; discards the
post-boot warmup request and records the engine's 400 body (which names the
exact token overage) instead of dying on it.
trial-frank-dragoon.sh — the guarded candidate window: fetch gate, verify,
stop+mask, census, boot through a transient FRANK_DRAGOON_ENV_FILE drop-in,
battery, trap-guarded restore.
Quat3rnion/halogen-qwen3.8-flash-next-v2-abliterated is now the NATIVE
2026-10-11 v2 conversion of orcarouter's pinned BF16 (66,687,678,432 B,
LFS sha256 b5bd925f...) — 32 bytes smaller than the writer transplant it
used to ship under the same filename. The n-gram table, vision tower and
tokenizer are the stock peonist companions, hardlinked out of the swift15
payload dir: the release publishes only the main file and no source-matched
pair exists, so the engine's own tools are the referee (niah 18/18, ppl
3.968 with mean KL 0.155 against the swift15 logits, 2026-10-17).
Memory saving, per the card: HALOGEN_KV_POOL_POSITIONS 524288 -> 262144,
one pool behind both slots (boot ledger: 7.2 GiB, where the old pool
measured 14.6). HALOGEN_CTX stays 262144 so the per-request ceiling and the
harness contract (contextWindow 217088) do not move; only two MAXIMUM-length
conversations lose the ability to coexist.
Measured against the swift15 boot on identical traffic: cold prefill
+12.6 %, decode 50.0 vs 46.8 t/s, NIAH 18/18 vs 17/18, PPL +2.3 %, thinking
tokens x1.9 (the finetune's reasoning compression is gone), refusals 0/5 on
both.
Quat3rnion/halogen-qwen3.8-flash-next-v2-abliterated @main replaced the file
on 2026-10-11: a complete HIP native-v2 conversion of orcarouter's pinned
BF16 (66,687,678,432 B, sha256 b5bd925f), not the r1/r2 writer transplant
the filename used to carry. The release publishes ONLY the main file, and its
card states that the stock peonist n-gram/vision companions are NOT established
as interchangeable (source-derived table would be 51,200,246,084 B against
peonist's 51,200,246,144 B), so the script hardlinks the verified peonist trio
into the new payload dir — a hardlink is a regular file, so the serve gate's
no-symlink law holds and 47.7 GiB is not copied — and leaves the companion
question to the engine's own verify/ppl/niah tools plus the boot ledger.
- serve script: header + every fallback default re-keyed to the Bonsai-2 trio
(vendor LLAMA_BIN, PQ2_0 MODEL_FILE, Q8_0 MMPROJ) and CACHE_TYPE_*:=q8_0 —
a defaulted run must land on the shipped shape, never on the retired 9B/f16
pair (the trial window proved a silent default is how a swap goes wrong)
- unit Description now names the payload + KV dtype that `systemctl status`
will show during the next incident
- OpenWebUI decoration-row description re-keyed, with the law that the row is
live DB state a swap must follow into the database
- env: Bonsai2 payload pair, CACHE_TYPE_K/V=q8_0 (f16 OOMs at 2x200,704:
34 KiB vs 64 KiB/token over the 16 full-attention layers), measured boot
ledger 22,664 used / 1,461 free incl. the 873 MiB mmproj that lands AFTER
the engine fit check, q4_0 fallback rung, CRACK one-line MODEL_FILE swap
- deployer: runtime/model/mmproj gates re-keyed to the PrismML vendor build
(the only loader for GGML tensor type 142 + prism.hadamard.* metadata)
- seam: payload trio pinned; KV argv assertions now derived from CACHE_TYPE_*
in the env file — the wrong-var-name f16 fallback that produced the first
failed trial window can no longer pass silently
Model swap on the radiance endpoint: MODEL_FILE moves from the in-house
qwen4exp-swift15-w4dn8-i8-hc8m.rad (116,416 MiB) to the engine author's
published qwen3.8-next-flash-fp8-iq4r-moe.rad (116,322 MiB, official
Qwen3.8-Flash-Next bf16 -> 4-bit non-uniform codebook experts + int8
trunk, card KL 0.0789 / top-1 92.85 % vs bf16). NO argument changed:
25600/5120 vram pair, 44928 host pool, fp8 KV, 3 x 200000, MTP depth 2,
wht6, MNBT 2048, froggeric v22.5 — the 2026-10-13 shape carries verbatim.
New fetch script scripts/download-frank-archon-iq4r.sh (hfget + exact
byte count 121,969,901,568 + LFS sha256, --verify-only default, ungated,
no token on the command line). Unit/serve/deployer lore re-keyed; the
seam pins the new payload id + sha256 and keeps the w4dn8 rollback
provenance. Homepage card points at the publisher repo.
Rollback: revert this commit (w4dn8 .rad stays on disk). The fp8-native
rad was deleted from dankserver 2026-10-15 (Frank's one-time rm, volume
for this payload); rebuild path documented.
Measured, 512 tokens, greedy, cold, at the deployed 3 x 200,000 fp8 shape:
prefill 22,208 -> 3,397, 24,000 -> 3,290, 25,600 -> 3,290 t/s at 43.5k —
the third slab step bought nothing, so the slab is a threshold whose cliff is
the 2026-10-11 arm at 26,071 (4,379), not a slope. Decode ladder: 99.0 t/s
single-stream at 43.5k, 93.3 at 197k (nearly window-free), and 3 concurrent
streams 113.9/43.3/10.9 = 18.4 t/s aggregate — the slot count is paid in
decode concurrency, the axis the shape hunt never measured because probe leg
B generates 24 tokens. probe-frank-archon-kv.py gains leg D (N-stream decode,
512 tokens, per-stream + aggregate) so no resize is priced on prefill alone.
Measured A/B on the shipped pair 25600/5120: A-cold 43.5k prefill 3,290 ->
3,584 t/s (+8.9 %) and the 3-way 197k aggregate 3,051 -> 3,589 t/s (+17.6 %,
wall 193.7 -> 164.5 s). The prediction I shipped with the arm -- slab is a
threshold below ~26 GiB, so 1,600 MiB buys ~0 -- is falsified by the measurement
it asked for. Two slab points inside noise of EACH OTHER (22,208 / 24,000) are
not a plateau. The cliff is above 25,600: 26,071 measured 4,379 t/s.
Ceiling re-derived from the engine's own budget table (card 32,624 MiB, pair at
29,120 settling 2,058 MiB free against its 209.19 MiB floor => 30,969 claimable).
The 29,120 bound was my 1 GiB \`keep free\` reserve; the seam now derives the
number, so the next resize starts from the engine's arithmetic.
Ledger for the record: slab 16,235 -> 17,548 slots while the pinned host pool
hands the same experts back (8,893 -> 7,580) -- the transfer is the mechanism,
the pinned pool is the slow path. Capacity contract live-verified at 3 slots:
3 x 199,984-token prompts accepted (599,952 tokens), fourth refused as too
large to ever fit.
Probe law learned: the oversized-prompt refusal names the prompt's token count
BEFORE prefill (free calibration) and fires on prompt tokens ALONE, not on
prompt + max_tokens; the window is ~514,286 chars at 2.5714 chars/token.
The 2 GiB/card that \`rocm-smi\` shows free is not idle: the engine's budget
table calls it driver context + the 896 MiB activation arena + 96 MiB headroom,
and says the rest is `throughput on the floor`. Its own settle line is the
number that matters — `card 0 settled at 2.01 GiB free against 209.19 MiB (the
96.00 MiB --gpu-headroom-mib and 113.19 MiB kept for the static region)` — so
1,849 MiB above the ENGINE's floor is claimable and the real ceiling is 30,969
MiB, not the 29,120 my 1 GiB reserve invented on 2026-10-11.
Shipped pair 25600/5120 = 30,720: +1,600 MiB of slab, ~458 MiB above the floor.
The slab is already saturated at 24,000 (mover took all of it, 451 KiB slack),
so this is the only direction left — KV is at its 3-slot floor (5,063 needed of
5,120).
Named prediction, because today's two points say the slab is a THRESHOLD and not
a slope (22,208 -> 3,397 t/s, 24,000 -> 3,290 t/s at 43.5k single-stream): this
arm buys ~1,600 resident expert slots and ~0 prefill. The point that moves
prefill is 26,071 (2026-10-11: 4,379 t/s), fundable only at two slots. Prefill
moving > +-3 % here falsifies the threshold reading.
Seam: the ceiling gate now derives from the engine's budget table (30,969) and
is the single owner of the pair arithmetic; the duplicate I put in the shape
test is removed.
Frank's call: three slots, and spend the KV the third slot doesn't need on
resident experts. `--max-num-seqs 3`, `--vram-kv-mib 6912 -> 5120`,
`--vram-weights-mib 22208 -> 24000` — the pair is still exactly the measured
29,120 MiB/card ceiling. The pool funds 608,027 tokens against the 600,000-token
contract (3 x 200,000, +1.3 %), and the slab gains 8.1 %.
The direction is the 2026-10-11 A/B: prefill is slab-limited, KV beyond the
contract buys nothing. At fp8 the slab curve measured on this host is
22,208 -> 3,397 t/s and 26,071 -> 4,379 t/s at 43.5k, i.e. ~0.25 t/s per MiB, so
24,000 predicts ~3,850 t/s single-stream. Falsifiable; the probe is the check.
The cost is named, not hidden: a FOURTH full-height conversation no longer fits
(4 x 200,000 needs 6,601 MiB of pool) and is refused as too large to ever fit —
radiance never truncates. Reverting to four slots is one env pair.
Seam: two new arithmetic gates, because both failures are boot-then-400 shape
faults that no flag check catches — the pair must respect the per-card ceiling,
and the pool must back seqs x max-model-len at the dtype's measured rate plus
the dtype-independent terms (110 MiB a sequence of recurrent state, 448 MiB of
prefix checkpoints).
a9da65e put the raise inside the serve script, which runs as `User=dank`: the
knob is root-owned 0644, so `-w` is false and the write silently failed — the
gate printed its refusal on a cap it had just failed to move. The privileged
path is systemd's `+` prefix: `ExecStartPre=+scripts/ttm-preflight.sh --raise`
runs as root even under `User=`, before ExecStart, so `Restart=always` can
never outrun it. The lib is now also executable (`--raise` sources the env file
and verifies), and keeps a `sudo -n` fallback for a human running the gate from
a shell.
Seam: the refusal cases run with a PATH that has `cat`+`bash` and NO `sudo`
(root writes a 444 file, so a passwordless dev box would have tested nothing),
and both now assert the gate's own exit 1, not merely non-zero. New seam covers
the unit line and the `--raise` entry point end to end (writable knob -> raised,
no args -> usage exit 2).
The TTM_TARGET_MIB guard I added on 1a6a78e was written `:"...` (no space), so
bash parsed the expansion as one word and executed `cmd :204800` as the
command `:204800`: exit 127 `command not found` on every start, before the
ttm gate ran. Caught on the deploy: the unit crash-looped 34 times in ~4 min
with exactly that journal line.
probe-frank-archon-kv.py now reads the slot count, window and KV dtype from the
deployed env on the host, so probe B exercises the capacity contract that is
actually live (4 x 200,000 today) instead of a hardcoded pair, and each probe's
marker starts with a unique word because show() scores the hit on that word.
The 2026-10-12 23:17 power failure reset /sys/module/ttm/parameters/pages_limit
from the manual 2026-10-05 raise (200 GiB) back to the kernel default (half of
RAM = 125,988 MiB), below the pinned-RAM ledger floor of 150,000 MiB. The serve
script's preflight named the refusal and exited 1; Restart=always re-ran it
6,919 times over ~20 hours and frank-archon served nobody.
scripts/ttm-preflight.sh (sourced by the serve script) now RAISES the knob
itself before refusing: the unit is root, the knob is a 0644 root module
parameter, and the cap is a runtime knob whose default is half of RAM — so a
manual sysfs write is not a repair procedure, it is a countdown to the next
reboot. FRANK_ARCHON_TTM_TARGET_MIB=204800 records the host ceiling; the gate
re-reads the cap after writing and only then refuses, so a knob the kernel
will not move is still a named failure.
Shape change (Frank's call): FOUR decode slots x 200,000 tokens with fp8
(quantized) KV. The pool must back seqs x max-model-len, so fp8 is what funds
the fourth slot: at bf16 the four slots cost 11,679 MiB of pool and would cut
the resident expert slab to 17,441 MiB (-33 %); at fp8 they cost 6,601 MiB, so
the pool stays 6,912 MiB (predicts 843,566 tokens), the VRAM pair stays at the
measured 29,120 MiB/card ceiling, the weights slab and the 44 GiB host pool are
untouched, and the fused TP all-reduce rides 4 x (1+2) = 12 rows, inside the
25-row bound where the wht6 wire is bit-exact.
Runbook section `2026-10-11 — unquantized KV at 2 x 200,000`: the dtype law
(only bf16|fp8, no K/V split; the QSA block key is dtype-fixed, which is why
bf16 costs 1.89x and not 2x), the pool formula that reproduces both live
boots within 0.1%, the seqs-3-to-2 provenance (the tcclaviger fork's
torch.compile constraint, not radiance's), the boot ledger, and the A/B.
The A/B, fp8 control arm through a transient FRANK_ARCHON_ENV_FILE drop-in:
prefill -29% (4,379 -> 3,126 cold), TTFT at the full window +25% (the 2 x 197k
pair: 110.5 s -> 138.1 s), decode inside the +-3% band on two of three passes,
draft acceptance UP 0.661 -> 0.740. The price is prefill, not decode. Open
call for Frank: keep exact KV, or take the cheap fp8 2 x 200k fit (pool 3,648).
Also corrects two stale index entries (w4dn8 said STAGED, fp8-native said
ACTIVE), adds the worklog entry, and commits the verification tool as
scripts/probe-frank-archon-kv.py — it carries the probe law that cost two
false negatives: radiance answers in reasoning_content, so a chat probe with a
small budget reads as empty content + finish_reason stop on a healthy server.
Tests 368/368.
The KV pool at the fp8 shape backed 386,904 tokens — 1.93 sessions of the
served 200,000 window, so a second full-height conversation preempted to the
host/disk prefix tiers. Moving the cache to bf16 (the engine's only
unquantized rung; `--kv-cache-dtype` takes bf16|fp8 and has no per-tensor
K/V split) doubles what a token costs: 7.3125 -> 14.625 KiB per rank
(kv_attn 26.00 -> 52.00 KiB per 4-token page + kv_qsa_bkey 3.25 -> 6.50).
Pool 3543 -> 6912 MiB (437k tokens, >= 2 x 200,000), and the weights slab
pays for all of it: 26071 -> 22208 MiB, the pair pinned to the measured
card ceiling 29,120 MiB (card 31,856 - 1,710 consumed outside both budgets
- 1,024 free). The resident set that leaves VRAM spills into the pinned host
pool, so 40960 -> 44928 MiB, and the prefix host tier doubles with the entry
size (58.50 -> 117 KiB): 4096 -> 8192 MiB. Pinned ledger ~147 GiB, ttm floor
140000 -> 150000 (cap 200 GiB).
--max-num-seqs 3 -> 2: two slots x the window is the capacity contract, and
the "seqs 2 cannot boot" law is the retired tcclaviger fork's torch.compile
query_start_loc constraint, not radiance's (radiance booted seqs 2 in the
fp8-native era; the 2026-10-06 sweep had 2 and 3 tied inside +-3%). Bonus:
the fused all-reduce is now declared over 2 x 3 = 6 rows, under the 25-row
bound at which the wht6 wire is bit-exact.
Cost, paid twice on this engine: idle KV VRAM is also the expert cache's
spillover (lends 2.69 GiB + 740 MiB of arena), so the pool steals weight
residency from the slab AND the lendable cache. Priors: 2.14 GiB of stolen
weights cost 2.5% decode (MAX_MODEL_LEN 262144), the EVC ladder's other
direction cost 8.9%. Prediction -3..-8% vs the fp8 control (136 t/s at 44.5k,
acceptance 0.7246); the deploy boot ledger and live probes are the referee.
Seams: MAX_SEQS/HOST_POOL/KV_DTYPE/PREFIX_HOST literals follow the shape, and
two new invariants guard it — the KV pool must back seqs x max-model-len at
the declared dtype (an engine hard limit: an over-long request is accepted
then refused as too large to ever fit; the formula reproduces the fp8 boot's
386,904 to within 0.1%), and the VRAM pair must stay inside the ceiling.
The 2026-10-09 wire aliases become the service names everywhere:
systemd units, serve scripts, env files and FRANK_* variables,
deploy-plan/deploy-remote service keys, container names, Prometheus
job names + alert rules, Grafana dashboard queries, bats seams,
runbook filenames, homepage/AGENTS lore, and historical prose.
Rollback tags renamed in lockstep. Residue in the tree: zero.
Retired frank-medium/small/npu names are history and stay.
Known casualty: scripts/patches/*.patch context lines were rewritten;
those retired-era vLLM-fork patches no longer apply byte-exact.
The 2026-10-08 rename moved the harness model ids and the homepage cards;
the served aliases on the wire stayed frank-primary/frank-large/
frank-secondary. This closes the gap: the alias each endpoint announces is
now the family name, while the services, units, scripts, env files and
runbooks keep the old names (container identity stays stable).
- scripts/frank-secondary.env ALIAS, scripts/frank-large.env
HALOGEN_MODEL_ID, scripts/frank-primary.env FRANK_PRIMARY_SERVED_NAME
(+ the two serve-script defaults/gate messages)
- seams re-pinned: frank_secondary/frank_large/frank_primary unit bats
- fleet harnesses retargeted: bench-probe, deep-probe, calib-push,
fleet-probes-swift15, autoresearch(+measure), bench-threads-batch-http,
bench-decode
- shared/homepage/docs.html: api-id annotations and the "switchover will
rename the ids" footnote removed — the page's names are now the ids;
opencode/Pi/curl examples send frank-zealot
- dankclaw/hermes/config.yaml.example: frank-large -> frank-dragoon and the
live 262144 context (Hermes routes by served id; see Hermes.md)
Consumers with live state re-keyed out-of-band in the same window:
OpenWebUI (model row, access_grant, openai.api_configs model_ids) and the
Hermes config.yaml.
The model cards become the StarCraft trio: frank-large -> frank-dragoon,
frank-primary -> frank-archon, frank-secondary -> frank-zealot (family
names only; endpoint/API ids unchanged). The three cards fuse into one
Model Information component: per-unit Protoss accent colours and hex-
clipped unit art, 15px type, clickable StarCraft unit-voice quotations
(original .ogg files from the fandom wiki, served from /sfx/), supervision
notes (dragoon/archon safe for autonomous use, zealot tight supervision),
compaction-model recommendation for frank-zealot, and uncensored-model
legal notes rendered as hover tooltips instead of wall text. docs.html
carries the same renames plus api-id mapping.
App-card updates: Seerr, Calibre-Web, RomM, Jellyfin set to WIP: DOWN
with Open links removed (new wip "down" badge state); Copyparty notes it
serves all other media apps content including WIP apps; Gitea and Plex
noted as needing logins separate from SSO; SearXNG noted as obfuscated,
not anonymous (upstream API keys tie to Frank).
Plumbing: Caddy @data allowlist grew /sfx/*; deploy_homepage copies sfx/;
homepage_config.bats pins the new probe names, the component fields,
art/voice-file presence, and the wip-down schema.
The harness clamps max_tokens to min(45056, contextWindow - est - 4096),
its chars/4 est under-counts the rendered prompt by ~3 % (measured:
152,938 -> 157,388 and 158,906 -> 165,301), and compaction fires only at
85 % -- so with window == engine ceiling 200192, every request in the
est ~146k-184k zone was a guaranteed 400. HALOGEN_CTX=262144 (native) +
KV pool 524288; harness window 217088 = ceiling - 45k absorbs the
estimate error (worst-case request ~219k <= 262144). models.yml updated
out-of-repo: contextWindow 217088, maxTokens 45056.
- style.css: green phosphor -> amber CRT-AMBIOS palette; docs styles for
screenshots, code blocks, key warning (full + bare mode)
- index/admin/docs: VT323 from the Google Fonts CDN
- docs.html: rewrite #chat/#models for the live fleet (flash-server v2,
radiance native FP8, Qwen3.5-9B Q6_K); new #api-access with OpenWebUI
key walkthrough + Opencode and Pi configs; three OpenWebUI screenshots
- services.json: AI card details match the current fleet
- deploy-remote.sh: ship shared/homepage/img/ to /srv/homepage/
- docs: Homepage.md theme wave + OpenWebUI.md picker access-control law
rad_arch_declare rejects the split tier (down block-fp8 + gate_up w4): the
qwen4exp_fp8 kernel serves one encoding per expert table (w4nl64a8h) with only
whole bf16 experts beside it. Recipe now sends gate_up AND down through the
GPTQ w4 rule (down calibrated by the Grams, gate_up uncalibrated), drops the
down-fp8 rule, and lowers the size floor 126000->108000 MiB for the ~19 GiB
lighter container. Convert script: checkpoint lives under models/ukisai/, and
the in-container path must keep the subdir (basename stripped it).
Deploy verified: nvidia-smi 21,436 MiB used / 2,689 free at ubatch 2048 -
the MiMo-priced projection (~21,430/~2,700) landed within 11 MiB, and the
verbosity-4 ledger confirms the KV law to the MiB (CUDA0 KV buffer
12,544.00 MiB = K(f16) 6,272 + V(f16) 6,272 = 32 KiB/token over the
401,408 pool), compute buffers 1,136 MiB CUDA0 + 848 MiB host, mmproj
879 + 248 MiB. Playwright gate: 33 passed.
Two lessons recorded: the 203/EXEC boot loop my shebang-less rewrite
caused (journal signature + the head -1 check), and that llama.cpp's
"n_ctx_seq < n_ctx_train" line is informational when the context is BELOW
the trained window - the Gemma4 drafter warning was the opposite case.
The unit's Description named the Swift-1.5 era (and said 2x131072/q8_0/
draft-mtp); it now names the Q6_K shape. Cosmetic, so it is NOT deployed
as its own restart: the next frank-secondary deploy relinks it.
Co-Authored-By: omp <omp@franksdank.cloud>
Capacity trade: the 9B dense-hybrid (qwen35: 32 layers = 8 full-attention
+ 24 Gated DeltaNet, 4 KV heads x 256 dim, native 262,144, MTP-less,
natively multimodal) costs ~7 GiB of weights where the 26B MoE cost
~16.6 GiB. The freed room buys back the second conversation slot and
exact KV: --ctx-size 401408 --parallel 2 (two 200,704 windows), f16 K +
f16 V, MTP off (the file carries no nextn/mtp tensors - flag-only enable
is impossible), vision via the publisher's BF16 mmproj, ubatch 1024
(the CUDA decode-optimal value), verbosity 4, publisher sampler recipe
(0.6/0.95/20/0) in models.yml.
Hosts with the MiMo-V2.6-Distill Q6_K boots at this geometry: 21,276 MiB
used / 2,849 free at ubatch 1024, 40,960 tokens of fp16 KV at ubatch
2048. Compaction stays routed here; window math is unchanged (maxTokens
16,384 keeps the worst-case harness request inside the slot).
Gemma4 payload (Q4_K_M + mmproj + MTP assistant) stays on disk as the
rollback; its runbook is retired in place, keeping the attention_k_eq_v
pair-billing law.
Co-Authored-By: omp <omp@franksdank.cloud>
Verbosity 4 paid for itself: the engine's own llama_kv_cache lines give the
exact accounting — the cache bills the K+V pair TWICE under attention_k_eq_v
(6.125 B/elem at f16 K + q8_0 V), reproducing 3,001.25 MiB (5 full layers at
n_head_kv 2) + 306.25 MiB (25 SWA rings at n_head_kv 8, window 1024) to the
byte. The earlier "bill K-only" note was wrong; full layers have 2 KV heads,
not 8.
Deployed boot: 21,868 used / 2,257 free — the f16-K raise (+2,161 MiB KV) is
funded by the ubatch 2048 -> 1024 cut (~616 MiB of compute buffer) plus
variance, so exact K came essentially free. Measured decode 139.9 t/s temp 0;
vision probe returns Red at prompt_tokens 151.
Frank
