Prime-approved switch of the fleet STT seat (fv-ml1 :8300, LiteLLM ext-stt/
whisper-1, caller talk) from the sherpa-onnx int8 seat to arm B-bf16w of the
2026-09-30 A/B (docs/pfi/parakeet-seat-ab-2026-09-30.md): p50 33/36/42/71 ms
vs the old seat's 187/308/626 measured on the same card today, WER 1.965/3.026
vs the A/B floor 1.97/3.09. All three seat defects fixed: 12-min file 200s
(windowed at 360 s after a GPU 0 OOM on one whole request; the A/B's own
long-form method), no pause truncation, no long-form dropout.
GPU 0 room: gen-small --gpu-memory-utilization 0.48 -> 0.36 (0.46 and 0.40
refuse their boot check; cyberprev+voices hold the card). Its KV is byte-
pinned, so the boot log is token-identical: 670,142 tokens / 2.56x before
and after. Seat rests 2,088 MiB; GPU 0 keeps ~1.9 GB free.
Two runtime landmines documented in the README: NeMo's attention mask is
materialised T x T even under local attention (hence the window), and
httptools 0.8.0 writes a NUL into the HTTP status line that httpx — i.e.
LiteLLM — rejects, so the image ships plain uvicorn with --http h11.
Old seat stopped, not removed: docker stop parakeet-nemo && docker start
parakeet is the rollback.
License: NVIDIA Open Model License (accepted by Prime 2026-09-30); note in
stacks/parakeet-nemo/README.md.
A/B of the live STT seat (fv-ml1 GPU 0, sherpa-onnx int8 v3) against
nvidia/parakeet-unified-en-0.6b, measured on GPU 3 with the seat's own image,
k2-fsa's published unified int8 export, fp32/fp16 exports made with k2-fsa's
recipe, v2 int8, and NeMo 3.0.0 (fp32, bf16 autocast, bf16 weights).
- Seat int8 graph runs on one CPU thread (cpu/wall 1.00, GPU 2-9%).
- unified-en under NeMo: -121/-234/-530 ms vs the seat at 1-3/3-8/8-20 s
(paired, n=120/bin; floor <=6 ms; +50 ms positive control reads +52-54).
- unified-en WER lower in every runtime: -0.7 pp clean, -1.5 pp other,
-3.2 to -4.4 pp AMI (paired CIs exclude 0).
- Seat defects found: hard 400 s input ceiling (HTTP 500), truncation after
a quiet 1.5 s pause, and severe long-window dropouts (int8 v3 only).
- B-bf16w needs +0.8 to +1.5 GB over the seat's 1,690 MiB on GPU 0.
Raw requests, hypotheses, manifests and the full harness under
services/parakeet-ab-2026-09-30/. No deploy; live seat untouched apart
from 240 light test requests.
Prime: Scriberr gets the basic fix, v3 stays (no NeMo 3.0.0 surgery). 0002 moves
from proposed/ into the carried set; scriberr-rebuild now applies 0001+0002 by
default (suffix dropout2) and its memory budget becomes a 5,600 MiB regression
guard (Scriberr is on GPU 3). Live on fv-ml1 1602: scripts rewritten from the
patched embed, a 20-min file at 5,502 MiB with retried_gaps reported.
The ~6.5-9 s first-call-per-bucket autotune lived in the container's writable
layer and died on every recreate. 0.1.3 creates /tmp/triton-cache in the image
owned by 10001 so the named volume intern-decision_triton-cache inherits a
writable mount point, and compose mounts it.
Bucket model PROVEN, not inferred: 2,048-token buckets, 16 up to 32,768. After
one warmed call per bucket, 12 random sizes across 8k-32k were all warm (worst
2.09 s); cold entries cost 6.5-9 s. Full cold warm-up 109 s; warm re-run 17 s.
scripts/intern-decision-warmup: one noul call per bucket, MAX_TOKENS from
/health, two-point live calibration of the tokenizer's linear token model (a
single probe overcorrects and the aim oscillates around the bucket edge),
per-bucket wall times, non-zero exit on a missed bucket. Run it after an IMAGE
CHANGE only; the volume carries ordinary recreates (measured: force-recreate,
then a warmed 32k call answered in 2.11 s).
Acceptance on 0.1.3: JevBench 202/231, hard 83/111, 0 diffs / 924; warm 32k GPU
1 peak 15,218 MiB (budget 15,220; a COLD autotune touched 15,224 once, README
caveat); /decide answers. Artifacts in the acceptance dir.
Prime's ask (via the coordinator): investigate the "Parakeet skips
stretches of speech" finding, including other Parakeet weights.
Investigation only; nothing deployed.
Against ground truth (official SCOTUS transcript, Gutenberg #38916) the
drops are real: production v3 loses 140 / 66 clean words per transcript on
the two public files and ~50 on each private one (Whisper-referenced,
Canary-confirmed; adjudicator 129/129 correct on the calibration). Cause:
the v2/v3 0.6B weights collapse deep inside long full-attention windows;
the encoder output is degraded, the audio alone transcribes fine, and
1.1B TDT/RNNT/CTC and CTC-0.6B never do it. Decoding (CUDA graphs, greedy
variants, max_symbols, beam), slice length, local attention, loudness,
resampling and a noise floor do not fix it. Controls: A-vs-A, silence
positive control (>=15 words 36/36), null control, bootstrap floor.
Proposed patch 0002 re-transcribes >=3 s stretches where the audio holds
speech but no word came out (-80 to -90 % lost words on all four
recordings, lower WER, no invented text, +10 MiB) and adds an explicit
PARAKEET_MODEL_PATH with the loaded model recorded in JSON and ModelUsed.
Reviewed at high effort, all findings fixed; built and tested as
scriberr:local-blackwell-a353078-dropout2, not deployed.
scriberr-rebuild: --patches takes DIR[:DIR...]; embeds and seam-checks
both Parakeet scripts (seam-check --standard for the short-audio one).
Prime 2026-09-30: move scriberr to GPU 3 and extend the Jev endpoint to 32k tokens.
Scriberr holds 0 VRAM idle; verified a 20-min job on GPU 3 at 5,496 MiB. With GPU 1
freed, intern-decision's measured card peak at MAX_TOKENS=32768 is 15,220 MiB against
a 15,437 MiB budget (n=3, 1 and 16 questions); 32,769 tokens is refused 422 up front.
JevBench v1.2.16 via /v1/systemone unchanged: 202/231, 0 diffs vs the bench.
Scriberr moved to fv-ml1 GPU 3 on 2026-09-30, so the card is no longer idle when the rebuild runs. The memory stage now requires >= 20 GB free instead of an idle card (GPUs 0-2 still fail that) and attributes the peak only to the host PIDs of its own container, captured with docker top alongside the 0.2 s nvidia-smi samples. Verified with five processes on the card: peak 5,496 MiB, identical to the exclusive-card figure.
Audit finding on 0.1.1: a Jev client that always sends an images array with no
images was rejected for nothing. Only a non-empty value is 422 now. Also aligns
the contract's response example with the wire (model is the name@revision
string, not an object).
Re-accepted live: images []/null -> 200, ["a.png"] -> 422; JevBench all 202/231,
hard 83/111, 0 changed rows across the bench's r1..r4 (924).
Straight passthrough to the checkpoint's own DecisionEngine.predict — never the
semif mapping, whose different prompt would change the answers. Reuses the one
inference thread, bearer auth, MAX_QUEUE, VRAM cap and error envelope; no new
concurrency. 1..16 questions in ONE call (never chunked: Jev questions share a
prompt); images 422; over MAX_TOKENS 422 before the forward. /health advertises
the surface. Response 'model' is a string name@revision (JevBench's runner
hashes it; a dict broke its manifest step).
Acceptance on the live service (see acceptance/systemone-2026-09-30/): JevBench
v1.2.16 typesafe adapter over the 231 public items scores all 202/231, hard
83/111, with 0 changed answers across all 924 rows of the bench's own r1..r4;
controls 401/422x3 (token boundary proven at 7168 pass / 7169 refuse); GPU 1
per-process peak 9,866 MiB under the largest accepted request (budget 9,876);
/decide/shared positive control unchanged. 120 tests green.
Deployed 2026-09-30 1211 PT by pointing SCRIBERR_IMAGE at the patched tag (.env backed up as .env.bak-20260930-pre-slicer1; rollback is the unpatched scriberr:local-blackwell). PrepareEnvironment rewrote the env's parakeet_transcribe_buffered.py from the embed (sha256 matches the patched source). One live run on GPU 1 beside intern-decision peaked at 5,496 MiB. Memory records the open Parakeet mid-chunk dropout finding and the held upstream PR.
Carry patches/0001 on our Scriberr build (upstream a353078): adjacent
buffered chunks overlap by 4 s inside --chunk-len and hand over at a word
both chunks transcribed alike, instead of cutting at fixed marks with no
overlap. Pause-aware cutting is included as an opt-in (--pause-search);
it measured neutral once the stitch was right. The Go<->Python CLI and
JSON seam is unchanged.
Bench (4 recordings, 118 min, 3 cut placements each, against a no-cut
whole-file reference; metrics only, private audio stays on fv-ml1):
cuts with an error within +-3 s fall from 52% (93/179) to 22% (41/184)
against a 19% background; floor +-0.08. Positive control: upstream's
cutter +0.33 over background. A-vs-A byte-identical in-process and
across CLI processes. Peak GPU memory unchanged at 5,496 MiB (n=3).
Also found: Parakeet skips runs of >=10 words mid-chunk with any
slicer, upstream's included; not addressed here.
scripts/scriberr-rebuild clones a pinned upstream sha into a new
/opt/docker/src dir, git-apply-checks the patches, builds a distinct
tag, and checks embed, unit tests, the JSON seam (scriberr-seam-check.py)
and the memory budget on idle GPU 3. Deploy stays manual. The upstream
PR is prepared under patches/upstream-pr/ and not opened.
intern-decision deployed 0941 PT (cap 9.0 GiB, MAX_TOKENS 7168). Live acceptance: positive
control 240/259 and Wyrd 79/84, bit-identical to the bench (0/560 rows, Δp 0); negative control
10/122/14; largest accepted requests 200 with no 503; per-process 8,812 MiB at rest and 9,866 peak;
GPU 1 Free 15,442 before and 6,581 after (lowest 5,569 under load). Latency from nh3-dev:
21 criteria 114 ms, 16 over ~3,900 tokens 238 ms. semif README banner now REPLACED with the
rollback; fv-ml1 GPU 1 note updated.
Both are required in compose because they are coupled: MAX_TOKENS is checked before the forward
pass, so an oversized call is a clear 422 instead of reaching the cap as a 503. Pre-deploy floor
is nvidia-smi Free >= 15,400 MiB on GPU 1 (card peak 9,876 + scriberr 5,496).
stacks/intern-decision: compose (GPU 1, :8033, hard VRAM cap as the single .env knob,
healthcheck, Homepage group 'AI - Eval & Retrieval'), .env.example and README.
dns: intern-decision.fv.internal -> fv-ml1 (synced to ana/esh/nh3).
acceptance on fv-ml1 GPU 3, 3 fresh processes: bit-identical to the Jev bench's native rows
(pooled 240/259, Wyrd 79/84, 0/560 flips, Δp 0), negative control 10/122/14, 0 flips across
restarts; largest accepted request 200 at a 10,134 MiB card peak under a 9.25 GiB cap; 503 and
recovery proven at a tight cap. GPU 1 deploy held: nvidia-smi Free on GPU 1 is 15,442 MiB.
- a failure while building the response (a non-finite number included) is a 500 inside the
envelope, never a 422 or a render crash outside it
- an engine ValueError keeps its message but is released and raised unchained, like an OOM
- the prompt is built (and the model's own validation run) before the forward
- the row cap is counted before any ordering is built
- /health reads a device name cached at load, so it makes no driver call off the inference thread
torch keeps CUDA state per host thread (cuBLAS handles and workspaces), partly outside the
per-process VRAM cap. Scoring on anyio's threadpool let 40 threads each create it: measured on
fv-ml1 GPU 3, +252 MiB outside the cap and +326 MiB inside, which pushed the process past the
10,300 MiB GPU 1 budget. Load, warm-up and every call now run on the same single thread.
Its dataclasses use postponed annotations and look their module up in sys.modules while the
class is built; importing by file path without registering it failed startup (closed).
Contract, service and tests (fake engine, no GPU). Scores through the checkpoint's own
inference.py (DecisionEngine.predict, sha256-pinned); maps semif decisions onto Jev choice
questions, packs /decide/shared into calls of at most 16, runs orderings in waves, and keeps
semif's error mapping, admission, body limit and hard VRAM cap. Deltas from semif-serve are
listed in the contract.
Measured Parakeet peak on a 35-min file (n=3 each, deterministic): 300 s
9,384 MiB, 120 s 6,510, 60 s 5,976, 10 s 5,634 (fixed floor); with
expandable_segments 120 s 5,496 and 60 s 5,502. Verified 5,496 under the
recreated container's own env. Transcripts: 95.8% word-sequence similarity
vs 300 s, diffs mostly casing/punctuation.
Operator ask relayed by brokkr-smithy-dev. Positive control (SemIf 187/231,
hard 0.613) reproduced exactly; negative control and a 4-restart noise floor
(0 flips) measured. On our replaced-baseline sets no candidate beats SemIf-with-
rotations beyond the ~4-pt floor; Intern-Decision-4B native matches it at one
ordering, is better on Wyrd, fits 9.7/10.3 GB and is 1.5-2.3x faster. JevBench
rank does not transfer. Raw per-item data kept out of git.
Both verdicts now build the n for the U11b data-deletion gate (three
consecutive PASS batches at off). First real run 20260930T090608Z via the
detached-worktree path: PASS, user median 0.83.
Stopped (not removed) 2026-09-30 0135 PT. Scriberr's Parakeet path hardcodes
5-minute slices; attention memory is quadratic in slice length, so a long file
needs >6 GB and hit CUDA OOM with SemIf resident (~6.7 GB free). GPU 1 now
81,806 MiB used. Durable fix (shorter scriberr slices) deferred to later.
For infra-hermes's daily batch through the U11 off window (operator ruling
2026-09-29 2340; mode revised to off 2026-09-30). Enforces worldtree-dev's
terms: flock plus a refusal while any harness batch runs, code under test =
demo's deployed sha (a detached worktree when the main tree has moved; refuses
on pyproject/uv.lock/packages drift the shared venv cannot honour), no git
commit, no retry. Exit 0 PASS / 1 FAIL / 2 harness error / 3 refused.
The 2026-09-28 expansion grew the one-page design to nine pages by
bolting a nav band under each hero. This makes the cluster one site:
- One site header on every page: the wordmark (home link) and a single
row of short, consistent labels (Applications, Systems, Hosting, AI,
Engagement, Track record, Contact). The nav no longer wraps into two
rows at desktop width or four on a phone.
- A real <h1> on every page (the title was a <p>), a skip link, and one
footer outside <main> on every page, home included.
- The contact page was unreadable: its section borrowed the home page's
dark contact-card id, so its copy was grey on near-black (2.3:1) and
its headings measured 1.0:1. It is an ordinary section now.
- Section numbers only where there is a sequence (home, engagement).
- The spec bar names all four services; it stacks cleanly on a phone.
- Contrast: --text-tertiary 4.3:1 -> 4.96:1, --accent-text 4.36:1 ->
4.74:1 on paper. Track-record labels drop tracked capitals.
- The grid behind the hero and the contact card is its own layer; the
contact card loses its left side-tab stripe.
- The stylesheet link is versioned (style.css?v=2026-09-28): CSS is
cached for an hour, and the new markup must not meet the old styles.
- README: nine pages, the shared structure, the CSS-version rule, and the
operator's anti-slop waivers (the hero's volt rule, Space Grotesk).
Copy is unchanged apart from nav labels (the brief's legal constraints
hold). Still zero JavaScript and zero external requests. Impeccable
detector (controls 7/0, static + rendered at 1280 and 390, light and
dark): 200 unwaived -> 54, all of them the two waived brand devices;
two identical runs. Operator: "ship the site cleanup, keep the rule and
the font".
Blender is now a mandatory stage in draupnir's pipeline (Prime, 2026-09-28), and draupnir asked
for eight add-ons from extensions.blender.org: SurfacePsycho 0.10.4, CAD Sketcher 0.32.1,
3D-Print Toolbox 1.4.1, STEP Importer 1.2.1, Bool Tool 2.1.0, LoopTools 4.7.7, MeasureIt 1.8.4,
3MF Import/Export 2.7.7.
- stacks/blender/extensions.lock pins each by version and archive sha256.
- scripts/blender-extensions sync builds fv-ml1:/tank/blender-extensions/5.2/system with Blender's
own install-file, pre-warms and byte-compiles it, checks a read-only enable, then swaps it in.
It refuses while the GUI or a blender-run job holds the old directory.
- conf/scripts/startup/fleet_extensions.py enables every package in the System repo: in a timer
in the GUI (after the prefs load), and as --python ahead of the caller's args in
blender-run --extensions (a failed enable exits 1 before the caller's script).
- It also patches SurfacePsycho's sp_overwrite_segment_selection from eval() to literal_eval():
the eval walked past MCP safe mode (control: unpatched ran code, patched refuses).
- blender-run: --extensions (bind mounts via --mount so a missing source fails instead of being
created); USER/LOGNAME set, which CAD Sketcher's getpass needs.
- compose.yaml mounts the repo read-only and the hook into the GUI container. NOT yet deployed.
- scripts/blender-probes/extensions_acceptance.py: one operator run per add-on, safe-mode
compliant. Headless 8/9 online and with --network none; CAD Sketcher sketching is GUI-only.
A Python audit hook saw no network/process events (positive control fired).
Single page grows to nine: custom app development, systems engineering,
hosting, and AI services each get a detail page, plus engagement,
track-record, contact, and privacy. Cross-linked nav and footer on every
page; brand CSS extended in place (still zero external requests, zero JS).
Motivation: an app-store registration review flagged the single-page site
as minimal content. This version is intended to be reverted after approval.
draupnir review: the remote job dir was keyed on the basename alone, so two
local dirs with the same name shared one remote dir, and a rerun inherited
stale files. The dir is now <basename>-<8 hex of sha256(abs path)>, and it is
mirrored with rsync --delete, confined to that one directory. There is no rm on
a computed path. Verified: a file deleted locally is gone from the rerun's
remote dir. The /defaults stderr line is documented as harmless noise.
scripts/blender-run launches each call as a docker run --rm of the Blender
image on fv-ml1 GPU 3, capped at 64g / 48 CPUs. It needs no desktop and does
not affect the GUI container's lifecycle. --job DIR stages a local directory
to /tank/blender/jobs/<name>/, runs Blender with that as the cwd, and copies
results back. It always passes --python-exit-code 1, because Blender otherwise
exits 0 when a --python script raises (measured).
Tested headless: Cycles GPU and CPU, EEVEE via EGL, Workbench, an STL
round-trip, and exit codes (3, 7 and 1 pass through). There is no STEP
importer. Written for draupnir's design work, and indexed in FLEETTOOLS with a
detail file.
The MCP server (mcp-for-blender 2.1.1, frozen requirements) runs inside the Blender
container. Its add-on is vendored at upstream 41a18432 (MIT) and started by a
startup hook. scripts/blender-mcp carries the stdio over ssh + docker exec, so the
add-on socket, which runs arbitrary Python with no auth, stays on the container's
localhost with no published port. It also runs there because viewport screenshots
need a filesystem shared by server and Blender. Telemetry is off and safe mode is
on. The hook also defaults Cycles to OptiX on GPU 3, because safe mode forbids
agents from touching preferences.
Verified end to end from nh3-dev: 36 tools; a GPU render of an agent-built scene;
a viewport screenshot; and safe mode refusing 'import os'. Blender left down
(on demand).
Uses linuxserver/blender (Selkies Wayland desktop, NVENC), digest-pinned. The
container sees only GPU 3, has restart "no", and runs only while in use,
because GPU 3 is the reserve card for a full-size vLLM seat. Web desktop on
:3001 with basic auth (vault fv-ml1/blender-web-password). /work is on /tank
and is not backed up; /config lives under /opt/docker (restic).
Acceptance: Cycles finds the card on OptiX and CUDA (sm_120 kernels ship in the
build). The heavy self-test renders in 3.54 s on OptiX vs 22.71 s on CPU, a
functional check with n=1. Web auth answers 401 without credentials and with a
wrong password, and 200 with the right one.
The shim holds 10.0.50.47/24, which gives two equal connected 10.0.50.0/24 routes,
and ens18's wins. Host-to-HA traffic therefore left via the macvlan parent and was
dropped. HA lost MQTT to the broker on this host on 2026-08-19, 2026-09-21 and
2026-09-25 (the last lasted two days). This adds an ifupdown if-up.d hook that
routes 10.0.50.46/32 via macvlan-shim; /etc/network/interfaces is not edited.
Verified: the route resolves via the shim, the host pings HA, HA reaches :1883,
and HA reconnected to the broker. Diagnosis by ha-dev.
Also corrects the zigbee2mqtt acceptance note, which had wrongly reported HA as
connected.
Requested by ha-dev; approved by Prime in this session. Radio: SLZB-MR1U chip 0
(EFR32MG21, EmberZNet 8.0.2) at tcp://10.0.90.10:6638, adapter ember. A fresh
network was formed on channel 25, PAN 0xCFF4. The broker is reached as mosquitto
user zigbee2mqtt, with HA discovery on homeassistant/. The frontend on :8099 is
token-protected.
State and the network key stay host-only in /opt/docker/data/zigbee2mqtt
(root 0700, restic). The repo carries only compose, .env.example and the README.
configuration.yaml refers to the secrets as !secret.yaml, and those references
survived Z2M's v4->v5 settings migration.
SemIf's shared scorer trims one token at the state boundary. When an object
state's last value ends in ')', ';' or '}', the JSON that follows re-merges two
tokens back, so score_shared refused the request with 422. The engine now wraps
semif_phase1.shared._state_prefix to keep only the tokens the full prompts
share. Each row scores the same token sequence; only the prefill/suffix split
moves.
Startup proves the fix is in effect, not just installed (heid bug hunt SKAL,
folded). It checks that the hook is callable and is what score_shared resolves,
that an ordinary state keeps upstream's whole prefix, and that a merge-prone
state scores through the shared path.
Real tokenizer: 154 states, 23 refused before and 0 after, with no ordinary or
authored144 prefix changed. Acceptance: 144/144 parity. Shared vs direct 71/72;
the miss is a bf16 tie that flipped across a plain restart (see README).
Against talk /face's guided pose (first paragraph 246 ms median), SemIf in
parallel adds 32 ms and SemIf-first adds 94 ms (n=72 each, noise floor 16.5 ms).
Removing the pose header saves only ~31 ms, and SemIf shares GPU 1 with the LLM.
Acceptable pose 67% vs 92% on clear-emotion lines, and the mood carried through
mundane follow-ups 7/15 vs 14/15. SemIf gestures far less (13% vs 58%).
README: rotations cost options^2 in suffix tokens, and /decide/shared returns
422 when an object state's last value ends in ) ; or }.
The comment claimed 'gateway-key-gated'; v1.97.0 actually gates auth=true
passthrough routes on per-key metadata allowed_passthrough_routes (OSS path,
not Enterprise), 401 unauthenticated. Grants applied live via /key/update for
all-agents-local and the worldtree gateway key; comment-only change here,
picks up with the next conf deploy.
Harness consumer_fit.py runs a consumer's per-turn decisions over hand-labelled
cases, with rotations, a content-free null control and tagged positive controls.
Cicada: an input-only 'does this earn a visible reaction?' gate scored 30/31
with descriptive options and 19/31 with terse yes/no options. Scoped by
Cicada's 2026-09-20 ruling (affect is emitted once, no mood-ring classifier).
Wyrd: on 3 real seed graphs, the first place-change wording failed its
positive controls (1/6 moves). A location-anchored rewording scored 21/21, and
exit selection scored 18/21. Semif fits the choice, not writing the node.
Order averaging (Prime, after the 739aa03 spike):
- A decision may set orderings: rotations|all (all only for <= 4 options). Every
ordering goes to the engine in one shared batch.
- The reply keeps each native result and adds combined {probabilities (log-mean),
top, agreement, spread}.
- Through the service on SemIf's labelled sets (252 rows): 78.6% -> 88.1%
(group-bootstrap 95% CI +5.1..+14.3). Unanimous agreement is 94.5% accurate.
Fast kernels: flash-linear-attention 0.5.2 and causal-conv1d 1.7.0 are now the
default build. A/B on the empty GPU 3:
- parity with upstream went from 142/144 to 144/144;
- a ~2k-token /decide went from 169 to 92 ms server-side;
- short 3-rotation batches cost ~3-6 ms more.
triton builds a C shim at runtime, so the image carries gcc. Without it the
warm-up failed and startup failed closed.
Heid bug-hunt panel (4/4 arms, thread 01M3H3F4RR7XBP90KQ3A39H4SX), folded:
- Startup validation: VRAM cap 0 no longer means uncapped (C1); limits must be
>= 1 (S1); the token must be visible ASCII (S2); the calibration file must
exist and parse, with T in [0.05, 20] (S8, and C3's NaN leg).
- The body limit is checked before a chunk is kept, and a Unicode-digit
Content-Length no longer crashes (C2, S3).
- Failures while building the response now get the 500 envelope (C3).
- 429 busy past SEMIF_MAX_QUEUE requests in progress (C6).
- The engine releases memory on every non-validation failure, unchained after
gc; an empty OOM message is handled; 'out of memory' RuntimeErrors map to 503
(C4, C5, S9).
- The entry point forces HF_HUB_OFFLINE (S10). README wording fixed (S5, S6).
- New guard tests close the gaps the arms' mutation grids exposed: early stop of
the body read, a shared-route lock, calibration pass-through, the gc cycle,
the exact caps, TorchEngine.load's arch and device checks, and the offline
entry point.
86 tests.
Deployed on fv-ml1 GPU 1: parity 144/144, OOM and burst release verified, shared
capacity 63/51/26/16 rows at ~140/520/1960/3900 prefix tokens.
Latency, measured from nh3-dev (3 runs x 20 per condition; network floor 31 ms):
- /decide short: 71 ms end to end, 38 ms server-side;
- /decide with a ~2,000-token state: 210 / 169 ms;
- shared, 3 rotations: 113 / 79 ms;
- shared, 6 orderings: 137 / 99 ms.
Qwen3.5's fast kernels (causal_conv1d, flash-linear-attention) are not installed,
so transformers falls back to its reference PyTorch paths. That is a speed lever,
and using it needs a parity re-check.
Averaging over option orderings, on SemIf authored144 + perturbations108 (252 rows,
72 groups):
- a single ordering scores 78.6%;
- log-mean over the 3 rotations scores 87.7% (+9.1 pts, group-bootstrap 95% CI +4.7
to +13.8);
- all 6 permutations score 88.1%.
Rotations capture nearly all of the gain. Rows where the rotations agree unanimously
(161) are 94.4% accurate; split rows (91) are 75.8%.
services/semif-serve is a FastAPI wrapper around SemIf's direct and shared torch
scorers (SemIf-OpenJev @ 23cf1f39, MIT). Upstream ships only a batch CLI. The
wrapper loads the pinned Qwen3.5-4B (851bf6e8, BF16) once from the offline HF
cache and returns SemIf's result dicts unchanged, with an optional per-workload
temperature-calibrated view. Contract: semif-serve.contract.md. Built with a
short contract, TDD (39 tests, fake engine and fake torch, no GPU) and a heid
bug-hunt panel (pending).
On the card:
- torch 2.10.0+cu128 with sm_120 kernels, which is SemIf's own stack;
- a hard 12 GiB VRAM cap.
Two defects surfaced only on the card, and each fix is covered by a test:
- 0.1.1: an OOM raised as a chained exception kept the failed request's tensors
alive (11.9 GiB after the 503). It is now raised unchained, after gc.
- 0.1.2: a large request left 12.6 GB reserved on the shared card. After each
call, reserved memory over the baseline + 512 MiB is now released.
Acceptance against SemIf's committed torch predictions (authored144):
- 142/144 same top choice; both misses are exact bf16 ties;
- 144/144 identical prompt hashes;
- deterministic A-vs-A;
- negative control 14/144;
- shared vs direct 72/72.
21 binary criteria over one state take 159 ms. The shared-mode capacity table
under the cap is in stacks/semif/README.md.
The Dockerfile installs dependencies from a manifest with the project version
blanked, so a version bump reuses the ~4 GB torch layer. Verified: 41 s rebuild,
dependency layer CACHED.
DNS: semif.fv.internal. Token: vault semif/api-token.
Prime bootstrapped infra-ops on vm-esh-nas with playbooks/bootstrap-infra-ops-user.yaml.
It got the fleet-pinned uid/gid 850, NOPASSWD sudo with log_output, the docker
group and a 0700 home. That let playbooks/restic-repository-file.yaml migrate
the last restic host: the live profile matched the repo's pre-change sha,
its units no longer carry the URL, its secrets are vaulted, and the live and repo
profiles now match (a5ea75ea). All eight restic hosts are clean.
The staged helper script is gone, both from Prime's home on the host and from the
repo. The docs that described vm-esh-nas as lkraven-only are updated.
docs/pfi/fleet-conventions.md § 3 pins infra-ops to 850, but the bootstrap
playbook let the OS allocate, so every host drifted (1001-2001). New
accounts now get 850 when uid and gid 850 are both free, and fall back to
OS allocation (with a note) when either is taken. The home is set to 0700
per § 1.2. Existing accounts are untouched (the step is gated on id).
Scope note: vm-esh-nas is added to the operator-granted ESH exceptions
(Prime, 2026-09-27).
resticprofile schedule copies env-file values into the generated units, which
are 0644, so RESTIC_REPOSITORY (the rest-server basic-auth password included)
was readable by every local user on every restic host.
New playbooks/restic-repository-file.yaml:
- derives /etc/restic/repository (root 0400) from restic.env;
- uploads the profile switched to repository-file, but only when the live
profile's sha matches the repo copy it was edited from (drift guard);
- checks the repository is reachable through the new profile (cat config);
- regenerates the units and verifies they exist and contain no rest:http.
Applied to ana-docker, fv-ml1 (configs/restic/ana-ml2), esh-docker-vm,
esh-vm-db, irv-ml1, nh3-dev and nh3-docker. An independent check across all
eight restic hosts (these seven plus esh-ml1) found 0 leaking units. nh3-docker's
scheduled unit ran a real backup afterwards (snapshot a29b889d). Each host's
URL and passphrase are vaulted as <host>/etc/restic/{repository,password}.
restic.env is kept (root 0600) because the per-host READMEs and the freshness
probe source it. A rotation must update the vault, restic.env and repository.
vm-esh-nas has no infra-ops account. Its in-place migration script is staged
for Prime to run with sudo, and its repo profile is pre-edited to match.
Also mirrors augaman-dev's 401df2d (compose header only). The config hash on
esh-ml1 is unchanged.
Both hosts had live improvements the repo never recorded:
- esh-docker-vm excludes ESPHome's 539 MB of PlatformIO cache (2026-09-14).
- irv-ml1 runs /etc/restic/arbo-checkpoint.sh before the backup, a non-fatal
SQLite online-backup of arbo's gallery DB. That script is added here too.
The live copies were correct and are the source for the next change. (Also fixes
the fv-ml1 augaman removal time to ~0130 PT.)
Prime removed the second instance after the v0.1.3 bench. esh-ml1 handles a face
in ~48 ms, sits in the house next to the cameras, and holds the verified
backup. fv-ml1's gallery was empty (0 identities). The container, gallery volume,
image, compose dir (with its .env), backup dir and build sources are removed from
fv-ml1. GPU_ID / CARD_SUFFIX stay in the compose for any future second host.
vh