Adds the Yarros-side evaluation the training exists to justify: does the adapter
move arbitrary prose toward Yarros, and can the instruct arm still expand a beat
to a paragraph on direction. Yarros-flavoured voice prompts (modern/neutral/
romantasy tiers so any Yarros voice in the modern tier is adapter-attributable,
not prompt-supplied) and a Yarros-register beat SYS on the chat generator.
voice_distance.py is the honest slice of adjudication that needs no seat: Burrows's
Delta over character bigrams against held-out Yarros. Its first cut mis-framed the
noise floor — it used the same-author distance (held-out vs itself) as the
between-arm significance threshold, which is the target, not the threshold. Fixed
to the measured floor: the within-arm seed spread, which is this metric's sampling
variance at this sample size, computed from the two seeds already generated rather
than assumed.
Result on the built corpus, ordering base-125-tuned < instruct-tuned <
base-unadapted, both adapters clearing the 0.046 measured floor (base +0.157,
instruct +0.076), and the ordering corroborating the independent held-out loss
ordering (Base below Instruct). One seed-pair per arm, so it corroborates rather
than settles; the full frozen adjudication still needs a romantasy control panel,
a second seed, and the gen seat for the beat-incumbent leg.
The recurring incident-time failure: during an outage the session reaches for
althing to notify infra-ops, and since ALTHING_HANDLE is infra-ops the send mails
itself, the herald pokes about new mail from infra-ops, and the session triages
its own words as a peer report. Documented the identity, the check
($ALTHING_HANDLE before addressing an ops message), and the one legitimate use of
the infra-ops thread during an incident (a durable memo to a future infra-ops
session on recovery, not a page to a peer).
Records the comparison (Sentinel-R3 is a real agentic-pentest SFT on the same
Qwen3.8-27B base vs M.O.G.-SEC's persona-on-stock-weights, on a tool surface that
matches our own harness), the HF check (M.O.G.-SEC repo unchanged), and the graft:
Sentinel ships no MTP head, so the verbatim base head was grafted in, which is the
correct lineage but leaves acceptance unverified because the head now sits on an
SFT-finetuned body rather than stock weights.
The quant is done and structurally verified; serving and the acceptance/A-B gate
are blocked on GPU space and are the operator's next call, along with the
proprietary-license posture. Detail in the artifact's PROVENANCE.txt.
VRAM said yes and the box said no. Two 4B LoRA jobs fit easily -- 36 of 121 GiB
with 98 free -- but both dropped from a 37.10 s/it solo baseline to ~85 s/it,
2.29x each, so combined throughput fell below serial. The GB10 is past its
memory-bandwidth roofline and pays a contention penalty on top of the split.
Worth keeping because the intuition it corrects is a common one: free GPU memory
is the number everyone reaches for, and on this box it does not predict whether a
second job is affordable. The control is what makes it a finding rather than an
anecdote -- killing the second job returned the first to 37 s/it on the next step.
One arm is not a comparison. On the Brontë corpus Base won held-out loss (2.814
against the instruct arm's 2.908) while instruct won instruction-following, and
that trade is the open question for Skaldsong. It has to be re-measured on this
corpus rather than carried over: BabyYarros is 12% larger and contemporary.
Two entry points because the right one depends on a measurement rather than an
argument. The box has 98 GiB free against an 18.4 GiB training footprint, so VRAM
is not the constraint -- but VRAM was never the binding constraint on a GB10 that
onboarding measured at 6x slower than ana-ml2 where compute predicts 2.7x, and
where batching was not a throughput lever. If the box is already at its bandwidth
roofline, a second job splits the same bandwidth and buys nothing.
So: launch-yarros-4b-base.sh takes an explicit --allow-shared-gpu that bypasses
the GPU-clear guard, and chain-yarros-4b-base.sh waits on the instruct run and
refuses if it produced no adapter. The bypass is an argument and never a default,
because the guard's normal job -- stopping a chain from firing into a live run --
is what keeps one lost run from becoming two. A shared-GPU launch stamps its own
log with a warning that its s/it is not comparable to a solo run, since the
harness is part of the number.
Both carry the same gate guard as the instruct arm: refuse to start unless the
leak gate report on disk says PASSED.
Records what the gate cost to build and what it found: a second typography defect
(small-caps epigraphs and drop caps, which the D1 "no unwrap needed" note was right
about in the wrong direction), back matter inside the prose of all five works naming
real people, a possessive-only entity that was unrenamed and unreported at once, a
character lost to an in-book article typeset in lowercase, and a phrase class the
unigram scan structurally cannot see.
Also records the two things that will matter next time: every new detector flag is
opt-in and the Brontë map was re-derived and diffed after each change, because one
intermediate version silently dropped four real names from Shirley; and the gate's
sensitivity floor is part of its result, not a footnote.
The gate is new. There was no committed instrument for "does any of the author's
own proper nouns survive the rename" -- the Brontë number was produced by hand
-- so leak_gate.py is now that instrument, and it runs both directions every
time: the same scan over the unrenamed source as a positive control, and a nonce
string as a negative one. A detector that only ever sees renamed text cannot
distinguish absent from blind.
Run against BabyYarros as built it reported 212 surviving entities, not the 86
recorded earlier, because it scans the whole corpus rather than each work
separately and it counts the sub-threshold entities rename never looked at.
Three findings came out of closing that.
The corpus had a typography defect of its own. The D1 notes correctly say no
unwrap was needed; a different defect was there instead. The Empyrean books set
their chapter epigraphs in small caps and the extractor rendered the run as
uppercase while leaving the large initial as a separate token, so the corpus
carried "M AJOR A FENDRA'S G UIDE TO THE R IDERS Q UADRANT" -- 106 lines, ~700
splits -- plus 52 drop caps like "T he flight field". That is where the entities
called IDERS, UADRANT, NAUTHORIZED and seventeen bare single letters came from.
A split initial next to an uppercased run is enough to recover the original
mixed case, so the restore is exact rather than approximate: a word with a split
initial was capitalised, an all-caps word without one was lowercase.
Back matter was inside the prose. The builder splits on chapter headings and
nothing follows the last one, so every work carried its acknowledgments,
newsletter pitches and cover-artist credits -- 4,555 words naming the author's
agent, editors and children, in a corpus whose entire purpose is that no
identifiable name survives.
And the gate passed at 0 of 314 while Afendra was still in every copy. The name
never appears unpossessed, so it keyed as an apostrophe form, and rename and the
gate both skip those as contractions -- unrenamed and unreported at once, which
is the worst failure shape available. Baxter escaped a different way: wilder
renders an in-book news article entirely in lowercase, putting the cap/lowercase
ratio at 0.13 against a 0.05 bar.
Then a second class the unigram scan structurally cannot see. Riders Quadrant,
Flame Section, War Games and Fourth Wing -- the book's own title -- are built
from ordinary words the detector correctly refuses to call names. The gate now
audits recurring capitalised 2-3grams against an explicit allow list, and
rename applies a phrase map after the entity pass.
Every new detector flag is opt-in and off by default, and the Brontë entity map
was re-derived after each change and confirmed identical in keys, surfaces and
every field. The stoplist was built by reading each surface in context, which is
why it is short: Violence is Xaden's nickname for Violet, and Continent,
Presentation, Barrens, Originals, Montserrat, Athena, Aura, Curator and Sage are
all in-world. A plausible-looking guess would have excluded most of them.
Final: 0 of 325 entities and 0 of 91 audited phrases survive in any of 30 copy
files, both controls passing. The sensitivity floor is stated in the gate's own
output -- 3 occurrences for a name, 5 for a phrase -- because a negative without
one is unfalsifiable.
Located the source: five Rebecca Yarros works in the Kvasir licensed library, with
rights recorded as gated. Built D1 at 208 chapters and 780,744 words, which is 15%
larger than the Brontë corpus. No unwrap step was needed because Kvasir's cleaner
already emits flowing paragraphs, so the hard-wrap defect that cost a re-cut on
Brontë does not exist here. The alphabet was re-derived rather than inherited: 23
non-ASCII letters across three forms, against F02's 4 on a smaller sample. Same
ASCII-fold conclusion from a different measurement, which is the reason to re-derive
per corpus.
The interesting finding is a new pathology. In a rotating first-person POV corpus,
every book's narrator gets the wrong gender. Measured against six names verified in
the text, the pronoun resolver called Violet male, Leah male and Landon female --
three of eighteen wrong, and all three are the narrator of the book where they were
misgendered. A narrator is "I" in her own book, so her name appears mostly inside
the other lead's dialogue surrounded by his pronouns. This is Brontë's "Jane called
male" amplified by rotating POV. Title-first resolution, which fixed it for Brontë,
is nearly blind here because contemporary romance uses given names rather than
honorifics. What works is the POV header: resolve each name from the chapters it
does not narrate. Validated at 9 correct, 9 held, 0 wrong against the previous 7, 8
and 3 wrong, and the instrument refuses to write unless it beats what it replaces.
Re-pointing rename.py surfaced three bugs, two of which would have silently
corrupted the corpus. Gender came only from honorifics and the entities file's
gender field was ignored, so the POV fix had no effect until wired through; that
took wilder from 1 gendered entity to 13. The pool labels were hardcoded in a print
statement, so any non-Brontë preset crashed. And the collision-filter log claimed
it dropped names colliding with Brontë entities regardless of which corpus it
filtered against -- the logic was right but the message named the wrong corpus,
which is how a reader later concludes the filter ran on the wrong thing.
D3 is blocked and nothing has been trained. The leak gate shows 86 of 232
renameable source entities surviving where the Brontë run reached 0 of 203. It
decomposes into detector false positives that need a stopword filter rather than
renaming, genuine misses among worldbuilding proper nouns, and a third class whose
cause is not yet established. Training before the gate passes means fitting
in-copyright text with 86 identifiable source entities intact, in a corpus F02
already flagged as small enough for leak to be a real concern.
Trained the same corpus onto Qwen3-4B instruct rather than -Base, with seed, steps
and token count held so the carrier is the only variable. The chain picked
checkpoint-150 by loss automatically, applying the lesson that cost a re-cut on
4B-Base.
The central risk did not materialise. The assistant prior did not block the voice:
curly quotes land at 16 of 18, identical to the 4B-Base tuned arm, against 1 of 18
on the unadapted control, and task-leak is 0 of 18 where the base carrier leaked 4.
Instruction-following also survived raw-text training -- 10 of 10 on-beat through
the chat template, the same as the untuned control.
The cost is length discipline rather than comprehension. In-band dropped from 10 of
10 to 6 of 10 and the median went from 124 to 140 words. Training on Victorian
prose made it wordier, which is a soft degradation and not a break.
Held-out sits at 2.908 against 4B-Base's 2.814, and it plateaus without turning
where the base carrier overfit at step 75. The assistant prior competes for
capacity, so the instruct carrier absorbs less rather than overfitting more.
What raw-continuation training does not fix is the plot furniture. The tuned
instruct arm renders the beat and then drags the referent -- "He licked her clean...
my master thus, my husband thus", turning the dog into a man, because the corpus is
about masters and husbands. Another beat ran to 247 words and gave the narrator a
list of duties. That is precisely what instruction-pair training addresses, since
pairs teach render-this-and-stop where continuation teaches keep-writing. The probe
de-risks the instruction-pair path without substituting for it.
One metric note against future misreading: ran_on reports 10 of 10 on both arms and
is uninformative on this job, because a single paragraph contains no blank line for
it to find.
Skaldsong's requirement is beat-sentence in, paragraph out, stitchable. That is
narrower than free-form continuation and it fails differently, so it was tested
directly rather than assumed from the earlier prompt work.
The adapted 4B carrier cannot do it. Ten prompt formats across three seeds each --
bare, paragraph-break, labelled, epigraph, one-shot, one-shot unlabelled,
three-shot, "it happened in this way", "I remember the whole of it", and the
letter prompt's label-and-begin move applied to a beat -- produced thirty samples
and not one that reliably renders the beat. They drift, they frame, or they
truncate. The cause is structural rather than a shortfall of the adapter: "write a
paragraph about this sentence" is an instruction, and a completion model has no
mechanism for "about". It continues the text it is handed.
Two formats failed worse than drifting and leaked pretraining task data. The
paragraph-break format emitted an NLI multiple-choice item, and label-begin emitted
a grammar-correction exercise. A standalone sentence followed by a blank line looks
exactly like a dataset entry, and style adaptation does not remove base-model task
artifacts.
The same five beats through the gen seat with a style instruction and no adapter
landed all ten samples inside the requested 90 to 140 word band, every one on its
beat, with no drift into a following scene. The voice, however, is generic literary
pastiche -- abstract-noun-heavy, fond of aphoristic openers, and over-written --
rather than Brontë.
Voice without direction against direction without voice. The product needs both,
which makes rebuilding the corpus as instruction-response pairs on an instruct
carrier the evidence-backed option rather than more prompt cleverness. This applies
to Yarros identically, because the carrier question is orthogonal to the author, so
the next corpus should not re-run this experiment.
Stitching also has its own failure mode, shown in the booth's third panel: the
paragraphs were generated independently and the point of view slides between them,
so by the fourth beat the narrator is both watching the girl carry the animals and
carrying them herself. A real stitcher has to feed prior paragraphs back as
context, which means the instruction-pair corpus wants multi-paragraph continuity
examples and not only isolated beat-to-paragraph pairs.
Two operator prompts settled the question the handoff prompt opened.
The winner labels the artifact and then begins it: "This is the letter I wrote
verbatim, my two short paragraphs, detailing the time I saw the mangy gray dog
meet and then lovingly and tenderly lick a calico kitten: Auntie, You'll never
believe what I saw--". Two of three seeds delivered the actual event in first
person, and one is the strongest output of the sweep -- the dog following her to
the mill, the kitten caught in a hedge, the dog carrying it out and laying it in
her lap before licking it. Coherent arc, correct subject, no gloom-override, no
meta-frame. It also reproduced the Gutenberg underscore-italics convention.
It works where the handoff failed for a specific reason. The handoff could be
satisfied by narrating compliance, because the letter did not yet exist. Naming the
artifact and starting it leaves nothing to narrate around, so the only continuation
is the artifact.
The chapter-heading prompt refuted a hypothesis of mine. I predicted that rendering
the title in the corpus's own conventions would make it land harder than the
operator's inline form. It did the opposite: both corpus-form seeds ignored the
title and opened unrelated scenes, while the inline form at least finished the
heading and wrote a chapter about the story. Corpus chapter titles are short and
decorative, so a long descriptive one in that slot reads as decoration to skip. A
label only instructs if the model treats that slot as load-bearing.
That prompt also surfaced an unnoticed consequence of the D2/D3 rename pipeline:
given "Alice", the adapter produced "Alexander the Alexander, as he was known in
Little London". The corpus was entity-renamed from a French/English pool, so the
adapter learned that character names come from that pool and rewrites outside names
into it. Callers cannot reliably name their own characters at prompt time. That is
not a defect of the rename, which exists to prevent memorising Brontë's cast, but
it is a usability constraint worth stating.
The 4B tuned arms now come from checkpoint-75, the actual loss minimum at 2.813826
confirmed from loss-series.json rather than from reading the training log. Only the
tuned arms needed re-cutting; the base arm never touches the adapter.
One surprise in the comparison: step-75 and end-of-run differ on typography rather
than voice. Curly quotes read 16 of 18 against 17, collapse is 0 either way, but
the hard-wrap ratio is 0.33 at step-75 against 0.12 at the end. Further training
washes the residual line-break habit out while held-out loss gets worse, so best
loss and best typography are different checkpoints. Neither is near the original
0.85 defect and the corpus's own residual from preserved verse is 0.25.
The operator's second prompt embeds the instruction inside the fiction: Abernathy
tells the tale badly and then asks the narrator to retell it in a few short
paragraphs. Across six seeds the model acknowledged the handoff every time and
never once performed it. It wrote "I told it, briefly, to his satisfaction", and
"So I wrote it out, and kept it in my pocket-book", and one seed negotiated the
brief in character, objecting that it could not condense the story into a few short
paragraphs because there were too many points to touch.
The reason is structural rather than a failure of the adapter. In a novel, "she
retold the story" is an ordinary sentence, so the likeliest continuation of a
request is narration of compliance rather than compliance.
A trailing paragraph break did shift the behaviour -- one seed opened in the
narrator's own quoted speech -- so typography is a real lever, just not a
sufficient one. That is direct evidence on the instruct question raised earlier: if
the product is asking for a scene and getting the scene, no amount of in-fiction
framing substitutes for a post-trained instruction-follower, which favours
rebuilding the corpus as instruction pairs over more prompt cleverness.
The three rungs now sit on the same unwrapped corpus, seed, step count and token
count, so carrier size is the only variable. Held-out loss reads 3.329 at 0.6B,
3.018 at 1.7B and 2.814 at 4B -- deltas of 0.311 and then 0.204, diminishing but
still real.
4B answers the question the rung existed for. Scene-level continuity holds: on the
office-kitchen prompt it produces a named character with motivated dialogue, a
spatial layout the narrator navigates, and a physical description, all in one
passage, where 1.7B wrote pretty but eventless prose about opening doors and
looking at stars. On the letter prompt it opens the letter, promises to quote it,
and then quotes it across a paragraph break. Voice saturation is also the best of
any rung: curly quotes 17 of 18 against its own base arm's 1 of 18, and collapse 0
of 18 against 4 of 18.
Two findings that change earlier conclusions.
4B is the first rung to overfit inside one epoch. Its series runs 2.832, 2.816,
2.814, 2.820, 2.824, 2.825, 2.825 -- a minimum around step 75 and then a turn. Both
smaller rungs plateaued without turning, so the optimal epoch count shrinks as the
carrier grows and my earlier "one epoch is right for this corpus" holds only for
the small end. The consequence is operational: the adapter directory holds the
end-of-run weights at 2.825 rather than the step-75 best at 2.814, and it exists as
a recoverable checkpoint only because save_steps was set. The voice arms were cut
from the end-of-run adapter, so the booth understates 4B slightly.
The tone-override also appears to close. On the operator's frame prompt asking for
a wonderful story, 1.7B held the frame on every seed but killed the animals on two
of four; 4B kept them alive on both seeds, and one of them had the narrator doubt
the story he was told and supply a parallel childhood memory to explain the doubt.
That is a narrator with an interior position on the tale. Two samples per arm, so
directionally right rather than established.
Rung 3 of the sweep is training: Qwen3-4B-Base, one epoch, seed 4919, the same
unwrapped corpus, 159 steps at roughly 37.8 s/it for about 100 minutes. Carrier
size remains the only variable across the three rungs. A two-arm voice test is
chained behind it, gated on the adapter existing, so the rung is readable when it
lands rather than needing a second round trip to become informative.
The finding worth keeping came from an operator prompt rather than from my test
set. Asked for a wonderful story about a stray dog licking a calico kitten, framed
as reported speech, the 1.7B adapter held the frame across all four seeds and was
unmistakably Victorian throughout -- and two of the four turned the premise into
animal death, with the cat licking the puppy to death and the narrator's
interlocutor answering that he wished they were all dead.
That is not a coherence failure. It is Charlotte Brontë's preoccupations arriving
alongside her sentences: Jane Eyre opens on a beaten child, Helen Burns dies, and
Villette is saturated with grief. Train 680k words of that into a small carrier
and a request for something wholesome reaches for mortality.
My nine test prompts were all emotionally neutral and structurally could not have
surfaced this. The operator's prompt found it on the first try, which is worth
recording as a property of the test set rather than of the adapter.
The implication for the regime is that "voice transfer" includes tone and subject,
so wanting the voice without the gloom is a corpus-selection or prompt-framing
problem rather than a training-length one.
Also observed: one seed closed its anecdote and began "CHAPTER XIX. THE CHILD'S
CLOSET", having learned book structure unprompted. The base control on the same
prompt went modern and essayistic, so the shift is the adapter rather than the
carrier.
Both rungs now sit on the same unwrapped corpus with seed, steps and token count
held, so carrier size is the only difference and the effect is attributable:
held-out 3.329 at 0.6B against 3.018 at 1.7B, a gap of 0.311 nats. The chained
0.6B rerun closed the confound the unwrap opened.
Two things in those numbers need stating or they will be misread.
First, the original wrapped-corpus 0.6B reached 3.172, which looks better than the
unwrapped 0.6B's 3.329 and is not. Different corpus means a different held-out
set, and the wrapped version's 5.7% newline tokens are near-deterministic after a
70-character line, so they deflate the loss with cheap wins. Removing them removed
the easy tokens. It is a measurement artifact, not a regression.
Second, a correction to my own earlier claim: I twice described the 0.6B run as
still descending and undertrained at 3.172. Its series reads 3.176, 3.173, 3.172,
3.172 -- it flattened. All three runs plateau, so one epoch is about right for
this corpus rather than short.
The three-way booth puts 1.7B base, 1.7B tuned and 0.6B tuned side by side on the
same nine prompts and seeds. The base arm is the control that matters: curly
quotes go 0 of 18 on 1.7B base to 15 of 18 on 1.7B tuned, and worksheet-or-
explainer collapse goes 3 of 18 to 0 of 18, so the shift is the adapter rather
than the larger carrier. Hard-wrapping fell from 0.85 to 0.18, confirming the
corpus unwrap carried through into the adapter.
Sense partially returned. The 1.7B arm produces locally coherent sequential
Victorian prose where the 0.6B produced word salad, but scene-level continuity
still breaks mid-passage.
One observation held loosely: curly quotes are slightly lower at 1.7B than 0.6B,
which would fit a bigger model's stronger priors resisting the adapter at the same
rank. That is untested and is not offered as established.
Operator: "loading up the context killed sec again." That reproducer is what
finally made the failure legible, and it showed the previous four fixes had all
been aimed at the wrong quantity.
What the KV pool can hold and what the card can process at depth are different
numbers. Cutting context 420k to 384k to 320k, pinning the KV in bytes, and
dropping the prefill chunk from 16384 to 4096 all sized the pool. The crashes were
governed by the transient needed to process a prefill chunk against a quarter
million tokens of resident KV, which scales with depth and not with pool size.
Each change helped and none fixed it.
Bisected against the real reproducer, with a non-repeating prompt because prefix
caching would let a repeated one hash to cached blocks and never prefill deep:
113,247 prompt tokens SURVIVED (27 s)
200,088 prompt tokens SURVIVED (174 s)
~285,000 prompt tokens ENGINE DIED, HTTP 500, container restarted
The sustainable ceiling therefore sits between 200k and 285k with gen idle, and
gen shares the card with its load uncontrolled, so 163,840 takes about 20% margin
under the proven-good depth rather than sitting at the measured edge.
The ceiling's purpose is the refusal. Verified after the change: a 149,073-token
request serves in 41 s, and requests at both 200k and the ~285k depth that killed
the engine now return a clean 400 naming the limit in under a second with the seat
untouched. A seat that refuses what it cannot serve is strictly better than one
that dies trying.
Concurrency went 1.03x to 2.09x. The compose header's "served at native 262K" was
never actually deliverable on a shared card; it had simply not been exercised at
depth until today.
The probe is committed rather than described, so the ceiling can be re-measured
when the card's tenancy changes.
Operator: "start the 1.7b training."
The 0.6B adapter learned the Gutenberg transcription's ~70-character line breaks
along with the prose -- its output wrapped at a mid-length-line ratio of 0.85
against the base model's 0.00. That is typography rather than style, and every
further rung would have inherited it, so the corpus is reflowed before rung 2
rather than after the sweep.
The reflow joins 57,430 of 85,380 paragraph blocks and keeps 27,950. Verse is the
hazard a blind join would destroy, so the decision is per block by median line
length: blocks whose lines cluster near the wrap width are flowed prose, blocks of
consistently short lines keep their breaks. Every kept multi-line block in the
sample was genuinely verse with its lineation intact. No line ended in a lone
hyphen, so the space-join could not split a word across lines. The acceptance
check is content identity -- " ".join(text.split()) byte-identical before and
after -- and it passed on all 852 records, proving only whitespace changed.
Concrete cost of the old defect: 5.7% of the training budget was newline tokens.
The same words pack to 5,210,112 tokens unwrapped against 5,525,504 wrapped.
The 1.7B run is live at 159 steps and roughly 18.7 s/it. Everything but the
carrier and the corpus is held from the 0.6B run: seed 4919, rank 32, lr 1e-4, seq
4096, batch 1 by accum 8, one epoch, eval and save every 25 steps so the minimum
is located rather than assumed.
That corpus change is a second variable and it is named as one. A 0.6B-vs-1.7B
comparison is descriptive, not attributable, until the chained 0.6B rerun on the
same unwrapped corpus lands behind it -- gated on the 1.7B actually producing an
adapter, because a chain that fires on failure turns one lost run into two.
"Did sense come back at 1.7B" is a within-arm reading and survives the confound;
any between-rung delta does not.
The original wrapped corpus is untouched, so the 0.6B run's pinned corpus sha
3959036cf851bf62 stays reproducible.
Operator: "yes, pin the kv and take it to 320k."
The real finding, which took three crashes and two failed attempts to reach:
--gpu-memory-utilization does not bound actual usage. It sizes the KV
calculation, but peak activation is measured at profiling time and real
long-context work exceeds the profile. vLLM's own budget line showed mog-sec
running 0.9 GiB over its 47.48 GiB reservation -- 26.44 consumed plus 3.53 peak
activation plus 0.89 CUDAGraph plus 17.52 KV equals 48.38 -- and gen was over by
0.33 on the same card. That overage came out of the shared card's slack, which is
what kept OOMing after the utilization drop.
The fix is the one vLLM printed itself: --kv-cache-memory=17697765376, its own
recommended figure to fit inside the requested budget. Same discipline erp-seat
already uses, and for the same stated reason -- an explicit figure is
reproducible where a ratio silently yields a different cache depending on what
else is resident at start time.
The KV pin and the context length are coupled. 16.48 GiB yields about 383,730
tokens, so a 393,216 max_model_len falls under the 1.0x floor and vLLM refuses to
start rather than crashing later; pinning the KV while keeping 384k was never an
available combination. 327,680 leaves 1.15x, up from 1.03x.
Verified: the engine now logs "reserved 16.48 GiB memory for KV Cache as
specified by kv_cache_memory_bytes config and skipped memory profiling", KV
375,901 tokens, GPU0 down to 90,561 MiB from 91,313, RestartCount 0, and both sec
and sec-reasoning return 200 through the gateway.
Also records the BabyBronte eyeball A/B, whose result is the operator's own: the
voice transferred and the sense did not. Curly quotes went 1 of 18 to 18 of 18
and worksheet collapse 3 of 18 to 0 of 18 between arms. That voice is separable
from coherence at 0.6B is the premise the lightweight-adapter regime rests on, so
this is the informative outcome rather than a disappointing one. A corpus-prep
defect surfaced with it: the tuned output is hard-wrapped at about 70 characters
because the Gutenberg source kept its line breaks and the adapter learned the
typography too.
Cost: 320k of context instead of 420k, on a seat whose crashes happened at 151k.
sec/sec-reasoning crash-bounced twice in ten minutes, not once in thirteen days.
My earlier read of "rare, not chronic" came off a RestartCount of 1 and was
wrong; the operator pushed back and the second and third failures arrived while
that recommendation was still on screen. The memory entry making that call is
replaced rather than left standing.
Cause is unchanged from the first diagnosis: mog-sec at 0.52 plus gen at 0.38
reserve 0.90 of the card, leaving about 4.6 GiB, and vLLM's utilization figure
covers weights and the KV pool but not transient activation memory. A request
about 151,700 tokens deep scheduling a further 15,700-token chunk asked for 1.04
GiB with roughly 600 MB free.
Dropping utilization alone does not work, and fails in a worse way: a single
420,000-token sequence needs 17.88 GiB of KV, and at 0.50 the pool is 17.4 to
17.5 GiB, so vLLM refuses to start at all and the seat crash-loops during startup
instead of during a request. The context length and the crash were directly
coupled -- 420k was only reachable at the utilization that left no transient
headroom. So both moved: 0.50 and 393,216.
384k rather than vLLM's suggested maximum, deliberately. It estimated 406,352 on
one boot and 409,840 on the next, because the available-KV figure drifts about
0.1 GiB boot to boot; pinning the edge value fails to start on an unlucky boot.
393,216 sits 3% under the lower estimate and leaves roughly 0.7 GiB of the pool
unspent, which is the transient headroom the change exists to buy.
Verified after: KV 405,612 tokens, concurrency 1.03x at 393,216, and both sec and
sec-reasoning return 200 through the gateway.
num_speculative_tokens is documented as NOT the lever. The crash window logged
17.6% draft acceptance with positions 5 through 7 at 1.5 to 4.9 percent, which
reads as an obvious cut from 7 to 3; across 180 samples the median acceptance
length is 3.12 of 7 and median draft acceptance is 30.4%, so the crash window sat
near the minimum and cutting would cap the workloads accepting nearly the full
draft.
Cost: 384k of context instead of 420k, an 8.5% reduction on a seat whose crashes
were happening at 151k.
sec/sec-reasoning on ana-ml2 :8019 took a CUDA OOM on GPU0 at 13:20 PT and
auto-recovered in three seconds. First fatal error since the container was
created on 2026-08-28. A request already 151,728 tokens deep scheduled a further
15,696-token chunk, needed a 1.04 GB allocation, and found about 600 MB free.
EngineCore died, the API server exited 0, and restart: unless-stopped brought it
back; the dockerd journal records manualRestart=false, so nothing restarted it by
hand.
Not caused by the same-day char-rp-fast swap. That work was entirely on GPU1 --
erp-seat is pinned device_ids ["1"] -- and GPU0 read 93,257 MiB both before it
started and after it finished.
The standing condition is that GPU0's two seats reserve 0.38 plus 0.52 of the
card between them, leaving about 4.6 GiB, and vLLM's utilization figure covers
weights and the KV pool but not all transient activation memory. Long-context
prefill chunks with speculative decoding live in what is left.
Also recorded as a specimen: the crash window logged a 17.6% draft acceptance
rate with positions 5 through 7 accepting 1.5 to 4.9 percent, which reads as an
obvious case for cutting num_speculative_tokens from 7 to 3. Across 180 samples
of the same counter the median acceptance length is 3.12 of 7 (range 1.83 to
6.75) and the median draft acceptance is 30.4% (range 11.9 to 82.1). The crash
window sat near the minimum. The single window pointed the opposite way from the
distribution, and acting on it would have capped the workloads that were
accepting nearly the full draft.
No configuration changed. One self-healing failure in thirteen days does not
justify shrinking a KV pool on a seat configured for 420k context.
Operator: "replace that a4b moe over pfish-6 -- remove the pfish-6 alias and
create an alias for char-rp-fast."
G4-MeroMero-26B-A4B-it-uncensored-heretic-NVFP4A16 is live on ana-ml2 :8021 under
its own served name, behind the new gateway alias char-rp-fast. Pfish-6 is gone
from the gateway and now returns an explicit 400 rather than a substitution; 0 of
17 LiteLLM keys scoped it, so nothing was orphaned. The compose project name stays
erp-seat because asset-engine derives seat liveness from it.
The first quant of that A4B served NaN and passed its healthcheck doing it. It was
built with the dense v2-31B recipe, whose ignore list has no router regex, so all
30 MoE routers were quantized to 4 bits -- and a 4-bit router changes which experts
run rather than degrading them. Quant rc=0, healthcheck green, correct KV pool,
correct served name, and every completion returned finish_reason=length with the
full token count and content: null. The model was emitting a full budget of tokens
that decoded to the empty string. Raw /v1/completions was empty too, ruling out the
chat template and the reasoning parser. The signal that named it was logprobs:
vLLM refused to serialize the response, "Out of range float values are not JSON
compliant: nan".
The lesson is about the control rather than the router. That tree had already been
structurally diffed and passed -- against a verified-good DENSE quant of the same
Gemma-4 family. A dense model has no routers, so the single thing that was wrong
was the single thing the control could not distinguish. Diffing instead against
Pfish-6, a known-good quant of the same 26B-A4B MoE, gave it in one line: 222
ignore entries against 252, the 30 missing being layers.N.router.proj. A positive
control is only worth what it can distinguish, and "same family" is not "same
architecture class".
Re-quantized with the MoE recipe, whose dry-run asserts 11,520 expert Linears and
refuses a router in the quantize set before any GPU time. The live seat then passed
prose with no channel-prefix leak, a solid-colour image read correctly, an auto
tool call parsed, finite logprobs, and KV 534,649 tokens / 2.04x carried over from
Pfish-6 unchanged. The broken tree is parked on ana-ml2 as
...-NVFP4A16.BROKEN-routers-quantized-20260910.
Section 4.4's temp port was not reachable: 15.9 GiB of weights plus KV plus
multimodal encoder-cache profiling does not fit in the ~19 GiB free beside GPU1's
six other tenants -- 0.20 utilization refused admission, 0.185 OOM'd in encoder
profiling. The substitute was reversibility and ordering: named .env backup, prove
the seat on its real port while no alias points at it, move the alias last. That is
why a NaN-serving seat never reached a consumer. The seat was down about 16 minutes
across two attempts; no consumer saw a broken alias.
Playbook gains the router-quant failure signature and the control-class rule in
3.15, and a logprobs check in 4.4. seat_verify.py carries that check as check 6.
Quality is NOT established: no RP eval, no long-context check, no A/B against
Pfish-6 or char-rp. Samplers are the author's card values, untuned here.
The v2 dense quant had failed four times. Attempt 5 lands it at 19 G.
The blocker was not what it looked like. `AmbiguousGlobalPerLayerAttributeError`
on `head_dim` read as a malformed upload -- DogOnKeyboard's config carries a
`per_layer_config` key zerofata's canonical one lacks -- and the standing fix was
to force `allow_global_per_layer_attribute_access=True`. Both halves were wrong.
`pip install llmcompressor==0.13.0` downgrades transformers 5.16.1 -> 5.14.1. The
config was serialized by 5.16.1, which materializes `per_layer_config` from
`global_head_dim` + `layer_types`; 5.14.1 carries the heterogeneity guard but not
the gemma4 resolver. Under the image's own transformers the same config loads
fine. `:latest` was also re-pulled during attempt 4 and no earlier run, so the
toolchain moved mid-diagnosis. Two things separated "malformed upload" from
"moved toolchain": reproducing the real failing call (a bare AutoConfig load does
not reproduce it; the trigger is reached through AutoTokenizer) and keeping
zerofata's canonical tree, quantized cleanly on 2026-08-21, as a positive control.
The fix drops `per_layer_config` rather than forcing global access. It is exactly
redundant -- keys are precisely the ten full_attention layer indices, sole value
(512, 4), verbatim the global fields -- and forcing instead would make
`config.head_dim` answer 256 to the callers building the 512-wide layers.
patch_perlayer.py re-proves that redundancy at apply time and refuses if it ever
stops holding.
Verified on the tensor table rather than the exit code: the output is identical
family-for-family and count-for-count to the August canonical quant, with 356
BF16 vision-tower tensors preserved and input_activations=None. A GPU-free load
leaves 0 tensors on meta and generates coherent prose. The section 4.4 serve test
has NOT run -- GPU1 has 19.9 GB free against 19.5 GB of weights, so it needs a
live seat displaced.
Also fixes the A4B output, which had a truncation cap baked into its tokenizer
(max_length 8192) from being quantized with the calibration corpus.
Playbook gains section 3.17 for the pinned-transformers class and sharpens 3.16
to say drop the dataset outright for any A16 scheme.
Ships the two pending code changes alongside the memory that describes them:
train_voice_lora.py gains --eval-steps/--save-steps (the 3-epoch pilot overfit
with per-epoch eval and save_strategy="no", so the minimum was neither visible
nor recoverable), and the althing post-office compose is pinned to 3.6.2.
Index rewritten: in-flight compressed from 176 lines to the live state, three
detail files added for R49 D1-D3, the MeroMero four-failure chain, and the
althing rollout. Seven closed pre-08-27 entries archived; the guards held back
the rest because they carry open deferred pointers.
The LossLog callback filtered on `"loss" in logs`. Trainer emits eval under
`eval_loss` with no `loss` key, so every eval record was discarded and
loss-series.json showed zero eval points while the log a metre away carried
`eval_loss: 3.198`. An artefact that omits data which demonstrably exists reads as
"no eval was run" rather than "the collector dropped it", which is the failure mode
that costs someone a re-run.
Collector now accepts either key. Seed 2 gets it from launch; seed 1 is already
running with the old code, so recover_eval_series.py parses the eval records back
out of its log -- the data was printed, not lost -- and the chain runs that
recovery before cutting the generation arms, so the artefact is complete before
anything reads it.
The two seeds therefore differ in logging code but not in training math: the
callback only affects what is recorded, never what is computed, so the weight
trajectories remain comparable. Noting it because a difference between the two
arms whose spread sets the decision threshold is worth stating even when it is
provably inert.
brokkr-smithy corrected H02's incumbent naming: qwen3.6-35-a3b-heretic was retired
from the gateway roster on 2026-08-15 and is not what Skaldsong would call today.
Verified against the gateway and the seat itself -- alias `gen` resolves to
hosted_vllm/qwen3.8-27b-uncensored on ana-ml2:8015, container vllm-gen, 262,144
ctx. The arm targets that.
24 records, style-prompted on the same prompts and sampler as the other arms.
Alias resolved at run start AND end and confirmed stable across the run, per the
fleet rule that an artefact records the backing model rather than the alias.
Two things recorded rather than glossed:
The harness is NOT matched to the other arms and the artefact says so. Base and
adapted arms are local transformers on gx10; the incumbent is a served NVFP4 27B
reached over the gateway, and it is an instruct model receiving a style
instruction where the others are base models receiving none. That asymmetry is the
comparison H02 asks for -- prompted imitation against trained voice -- but it must
not be reported as if the harnesses were identical.
The gateway echoes the ALIAS in each response's `model` field, so a row read on
its own would have recorded "gen" as provenance -- the same class of mistake that
inflated an exposure count 4.7x on this fleet. Rows now carry
alias_echoed_by_gateway beside backing_model_resolved and its date, and the
generator was fixed at source rather than only in the emitted file.
Sanity: median 392 completion tokens, zero records opening with markdown or
meta-commentary, output reads as continuation prose. The style prompt was written
to be a fair incumbent rather than a strawman, since this arm is what the adapter
must beat.
Entity resolution, deterministic rename augmentation, packing and the pilot
trainer. Qwen3-0.6B-Base is training now: 507 steps, 11.2 s/it, ~1h35m.
D2 -- gender resolution is TITLE-FIRST, and that is a change from F02's method
rather than a port of it. F02 used pronoun proximity and recorded that it is
structurally blind to the first-person narrator, whose name appears mainly in
dialogue surrounded by other people's pronouns. Measured here, proximity called
JANE MALE -- the narrator of Jane Eyre and the single worst entity to get wrong.
Titles have no such blind spot: Miss Eyre, Mrs. Fairfax, Mr. Rochester, Madame
Beck, M. Paul, and a 19th-century novel is saturated with them. Measured: 16
entities resolved, zero wrong, every ambiguous case landing on HELD -- shared
family surnames like Helstone and Pelet genuinely belong to both a man and a
woman and hold as they should.
Held means ungendered, not unrenamed. A HELD entity is still renamed, from the
gender-neutral surname pool, because the operator's Yarros directive was "rename
all proper nouns" and holding a place leaks it -- Thornfield appears 100 times in
Jane Eyre and is as author-specific as Riders Quadrant was. Substituting a neutral
token makes no gender claim, so no gender claim can be wrong.
D3 -- pool is French + English per the operator, weighted per work by setting:
Brussels novels 60% French, Yorkshire novels 25%. Locales restricted to
fr_FR/fr_BE/en_GB/en_IE; en_US and en_AU carry modern surnames that are wrong
register for the 1840s. The pool is filtered against Brontë's own 75-letter
alphabet, so French accents stay and Czech/Latvian marks do not.
Two collision defects found by running the leak gate rather than trusting it:
`Burns` and `Marie` were drawn as replacements while being Brontë characters --
F02's collision filter was built against Yarros and does not carry -- and then
`Pierre-Yves` passed a whole-string filter while `Pierre` (Mademoiselle St.
Pierre) is a Villette character. The filter now compares by COMPONENT. Final gate:
0 of 203 source entities survive in any of 24 copy-files.
Trainer records what the run RESOLVED to rather than what it requested -- attention
implementation, dtype, device, corpus sha and harness cleanliness are read back off
the live objects. transformers 5.x has dropped warmup_ratio, caught by reading the
signature after the first launch failed on it; the 3% warmup is computed into
warmup_steps instead.
scripts/r49-corpus/{build_corpus,verify_corpus}.py; corpus staged at
gx10:~/r49-corpus/. Catalogue ids verified against gutenberg.org's own search
rather than recalled. Charlotte only -- the Bell poems are co-authored and the
Gaskell biography is a different hand, so neither belongs in a single-voice corpus.
Jane Eyre 1260 · Villette 9182 · Shirley 30486 · The Professor 1028
680,291 words · 142 chapters · 950,974 Qwen3 tokens (1.40 tok/word)
alphabet 75 letters, 23 non-ASCII · round-trip lossless · 0 byte-fallback
All 11 acceptance checks pass, including both tokenizer legs run against the pilot
carrier itself. With a real denominator the projections tighten: at 6 rename copies
x 3 epochs = 17.1M tokens, the 0.6B pilot is 1.98 h.
THE ALPHABET INVERTS THE YARROS RESULT. Brontë writes French constantly -- Villette
is set in a French-speaking city, Jane Eyre has Adèle, The Professor is set in
Brussels -- so the corpus carries é 432, è 237, à 93, ê 79, ô 48 plus œ and æ. F02
measured Yarros at 0.0002% non-ASCII and derived an ASCII-fold for the name pool.
Under F02's own subset rule the Brontë pool may keep FRENCH accents and must still
exclude the Czech/Latvian/Slovak/Hungarian marks that never appear here. The fold is
per-work, and this is the first corpus where deriving it changes the answer.
Typography was inconsistent across works and it was the transcriber, not the author:
Shirley uses straight quotes and `--` with zero em-dashes while Jane Eyre and
Villette use curly and em-dash. Normalised toward what the text means.
Three defects, each found by running something rather than reasoning about it:
`Produced by` matched Brontë's own prose four times, which is the adjective-"minor"
shape again and is fixed by anchoring boilerplate patterns to line start; asserting
open/close quote counts must be equal is wrong, because 19th-century multi-paragraph
speech legitimately runs a surplus of opens, so the real error signature is that no
paragraph may begin with a closing quote; and The Professor's table of contents puts
two chapter names per line, so a bare regex returns 38 headings for a 25-chapter
novel and a minimum-gap filter still leaks its tail -- the rule that works is that
the body's "CHAPTER I" is the last one in the file.
Records the operator's pilot ruling: trial on Qwen3-0.6B-Base first, move up only if
it produces something useful.
CronList showed job 12bdea3c firing hourly at :37 with the verbatim "check on the
run ... inform brokkr when ready" text; CronDelete removed it and the list is now
empty. No crontab entry, user systemd timer, or background shell was re-seeding
it -- the only on-disk hits are conversation transcripts.
This is at least the second kill: the operator killed the same loop on 09-09
("kill the check, no training is running") and that session's handoff warned
against starting one. The warning was insufficient because a cron-fired prompt
arrives as an ordinary user turn with no marker, so it is indistinguishable from
the operator typing it -- three turns went into re-verifying an idle fleet and
answering a timer carefully.
The memory entry now says what to DO rather than what not to start: run CronList
before answering that prompt, and treat a second identical arrival as a signal to
check the cron list rather than answer again. Worth the prominence because the
instruction chain ends in three outward-facing acts -- an althing message to
Miranda, standing up a serving seat, and cueing brokkr -- every one of which
carries something false when no run exists.
uv tool install --force --reinstall from ~/development/althing (clean tree, HEAD
== origin/master 719bc5f, pyproject declaring 3.6.1). althing-core 3.6.0 -> 3.6.1;
herald PID 2457743 -> 3690918 with the old process confirmed gone; version read
back from the running install; postbox status reachable/push/alive; journal clean
since restart. postbox, althing-listen and althing-route moved on the same install.
Six boxes stay queued and the plugin hop (0.1.6) is separate and not done, so this
is deliberately not a "fleet done" claim. Reported to forseti per box.
Records one verification trap worth keeping: POKE_TEXT in post_office_herald.py is
a four-part implicit string concatenation, so a line-oriented grep for the new poke
sentence returns empty on a CORRECT 3.6.1 checkout. The first pass came back empty
for both the new text and the old, which reads as "new absent, old removed" -- two
wrong conclusions from one wrong pattern, detectable only because both cannot be
true at once. Check the effective value: reconstruct the concatenation or import
the module and print the constant.
Operator: "purge the merged models, keep run06 and the v6 quant." Eleven merges
removed with literal paths, one rm per line.
pfi-gx10 merged-run03c, merged-run04, merged-run05 3 x 49 GiB -> 145 GB
ana-ml2 merged-final, merged-run02, merged-run03,
merged-run03-s{025,050,075}, merged-test,
nvfp4a16-test 8 dirs -> 354 GB
gx10 382G->237G used, 632 GB free. /tank/erp-tune/serve 354G->8.6M with no
snapshots holding the blocks; the pool's raw FREE moved 4.80T->5.30T. Combined
with the earlier checkpoint purge, ~573 GB reclaimed tonight.
The check that made this safe: docker inspect on the live vllm-erp-seat shows it
binds /tank/aimodels ONLY, with model arg /tank/aimodels/erp-tune-v6-nvfp4a16, so
/tank/erp-tune/serve was never in the serving path. Also confirmed no container
mounts that tree and no process held a file open under it, reading /proc/*/fd and
/proc/*/maps rather than trusting an empty lsof. All eleven were run-1/2/3-era
Gemma-4 merges dated Aug 24-26; the keeper is dated Sep 8 and lives in a different
tree, so there was no ambiguity about which was which.
Kept and re-verified after the deletion: gx10 serve/merged-run06,
/tank/aimodels/erp-tune-v6-{bf16,nvfp4a16}, all eight adapters sha256-unchanged,
and the merge/quant tooling and logs under /tank/erp-tune/serve (8.6 MB) that
document how the artifacts were built.
The live seat never bounced -- Pfish-6 answered a real completion after the delete
with finish_reason stop and correct text, container still healthy at 4 h uptime.
relaunch-trial-seat.sh now names a deleted model; it was banner-marked RETIRED
rather than removed, because its flags carry the FlashInfer JIT/PATH trap and the
gpu-clear / never-pkill notes.
The operator asked brokkr-smithy to look for a smaller Qwen3.6/3.8; brokkr reported
none exists and that both 27Bs are the same hybrid kernel path one size up. Verified
independently against the HF API rather than adopted, and it holds -- with two facts
brokkr's note did not carry.
1. No official Qwen3.6 or Qwen3.8 below 27B. The listing is Qwen3.6-{27B, 35B-A3B}
and Qwen3.8-{27B, 2.4T-A95B, Flash-Next}.
2. Neither family publishes a -Base checkpoint AT ALL. Every Base newer than Qwen3
is Qwen3.5. Since the regime requires a non-instruct carrier, that rules the
3.6/3.8 lines out on its own, independent of size.
3. Qwen3.6-27B and Qwen3.8-27B both report model_type qwen3_5, 64 layers as 16 full
+ 48 linear-attention, vision tower present -- confirmed as stated.
So the dense Qwen3 ruling is not a compromise against a better available option; it
is the only dense Base option. Also recorded as insurance, not a proposal:
Qwen3-8B-Base and Qwen3-14B-Base exist, so the sweep can extend past 4B without
changing family if the carrier curve has not flattened there.
brokkr-smithy flagged that R49 F02's name-pool token splits were measured with the
Qwen3.5-2B tokenizer, so the dense-Qwen3 carrier ruling invalidates them. Measured
rather than left on their critical path; handed over as input to their re-check,
since the dictionary and the adjudication are theirs.
The multi-token property strengthens on the chosen carrier: pool multi-token
88.0% -> 90.3%, mean tokens 2.33 -> 2.46. A smaller vocabulary fragments more, so
Qwen3's 151,936 splits names into more pieces than Qwen3.5's 248,320. The operator's
requirement that names be multi-token, so the drafter reconstructs them from the
prefix instead of recalling one embedding, is better served after the ruling.
Positive control: the Qwen3.5 column reproduces F02's published figure on the same
pool and tokenizer (F02 89% / mean 2.35; here 88.0% / 2.33), so the instrument
recovers a known-true value before being asked about an unknown one. The pool is
deduped across locales, which reconciles male_given and female_given exactly
against the dictionary's own totals block.
Two operator rulings, 2026-09-09.
"purge intermediate checkpoints" -- seven checkpoints/ directories removed with
literal paths, one rm per line, after confirming none was a symlink and that
every run's final adapter/ is an independent real directory:
pfi-gx10 run-03c 11G run-04 16G run-05 9.2G run-06 9.2G = 45 GB
ana-ml2 run-01 12G run-02 12G run-03 5.9G = 29 GB
gx10 419G->374G used, 496 GB free. /tank/erp-tune 392G->363G with zfs list -t
snapshot empty, so the space is genuinely returned rather than snapshot-held. All
eight adapters re-verified by sha256 after the deletion, matching the values
recorded during the mirror. Merged artifacts deliberately untouched -- they are
not checkpoints, and the ~550 GB of superseded merges stays a separate call.
"use dense qwen3" -- the R49 H02 carrier sweep becomes Qwen3-{0.6,1.7,4}B-Base,
which overrides the Qwen3.5 arms H02 names; brokkr-smithy owns that file and was
told directly. Qwen3-4B-Base staged and benched to complete the family:
Qwen3-0.6B-Base 0.616 B 1.707 s/step 2,399 tok/s spread 0.6%
Qwen3-1.7B-Base 1.755 B 2.895 s/step 1,415 tok/s spread 0.8%
Qwen3-4B-Base 4.089 B 5.714 s/step 717 tok/s spread 0.3%
The dense 4.089 B carrier still trains 33% faster than the hybrid 0.765 B one.
Projected per voice 2.7 / 4.6 / 9.1 h; the three-arm sweep at two seeds is ~33 h
of GPU, ~10 h if H03's corpus floor holds. The three Qwen3.5 checkpoints stay
staged so the decision is reversible behind an fla install.
Also recorded: verified at 22:45-22:48 PT that nothing is training on gx10,
ana-ml2, nh3-dev or irv-ml1, and that brokkr's own run07-gate close states
"Nothing is owed. No battery to run." Run 7 has no servable artifact left. And a
correction to a standing lesson -- the bracketed-class trick does not defeat a
wrapper's argv, since the invoking shell's command line carries the literal
pattern; observe the artifact instead.
The base-viability pre-flight had three checks (fits / MoE expert mapping / LoRA
support) and would have passed Qwen3.5-0.8B-Base clean while it trained 2.6x
slower than a dense model 2.3x its size. Check 4 closes that: read `layer_types`
for a linear_attention majority AND probe for mamba_ssm / causal_conv1d / fla /
kernels. It is the intersection that is slow -- a hybrid shape with the kernel
present is fine, a dense shape does not care.
Carries the measured table (gx10 GB10, n=10/arm, spreads 0.6-2.6%), plus the two
things a hybrid Base checkpoint brings that a dense one does not: a vision tower
and MTP head that target_modules="all-linear" would train on text, and the module
rename that AutoModelForCausalLM introduces relative to the vLLM serving class;
and unsafe cross-document packing, since SSM state ignores the attention mask.
Section heading corrected from "three greps" to "four checks". The example was
made runnable and verified on the box rather than shipped untested.
Prep for the BabyBronte / brokkr-smithy R49 author-voice adapter regime, plus
the operator's "keep the adapter" ruling made durable.
Measured on pfi-gx10 (GB10, sm_121), n=10 per arm after 3 warmup steps, seq
4096, LoRA r=32 on q/k/v/o + MLP, bf16, sdpa, grad-checkpointing on:
Qwen3-0.6B-Base dense 0.616 B 1.707 s/step 2,399 tok/s
Qwen3-1.7B-Base dense 1.755 B 2.895 s/step 1,415 tok/s
Qwen3.5-0.8B-Base hybrid 0.765 B 7.581 s/step 540 tok/s
The dense 1.755 B carrier trains 2.6x faster than the hybrid 0.765 B one on 2.3x
the parameters (~6x per parameter), with more LoRA modules adapted (196 vs 96).
Spreads of 0.6-2.6% put instrument noise an order of magnitude below the effect.
Cause: Qwen3.5 is 18 linear-attention (SSM) layers to 6 attention, and no fused
linear-attention kernel is installed on the box. Grad checkpointing is not the
culprit (19%, and saves 2.6x memory). Batching is not the lever for either
family -- both sit at this box's roofline at batch 1.
Projected per voice on a Brontë-scale corpus: dense 0.6B 2.7 h, dense 1.7B
4.6 h, hybrid 0.8B 12 h. The hybrid would take longer than the 7 h 26B-A4B tune
the regime exists to replace, so the carrier family is now an open decision with
a recommendation for the dense Qwen3 line -- the design doc's original pin.
Two further Qwen3.5 findings, both measured rather than read off the config: the
Base checkpoints ship a vision tower (153/297 model.visual.* Linear tensors that
target_modules="all-linear" would train on text) and an MTP head, both dropped
for free by loading through AutoModelForCausalLM -- which renames modules
relative to the vLLM serving path, so adapter binding needs the
sampled-target-changed check on the serving side; and cross-document packing is
unsafe because SSM state ignores the attention mask, breaking the per-copy
name-consistency invariant the design doc calls sacred. Neither exists on dense.
Adapter disposition, per the operator's ruling: all five gx10-resident ERP
adapters (run-03c/04/05/06/07) mirrored to ana-ml2:/tank/erp-tune/run-<N>/adapter
matching the layout runs 01-03 already used, byte-totals identical both sides and
sha256 matching on every adapter_model.safetensors. /tank/* is deliberately
excluded from ana-ml2's restic sources, so the profile gains one documented
carve-out for /tank/erp-tune/run-*/adapter, verified by resticprofile --dry-run
to expand to exactly those eight paths.
Nothing is training and nothing is queued.
Operator ruling: run 6 is declared Pfish-6 and is the standing seat; run 7 is
retired (its gate failure was the detector bug fixed in cc42d76, but the run was
independently poor).
- served under its TRUE name Pfish-6; LiteLLM alias trial -> Pfish-6
- ana-ml2 :8021 and pfi-gx10 :8098, both at 262,144 ctx, same artifact
- SPEED: moe_backend=flashinfer_cutedsl PROVEN unusable here (engine init fails,
'kernel does not support current device'); Marlin is correct for a weight-only
scheme, and vLLM's 'no native FP4' warning is about the scheme not the card
- the real lever was max-num-seqs 8 -> 32: n=8 1269 tok/s and n=16 2170 tok/s,
3.2x the old ceiling, with single-stream latency unchanged
- head-to-head measured: ana-ml2 is 4.1x FASTER than the GX10, not slower
vh