Correcting a claim I made from a bad query: Q4_K_S does exist for this finetune.
`mradermacher/Qwen3.8-Flash-Next-Uncensored-i1-GGUF` ships the large quants as
`.partNofM` fragments — Q4_K_M 37.00+37.00+36.97 = 110.97 GiB (which is how the
deployed file was assembled) and Q4_K_S 35.00+35.00+34.09 = 104.09 GiB — while
the smaller rungs are single files. A tree listing filtered on `.gguf` makes
Q4_K_S invisible, which is exactly how I concluded it did not exist.
Also records the fetch trap that hid it for a moment: `curl -o FILE URL` exits 0
on 401/403, so a gated repo produced three 153-byte "shards" with identical
sha256 values and a cheerful DOWNLOAD_DONE in 20 s for 104 GiB. -f plus the
Hub's declared size plus the LFS sha256 is the standard the fleet's own
download-frank-medium-gguf.sh already meets.
The log mtimes plus each run's own duration pin it: the first 'invalid internal
status' in the user journal is 18:58:30, while the F bench (launched 18:54:31)
was still loading, and the preceding llama-cli placement probe had just spilled
3.2 GB of '> ' prompts. The note had said 'after the experiment'; that was wrong
and is now recorded with the timestamp it can be checked against.
I had written 'memory pressure from the 116 GiB iGPU runs' as the cause of the
pause-process death. journalctl -k records no OOM kill for that window and the
kernel says nothing else about it, so the claim was a guess wearing a finding's
clothes — in AGENTS.md, the x4 note, and the log. All three now state the
trigger is not established, and the note's 'co-tenants' patience' bullet no
longer asserts that causality either. AGENTS.md also gains the real lesson:
journalctl -k is the check, and an unknown cause does not need a story.
The ledger number in the note was pre-reboot. Same model, same pool, no clamp —
but the host was measurably more fragmented when it came back, which is the sort
of thing that decides whether the next co-tenant fits.
Asked whether Vulkan on both devices unlocks the split. It does not: the
per-tensor design aborts byte-identically, because the gate is not the driver
mix but `--override-tensor` — `has_tensor_overrides()` is a required term of
the scheduler's `pipeline_parallel` flag (llama-context.cpp:433, the only place
in the tree that reads it), and that flag sets `n_copies` in
`ggml_backend_sched_new`. With `n_copies == 1` a pre-allocated weight on a
non-primary device has no backend and the graph aborts.
Chasing it exposed a harness artifact that invalidated the earlier conclusion:
llama-bench splits a device LIST with `/` and SWEEP values with `,`, so
`-dev CUDA0,Vulkan1` was two single-device runs all along (and `-ts 0.85,0.15`
two sweep values, i.e. device 0 takes everything — which is where the bogus OOM
came from). With the correct syntax two-device splits work in both flavours:
pure Vulkan -dev Vulkan1/Vulkan0 -ts 0.85/0.15 -> 27.55 tok/s @64k (12.8 GiB on the 3090)
cross-vendor -dev CUDA0/Vulkan1 -ts 0.15/0.85 -> 27.23 tok/s @64k, pp512 268.5 (15.0 GiB)
iGPU alone 25.77 tok/s @64k, pp512 231.8
+7 % decode, +16 % prefill — real but not a regime change, and it costs the
3090 ~15 GiB, which this fleet spends on its only vision endpoint. The design's
premise still fails: whole layers carry their experts (1.62 GiB each), the 12
attention layers are spread every 4th, and the 32.8 GiB PLE must stay on the
main device, so KV cannot be consolidated on the 3090.
Also records the evening's incident (a dead rootless podman pause process took
all container management down; `podman system migrate` panics in 5.7.0; reboot
repaired it) and the netavark host-address lesson from the hermes 502.
The design we sketched (trunk + KV on the 3090, MoE experts on the iGPU) is not
expressible in mainline llama.cpp. Per-tensor cross-backend placement aborts in
the scheduler -- "pre-allocated tensor (blk.0.ffn_down_exps.weight) in a buffer
(Vulkan1) that cannot run the operation (NONE)" -- and both per-layer
cross-backend forms fail at load. CUDA + Vulkan cannot serve one model; the CPU
is the only fallback side that works, which is exactly why -ot exps=CPU runs.
What the same build does instead is the surprise: iGPU-only (Vulkan/RADV, PLE in
page cache) gives 32.64 tok/s shallow, 25.77 tok/s at 64k depth and 456.7 tok/s
prefill, with the 3090 idle at 0 %. That matches halogen's deployed ~25 tok/s --
but commits 116 of the host's 121 GiB against halogen's 95.4 held with 14.7
spare, so it is not deployable beside the co-tenants. Throughput parity, memory
bankruptcy: the fleet's two-endpoint split stands, now on evidence rather than
preference.
New page LLM/x4-pair-experiment.md carries the tensor inventory, the results
table, the two scheduler limits, reproduction commands, and the traps found on
the way. Both index pages also stop calling halogen retired -- the ik swap's
stale claims were still sitting in them.
ngram-mod is draftless (added to the spec config vector unconditionally, while
every model-backed drafter needs ctx_dft), so unlike MTP it can share the
process with DFlash2. Chained at the fleet's 48/64 window, over three shapes:
prose 1k/512 dflash 43.77 | ngram 17.72 (0 drafts) | chain 43.65
echo 1.5k/400 dflash 76.82 | ngram 102.50 | chain 200.58
structured/400 dflash 91.08 | ngram 17.80 (0 drafts) | chain 91.22
identity dflash 70.37 | ngram 17.77 | chain 70.40
On prose the chain is byte-identical to the drafter alone (same 119 drafts/49
accepts) -- that is what makes it a safe default -- and on prompt-echo traffic
it is 2.6x. VRAM is unchanged: the matcher is a 64 MiB host-RAM hash table.
The window sweep says keep the long one: 48/64 beats 16/32 (161.0) and 8/16
(113.8) on echo, and shortening it costs prose (43.9 -> 36.0) because the
matcher then proposes on ordinary fragments and the rejections still have to be
verified. That is the 'ngram hurts' result this fleet hit twice before.
Two documentation traps recorded with it: ngram-mod prints NO startup banner in
this build (its drafts are only visible behaviourally), and the harness's
'structured' shape is a negative control because its lines differ.
Four configurations of the same binary, identical prompts at temperature 0, the
unit stopped between runs, on the deployed payload:
none 17.7 tok/s card1 26.07 GB
draft-mtp n=3 32.4 tok/s 58 % accepted card1 29.35 GB
draft-mtp n=7 26.7 tok/s 33 % accepted card1 29.65 GB
draft-dflash 7 43.5 tok/s 41 % accepted card1 29.42 GB
DFlash2 is worth 2.46x over autoregressive and 34 % over the best MTP, and all
four produced IDENTICAL greedy text, so the 'lossless verification' claim in the
runbook is now a check rather than an assumption. MTP is not the cheaper option
either: its head loads ~3.3 GB against the drafter's 2.06 GB file, and deeper
drafting loses (n=7 accepts a third of its proposals, n=3 more than half).
The two CANNOT be chained, and the failure is structural rather than tunable:
'spec-type draft-mtp,draft-dflash' dies at load with 'context type MTP requested
but model doesn't contain MTP layers', because a chain has ONE draft context and
MTP builds against whatever model that context holds -- here the DFlash2 file.
The serve script therefore passes --spec-draft-model in the dflash branch ONLY,
and the seam now asserts exactly that (the wiring whose violation is this
error), plus that the deployment stays on the measured winner.
scripts/bench-spec-ab.py makes the comparison reproducible: the benchmarking
policy wants the runtime's own harness for this, and the numbers above were
otherwise unreproducible one-offs.
The loader prints no per-device ledger at this verbosity, so the claim is
measured rather than read: +131072 tokens of context added 4.83 GB to one card
and 4.84 GB to the other (both carry their own attention layers' KV), and
neither card's total usage could hold the 31.46 GB file. Worth writing down
because the retired shape WAS two full copies — that is what a replica is, and
it is why the 29.8 GB FP8 checkpoint could never run there.
The LLM index still opened with vllm-radiance as frank-medium's runtime and
drew the replica pair in its topology diagram; the cooperation page's thesis
(collectives are impossible, so the answer is hardware) now has to name the
path the fleet actually took: llama.cpp's layer split, which launches no RCCL
kernel on either card. That reframes the M.2 riser as a means to TENSOR
parallelism rather than a means to using both cards at all.
Two families of ~30k-token turns ran against the deployed endpoint: 24.8 s and
51.4 s for the first turn (the second waits for the slot), 30.7 s and 8.0 s for
the second. Against the prefill table that is ~25 s of cold prefill plus the
wait, so the honest reading is that ONE slot is now the limiting resource for
concurrent agent loops — the trade this deployment makes for giving a single
conversation the whole 262144-token window. --parallel splits the window; it
does not add throughput.
The harness header still described the retired two-replica proxy ('the deployed
routing pins such a conversation to one replica', vLLM's prefix-cache key). It
now says what is true: one endpoint, one slot, concurrent families queue, and
the old routing numbers are history in the runbook.
Context: 262144 was measured before it was deployed -- 24.72 GB on GPU 0 and
29.14 GB on GPU 1 of 34.21 GB each, ~5 GB of headroom on the busier card, and a
140,069-token prompt prefilled through the single slot in 134 s to prove the
window is real. The 114688 wall was the vLLM predecessor's arithmetic, not this
runtime's.
Throughput, measured through the service: 1.0-1.4k tok/s prefill from 4k to
140k, 43.8 tok/s decode on a 512-token generation, 41-71 % DFlash2 acceptance,
and the prefix cache re-prefilling 4 of 998 tokens on a repeated prompt.
The signals move with the runtime, so the fleet's rules and dashboard are
re-cut rather than renamed: vllm: series do not exist any more, and two of the
old rules watched things llama.cpp does not export at all (KV-pool fill, p95
latency). Those are retired with the reason written down, and LlmSpecDecodeOff
takes one of the slots -- a lost drafter is the failure no client can see. The
observability seam now fails if any rule names a series the endpoint does not
export, which is what would have caught the stale vllm: expressions.
omp models.yml (outside the repo) is updated in the same pass: contextWindow
262144, and the full reasoning-effort map, because Qwen3.8's template answers
500 on 'high' -- verified against the live endpoint, low/medium/xhigh answer
200. Subagents run at high.
The runbook for the new deployment, plus the sweep of everything that still
described the vllm-radiance replica pair: the vault index and its ACTIVE
marker, the superseded banner on the radiance page, the pointer from the
ROCmFPX page, the homepage card, the Grafana dashboard description and the
observability page's scrape table.
The new page carries the two facts a future agent needs and the flags do not
say: llama.cpp's layer split is what lets BOTH cards serve ONE model without
any collective (vLLM's PP handoff is an RCCL send/recv, which is the primitive
this host cannot dispatch), and the rocBLAS Tensile-path trap that makes a
healthy-looking build fail on the first request.
halogen-flash-server is ACTIVE again with its 2026-09-16 measured ledger (budget
33.5 GiB against a need of 27.2 — no clamp — and 14.7 GiB left for everything
else), ik_llama.cpp is retired with its measurements kept as the record of why,
and the new frank-small runbook covers the -mtp file, the 64 KiB/token KV
arithmetic, the MTP contract, and the reasoning_effort trap that cost me every
subagent on the endpoint.
AGENTS.md: topology, Active Services, runbook index, and three new llama.cpp
lessons (MTP is a property of the file; --jinja is what `tools` needs; a model's
template can reject the harness's thinking level). Three pre-existing defects
found while editing it: a truncated sentence mid-file, a harness path that does
not exist on either server, and the benchmarking stop/mask recipe still calling
frank-large a root unit.
Truthfulness notes worth keeping: the co-tenant stall NOTE prints again, so the
claim that it "no longer prints" is gone and the three observed block counts are
recorded instead of a guessed threshold; and the Observability page says out
loud why frank-large has no scrape job rather than leaving a silent absence.
New page: LLM/Dual-R9700-Cooperation.md — registered in the vault index, with
the measured evidence rather than the theory.
**The answer.** Card #2 sits behind the B650 chipset, whose bridges advertise
`AtomicOpsCap: Routing-` (read-only), so ROCr refuses to fill the hostcall
buffer RCCL's kernels declare and no RCCL op can dispatch on it. Two paths fix
it: an M.2→PCIe x4 riser into the free, CPU-attached M.2_2 slot (~$30; per
ASUS's spec page M.2_1 and M.2_2 are CPU-attached, PCIEX16_2 is chipset in x1
mode, and there is no x8/x8 bifurcation on this board), or a board whose CPU x16
bifurcates to x8/x8. Bandwidth is not the constraint: ≈10 MB per token-step ≈
1.4 ms at Gen4 x4.
**New measured results.** A bare 2-rank torch.distributed run: rank 0 (CPU-lane
card) passes broadcast/all_reduce/barrier; rank 1 (chipset card) fails all
three, and `send`/`recv` fails too — which kills pipeline parallelism as an
escape, since vLLM's PP handoff is exactly that primitive.
**Correction.** I previously said no software path exists. That overshot. One is
open and unverified: recompiling RCCL so its kernels stop declaring a hostcall
buffer (`-mprintf-kind=buffered` is the claim). Counter-evidence recorded: RCCL's
device source contains no live printf/assert/__assertfail, so the trigger is
unexplained, and the claim traces to search summaries, not a primary source. The
page carries the two-command test that would settle it. Also corrected: 0.9.3 is
radiance's version, not vLLM's — the base is vLLM 0.27.1 / torch 2.11.0 / ROCm
7.14, and its shipped RCCL exposes no `ncclDevComm*`, so torch's symmetric-memory
path cannot run.
**Rule added to the benchmarking policy**: it covers ANY GPU experiment, not just
benchmarks — including a bare 2-rank probe. Running one against the live service
on 2026-09-16 preceded the host going hard down (no ARP, no ICMP, power cycle
required); a hung AQL dispatch wedging the driver is the plausible mechanism.
The policy also now says how to tell off from wedged: `ip neigh` from a peer.
Retiring dankclaw/compose.yaml and adding alert rules invalidated claims across
the repo, including in the rule file agents read first. Clean cutover includes
the prose.
**Tests that had gone hollow.** Four assertions grepped dankclaw/compose.yaml
for a service name. With the file gone they passed by *absence* — a gate that
can no longer fail. Each now asserts the real invariant ("no dankclaw compose
file competes with the quadlet units") and keeps a live check for a re-added
entry, so it still fails when one appears: verified by recreating a compose file
with a comfyui service and watching three tests go red, then removing it.
**AGENTS.md, which was wrong about ComfyUI.** It said ComfyUI runs CPU-only
("no ROCm wheels for gfx1151, CUDA images wrong"). That was true in the
compose era and is false now: the restored unit runs with /dev/kfd + /dev/dri and
must NOT pass --cpu. An agent trusting the rule file would have re-broken it.
Also corrected: the observability paragraph ("RUNNING BUT UNCONFIGURED, no
datasource, no rules" — all three have been configured since 2026-09-16), the
Active runbook row that still named halogen for frank-large, halogen's absence
from the Retired table, the benchmark policy naming halogen's harness, and the
seam list.
**Unit and page claims.** prometheus.container said "no alerting rules anywhere
in this stack"; node-exporter.container said Dankserver runs from its compose
stack; Prometheus.md documented three jobs as working (they scrape HTML: they
can never report up) and said no dashboard existed; index.md and the three
authentik service pages pointed at the retired file; Authentik.md's upgrade
instruction told you to bump an image tag in it.
One bad edit of my own on the way (a stale line number clobbered a See Also
entry in Prometheus.md) — caught by reading back, fixed, and the vault link
checker now proves the cross-references hold.
The vault is the project's source of truth, so a link to a page that does not
exist is a promise the repo cannot keep. Three were dangling:
- Dankclaw/Gitea.md pointed twice at dankclaw/compose.yaml — **which this
session retired**. The clean cutover obligation from my own change: the page
now describes the unit that publishes 3000 and the canonical Caddy unit, and
records why the compose file is gone.
- Dankclaw/Hermes-Dashboard.md linked [[LAN DNS]]; the page is [[LAN-DNS]].
- LLM/frank-medium-radiance-mxfp4.md linked [[Observability on Dankclaw]]; the
page is [[Observability]].
scripts/check-vault-links.py + a seam keep it that way. The checker models
Obsidian's rules, which took three corrections to get right — and each was a
false positive that would have sent me editing healthy docs:
1. Wikilinks are case-insensitive. A case-sensitive matcher reported 45
breaks; 39 were `[[qBittorrent]]` vs qbittorrent.md.
2. `[[folder/Note]]` resolves by path suffix, not by stem.
3. `%20` in a markdown link is URL-encoding, not a filename.
Validating the linter against the resolver it models is the prerequisite for
trusting any of its output — the same discipline that caught a bad assertion of
my own earlier tonight.
The board showed mariadb, valkey and Postgres as "down". None of them is down:
the checker runs on Dankclaw and can reach that host's loopback, its container
names, and published LAN addresses — while all three cards probe
container-internal database ports (192.168.0.235:3306, 127.0.0.1:6379/:5432).
They were red by construction, which is how a status board stops being read.
Publishing a database port on the LAN to make a badge green is not a trade this
fleet makes, so they are wip with the reason in their description, and the apps
that front them carry the status. Gluetun (WireGuard, UDP) was the pre-existing
example of the same call.
The rule is now written down in the Homepage doc: a card must be probeable from
the monitor's host or be wip. That is the same lesson as the Prometheus jobs
that scrape HTML pages and can never report up — an always-red signal teaches
the operator to ignore the surface it lives on.
Two gaps behind the same near-miss.
1. **No spec covered hermes.franksdank.cloud** — the one guarded vhost whose
upstream is a host address (`reverse_proxy 192.168.0.143:8787`) rather than a
container name, because its service is host-networked. That is how a
loopback-only bind could take the vhost down with a 502 and leave every gate
green. tests/e2e/specs/17-hermes-vhost.spec.ts asserts what a dead upstream
violates: no 5xx through the guard chain, no Caddy error page, and no app
bytes to an unauthenticated raw client. It does not assert SSO-silence, since
this app has its own password login.
2. **Specs are matched by number prefix, not globbed.** A spec file whose prefix
is in no project silently never runs — which is exactly what happened to the
new one until it was registered. The prefix is registered now, the comment in
playwright.config.ts says why, and tests/e2e_scaffold.bats fails if any
spec file matches no project.
The hermes-webui seam assertion is corrected too: it asserted the loopback bind
I wrongly imposed, and now asserts the real contract — the unit's bind must serve
the address the Caddyfile fronts it on.
ComfyUI's page described a compose service and a systemd unit that shelled out
to `podman compose up`. It now documents the Quadlet, the keep-groups + udev
mechanism (verified inside the running container: torch sees Radeon 8060S), the
9,146-restart incident that forced the move, and the trap that caused it — a
stale container outlives the compose edit that should have replaced it.
The observability page gains the audit result: four jobs read "down" and three
of them always will, because they scrape HTML pages or a /metrics that does not
exist. The page names each error, says plainly that none is an outage, and
refuses the tempting fix (deleting the job makes an absent scrape look like a
healthy one) in favour of the remedy it already recommended: a blackbox exporter
in probe mode, and Gitea's metrics enabled or the job repointed.
The page said "all five are rootless Quadlet units on Dankclaw", which stopped
being true when Dankserver's node-exporter was restored to a unit. It now names
both hosts and both deployments, records why the second one happened (the same
compose-only reboot argument, a day and a half of missing host metrics, and a
wildcard bind), and states the exposure rule for a two-listener fleet with the
`ss` evidence that the wildcard is gone.
Every measurement behind the frank-medium runbook was taken with scripts that
lived in /tmp on two hosts — reproducible only until the next reboot, and
invisible to anyone reading the repository. This moves the useful four into
scripts/, where this repo already keeps its bench harnesses:
bench-vllm-corpus.py code-shaped corpus from the repo's own sources, because
synthetic random prompts misreport draft acceptance
badly (2.9-3.6 tok/update against ~5.8 on code)
bench-vllm-serve.sh concurrency sweep; encodes the two traps that produced
silent garbage — the tokenizer argument, without which
the client dies resolving "frank-medium" on HF and
reports no numbers at all, and the --ignore-eos
semantics that must not be compared across
bench-turn-latency.py per-turn TTFT for an agent-loop shape; the instrument
that decided the routing policy, and the one that shows
cold-vs-warm prefill directly
bench-llm-soak.py sustained multi-family load; what found the capacity
wall (three 30k-token families, 75% cold re-prefills,
zero failures)
Each header carries its traps and the interpretation; the runbook's Benchmarks
section now points at them instead of describing methodology with no tooling.
Also adds the missing half of the script seam: scripts/*.py are compiled, not
just scripts/*.sh syntax-checked — a harness with a syntax error could otherwise
ship unnoticed until someone needed it during an incident.
Both python harnesses were smoke-tested against the live fleet before commit:
the corpus builder wrote 15 prompts, and the turn harness measured 12.5 s cold
against 1.4 s warm on the same prompt — the cache behaviour the runbook claims.
Both pages claimed there was no alerting; there are six rules now, so they say
which rules exist, that promtool and /api/v1/rules both confirm them, and — in
the same breath — that no rule reaches a person, because this stack has no
Alertmanager and no contact point. The dashboard's ALERTS panel is named as the
only place an alert becomes visible without opening Prometheus.
The page said Grafana has zero datasources and zero dashboards; that is now
false, so it says what exists (Prometheus datasource, LLM fleet dashboard), how
it was verified (the API — a provisioned dashboard does not show up in the
legacy dashboard table on this Grafana version, which is a trap worth recording)
and what is still missing (nothing alerted on, Loki unread, correlation manual).
The section heading drops "UNCONFIGURED" because it no longer is, entirely.
The capacity work in this runbook was done with ad-hoc probes; the same signals
now have history in Prometheus, so the page says which series to watch and
admits the honest gap — this stack has no alerting, so a saturated pool shows up
on a graph after the fact rather than paging anyone.
Two more hypotheses, both measured on the same three-family workload (one
~80k-token family plus two ~50k-token families, each turn re-sending its
document):
1. Long-prefix caching is fine. One card, idle engine, a ~50k-token prompt
sent three times: 22.3 s cold, then 1.4 s, 1.4 s — with a varying tail too
(22.4 / 1.41 / 1.41). So the engine's lifetime "Prefix cache hit rate: 29%"
is an average across every prompt that replica ever saw, not a verdict on
any one request, and the pinning policy is safe.
2. Serialising large requests does not fix the overload. Admission control
(at most one >16k-token request per card, small requests untouched) gave no
clear win on small samples — heavy 25.4 s -> 48.2 s median, mediums
34.7/45.2 s -> 22.7/22.6 s, cache-hit rates 14-33 % both ways — so it was
reverted like the placement change before it.
The arithmetic now closes: with three families on two cards, one card holds two
families and must carry 2 x 50k of cached prefix plus the 50k of the sequence
generating — 150k against 122,636 tokens. Those co-tenants evict each other
regardless of concurrency limits. The wall is capacity, and the runbook says so
with the numbers.
The obvious next lever for cache thrash was placement: send a new conversation
to the card carrying less context instead of alternating. Implemented it
(decayed prompt-token estimate per replica, continuations still pinned) and
A/B'd it against the counter on the workload it targets — one ~80k-token
family plus two ~50k-token families, each turn re-sending the same document:
counter heavy 13 turns median 25.4 s (23% cached)
medium 34.7 s / 45.2 s (14% / 15% cached)
balanced by load heavy 5 turns median 46.7 s (20% cached)
medium 23.4 s / 34.3 s (0% cached, small sample)
No improvement, so the change is reverted and routing.py keeps the counter.
The reason is arithmetic: a running request's live sequence and its cached
prefix draw on the same pool, so a card running two 50k-token generations has
already spent its 122,636 tokens before any cache entry is considered.
Placement chooses which card is over capacity, not whether one is.
Third hypothesis measured and rejected tonight (with SPEC depth kept and CHUNK
disproved); the runbook records all three so the next operator starts from
evidence rather than from these ideas looking obvious.
Rebooted dankserver unattended and verified without any manual step: all three
units active 59 s after boot, both replicas healthy ~3.5 min after boot, real
completion through :7770 — the failure mode this repo has been bitten by before
(a service that only ever ran from compose and stayed dark after a reboot).
Records why it works: [Install] WantedBy=default.target wired into the
generator's default.target wants, plus Linger=yes, plus a disk-backed compile
cache so a cold boot needs no warm-up beyond the model load.
A 25-minute soak with three concurrent large-context agent loops (each turn
re-sending a ~30k-token document, growing history) came back clean — 155
requests, 155 answers, no errors, no failover, both replicas steady — but only
25% of turns hit the prefix cache, the rest paying a 13.9 s median re-prefill.
That is the capacity budget showing: 122,636 KV tokens per card is about two
30k-token conversations with their cached prefixes, and a third family on the
same card evicts.
Chasing the obvious lever disproved it. Available KV per card is 5.44 GiB at
CHUNK=8192, 4.22 GiB at 16384 and 4.16 GiB at 4096 — where the engine refuses
to start. The budget is not monotonic in the chunk size, so 8192 is a sweet
spot rather than a compromise and freeing KV needs another card, not a knob.
Upstream ships dflash depth 7 as a content-dependent default measured for TP=2
on their corpus. Re-measured on one card here, two runs per setting, 16
repo-derived prompts at 512 output tokens:
SPEC=7 conc1 46.2 tok/s, TPOT 20.8 ms, 3.86 tok/update
conc4 118.7 / 127.0 tok/s
SPEC=5 conc1 41.5 tok/s, TPOT 23.1 ms, 3.41 tok/update
conc4 136.1 / 129.4 tok/s
Step time is nearly identical at both depths (~79 ms vs ~80 ms single stream):
7 wins by landing more tokens per step, 5 wins at conc 4 by ~8% because more
narrow verify steps pack better. Depth 7 stays — this fleet is interactive
agent loops where per-turn latency is what is felt — and the runbook now
records the numbers and the one-line switch for a batch-first workload.
Raising the prefill chunk to 16384 grows the arena enough that the KV cache
falls to 4.22 GiB against the 5.21 GiB a 114688-token context needs, so the
engine refuses to start. Measured single-card prefill at 8192 is ~2.5k tok/s
(~15k-token prompt, 5992 ms TTFT); the trade would mean cutting context for
unmeasured prefill gain, so the default stands.
The two replicas behind :7770 were on plain round-robin, which throws away
prefix-cache hits: an agent loop re-sends its system prompt and file context
every turn, and alternating replicas means every other turn pays a cold
prefill. Measured with 90k-token contexts, warm-turn TTFT was 2676 ms.
Pinning everything to one replica is not the answer either — a cold batch
(the shape a benchmark or a burst of parallel subagents produces) then loses a
card's worth of prefill: 16 requests took 30.1 s pinned against 18.6 s spread.
So the policy routes by request shape. A continuation (more than one message —
the agent-loop signature) is pinned to the replica that served its prefix;
anything else spreads across the cards by counter. Measured: 19.97 s for the
burst and 1095 ms warm-turn TTFT — the best of both, without a regression on
either workload.
The policy lives in proxy/routing.py, separate from the HTTP plumbing in
proxy/affinity-proxy.py, so it can be tested without aiohttp, a network or a
GPU. tests/frank_medium_proxy.bats covers it, and it earned its keep: the
first version hashed fresh prefixes to pick a card, which sends every member of
a same-prefix burst to one card (29.8 s, measured), and a stale index in the
pin path returned 500 on every continuation. Both are now assertions.
ROUTING=roundrobin and ROUTING=affinity stay available to re-measure the
choice on this host; the default is adaptive.
The radiance MXFP4 stack cannot run TP=2 on this host. RCCL's collective
kernels carry a hidden hostcall-buffer parameter, and ROCr refuses to
dispatch them on a device without PCIe atomics: the AQL dispatch fails with
hipErrorIllegalState ("the operation cannot be performed in the present
state"), which is what every two-rank collective hit. The second R9700 sits
behind the chipset bridges 05:00.0/06:09.0, which advertise
`AtomicOpsCap: Routing-` — a read-only capability, so no software can enable
atomic routing on that path and the driver's
pci_enable_atomic_ops_to_root() fails for that card. The first card, on CPU
root port 00:01.1, has atomics enabled.
Ruled out: kernel version (7.0.0-28/-30/-31 identical, after a full upgrade),
iommu=pt, P2P hardware (45.3 GB/s between the cards), the shipped librccl (a
from-source rebuild reproduces the failure at the same call path), vLLM and
pynccl (a bare two-rank torch.distributed allreduce fails the same way), all
NCCL transport knobs, and single-rank tests (NCCL short-circuits at
world_size=1, so one-card success proves nothing).
Deployment: three rootless Quadlets — frank-medium-a (:7771, card 1),
frank-medium-b (:7772, card 2), both TP=1 at 114688 context (262144 and
131072 do not fit one card's KV) with vLLM-profiled KV — and
frank-medium-proxy (:7770), a round-robin that preserves the fleet contract
of one endpoint and one served model name. frank-medium.container stays as
the TP=2 target for a host whose second card is atomic-capable; the deploy
handler does not install it and trashes any stale copy.
Measured through the proxy (vllm bench serve, random dataset, --ignore-eos):
conc 1 42.8 tok/s, conc 4 147.6 tok/s, conc 8 230.8 tok/s, 32k prefill at
~2.8k tok/s per card, TPOT p50 23.6-28.8 ms.
Also: healthcheck reads PORT so one probe serves all three units, the hermes
example and the harness models.yml drop to the real 114688 context, and the
runbook records the full evidence chain.
- Runbook with the measured envelope (prefill ~192 t/s at the VNNI ceiling;
decode ~3.8-4.3 t/s gated by 100 backend splits/token, GPU 95% idle), the
root-cause evidence chain, tuning levers, and co-tenancy traps
- AGENTS.md: fleet topology row + Active Services section for the ik
deployment; halogen demoted to rollback; bench policy unit type fixed
- LLM + root indexes relinked (frank-large-ik-uncensored ACTIVE,
halogen-flash-server retired)
New runbook page serverAppConfigs-docs/LLM/frank-medium-rocmfpx.md records
the hardware swap, the self-contained gfx1201 build, the measured layer-vs-
tensor decision (llama-bench multi-device syntax gotcha included), the
tensor-mode draft abort, served-path numbers, and the post-cleanup host ROCm
posture. AGENTS.md, README, docs indexes, Dankserver System Specs,
deployment/Hermes/OpenWebUI/Plex/Stack pages follow the rename;
freetoken-frank-small is marked RETIRED and kingjones-q8agent SUPERSEDED.
Work log entry added.
Predicted ~10.8 GiB KV / ~33.4 need / ~10.5 host-left; the boot ledger
printed ~11.0 / ~31.5 / 100.0 GiB held with 9.1 GiB + 1396 contiguous
2 MiB blocks left. Stall NOTE no longer prints at this footprint. Also
documents the ~45-60 s first-request PING silence after a restart (the
healthcheck's HealthStartPeriod covers it; do not restart the unit).
The 09-12 doubled pool fit only at reserve 15, leaving 7.1 GiB / 310
contiguous 2 MiB blocks for the ~28 co-tenant containers and making the
engine print its minutes-long-stall warning on every boot. 401,408
positions (~10.8 GiB KV) hands ~3.4 GiB back while keeping MAX_TOK=32768
and reserve 15: modelled need ~33.4 GiB vs budget 38.5, so the fit check
should not clamp. Capacity becomes one full 262k conversation resident
beside a ~139k one, or three concurrent 131k chats (three slots).
Docs: AGENTS.md fleet table + halogen section, runbook flags table /
co-tenancy tense / admission-arithmetic paragraph updated for the trim.
Answering "does the FreeToken runtime use sampling arguments?" -- yes, and
nothing in our config revealed it. serve-frank-small.sh passes no sampling
flags; the behaviour is the runtime default:
freetoken/server/args.py:31 sampling_defaults: str = "model"
which fills temperature/top_k/top_p from the checkpoint's
generation_config.json -- this one ships do_sample:true, temp 1.0, top_k 20,
top_p 0.95. Proved live: three identical requests with no sampling params
produced three different reasoning traces (487/540/577 chars).
The asymmetry with frank-large is the part that will bite someone: a silent
client samples on frank-small and decodes greedy on frank-large, so the same
request is reproducible on one endpoint and not the other. Documented with the
determinism options (explicit temperature:0, or --sampling-defaults none).
Also recorded two request-level traps found during the probe: thinking eats the
whole max_tokens budget and returns empty content with finish_reason length,
and the API accepts unknown fields with a 200, so a 200 proves nothing about
whether a key was understood.
Ledger confirms the change landed: "3 slot(s) over one 524288-position pool",
"slots: 1 -> 3 POOLED, 111.0 MiB each", held 103.5 -> 103.4 GiB.
Recorded the oversubscription nuance that the old row missed: the slots were
never the binding constraint on context. 4 x 262144 exceeds the 524288 pool,
so admission was pool-bound all along -- two full-length conversations, or
four at 131k. Lowering the cap trims simultaneous requests, not capacity.
Raised /swap.img from 8 GiB to 64 GiB on Dankclaw. Host-level state with no
repo file, so it was invisible to the repo until now: fstab line 12 drives the
generated swap.img.swap unit, ext4, fallocated hole-free, swappiness 60,
zswap off. 748 GiB still free on /.
The reusable part is the ordering. The old file held 4 GiB of live pages and
MemAvailable read 73 GiB while only ~5 GiB was genuinely free -- the pinned
model weights are counted as reclaimable page cache, so the headline lies by
the size of the model. Swapping off before the replacement is online would have
pushed those pages into RAM that does not exist. Creating and enabling the new
swap first let them migrate instead: swapoff took 65 s and the host never
wobbled.
New page serverAppConfigs-docs/Dankclaw/Swap Configuration.md, linked from
the index, plus a pointer in the AGENTS.md memory-model section so nobody
reasons about headroom without checking swapon.
Old 8 GiB file staged at dankclaw:~/.AGENTTRASH/20260912-swapfile-8g/
pending approval to purge.
Seven restarts measured. New section "The reserve cannot buy air": reserve 16
was deployed and held exactly what 15 holds — 103.5 GiB, 7.1 GiB left to the
host. The flag is a term in the fit-check inequality, not a reservation, so it
has only two regimes: below the pool's requirement it changes nothing
observable, above it the pool silently disappears.
Sizing rule now defers to the engine's printed requirement (budget >= 36.5 for
524288 at max_tok 32768, from "need ~35.0 GiB plus 1.5 of margin") over my
modelled 37.0, and the ledger table carries the reserve-16 row.
Ledger confirms the doubled pool is actually reserved (524288, 13312.0 MiB)
at max_tok 32768 with no clamp and no WARNING: 38.5 budget vs 37.0 needed.
Recorded the cost honestly. The server now holds 103.5 of 121.2 GiB and the
engine reports only 310 contiguous 2 MiB blocks / 7.1 GiB left for everything
else (24-reserve baseline: 2081 blocks / 12.1 GiB). All 28 co-tenants stayed
up, but the engine volunteers a new note: when memory runs out, frank-large
AND its neighbours "can stall for minutes at a time at full CPU with no disk
activity and no output, which is not a crash and needs no restart". That is
the least diagnosable of the three failure modes seen today, and `free` hides
it — MemAvailable still reads ~74 GiB because the pinned weights count as
reclaimable page cache.
Sizing rule now expresses the budget as a function of reserve, shows both
endpoints (24 -> 29.5, 15 -> 38.5), and the max_tok ceiling is labelled as
reserve-conditional (<= 18,130 at 24; <= 35,700 at 15).
Deployed HALOGEN_MAX_TOK=32768 by operator call (upstream default, full prefill
speed) and confirmed the prediction from the ledger:
kv pool: 524288 positions need ~35.0 GiB ... LOWERING THE POOL TO 262144 (~27.8 GiB)
memory: 68.0 weights + 7.2 KV pool + 21.1 working = 96.3 GiB in all
so the effective pool is 262144 again and the doubling this task set out to do
is not in force. Healthy, completion verified. Left the pool requesting 524288
deliberately: with KV_POOL_FIT on that is a grow-if-RAM-frees-up declaration,
not a mistake.
Two corrections to what I wrote earlier today:
1. The engine's sizing block, and its "WARNING even 262144 ... host RAM cannot
spare it" line, print ONLY when the fit check had to clamp something. The
09-11 baseline boot prints no kv pool: sizing lines at all. So the absence
of a warning in old logs was never evidence of headroom.
2. The budget needs a vision term. Actual need is pool + arena + 1.97 GiB of
vision tower, not pool + a 1.5 GiB nominal margin. That makes the ceiling
for a doubled pool max_tok <= 18,130 rather than 19,115, and it means the
32768/262144 corner this box has run in since 09-11 is ~0.3 GiB OVER its own
stated budget -- and has been the whole time. It boots and serves with
12-14 GiB genuinely free because the 24 GiB reserve is generous. A soft
floor, not a wall.
The arena model also reproduces upstream's three published "on the device"
figures exactly (35 / 42 / 41 GiB), confirming they are pool+arena totals at
the default max_tok and not pool-only -- which is the misreading that made the
doubling look affordable in my first estimate.
Raising HALOGEN_MAX_TOK to 65536 did not shrink the pool, it OOMed the boot:
the arena alone (37.3 GiB) exceeds the 29.5 GiB device budget, so even the
minimum pool faulted and dmalloc failed at 39.97 GiB. Rolled back to 16384,
which restored the 524288 pool and a clean completion.
The three measured arenas are exactly linear in the flag — 3.80 GiB base plus
8.40 GiB per 16384 tokens, predicting all three to 0.1 GiB. That gives a real
sizing rule instead of guesswork, and it says the deployed 16384 is the
CEILING, not a conservative choice: 29.5 - 1.5 margin - 14.4 pool leaves 13.6
for the arena, i.e. max_tok <= 19,115.
The corner check falls out of the same model: set max_tok back to 32768 and
only 7.4 GiB remains for the pool, ~269k positions, which is precisely the
262144 config this host ran before. The two corners of the feasible box are
the only two points it has ever served.
No functional change — the quadlet edit is comments only; the running service
is already at 16384 from the rollback deploy (258cccb).
The first draft of these notes estimated the doubled pool at 103.5 GiB
held / ~5 GiB host remainder. Both numbers were wrong — they costed the
KV alone and ignored that the arena shrinks when MAX_TOK drops. Measured:
pool 262144 / max_tok 32768 -> 96.3 GiB held, 12.1 GiB host left
pool 524288 / max_tok 16384 -> 94.6 GiB held, 13.8 GiB host left
The doubled pool is 1.7 GiB cheaper overall. Also records the failed
first attempt verbatim, since the silent-revert failure mode is the thing
a future operator most needs to recognise, and adds the HALOGEN_MAX_TOK
row to the deployed-flags table (the "baked 32768" line it overrode was
stale).
HALOGEN_KV_POOL_POSITIONS 262144 -> 524288, which is upstream's own
default (2 x HALOGEN_CTX). The old pool was the binding constraint on
concurrency, not context: a request reserves prompt + max_tokens against
the pool, so with pool == CTX four concurrent 65k chats were contending
for one full-length conversation's worth of positions. Two full-length
conversations, or four at 131k, now stay resident.
Cost is ~7.2 GiB more device RAM (measured 28 KiB/position), taken from
the page cache the 47.7 GiB n-gram lookup table reads through. Live
startup ledger for the prior config: 68.0 weights + 7.2 pool + 21.1
working = 96.3 GiB held, 12.1 GiB left on the host. The doubled pool
puts the remainder near 5 GiB against a ~4.3 GiB co-tenant working set,
and cache at ~69 GiB -- still above the lookup table, so the table can
stay hot. Compensating lever if long prompts start crawling:
HALOGEN_MAX_TOK=16384, which frees 8.3 GiB of prefill arena for ~8%
of prefill throughput.
Docs updated to match, including the admission-trap paragraph, which was
only true while pool == CTX.
Frank