The env file first landed with ESPERANTO_TTS_TIMEOUT equal to SHIM_TIMEOUT=600.
Equality is a race: both budgets expire at the same instant, so the case the
number existed to prevent — the worker abandoning a synthesis frank-tts is still
finishing — is exactly what happens on the tail of a long line. Nested budgets
go inside-out, which is the rule scripts/frank-tts.env states for its own
LLAMA_TIMEOUT; the client waits one round-trip longer than the server, so a
frank-tts give-up arrives labelled as frank-tts's error. 660.
The seam's assertion moves from equality to `>` for the same reason, and now
also holds ESPERANTO_LLM_TIMEOUT under upstream's 600 s web-UI patience — above
it the failure surfaces as a UI timeout with no name on it. The scriptwriting
number itself stays at the app's 180 s default: frank-zealot answers an outline
or a transcript well inside it, which corrects the earlier commit message's
claim that both timeouts were raised past upstream's 300 s.
Page registered in the vault index (the docs-completeness seam fails an
unregistered page) and logged in the bundle change log.
What the page owes the reader: the provider contract (frank-zealot for scripts
with a /v1 base URL, frank-tts for voices with 1-4 speakers mapped to the
narrator1/2/3 voice packs typed as free text), why Test Connection and Discover
Models fail cosmetically against frank-tts rather than being a fault, the
encryption key as the precondition for saving any credential, and the honest
blocker — frank-tts is not deployed and nothing listens on 7772, so a generation
run writes its script and fails on the first voice.
Episode destination is stated as interim and cites Gitea issue #7 (human-only,
agents must not work it): the app keeps its named volume, one bind mount puts
finished MP3s under /home/dank/media/podcasts, and deleting that one line is the
whole rollback. Rollback section also covers the full undo, including that the
podman network only disappears once the units are gone.
The dragoon-window provenance bullet still said the zealot arm owed a
measurement — it does not: the numbers are recovered and in the fetch script and
runbook. Adds the frank-tts phase-1 entry (naming law, port 7772, voice-pack law,
384/384 suite, PR #4 merged as 49c4e19, and the measured Gitea merge-field law).
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.
Post-deploy probe on frank-dragoon: one 101,987-token prompt through
/v1/chat/completions took 81.5 s cold (1,251 t/s of prefill) and 0.3 s on two
repeats — HALOGEN_CACHE_INPLACE=1 keeps the KV rows in the pool, so doubling the
pool evicted nothing the prompt cache needed. The RAM price is still real:
MemAvailable 26 -> 22 GiB.
Census of the same engine's surface corrected the Benchmarking Policy, which
quoted llama.cpp instruments at a halogen endpoint: frank-dragoon answers
/health, /metrics, /v1/models and POST /v1/chat/completions and 404s /ping,
/props, /completion and /tokenize, and no response carries Server-Timing. The
/completion `timings` object and the ttft;dur= header are frank-zealot's; on
frank-dragoon the instrument is wall clock plus usage.prompt_tokens.
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.
Window 01:24-01:35. Byte-identical A/B (rebuild proved it: same 1,287,656 B corpus,
same billed token counts 59,024 / 21,568): cold prefill 1,585.1 -> 1,616.0 t/s
(+2.0 %), warm 208,531 -> 214,131, needle hit/hit both boots. The 13-16 % that
justified amd_iommu=off since 2026-09-11 does not exist on this host; the decode
delta (-9.7 %) is instrument noise -- three repeats of one request on the same boot
gave 31.8/46.9/47.8 t/s.
NPU precondition restored: amdxdna bound 0000:c6:00.1 (driver 0.7.0, firmware
amdnpu/17f0_11/npu_7.sbin), group 26 identity, PASID width 10, /dev/accel/accel0
root:render with dank in render. No endpoint deployed, so frank-npu stays retired
-- only its rationale changes. Gates: 0 failed units, 32 groups (31 identity + the
SD reader's DMA quirk), iGPU c5:00.0 identity with ATS, dragoon ledger unchanged
(74.1 GiB, no clamp, listening at 39.8 s).
Deferred on purpose: the pp_dpm_fclk pin (changes iGPU timing for everyone, nothing
co-runs yet) and PerfOpt (7.0.0-38 has zero perfopt symbols, 26.04 caps at 7.0.0-38.38).
Re-keyed: decision record banner/front matter/verification/revert + new execution
section, frank-npu rationale, AGENTS key lesson, both index pages.
Reading patch v2 1/2 verbatim: amd_iommu_enable_perfopt() rehoming needs the
device ALREADY on the identity domain and switches ATS/PRI/PASID/GCR3 off via
the skip_caps path. So an iGPU in identity mode still pays a 1:1 page walk plus
those four capabilities until PerfOpt exists -- iommu=pt recovers most of the
gap, PerfOpt the last part. The first pass claimed pt alone already avoids the
cost; that was wrong and is retracted here and in the posted issue comment
(edit 23).
Also re-keys frank-npu.md, which carried the dead law 'iommu=pt is not a middle
ground' as its retirement rationale: identity mode + SVA is exactly the revival
path (amdxdna binds its own IOPT_TYPE_SVA domain), and the cost is single-digit
percent on this silicon class. Procedure step 2 now demands a local off-vs-pt
measurement instead of inheriting the vendor's 13-16%, whose slow-arm boot line
our record never identified.
The full-suite re-run after the reset fix came back "1 flaky": 12-crawl reached
plex.franksdank.cloud, the guard cleared, and Plex Web — signed out in the
qa-e2e profile — handed the browser to its own hosted sign-in surface, so the
hostname never settled on the vhost (60 s poll, Expected plex.franksdank.cloud /
Received app.plex.tv). settleThroughGuard now takes a HANDOFFS landing list per
vhost; only plex.franksdank.cloud gets one, every other host still has to land on
its own vhost, and the guard assertions (Anubis cleared, no authentik uidField,
status < 500) are unchanged. Worklog and E2E gotcha 6 carry both symptoms: the
ERR_ABORTED reset and the hand-off.
Deploy verification for fcd7630 (22:20): healthy in ~12 s, e2e 33/33, and the
boot ledger settles the shape — 62.9 GiB weights locked (0.8 vision) + 7.2 GiB
KV in ONE 262,144-position pool (6,656 MiB, two slots) + 5.7 GiB working =
75.8 GiB in all, n-gram 47.68 GiB read from disk, prompt cache ON (12 entries),
MTP depth-2, no clamp line. Live probes: 27*59 -> 1593, tool call, red PNG ->
Red. The two KV figures are reconciled (engine preflight FORMULA 7.4 GiB
includes block scratch + per-slot O(1) state; the RESERVATION is 7.2).
Also corrects two stale claims: the retired swift15 rows named 200,192 ctx
(raised to 262,144 on 2026-10-07 inside that same era), and the OpenWebUI
frank-dragoon row's description is now re-keyed in the live DB (backup
webui.db.bak-20261017b).
The Swift1.5 w4dn8 .rad, the GLM-4.7 GGUF, and every derived model cache
were purged from dankserver (trash-staged, emptied with approval; 763 GiB
free). Rollback of the iq4r swap is now an HF re-download of the
byte-exact frank8465 mirror (LFS sha256 09dd9306...). Corrects the w4dn8
runbook header (still said LIVE) and rebuilds both Rollback sections;
log.md carries the purge record.
- production boot reproduced the census (22,664/1,461, fit 21,439 leaving
2,119 >= 1,872, KV 13,328.00 MiB at n_ctx_seq 200,704, n_slots 2), live
probes green (pong + reasoning_content, tool_calls, 512px -> Red, 75.8 t/s),
e2e 33/33
- records the two live-path stragglers the swap had to catch: the serve
script's retired-payload DEFAULTS and the OpenWebUI model row description
(live DB state, re-keyed with a backup, no forced restart)
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.
Probe on the live service: four concurrent 196,960-token chats, 4/4 exact
marker retrieval, 787,842 tokens in 238.4 s (3,305 t/s aggregate prefill = the
single-stream ceiling, so the four streams shared one pipeline and the KV pool
never ran short); prefix-cache repeat answered in 1.0 s with cache_n=43008.
Recorded with the caveat that 8 generated tokens do not make a decode rate.
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 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.
23,120 used / 1,005 free at boot, 763 settled after a vision probe
(encoder buffers resident). Rung cost 746 MiB, not the transferable
616 prediction. Prefill baseline 1,209 t/s @ 3,760 recorded; decode
unchanged as the law says. 2048 = ceiling for this shape.
Cross-cutting rename of the current llama.cpp endpoint on the RTX 3090
(port 7771) from frank-small to frank-secondary:
- Service trio: scripts/{frank-small.env,frank-small.service,
serve-frank-small.sh,download-frank-small.sh} → frank-secondary.*
- Test: tests/frank_small_unit.bats → tests/frank_secondary_unit.bats
- Runbooks: serverAppConfigs-docs/LLM/frank-small-*.md →
frank-secondary-*.md (9 files); the FreeToken runbook stays
freetoken-frank-small.md (it documents the retired CUDA-era service)
- Identifiers: FRANK_SMALL→FRANK_SECONDARY, llm_small→llm_secondary,
LLM_SMALL_*→LLM_SECONDARY_*, deploy_llm_small→deploy_llm_secondary
- Historical references to the OLD frank-small (FreeToken/CUDA on the
3090, and the Dankserver service that became frank-medium) are
preserved verbatim — they describe what the service was called
at the time.
The 2026-10-06 cutover crash-looped frank-large 8 boots (03:09:39-03:11:42):
the engine's pin gate refused to start while host MemAvailable was ~74.7 GiB,
3.34 GiB short of the 78.11 GiB it needs (62.11 GiB checkpoint + 16 GiB floor).
It self-resolved in ~2 min when the dying 27B engine freed RAM, but left the
host with only ~1.8 GiB MemAvailable - a restart would crash-loop again.
serve-frank-large.sh now checks MemAvailable against the pin gate BEFORE
podman run and WAITS up to FRANK_LARGE_RAM_WAIT (default 600 s) for the RAM
to clear, so a transient dip boots once instead of crash-looping; if it never
clears it refuses loudly. The 16 GiB floor is the engine's (not
HALOGEN_HOST_RESERVE_GIB=20, a separate overall-budget knob).
frank_large_unit.bats: FRANK_LARGE_MEMINFO seam + a refuse test. Runbook +
log: the pin-gate law, the incident, and the fix.
The 17:47 Ornith-1.5-9B swap (77decb5) updated frank-small.env and
serve-frank-small.sh but left two consumers on the MiMo-Q6K paths:
deploy-remote.sh's LLM_SMALL_MODEL_FILE/LLM_SMALL_MMPROJ defaults (the
payload-existence gate checked files the unit no longer loads) and
frank_small_unit.bats (stubs + env assertion), which held the whole
bats suite red. Both now point at the Ornith-Q8 payload; docs updated
with the conditional frankbook e2e project.
The AGENTS.md lesson "`--ubatch-size` has an OPTIMUM on this stack" was load-
bearing and over-general: it was measured on the 2x R9700 layer split, and the
same knob on the CUDA 3090 behaves the opposite way (curve climbs monotonically
to 2048). That difference is now the reason frank-small ships ubatch 2048, so
the lesson names the backend it was measured on and points at the counterpoint.
Adds the 2026-10-05 boot-pair A/B to the runbook (+4.6 % prefill at 8,928
tokens for 616 MiB, decode flat, the sub-ubatch prompt as the control), marks
the superseded 21,276/2,849 VRAM row, and logs the day. The old "ubatch 2048
failed the scratch boot" line stays visible as what it was: a predecessor-shape
observation, not a law.
Record the live numbers and the UTIL 1.0 verdict (unbootable: worker init
demands free > budget; ValueError on both ranks before a single weight loads).
Driver used 30.97 of 31.86 GiB; the 0.89 GiB/card margin held through a live
4,096-token generation at ~38-39 tok/s.
Live after the 1ac4d48 deploy: 'expert_precise_memory': True, runtime overhead
5.25 GiB, expert cache space 17.22 GiB / 7522 slots, swap region 13.7 GiB /
5986 slots, GPU KV cache 398,658 tokens / 1.99x, pool 3,150,000,000, MNBT 1024
-- the probe's projection is now production.
Compile cache proven on consecutive boots of that shape: cold 'Compiling a
graph for compile range (1, 1024) takes 179.24 s' / 'torch.compile took
188.26 s'; warm 'Directly load the compiled graph(s) ... from the cache, took
2.143 s' / 'torch.compile took 3.97 s in total' / 'init engine ... took 41.13 s
(compilation: 3.98 s)'. Host dir 1.4 GB / 4,707 files. Time to /health stays
payload-bound (490 s warm), so the claim is ~184 s of inductor/AOT build per
restart, not a proportional boot cut.
Tests: 367/367.
Measured by boot probe on the shape-D base (production stopped, candidate
booted AS the service), plan numbers read off the engine's own log:
runtime overhead 7.60 -> 5.25 GiB (precise gfx1201 TP2 registry)
expert cache 14.87 -> 17.22 GiB
slots 6,496 -> 7,522 (+1,026, +16%)
swap region 13.7 GiB / 5,986 slots
boot healthy in 420 s, no OOM
'--expert-precise-memory': True also appears in non-default args, so a silent
argparse drop is visible. The pool is NOT cut (option A) because the pool is
not an arena lever: 3.15e9/rank keeps shape D's measured 1.99x concurrency,
and a 2.2e9 arm probed in the same window (277,686 tokens, 1.38x) is recorded
as the conservative fallback.
Boot-cost bug found while diffing Frank's 3xR9700 recipe: the image ships no
/root/.cache at all, our boot writes 1.1 GB of vLLM AOT/torch_compile
artifacts (4,665 files) plus 229 MB of comgr JIT kernels into the ephemeral
layer, and podman rm -f discards them — so every deploy/restart/reboot re-pays
'torch.compile took 181.20 s in total' (173.19 s of it one graph). Mounting
FRANK_PRIMARY_CACHE_DIR at /root/.cache persists both and hides nothing.
GPU_MAX_HW_QUEUES=2 added at Frank's direction (real HSA knob); not A/B
measured. Retracted my own recipe-diff claim that cudagraph_capture_sizes
needed pinning: capture is 3 PIECEWISE + 2 FULL graphs, 3 s, 0.31 GiB.
Tests: 367/367.
`deploy.sh <target>` with no `--service` also walks the per-service revision
markers and restarts everything behind HEAD whose paths changed — a drain, not
a sync. That is how the latent plex gate bug surfaced: the backlog fired all
at once and plex's http:// readiness probe, which can never succeed against a
PMS that drops plain HTTP off-loopback, burned its full 600 s and aborted the
run with every later service left pending.
Both facts recorded: AGENTS.md gains the drain law under Deployment, log.md
gains the plex entry with the evidence chain.
index.md intro and mermaid nodes still described 4 x 262144 fp8/bf16 and
frank-small at 2 x 262144; the active-deployment list had no frank-primary
entry at all. Fixed both, plus a malformed GLM list item.
log.md gains the shape-D entry (the SEQS=2 constraint, the arena-lever
measurement, the 2x2 probe table, the measured 1.99x), and the frank-small
entry's stale consequence bullet now points at the resolution.
frank-small runbook: the compaction-fit note is corrected - the 200704 window
moved frank-primary's trigger to 140,492, which fits a 180,224 slot again.
Runbook, both indexes, AGENTS.md (topology row, endpoint section, three new
Key Lessons, one stale SGLang label in the bench list) and the bundle log now
describe the deployed shape: pool 540,672, 180,224 tokens per conversation,
f16 K + q8_0 V, mmproj vision, 21,684 used / 2,441 free of 24,576 MiB, 106.8
tok/s decode, 3,814 t/s prefill.
The claims that mattered:
- three contexts buy 180,224 tokens per conversation, not the native 262,144,
and a longer prompt is a hard 400 - so models.yml contextWindow is 180224
and frank-primary's 183,501-token compaction request no longer fits here
(forward pointer added to the retired SGLang runbook, which asserted it)
- the vision tower bills ~1.4 GiB for a 918 MB file, established by subtracting
two measured boots, not by a formula
- /v1/chat/completions ignores the legacy top-level 'images' field: the model
invents a color; only the image_url content part reaches it
The first crash-looping boot (51 restarts): serve received the host path
/home/dank/models/... but the models tree is mounted at /models inside the
container, so transformers failed 'Repo id must be in the form
namespace/repo_name' before the engine even built. Same conversion the spec
draft already used. Pinned by bats. Root docs index links the new runbook
(seam required it) and the work log records the crash-loop plus the 52 G
payload reclaim.
New runbook carries the roof math, the measured three-arm spec table, the
greedy-text numerics trap, the reasoning-field contract change and the
rollback tag. LLM index and AGENTS.md fleet table, bench policy and runbook
index follow; the frank-small index entries catch up to the SGLang cutover
(entry was missed on 2026-10-01). Work log records the cutover.
bf16 at 2 x 262144 fails with and without --expert-precise-memory. The real
error (captured from the journal this time, container fixed via journald
rather than podman logs):
RuntimeError: Worker failed with error 'Constraints violated
(L['query_start_loc'].size()[0])! You marked it as dynamic but your code
specialized it to be a constant (3)'
- a torch.compile/dynamo dynamic-shape violation, NOT an OOM. Ruled out a
stale compile cache: ~/.cache/vllm holds only modelinfos (16K) and each
podman run gets a fresh container filesystem.
The deltas from the last shape measured serving (fp8 3 x 262144 / MNBT 2048 /
792,069 tokens / 3.02x / arena 6,248) are KV dtype (fp8 -> bf16) and SEQS
(3 -> 2). Restoring the known-good shape; bisect bf16 vs SEQS in a window,
with --enforce-eager as the bypass.
NRestarts climbed to 3, each attempt ~6.7 min then 'Deactivated successfully'
(API server drains after an EngineCore fatal error; Restart=always turned the
silent death into a visible loop). The arena is device-resident, so the budget
decides it: weights 6.28 + pool 6.80 (bf16 2 x 262144) + graphs 0.55 + arena
16.65 (precise overhead 5.25) + non-torch 1.45 = 31.73 of 31.86 GiB -> ~0.13
GiB free at boot, and graph capture is what allocates transiently.
Dropping the flag restores the 7.60 GiB default overhead: arena 14.31 GiB,
total ~29.4 GiB, ~2.4 GiB free. bf16 2 x 262144 (Frank's KV change) stays.
The serve script keeps the opt-in flag; the seam now asserts it is ABSENT.
Frank's call. Payload: ajgazin's abliterated Swift-1.5 UD-Q8_K_XL (29.3 GiB,
SWIFT OPEN LICENSE) -> unsloth/Qwen3.8-27B-GGUF UD-Q4_K_XL (16.35 GiB,
sha256 3f227079...b01e, vanilla/Apache-2.0). NOTE: the endpoint is no longer
uncensored — that was one of the reasons for the Swift payload on 2026-09-28.
New fetcher scripts/download-frank-large-q4.sh (no auth header: repo is public
and ungated; size + sha256 verified after transfer). The speculative drafter
and its fetcher are unchanged.
Shape: --sessions 4 --context 163840 with unquantized KV. gufo exposes no KV
quantization lever (execution_policy.hpp pins kv_cache_storage to kFp16), so
f16 is inherent.
KV arithmetic: f16 KV = 64 KiB/token (16 full-attention layers x 4 KV heads x
256 head_dim x 2 x 2B; the other 48 layers are linear-attention with a
per-sequence fp32 state). 4 x 163840 = 655,360 tokens = 40.0 GiB KV, each
session exactly 10 GiB. Budget ~62 GiB of the 114 GiB GTT.
Seam retargeted (FETCH -> the q4 script, TARGET_STUB, context/sessions pins).
Both of Frank's changes at once.
KV is bf16 (--kv-cache-dtype auto) at two full-size windows (2 x 262144).
bf16 measures 13,839 B/token (bf16 3x241664 boot: 11,047,010,304 B -> 798,280
tokens), decomposing as 12,276 attention + 1,563 mamba/side caches vs fp8's
6,138 + 1,563 = 7,701. 524,288 tokens needs 7.256e9 B; pinned at 7.3e9
(~527,500 tokens, 3,206 margin). Pool cost: 5.68 -> 6.76 GiB per rank.
--expert-precise-memory selects the fitted overhead (5.25 GiB) over the default
(7.60 GiB = fitted + compile/graph-capture headroom). It moves only the plan's
term, so the expert arena goes 6,248 -> ~7,278 slots (~1,025 more) at the
cost of boot-time capture slack. Overhead adds 0.22 GiB/1024 nbt above 4096, so
MNBT 2048 pays nothing.
Seam + docs + AGENTS row + log updated.
The gate failed twice in the same spec for two different reasons and then passed
33/33 on a manual re-run. Future sessions should re-run before treating the
DEPLOY INCOMPLETE banner as a real fleet fault.
Frank