The 320k pool (163840 per conversation over two slots) costs ~4.3 GiB of KV,
and the BF16 projector (888 MiB) lands on ROCm0. Fit at the deployed point,
computed, not estimated: ROCm0 29.19 / ROCm1 26.04 GiB of 31.86 -- 2.67 / 5.82
GiB free, both above the 1.5 GiB floor.
36:29 is one layer over the 35:30 balance on purpose: at this context the
balance point is 34:31 (0.11 GiB gap), and the operator moved a layer to ROCm0
to buy headroom on the drafter card. The gap is 3.15 GiB by construction, so
the seam's slack is that gap rounded up (3225 MiB) with the trade written down,
and the ctx cap moves 262144 -> 327680 with the rope caveat (the GGUF carries
no scaling; positions past the trained 262144 extrapolate on the base rope).
Vision: NO_MMPROJ=0, the serve script now refuses to start without the
projector when vision is on (a missing file is a deploy error, not a silent
text-only server), the download script fetches and byte-verifies it, and the
deploy handler's preflight checks for it. The projector is already on the host,
byte- and sha256-verified against the Hub.
6.44 / 6.90 is the state straight after load; after the service has served
traffic it sits at 6.23 / 6.64, with and without requests in flight. Both are
~5 GiB above the floor, but the runbook should not imply a load-dependent
difference that the two readings do not actually show.
New page [[frank-medium-batch-ubatch]]: the sweep's method (one invocation per
cell, a VRAM sampler per cell, 64k prefill measured AT 64k depth), the full
batch x ubatch matrix, the depth-0 control, the deployed-shape VRAM table, and
the limitations — including the one that cost me two runs, that amdgpu does not
return ~26 GiB at once, so a reload after a kill fails with "unable to allocate
ROCm0 buffer" and reads exactly like a configuration that does not fit.
AGENTS.md: the batch lesson now carries the measured curve (the peak at 512, the
10 % loss at 2048, the irrelevant logical batch, the invariant decode) instead
of the "~3.3 vs ~1.7 GiB" figure, which was measured at the bench's shape; at
the deployed shape the asymmetry is 2.41 against 1.30. Fleet row and the
frank-medium section now say 35:30 and batch 6144x512.
Runbook: deployed flags, the split history with the ubatch-512 table, the fault
note, and the verified VRAM (6.44 / 6.90 idle, 6.23 / 6.64 under two concurrent
requests).
Dual-R9700-Cooperation: the live capability table (card 0 ReqEn+ against card 1
ReqEn-, x1 effective link on card 1) and an honest label on the M.2_2 riser
premise — it is ASUS's spec sheet, not a measurement, and CPU-attached does not
imply Routing+ on this board (00:08.3 is CPU-side and Routing-).
log.md: the round trip recorded, newest first.
The batch/ubatch sweep (scripts/bench-batch-ubatch.sh, 23 cells of a 28-cell
grid, 3 reps each) says the ubatch knob has an OPTIMUM, not a plateau:
ubatch 256 384 512 768 1024 1280 1536 1792 2048
tok/s 1046 1068 1193 1135 1152 1151 1165 1148 1078
The curve is remarkably reproducible -- 512 came out 1192.88 / 1193.23 /
1193.28 / 1193.88 across batch 2048, 4096, 6144 and 8192, and 2048 came out
1077.38 / 1077.54 / 1077.59 -- so this is a deterministic property of the
shape, not noise. The logical batch is throughput-neutral (2048..8192 all
within 2 % at a fixed ubatch) and VRAM-neutral (the compute buffers are sized
by ubatch alone). Decode is 15.88 tok/s at depth 65536 on every single cell.
So the deployed 6144x1536 was both 2.4 % slower than the optimum and 3.71 GiB
more expensive. Measured at the deployed shape (262144, 2 slots, DFlash2):
36,29 ubatch 1536 -> 4.71 / 4.91 GiB free (what was deployed)
36,29 ubatch 512 -> 6.01 / 7.32 GiB free (ubatch alone)
35,30 ubatch 512 -> 6.44 / 6.90 GiB free <- deployed
The ubatch move freed 1.30 GiB on ROCm0 but 2.41 GiB on ROCm1, because ROCm1
holds the output layers and the drafter's graphs -- so it shifted the balance
toward ROCm1 by 1.31 GiB and the split had to follow it one layer. 0.46 GiB is
one DeltaNet layer's worth of asymmetry, i.e. the granularity floor: no split
can be tuned finer than a layer, which is why 34,31 is not better.
Re-fitted the seam with both ubatch coefficients measured at this shape and
re-anchored it on the new deployed point (it reproduces 6.44 / 6.90 to
0.01 GiB). Every neighbouring split still fails it: 36,29 -> 1.85, 34,31 ->
0.94, 37,28 -> 3.24, 1,1 -> 3.03, 39,26 -> 6.03 GiB.
Also replaced the unit test's literal `TENSOR_SPLIT=36,29` assertion with the
shape it actually cares about -- two shares that add up to the model's 65
blocks, with ROCm1 taking the fewer -- so re-measuring the balance no longer
breaks a test about parallelism. The numeric contract lives in the seam test.
pcie-link-test.cu (CUDA, pinned buffers, cudaEvent timing, both services stopped):
pinned H2D 6.72 GB/s and D2H 6.60 GB/s, i.e. 85 % of the Gen4 x4 ceiling. Async
is no better than sync and pageable sits within 1 % of pinned, so these copies are
link-limited rather than staging-limited — the usual "pageable costs 2x" does not
apply on this host. A 3 MiB slab costs 474 us, essentially 3 MiB / 6.72 GB/s plus
~27 us of launch.
That settles the expert-streaming question quantitatively: 1.70 GiB of active
experts per token (A3B @ Q4_K_M, 75 GiB expert mass) is 272 ms of link time ⇒
3.7 tok/s, 4.6 with a fifth of the experts resident. The same bytes computed in
host RAM at 256 GB/s cost 7.1 ms — the link is 31x slower than the memory behind
it, which is why llama.cpp's -ot exps=CPU path measured 20.46 tok/s while an
engine that streams loses by 3-5x. The port under the card is x8-capable and the
card is x16-capable, so the x4 is the adapter; doubling it still leaves streaming
well behind.
Also records FreeToken's own constraint: it keeps experts in host RAM with an LRU
GPU cache (exactly this streaming shape) and loads only modelopt block-FP8
safetensors for qwen3_5_moe, so Q4_K_M GGUF is unloadable there regardless.
The harness is in the repo now, so the runbook says how to run it and what the
tool will not tell you:
- / splits devices within one configuration and , splits sweep values, so a
legitimate --tensor-split 36,29 would have benched one card at a time.
- there is no -c: n_ctx = n_prompt + n_gen + n_depth, so no bench can decide
whether a configuration's buffers fit beside the deployed 262144 KV.
- a test that cannot allocate exit(1)s the process, hence one invocation per
cell.
- a shell's memlock limit (8 MiB here) is not the unit's (infinity), which
fails the load for the wrong reason.
- no --spec-* and no --parallel: decode figures are drafterless and slots are
never exercised; -d is what puts KV in front of the measurement at all.
Operator instruction: purge the cache so no run can serve a stale payload. The two
live payloads were cleared and then re-fetched rather than left absent — deleting
them without a re-fetch means neither endpoint starts.
halogen's re-fetch proved the point: its sidecars had been stale *in place*
(overlay.hgn 2,477,677,120 -> 2,572,466,560 bytes, overlay-speed.hgn
2,383,306,048 -> 2,478,095,488) while the 115.55 GiB checkpoint returned
byte-identical. frank-small's 27B and mmproj re-fetched to Hub-matching hashes,
unchanged.
Trap recorded: podman refuses a bind mount whose source directory is missing
(statfs ... no such file or directory), so frank-large needs /models to exist
*empty* for its first-start fetch to run.
Q4_K_S against the recorded Q4_K_M numbers: +2.1/+1.6/+2.6 % (iGPU-only pp512 /
pp16384 / tg64), +2.7 % tg32@64k, +4.1 %/+2.0 % for the two layer-split configs,
~0 on both pp512@64k, host RAM bottoming at 10.0 GiB against ~5 GiB. A memory
lever with a small speed dividend. IQ4_XS (97,473,155,360 bytes, sha256
b4787265...d0574f9d) is pulled and verified for the next comparison.
Also repairs an AGENTS.md edit that landed in the wrong section: the peer
session's commits shifted that file between my read and my write, so the
frank-large text briefly split the Model Config section.
`local b=$1 ub=$2 dir="$OUT/b${b}_ub${ub}"` expands every argument of the builtin
before the builtin assigns any of them, so $b was unbound and set -u aborted the run
four seconds in - after the service was already stopped. Two statements instead.
Found by running it under systemd-run: it exited non-zero with no cells run, which is
exactly what the per-cell TSV is for.
Question it answers: within the headroom the deployment has, how large may the logical
batch and the micro-batch be before the activation buffers stop fitting — and what does
each size buy in prefill?
- One llama-bench invocation per cell. llama-bench exit(1)s the moment a test cannot run,
so a single infeasible ubatch would otherwise take every later cell with it, and
infeasibility is one of the things being looked for.
- A VRAM sampler per cell, covering the load. Batch/ubatch sizes the activation buffers,
so a cell that looks fast because its allocation is smaller than intended has to be
distinguishable from one that is genuinely fast.
- The service is stopped for the sweep and started again from a trap, because a bench
beside a live service starves the same GPU.
- Run it from a transient unit: the deploy's unit sets LimitMEMLOCK=infinity for the ROCm
staging buffers and a shell on these hosts only gets 8 MiB.
- llama-bench separates -ts/-dev with SLASHES and has no -c, --parallel or --spec-*: the
context is n_prompt + n_gen + n_depth, and the drafter cannot take part.
Checking whether a download was still running with pgrep -a curl printed the
whole argv, Authorization header included, into the session transcript. The
token is not in the repo and nothing was committed, but the practice is wrong:
secrets belong in a -K config file or the environment, not in argv, and no
pgrep -a should ever target a command carrying an auth header.
The note described DSpark as the deployed drafter and 40:25 as the split; both
were true for a few hours. It now leads with DFlash2 at 36:29 and keeps the
excursion as the record it is:
- the drafter comparison table with DSpark's REAL size (1,455,376,576, not the
1,359,284,737 that went into the fetch script), and why it lost: no sliding
window, so its draft KV spanned the whole context.
- the split's two moves as one story: the batch pushed it to 40:25, the drafter
swap pulled it back to 36:29, with both measurement tables.
- the layer-size law (0.43 GiB DeltaNet, 1.43 GiB attention) as the reason the
balance is measured and not interpolated.
- the verified-behaviour section now carries the LIVE numbers: 4.71 / 4.91 GiB
free (4.43 / 4.61 under load, 0.18 GiB gap), the DFlash2 init lines, and
draft acceptance from both slots (0.63 / 0.48 on slot 1, 0.39 on slot 0).
- the fault note scopes its own evidence: 39:26 was loaded, not exercised.
DSpark came off after a day. It was the better-looking file (1.27 GiB against
1.92) but it declares no sliding window, and the draft context is built with
n_ctx = llama_n_ctx(ctx_tgt) -- so its KV spanned the whole 262144-token window
(5.0 GiB) and cost ROCm1 four layers of split to carry. DFlash2's window bounds
its draft KV at 2048 positions (~40 MiB), which frees ~4.3 GiB on the
draft-carrying card.
The balance point moved with it, so it was re-measured at the deployed batch
(6144 x 1536, 2 slots, 262144) rather than interpolated:
39,26 2.42 / 7.20 37,28 4.28 / 5.34 35,30 5.14 / 4.48
38,27 3.85 / 5.77 36,29 4.71 / 4.91 34,31 6.57 / 3.05
36:29 is the minimum (0.20 GiB apart) and leaves 3.2 GiB above the operator's
1.5 GiB floor on each card, so the full window stays. One -ts step is one layer,
and qwen35's layers are not the same size -- 0.43 GiB on a DeltaNet boundary,
1.43 GiB on a full-attention one -- which is why the table exists.
Also in here:
- the fetch script pointed at the wrong file (DSpark) AND declared the wrong
size for it: 1359284737 against the Hub's 1455376576, 96 MB short, which made
--verify-only fail on a payload that was in fact correct. Repointed at z-lab
DFlash2 Q8_0, 2056414816, and both files' sha256 were checked against the
Hub's LFS oids (host == Hub for both).
- the seam's fit was recalibrated for a window-bounded draft KV and for the new
point; it now passes 36:29 and fails every other split measured, including
1:1. Its ubatch term is an extrapolation and it cannot model contiguity, which
the comment says out loud.
- the serve script's dead dspark-only tunables (n-min, p-min, p-split, poll)
are gone; the build's defaults apply.
The note still described DFlash2, one slot at 262144 and a 1:1 split -- all of
which the day's work replaced. It now carries:
- the DSpark swap with the table that matters (1.27 vs 1.92 GiB, block_size 7 vs
8 so n_max 6 vs 7, a Markov head instead of the candidate selector, and NO
sliding window) and the rule behind the sizing: the draft context is built
with n_ctx = llama_n_ctx(ctx_tgt), so a drafter without a window caches the
whole target context -- 5.0 GiB here, against DFlash2's ~40 MiB.
- the failure that rule produces when ignored: load succeeds, /health answers,
and the first draft decode dies with an OOM at 1:1.
- why the split moved a second time: tripling the batch costs ROCm1 about twice
what it costs ROCm0, with the 38,27 -> 41,24 -> 40,25 measurement table.
- the deployed flags, the 2-slot shape (n_slots = 2, n_ctx_slot = 131072), the
DSpark init lines, and balanced 29.98 / 29.98 GiB cards.
- honest scoping on the older measurements: the throughput, drafter-worth and
ngram tables were taken on DFlash2, and the agent-loop numbers on one slot.
- the 39:26 page-fault signature as a thing to watch, marked as NOT an OOM.
The operator's revised spec: batch x3 (6144/1536), parallel 2, a 1.5 GiB per-card
floor, and the context maximised inside it.
Measured at 262144 with the old 38:27 split, 3x batch was brutally asymmetric:
ROCm1 (output layers + the drafter's graphs) paid ~3.3 GiB where ROCm0 paid
~1.7, leaving the draft-carrying card at 0.08 GiB free against ROCm0's 3.78.
One -ts step is worth ~1.5 GiB at this batch, so the balance is a measurement:
41:24 overshot (1.47 / 2.38), 40:25 landed at 1.92 / 1.96 GiB free.
That balance is also what maximises the context -- an unbalanced split would have
to buy its headroom out of CTX_SIZE, and CTX_SIZE is already at the model's
trained window (262144), now split over two slots of 131072. Verified under two
concurrent 1832-token prompts: 200/200, 1.87 / 1.86 GiB free, n_slots = 2,
n_ctx_slot = 131072, tool calls intact, and DSpark drafting on BOTH slots
(acceptance 0.38/0.22 and 0.52/0.43).
8192/2048 does not fit at this context at all: its 2020 MiB pp buffer fails to
allocate on ROCm1, and 4096/1024 fits but leaves 1.11 GiB -- both recorded in
the env's sizing table.
The unit seam now models both cards (share + draft + ubatch-scaled buffers,
calibrated to the rocm-smi figures to within 0.002%), asserts the 1.5 GiB floor
on each, and asserts the two cards stay within 0.5 GiB of each other. Both
assertions were confirmed to FAIL on a deliberately bad config (ubatch 8192, and
a 1:1 split) rather than pass by construction.
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 38:27 skew folded into the DSpark swap already does what step 3 asks: the
draft-carrying card sits 2.08 GiB above the other (5.49 / 3.41 GiB free) instead
of the ~9 GiB a 1:1 split would leave.
Chasing an exact balance is a mistake worth recording: 39:26 measured tighter
(3.86 / 4.69 GiB free) but that server died on its eighth request with SDMA0
UTCL2 page faults in the amdgpu driver -- "ROCm error: an illegal memory access
was encountered" -> ggml_abort, in the cross-card boundary copy that is the
layer split's hot path. 38:27 has since served the whole e2e gate and 12 varied
requests with no fault, so the residual 2.08 GiB gap is the safe side of this
trade. Comment-only: no restart needed.
Step 2 of the operator's six. --ctx-checkpoints 8 held per-slot KV snapshots;
0 is what llama.cpp ships, and the fleet's clients open conversations fresh.
DSpark declares no sliding window, so its draft context -- built with
n_ctx = llama_n_ctx(ctx_tgt) -- caches the whole target window: 5 layers x 2 x 8
kv heads x 128 dim x 2 B = 20 KiB/position, i.e. 5.0 GiB at 262144, where
DFlash2's sliding_window 2048 cost ~40 MiB. At a 1:1 split that put ROCm1 at
33.9 GB of 34.2 GB and the first draft graph capture died inside
common_speculative_impl_draft_dflash::process with "ROCm error: out of memory"
-- after the model had loaded and /health had answered, which is why the deploy
looked like a crash-loop rather than a sizing failure.
Handing three target layers to ROCm0 restores headroom on both cards at the full
window. Measured at 262144: 28.52 GB / 30.71 GB, i.e. 5.30 GiB and 3.26 GiB free
(was 2.7 GB free on the draft-carrying card with 1:1 + DFlash2). The unit seam
now models the busier card as a skewed share plus the drafter's full-context KV,
calibrated to the rocm-smi figure.
AGENTS.md gains the durable lesson: read sliding_window out of the drafter's
GGUF header before sizing a spec-decode deployment.
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.
DSpark (Q8_0, 1.27 GiB) replaces z-lab DFlash2 (1.92 GiB): same 5-layer
block-diffusion shape, Markov head instead of the candidate selector, and
block_size 7, so the draft depth drops 7 -> 6 (llama.cpp asserts
n_max <= block_size - 1).
The serve script's dspark branch never passed --spec-draft-model (only the
dflash branch did), so dspark was not deployable as written; the model file
gate now covers both types. Download script, env and the unit seam track the
new file and its exact byte count.
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.
hermes-webui is Network=host, so 8787 is a bare host-netns bind, and netavark
DNATs only *published* ports. From Caddy's bridge netns 192.168.0.143:8787 and
10.89.1.1:8787 are both refused, while host.containers.internal:8787
(169.254.1.2) answers 200 — measured live 2026-09-16 from inside the Caddy
container. The dashboard's published 9119 answers on that same LAN IP, which is
what made the old upstream look sound.
The vhost comment asserted "192.168.0.143:8787 is open" from that bridge; it is
not, so the claim is replaced with the measurement, and the hermes-webui quadlet
comment carries the same correction. The seam guarding this invariant accepted
any numeric upstream, so the LAN IP passed it; it now pins the measured-working
form and fails on the LAN IP (teeth verified by substitution).
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.
The fleet now runs draft-dflash,ngram-mod, so the rule needs to be written down
where the next agent will look: draftless types (ngram family) are added to the
spec configuration unconditionally, while draft-mtp/dflash/dspark/eagle3 all
require a draft CONTEXT -- therefore one context-backed drafter plus any number
of ngram types is legal, and two context-backed ones never are ('draft-mtp,draft-dflash'
dies at load with 'model doesn't contain MTP layers'). Plus the operational
trap: ngram-mod prints no startup banner, so a working chain looks unengaged in
the journal.
The port gate was written for the cutover, when the vLLM proxy held 7770. Once
frank-medium itself owns the port, a re-deploy waits 90s for a socket that only
its own predecessor is holding and fails: 'FAIL: port 7770 is still held after
90s', with the old llama-server still serving scrapes. Stop the unit first, then
sweep a stray manual llama-server (the shape a benchmark harness leaves behind);
pkill -x, never -f, so the deploying shell cannot match itself.
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.
Two real defects from patching a rewritten file by line number: the CUT that
removed a duplicate parse_configs also took write_env with it (NameError after
the first configuration, with the unit already stopped), and the restore step
lived after the loop, so a crash left frank-medium DOWN. The endpoint is the
reason the harness exists; it now restores itself in a finally and shouts if the
service does not come back healthy.
ngram-mod's match window (n_min/n_max) decides which traffic it can see at all
-- the fleet's 48/64 preset is tuned for long verbatim echoes and finds nothing
in a repetitive-but-short output. Config entries now take |KEY=VALUE overrides,
so the window can be swept in the same run as the configurations it applies to.
The first run asked for a verbatim repeat with thinking ON, so the model copied
the passage inside reasoning_content, content came back empty for all three
configurations, and the cross-configuration identity check compared '' to '' --
a pass that proves nothing. echo and structured now disable thinking via
chat_template_kwargs, the echo row carries a copy_ratio (share of answer lines
that really are passage lines), and an empty answer in any shape is a loud
warning rather than a silent pass.
'structured' is new and answers the case the fleet actually runs on: repetitive
OUTPUT with no prompt overlap (record lists, tables, config rewrites), which is
where an n-gram matcher should pay and an echo-shaped test cannot tell.
A relative serve-script path plus a cd into its own directory double-counts the
prefix, so every configuration failed to start ('scripts/serve-frank-medium.sh:
No such file or directory'). The safety net did its job -- the unit was restored
-- but the run measured nothing.
Unlike MTP, ngram-mod needs no draft context: common/speculative.cpp adds the
ngram types unconditionally while the model-backed drafters are gated on
ctx_dft != nullptr, so 'draft-dflash,ngram-mod' is structurally allowed (and the
config vector's fixed order -- ngram first, then the drafters -- makes CLI order
irrelevant).
The harness grows the two things this question needs: --configs, and an ECHO
shape. ngram-mod can only ever match text that is present in the prompt, so
measuring it on prose calls it useless and measuring DFlash on echo calls it
wasteful; both shapes now run, and the greedy text of each is compared across
configurations because speculation may not change answers.
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.
'cp -r SRC DST/' with a missing DST creates DST as a copy of SRC and drops a
directory level, so rocblas/library/* landed as lib/rocblas/* and hipblaslt's
arch subdirectories landed one level high. Nothing complains at load time: the
engine reads 29 GiB of weights, answers /health, and dies on the first request
with
rocBLAS error: Cannot read .../lib/rocblas/library/TensileLibrary.dat:
No such file or directory for GPU arch : gfx1201
Declare the destination and copy contents, both trees.
The vllm-radiance replica pair could not use both cards for one model: TP=2
needs a cross-card collective and the second R9700 cannot dispatch one. The
same wall blocks vLLM's pipeline parallelism, whose stage handoff is an RCCL
send/recv of the hidden state (distributed/parallel_state.py) -- a fact worth
writing down, because the obvious reading of the ask ('use PP to dodge the
atomics problem') is exactly backwards.
llama.cpp's layer split needs no collective at all: layers are distributed and
the boundary activation is a per-backend copy, once per forward pass. One
process, both cards, one model. The runtime is mainline llama.cpp at fb27a525
(the ref is pinned by DFlash2 AND qwen35 MTP; the ROCmFPX fork this host had
before has neither), built for gfx1201 against the private ROCm 7.14 toolchain
the earlier build already used, and bundled next to the binary.
Server: ajgazin Swift-Qwen3.8-27B-Uncensored UD-Q8_K_XL with the model's MTP
head in the file, plus the z-lab DFlash2 Q8_0 block-diffusion drafter at depth
7 (its block is 8, so 7 is the ceiling llama.cpp allows). Text only: vision
stays on frank-small.
Retired into the trash rather than deleted: the three vLLM quadlets, the
radiance entrypoint/healthcheck, the routing proxy and its tests, and the two
bench scripts that drove 'vllm bench serve' inside the retired image. The
deploy handler now stops them, waits for 7770 to be released, installs the
ROOT unit and refuses to start anything when the runtime or payload is
missing. The planner maps the scripts/ trio to frank-medium only, and the
scoping seam asserts the retired service dir is gone rather than merely
unmapped.
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.
The runtime for the next frank-medium shape is mainline llama.cpp at
fb27a525, built for gfx1201 against the ROCm 7.14 nightly the ROCmFPX build
already used on this host, with the runtime bundled next to the binary.
Two things that pin the ref: DFlash2 (ggml-org/llama.cpp#27342) and qwen35
MTP. Neither is in the ROCmFPX fork's tip (c49ebdb, 2026-08-22), and the
DFlash2 drafter cannot run without the first.
The fetch script pulls the target at UD-Q8_K_XL plus the DFlash2 draft at
Q8_0, resumable and byte-verified -- a truncated 31 GB GGUF is otherwise a
load-time crash.
A redeploy of a running frank-large holds 7770 itself, so waiting on the port
before stopping the unit would have made every subsequent deploy fail. Stop the
user unit first (a no-op on a first cutover), then wait, then install and start.
The 2026-09-16 halogen restore failed exactly here. The ik root unit stopped,
systemd logged "Deactivated successfully", pasta tried to bind 7770 about a
second later, reported "Address already in use", and the quadlet exited 126.
Only Restart=on-failure recovered it -- minutes later, which is not a deploy.
The handler now waits for every socket on the port to clear (listener, accepted
and TIME_WAIT) before installing and starting the quadlet, and fails loudly with
the socket table attached if it never does. The retirement also reports when the
unit file was already gone, because this host's `disable` deletes it.
frank-large has no /metrics to scrape any more, and scraping it would be red by
construction: the halogen engine publishes /health, /cache and /v1/* only. That
job is removed with the reason written down (the remedy is the blackbox
exporter this fleet already wants for the comfyui/services probes), and
frank-small's llama.cpp endpoint takes the slot at 192.168.0.143:7771.
Homepage: frank-large's card still described the ik runtime, and the new
endpoint had no card at all. Both fixed, with a loopback probe -- the checker
runs on Dankclaw, which is also what makes the card green-able.
Dashboard: the endpoint picker was sourced from a vLLM-only metric, so a
llama.cpp job could never be selected, and the prose named frank-large as the
llama.cpp endpoint. The picker now matches both runtimes' series and the
llama.cpp targets carry the job filter every other panel already used.
The seams assert the new contract including the ABSENCE of a 7770 target, so a
well-meaning edit cannot re-add a permanently-red probe.
Dankclaw's endpoints are re-split. frank-large is the halogen-flash-server
quadlet again -- the 2026-09-13 unit file was never touched by the two-day
ik_llama.cpp interlude, so the runtime was still in the repo and its image is
still in local storage -- retuned to the shape asked for: HALOGEN_CTX 200192
with a 400,384-position pool and two slots, exactly two 200k conversations.
frank-small is new: mainline llama.cpp on the RTX 3090 serving
ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF at IQ3_S, using the -mtp build (it carries
blk.64.nextn.*, which is what lets --spec-type draft-mtp run on the main
model's own weights instead of a second 12 GiB copy) plus the BF16 vision
projector, as one 131,072-token slot with f16 KV on port 7771.
The host is shared now, and that is what the halogen retune pays for:
HALOGEN_MAX_TOK 24576 (arena 20.6 -> 16.4 GiB, ~4-5% prefill) and the vision
tower dropped (1.97 GiB) hand ~6 GiB back to the host, and the reserve rises to
20 to keep the fit check honest. Vision moved wholesale to frank-small.
Retired to scripts/.AGENTTRASH/: the ik trio (env, service, serve script), the
Qwen-Sharp template only that runtime read, and four benchmark harnesses bound
to retired runtimes. Because retiring a file makes any test that greps it pass
by absence, the seams assert the archive exists AND the live path does not.
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.
I tightened HERMES_WEBUI_HOST to 127.0.0.1 on the reasoning that the compose
service had published loopback and the unit had widened it. That reasoning
missed the consumer: this unit is host-networked (it shares the CDP daemon's
loopback), so Caddy cannot reach it by container name and fronts the guarded
vhost hermes.franksdank.cloud with `reverse_proxy 192.168.0.143:8787` — recorded
in dankclaw/Caddyfile with the same explanation. With the app on loopback that
upstream is dead, so the vhost broke.
The wildcard bind is what serves BOTH consumers: Caddy on the LAN address and
the phone's SSH tunnel on the host's localhost. Controls are the auth chain in
front of the vhost plus the webui password. The unit comment now states this
instead of the incorrect 'as before' claim, and the near-miss is worth the
paragraph: the e2e gate stayed green because no spec covers that vhost.
Two things, found while checking whether the hermes trio still needed migrating.
1. **The phone-facing webui was exposed to the LAN.** Its unit runs with
Network=host, so there is no Podman port forwarder and the app's own bind is
the exposure. The compose service it replaced published
`127.0.0.1:8787:8787` (loopback); the unit bound `HERMES_WEBUI_HOST=0.0.0.0`,
and the LAN address answered 200 — password-gated, but on every interface,
which is not the documented "SSH tunnel is the sole ingress" posture. It now
binds 127.0.0.1, which serves the tunnel just as well (sshd dials
localhost:8787 from the host). The unit comment claiming "LAN-facing as
before" was wrong about "before", and is corrected.
2. **Dankclaw's compose.yaml is retired.** Auditing ownership showed every
running container on that host reports a PODMAN_SYSTEMD_UNIT label — the
hermes trio's units have owned their containers since 2026-09-12 — so the
three service blocks left in compose.yaml were dead weight that could only
create a duplicate container fighting a unit for a name. deploy_hermes was
already fully unit-based. The file is staged in .AGENTTRASH/; the planner
plans nothing for it, an unknown service on that target now errors instead of
silently no-oping through `podman-compose`, and DANKCLAW_COMPOSE_SERVICES is
empty.
The hermes-webui seam now reads the unit — it read compose, which is exactly how
the bind drifted while the test kept passing against a file nobody ran.
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.
ComfyUI had failed 9,146 times and had been down ~7 hours when this was found:
its container was created 2026-09-15 14:12, BEFORE the GPU declarations were
added to the compose service, and `restart: unless-stopped` kept restarting
that stale container — no /dev/kfd, no /dev/dri, no keep-groups — so it died
every time on `RuntimeError: No HIP GPUs are available`. Its "unit" was a
hand-written ~/.config/systemd/user/comfyui.service that shelled out to
`podman compose up -d`, which is not repo-tracked state and collides by name
with the Quadlet the generator produces.
- dankclaw/comfyui/quadlet/comfyui.container: image comfyui-rocm:gfx1151
(built from Containerfile.comfyui), /dev/kfd + /dev/dri, keep-groups,
the four comfyui-data volumes, loopback-only publish (an unguarded service
is internal-only here; the homepage card probes 127.0.0.1:8188), Restart=
always, WantedBy=default.target.
- the compose service is deleted — one owner per container name.
- deploy_comfyui builds the image, installs the unit, retires the compose-era
unit into ~/.AGENTTRASH, removes the stale container so the unit's own can
take the name, and waits on :8188.
- the healthcheck is a file, because the podman 5.x v1 generator splits
HealthCmd on spaces and an inline `python3 -c "..."` becomes a quoting
puzzle that fails silently.
The comfyui seam now reads the unit instead of compose for devices, command and
groups (and resolves the Containerfile from the deployer's declaration, not
from a directory listing).
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.
Dankserver's node-exporter has been dead since a reboot: `podman ps -a` showed
it `Exited (143)` and Prometheus' `node-dankserver` job has been scraping
nothing. It ran only from `dankserver/compose.yaml`, which is the failure mode
this repository's boot-persistence rule exists for (rootless Podman does not
restart bare containers after the stop that shutdown performs), and the compose
form bound the wildcard address — the unauthenticated-metrics-API-on-every-
interface posture already removed on Dankclaw.
- dankserver/observability/quadlet/node-exporter.container — mirrors the
Dankclaw unit, listening on exactly 127.0.0.1:9100 (host-local probes) and
192.168.0.235:9100 (Prometheus on Dankclaw), with Restart=always and
WantedBy=default.target.
- the compose service is deleted: one owner per container name.
- deploy_observability now installs from "$TARGET/observability/quadlet"; it
hardcoded the Dankclaw path, so a Dankserver unit could not be installed by
the shared dispatcher at all.
- planner: "node-exporter" joins the Dankserver service set, with a path rule
for dankserver/observability/**; the dispatcher's allowlist gains
dankserver:node-exporter.
Seams cover the silent halves: both listen addresses present and the wildcard
absent, the compose service gone, the allowlist entry present, and the planner
mapping the path to the service.
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.
dankclaw/comfyui/Containerfile is referenced by nothing — compose builds from
dankclaw/Containerfile.comfyui, deploy-remote.sh copies that one, and the
ComfyUI doc names it — and its content is an older iteration (torch
2.9.1+rocm7.12.0rc1 against the deployed gfx1151 base). It also caused real
damage: the comfyui test's directory-listing selector picked non-comfyui files
and reported a red suite for sessions.
Staged in .AGENTTRASH/ rather than deleted, so it stays recoverable and visible
until Frank decides it can go for good. This is also why the earlier test fix
had to resolve the Containerfile from compose rather than from the filesystem:
the directory is not the contract.
This test failed on a clean tree for many sessions, and the working belief
became "test 73 is a known failure, ignore it" — which is exactly what a
permanently red suite teaches. The cause was a selector bug: `find
dankclaw -name 'Containerfile*' | head -1` returned whichever Containerfile the
directory listing happened to put first (dankclaw/alloy/Containerfile), so the
assertion read a file that never claimed to be ComfyUI's. It would equally have
passed on a wrong file.
The test now resolves the Containerfile the way the build does — build.context
joined with build.dockerfile, both read from dankclaw/compose.yaml — so it
checks the file the deploy actually builds (dankclaw/Containerfile.comfyui, the
one compose, deploy-remote.sh and the ComfyUI doc all name) and follows a future
rename instead of guessing.
Suite: ALL TESTS PASSED (310), no exceptions to explain away.
The card probing 192.168.0.235:7770 — frank-medium's endpoint since the
radiance cutover — was still called "frank-small", a service retired on
2026-09-15. The probe target was right; only the identity was wrong, which is
the worse half: a green card named after a service that no longer exists teaches
the wrong thing about the fleet. It is now frank-medium, described as what it
actually is (Qwen3.8-27B Uncensored MXFP4, vllm-radiance, two R9700s behind one
endpoint), and frank-large's description stops crediting the retired halogen
runtime for work ik_llama.cpp does.
The seam added to homepage_config.bats pins the cross-file contract that let
this drift: a card's probe target and its name must agree with the fleet, and a
retired service must not keep a card.
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 scraped metrics have no rules over them, so every graph was read after the
fact by a human. This adds shared/prometheus/rules/llm-fleet.yml, loaded via
rule_files and installed by the same copier that ships prometheus.yml (a rule
file nothing loads evaluates nothing, and Prometheus reports no error for it):
LlmKvPoolSaturated kv pool > 90% for 10m — the 122,636-token wall
LlmRequestsQueueing waiting > 0 for 10m — MAXSEQS=4 reached
LlmPrefixCacheCollapse reuse < 20% for 15m with traffic — the thrash signature
LlmEndpointDown scrape target down 2m — both replicas gone
LlmLatencyHigh p95 end-to-end > 60s for 10m — what an operator feels
job:llm_prefix_cache_reuse:rate5m — recording rule shared by the alert and the
dashboard, so the two cannot drift
Every threshold is tied to something measured on this fleet, not to a guess,
and the file says plainly at the top that **nothing is notified**: there is no
Alertmanager here, so a firing alert reaches Prometheus' /alerts page, the
ALERTS metric and the dashboard panel added alongside it — and nowhere else.
Choosing how Frank actually gets paged is his call, not a missing config line.
Seam covers the silent failure mode: rules exist, the config references the
directory, the deployer installs it, and the dashboard graphs ALERTS.
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.
Grafana has been running with zero datasources and zero dashboards, so the
metrics scraped last commit were only reachable through the Prometheus API.
This wires the plumbing and one dashboard, both declared in
shared/grafana/provisioning/ and installed to /opt/grafana/provisioning by the
deployer (the unit already mounted that directory read-only, so nothing about
the container changes):
- datasources/prometheus.yml — Prometheus at http://prometheus:9090, the
container name, not the loopback publish, with a fixed uid so dashboards
reference it without a lookup.
- dashboards/llm-fleet.json — KV pool fill, queue depth, prefix-cache reuse
rate, e2e latency p95, and decode/prefill throughput, with an endpoint
variable covering both frank-medium and frank-large. Every series name was
checked against the live /metrics endpoint: this fleet already shipped one
comment naming a metric that does not exist, and a dashboard full of those
would be worse.
Dashboards stay editable in the UI (allowUiUpdates) because the panels are a
starting point, but the file is the source of truth and the deployer
overwrites Grafana's copy on every deploy.
Planner and seams follow the existing pattern: shared/grafana/** selects the
grafana service, and the tests pin the container-name address plus the
provisioning contract so a rename cannot silently empty the dashboard.
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.
vllm:prefix_cache_hit_rate does not exist in this build; the exported pair is
vllm:prefix_cache_hits_total / vllm:prefix_cache_queries_total, whose rates give
the hit rate. A comment naming a series nobody can query is worse than no
comment.
Neither LLM endpoint was scraped, and frank-medium exports 418 vllm: series —
including the ones this fleet argues about under load: kv_cache_usage_perc (how
full the per-card pool is), prefix_cache_hit_rate (whether pinned conversations
are being reused), num_requests_waiting (queue depth) and the request-latency
histograms. Everything diagnosed tonight by hand — 122,636 KV tokens per card,
thrash when three large contexts overlap, cold-vs-warm prefix reuse — is a
graph now, on the LAN addresses rather than the guarded vhosts.
frank-large comes along in the same change: llama.cpp exposes throughput
counters under llamacpp:, and watching one LLM endpoint without the other was
arbitrary.
The seam added to observability_quadlet.bats pins the cross-file contract that
breaks silently: if a scrape target drifts from the :7770 the units publish,
the graphs vanish with no error anywhere.
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.
The proxy trusted its health table, which the watcher refreshes every 10 s. A
replica that restarts — and a deploy restarts both — therefore produced a
window in which requests routed to the dead card got a 502 back, even though
the other card was serving the whole time.
A failed connection now demotes that replica immediately and retries the
request on another live one, before anything is written to the client.
vLLM's OpenAI endpoints are stateless, so re-sending a request that never
reached an engine is safe; a failure after streaming has begun is still not
retried. The failover is logged, because a silent one hides a card that is
down.
Measured with card 1 stopped mid-flight: six requests, six 200s, one failover
line in the log, no client-visible error. Before this commit that burst
returned 502s.
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 single-GPU replicas died with torch.OutOfMemoryError during KV
allocation (2.16 GiB requested, 116 MiB free, 29.34 GiB already held by
PyTorch) because the entrypoint blanked KV_MEM=0 into an omitted flag. The
image's patched KV sizing reads an explicit 0 as 'profile this shape'; with
the flag absent vLLM sizes the cache from --gpu-memory-utilization alone and
overshoots. Passing 0 literally is the shape the bench numbers were measured
on, and MAXSEQS gets its default back under set -u.
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)
Rootless Quadlet on dankserver serving
just1moremodel/Qwen3.8-27B-Uncensored-MXFP4-awq (the radiance stack's
designated alternate target; ships its fp8 MTP head, no rewrite step) with
the DFlash2-FP8 drafter at SPEC=7 on the 2x R9700s (TP=2, host network,
:7770, alias frank-medium unchanged).
- dankserver/frank-medium/: quadlet unit, container entrypoint (fork of
serve-mxfp4.sh's patch block at pinned clone commit 97353293), healthcheck
- dispatcher retires the root llama.cpp unit, re-tags the registry image to
the localhost alias, installs the quadlet, waits on /health
- planner maps dankserver/frank-medium/* to frank-medium; scripts-era
mappings removed
- ROCmFPX-era env/serve/unit/download scripts staged in scripts/.AGENTTRASH
with .rocmfpx suffix; seams updated (frank_medium_quadlet.bats new)
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.
The RTX 3090 left Dankserver; two AMD Radeon AI PRO R9700 cards (Navi 48,
gfx1201, 32,624 MiB each) replaced it. FreeToken/CUDA cannot run there at all,
so the `frank-small` service is retired and `frank-medium` takes the endpoint.
New service: charlie12345/ROCmFPX llama.cpp (built for gfx1201 against the
fork's self-contained ROCm 7.14 toolchain, so the host's stale /opt/rocm-6.3.3
never participates) serving kingjones777/Qwen3.8-27B-ROCmFPX-Q8_0-AGENT-GGUF --
a Q8_0_ROCMFPX_AGENT (ftype 115) checkpoint that stock llama.cpp cannot load.
Placeholder config: --split-mode tensor across both cards, 131072 context,
1 slot, f16 KV, MTP draft head on ROCm0.
deploy-remote.sh grows retire_root_llm_small(): the CUDA-era unit is stopped,
disabled and moved to ~/.AGENTTRASH/20260915-frank-small-retire/host-units
before frank-medium starts, because `systemctl disable` has been observed to
delete the unit file itself on this host (2026-08-26 incident).
The retired frank-small repo files (env, unit, serve script) move to
scripts/.AGENTTRASH/, and the planner, dispatcher, service allow-list and the
deployment test seams now name frank-medium.
Follow-up (same session, measured): llama-bench on the host with the correct
multi-device syntax (-dev ROCm0/ROCm1) overturned the tensor choice:
single R9700 pp4096 1160.9 / tg128 19.42 (d32768: 512 / 18.4)
layer 2x pp4096 1991.2 / tg128 18.76 (d32768: 859 / 17.6)
tensor 2x pp4096 271.8 / tg128 26.26 (d32768: 229 / 25.1)
This build has no RCCL, so tensor mode falls back to a "meta-backend
butterfly" allreduce and prefill collapses 7x. Served path with MTP:
layer 35.7-46.2 tok/s vs tensor 28.7. SPLIT_MODE=layer, draft pinned to
ROCm1 (the card holding the last target layers). Served prefix cache with
the draft head verified live: turn 2 reprocessed 22 tokens instead of 949.
The RTX 3090 left Dankserver; two AMD Radeon AI PRO R9700 cards (Navi 48,
gfx1201, 32,624 MiB each) replaced it. FreeToken/CUDA cannot run there at all,
so the `frank-small` service is retired and `frank-medium` takes the endpoint.
New service: charlie12345/ROCmFPX llama.cpp (built for gfx1201 against the
fork's self-contained ROCm 7.14 toolchain, so the host's stale /opt/rocm-6.3.3
never participates) serving kingjones777/Qwen3.8-27B-ROCmFPX-Q8_0-AGENT-GGUF --
a Q8_0_ROCMFPX_AGENT (ftype 115) checkpoint that stock llama.cpp cannot load.
Placeholder config: --split-mode tensor across both cards, 131072 context,
1 slot, f16 KV, MTP draft head on ROCm0.
deploy-remote.sh grows retire_root_llm_small(): the CUDA-era unit is stopped,
disabled and moved to ~/.AGENTTRASH/20260915-frank-small-retire/host-units
before frank-medium starts, because `systemctl disable` has been observed to
delete the unit file itself on this host (2026-08-26 incident).
The retired frank-small repo files (env, unit, serve script) move to
scripts/.AGENTTRASH/, and the planner, dispatcher, service allow-list and the
deployment test seams now name frank-medium.
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.
Concurrency cap trimmed to the real fleet size. Each slot is 111.0 MiB of
O(1) state, so this hands ~111 MiB back on a host sitting at 7.1 GiB left
for everything else.
Note the slots already oversubscribed the pool (4 x 262144 > 524288), so
admission was pool-bound regardless: at most two full-length conversations,
or four at 131k. This does not change how much context fits -- it caps
simultaneous requests at three, and a fourth now queues.
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.
Tested 16 live. Pool stayed 524288 (no clamp), server held 103.5 GiB and
reported "7.1 GiB left for everything else" at BOTH 15 and 16 — identical
footprint. HALOGEN_HOST_RESERVE_GIB feeds the fit-check budget only; it is not
an allocation, so raising it cannot give RAM back to the co-tenants. Its sole
observable effect above 16 is to take the KV pool away.
16 therefore cost half the fit margin (1.0 vs 2.0 against the engine's 36.5 GiB
requirement) and bought nothing. 15 is strictly better at the same cost.
Recorded the ceiling (16) and the real air levers in the unit comment: pool
262144 frees 7.2 GiB, max_tok 16384 frees 8.4.
Operator asked for 18. 18 is not available: the engine's own fit-check line,
read from a clamped boot rather than modelled, requires budget >= 36.5 GiB to
keep 524288 at max_tok 32768 ("524288 positions need ~35.0 GiB (plus 1.5 of
margin)"). budget(18) = 35.5, so 18 silently un-doubles the pool we just paid
the reserve for. 17 fits by 0.02, inside the noise of the 67.7 resident-weights
term. 16 is the largest reserve that survives the check with real margin:
37.5 budget vs 36.5 required.
Recorded the ceiling and both remedies in the unit comment. The reserve is now
spent, so a future stall is answered by the pool or the arena, not here.
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).
Neither lever alone could give both features. At reserve 24 the device budget
is 29.5 GiB and max_tok 32768's 37.0 GiB need always lost the doubling to the
fit-check clamp; at max_tok 16384 the doubling survived but prefill paid 8-9%.
Dropping the reserve to 15 moves the budget to 38.5:
121.2 host - 67.7 resident weights - 15 reserve = 38.5
need = 14.4 pool + 20.6 arena + 1.97 vision = 37.0 -> 1.5 GiB spare
So the payment is page cache for the n-gram lookup table rather than either
feature. 15 GiB, against a table of 47.7 GiB read through the file cache.
Two reasons this is a survivable risk rather than a cliff:
- The engine's own ledger reported only 12-17 GiB left at the previous 24
setting, so the declared reserve was never held literally; 15 may cost
less real cache than it reads as.
- The failure mode is legible: long prompts fault heavily against a busy
disk, and raising the reserve is the documented remedy. Not the silent
clamp, and not the 65536 no-boot mode.
1.5 GiB of headroom is thin and deliberate. Documented in the unit as a
matched pair with the pool: raise the reserve back to 24 and the doubling
quietly goes away again.
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.
Operator call. Prediction from the measured model: 524288 positions cost
35.0 GiB against a 29.5 budget, so HALOGEN_KV_POOL_FIT clamps the pool to
262144 (27.8). Rollback is a single revert; 16384 was the config that made
the doubling real.
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).
Operator call, measuring what the prefill arena actually costs against
upstream's do-not-raise warning. Rollback is one revert: the settled
config was 16384, which made the 524288 pool fit at 26.6 GiB.
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).
The previous commit's doubling was a silent no-op. HALOGEN_KV_POOL_FIT
measured the device budget at startup and lowered the pool straight back
to 262144, while the deploy still reported healthy in 12s.
The engine's arithmetic: 121 GiB host - 67.7 resident weights - 24.0
reserved for the n-gram page cache and the OS = 29.5 GiB of device
budget, and that single budget carries the KV pool AND the prefill arena.
At HALOGEN_MAX_TOK=32768 the arena is ~20.6 GiB, so 524288 positions
need 35.0 GiB against 29.5 available.
HALOGEN_MAX_TOK=16384 halves the arena to ~10.5 GiB, bringing the
524288 need to ~24.9 GiB, which fits with ~3 GiB of margin. Cost is
about 8-9% of prefill throughput, per upstream. This is the engine's own
prescribed fix, quoted verbatim in its refusal message.
Lowering HALOGEN_CTX was the other option the engine offered; rejected
because both clients declare 262144 as the context window, so it would
break the client contract rather than pay for the pool.
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.
Documents the root-cause chain for the marreta.franksdank.cloud login wall:
null invalidation_flow (outpost serde crash) -> missing flows -> and the real
blocker, ProxyProvider.application=None so the embedded outpost's
filter(application__isnull=False) never loaded the provider. Notes the
providers are live-only (no repo blueprint) as a provenance gap.
16-marreta.spec.ts (standard project) asserts the user homepage's Marreta
open link actually reaches the Marreta app through the full guard chain
(Anubis PoW -> Authentik SSO -> app), and explicitly NOT the authentik
login flow. Guards the 2026-09-12 incident where marreta-proxy was created
without a linked Application, so the embedded outpost never loaded it and
the homepage link parked on a login wall.
Wires spec 16 into the standard project testMatch. Verified passing live.
Adds the Marreta content-reader tile (http://marreta.franksdank.cloud,
probe 127.0.0.1:8082 via gluetun's publish) to the user grid so regular
authenticated users get a link to it. Registry-driven e2e picks it up
automatically.
The 2026-09-12 homepage outage slipped through because the live-fleet
Playwright suite was a manual 'MUST invoke' that deploys skipped. Wire it
into deploy.sh as the [3/3] gate: after a behavior-affecting deploy,
run tests/e2e/run.sh against the live fleet; a failure exits DEPLOY
INCOMPLETE. Add --skip-e2e as an emergency escape hatch. Update the
AGENTS.md deployment-completeness contract and flow diagram to match.
Verified: bash -n clean, --skip-e2e parsed, unknown flags rejected,
seam suite green (only the pre-existing comfyui test 73 fails).
Deployed live at 91e9410 (gluetun+marreta+caddy), authentik provider pk 13,
guard chain verified end-to-end. Documented that Laravel APP_KEY must carry the
base64: prefix — a bare base64 key 500s every app route with 'Unsupported
cipher or incorrect key length' while /up still passes.
The scoped runner (deploy-remote.sh) already knew marreta; the top-level
deploy.sh valid_service() gate did not, so --service marreta was rejected
before reaching the planner.
Verbose engine logging is the vendor's first-line evidence for load,
allocation and admission questions; shipping it on means an incident
report already carries the full startup account instead of needing a
restart-into-debug. HALOGEN_DMALLOC_LOG stays off (every alloc >64MB is
steady-state noise on a shared host). Doc updated with the rootless
podman-logs-vs-journalctl retrieval note.
Rewrites the [[Observability]] umbrella page and the five component pages for
what is actually deployed: rootless Quadlet units instead of compose services,
the loopback-only publish posture (node-exporter's LAN binding explained), the
loki healthcheck living in its image because the image has no shell, and the
Home=/root trap that killed the first grafana start.
Adds the note that matters most to a reader: the stack is installed, healthy
and UNCONFIGURED. Grafana reports zero datasources and zero dashboards with an
empty provisioning directory; Prometheus has no rule files and no Alertmanager;
Loki is configured but has no reader; node-exporter runs its default
collectors. Alloy is the only configured piece. Metrics and logs are collected
and retained, and nothing is visualized or alerted on.
AGENTS.md's boot-persistence gap note is replaced by the resolution.
The healthcheck added to the loki image was discarded by the build:
'HEALTHCHECK is not supported for OCI image format and will be ignored'.
podman builds OCI by default, and OCI has no HEALTHCHECK field, so the
container came back with no health at all — worse than the failing check it
replaced, because nothing reports the loss.
The build now passes --format docker, and the seam pins it so a future edit
cannot quietly reintroduce an image with no health field.
The loki unit came up 'unhealthy' after the quadlet cutover. Cause: a unit's
HealthCmd reaches podman as CMD-SHELL, i.e. /bin/sh -c, and the loki image is
distroless — no shell exists in it. The busybox command itself works (probed
by hand: /ready answers 'ready'); it was the shell wrapper that could not run.
podman does not accept an exec-form command through --health-cmd either: it
stores a JSON array as one literal argument.
The command therefore moves into dankclaw/loki/Containerfile as a
HEALTHCHECK ... CMD [...] exec form, which is what that image exists for — it
already ships the static busybox solely so a healthcheck has something to run.
podman reads the image's healthcheck when the unit declares none.
The first deploy of the grafana unit failed: 'stat /root/.config: permission
denied', exit 1, restart loop. Cause — the repo-root .env that the compose
service used to pick up is a gluetun env file, and it carries HOME=/root (and
the WireGuard keys). Pointing the unit's [Service] EnvironmentFile at it sets
HOME for the service, so podman looks for its configuration under /root.
grafana now reads /home/dank/Containers/grafana/env, the per-service file the
deployer seeds from dankclaw/.env with a single-variable filter — the same
shape caddy, pihole, copyparty, romm and seerr use. The seed is validated, so
a missing secret fails the deploy loudly instead of starting a container whose
OAuth login cannot complete. The value was copied to dankclaw/.env, which is
where dankclaw/.env.example already documented it.
prometheus, grafana, loki, alloy and node-exporter ran only from
dankclaw/compose.yaml. Rootless podman does not restart bare containers, so
they did not survive the 2026-09-11 reboot for the amd_iommu=off change and
had to be restored by hand — the same gap that had already cost the fleet
Caddy and Pi-hole, which is why the boot-persistence rule exists.
Each is now a rootless Quadlet unit under dankclaw/observability/quadlet/,
installed and dispatched by deploy_observability in deploy-remote.sh, with
the compose definitions retired so only one owner exists per container name.
Images, mounts, arguments and container names are unchanged, so Caddy's
proxies and Prometheus' own scrape config keep resolving.
Two posture changes, both about the unauthenticated surface:
* The three HTTP UIs published on 0.0.0.0 with no authentication of any kind
(Prometheus even with the lifecycle endpoint enabled). They now publish on
127.0.0.1: the browser path is the guarded franksdank.cloud vhost, Caddy
reaches them by name over dankclaw_default, and the host-side homepage
status timer probes loopback. Nothing legitimately used the LAN port.
* node-exporter bound :9100 on every interface; it now binds exactly its two
consumers, loopback (homepage timer) and the LAN address (Prometheus, which
runs in a container and cannot reach the host's loopback).
alloy gains a reachable health endpoint: its UI listens on the container's
own loopback and alloy 1.x refuses a config block for it, so the compose
service had nothing to probe — the homepage status card for alloy read 'down'
continuously. The unit moves the UI with a command-line flag and publishes
127.0.0.1:12345.
grafana keeps its Authentik OAuth client secret without receiving the whole
.env: systemd expands the value from a [Service] EnvironmentFile, which never
reaches the container.
Hermes hardcodes response_format=json_schema on its title call. No local
back-end implements constrained decoding — frank-small and frank-large both
answer 400 'response_format json_object/json_schema is not supported' — so
every session logged a title failure and fell back to a message-derived
title. The NPU endpoint that used to absorb this call was retired on
2026-09-11.
The call now goes out over the Responses transport (the only one that does
not forward response_format) with thinking off. A 64-token cap cannot hold a
thinking pass plus a title: with the default reasoning effort the model spent
all 64 tokens thinking and returned empty content, which is why disabling
reasoning is part of the fix and not decoration.
Verified against the live endpoint with the real prompt and the real
extraction path: raw {"title": "Friendly greeting"} -> title
'Friendly greeting' in 3.4s, where the previous configuration returned 400.
Rebooting Dankclaw for the amd_iommu=off change exposed five services that
exist only as compose.yaml services with no Quadlet unit (prometheus, grafana,
loki, alloy, node-exporter): they did not come back and had to be restored by
hand. The boot-persistence rule already requires units for them; this records
the concrete gap so the next reboot is not a surprise.
frank-large moves from the kingjones30/ROCmFPX llama.cpp build to
peonist-ai/halogen-flash-server 0.5.8 serving Qwen3.8-Flash-Next (w4b
.hgn 115.55 GiB + quality overlay). Upstream measures this engine at
roughly 4x the prefill throughput of the llama.cpp family on the same
gfx1151 silicon, and prefill dominates any prompt with real context.
Shape: a rootless Quadlet container (dankclaw/frank-large/quadlet/),
engine on container loopback :8730 (never published), OpenAI front-end
:8731 published as host 7770, HALOGEN_MODEL_ID=frank-large so no client
renamed anything. Context 262144, KV pool 262144 positions (~7.2 GiB,
conservative for a shared 121 GiB host), 4 slots, host reserve 24,
vision tower on, weights pinned, all other flags at vendor defaults.
Weights are fetched by HALOGEN_DOWNLOAD on first start into a writable
/home/dank/models/halogen.
The llama.cpp era (unit, env, serve script) is staged in .AGENTTRASH/
rather than deleted; the STRIX payload and the ROCmFPX build stay on
disk as the rollback target. deploy_llm_large becomes a user-quadlet
handler: it retires the old ROOT unit (moving its checkout symlink into
the home trash, never rm) and gates the first start on the download
rather than reporting a failure for a fetch that is working exactly as
designed.
frank-npu is retired in the same change. The kernel cmdline moves from
iommu=pt to amd_iommu=off for the 13-16% of prefill upstream measured,
and that takes the XDNA2 NPU with it: without PASID/SVA the driver
cannot open /dev/accel/accel0 (err=-19). Its five dependent routes are
repointed (Hermes vision -> frank-large, Hermes title generation ->
frank-small, harness vision -> frank-large, tiny -> frank-small, task ->
frank-large:medium) and its provider blocks are removed from both the
harness config and Hermes's live config. Audio input leaves the fleet
with no replacement; image input survives on halogen's vision tower.
Docs: new halogen-flash-server runbook, new IOMMU decision record, and
frank-npu rewritten as a retirement record. AGENTS.md topology, runtime
sections, benchmarking policy (frank-large is a user unit now) and the
harness model-config section are corrected.
The unit comment always claimed the webui drives the same agent state as
the gateway, but the image bakes HERMES_HOME=/opt/data and no unit
overrode it. Podman auto-creates a fresh anonymous volume at every
image-VOLUME point on container recreation, so each restart reset the
webui's config home to the image-default template (provider: auto ->
openrouter.ai with no API key), while the gateway read the corrected
/data/config.yaml. Pin HERMES_HOME=/data so both processes share one
config/state root and config edits propagate without a recreate.
Live config was still routing on pre-rename aliases (frank-default-large,
frank-default-small) after the 2026-09-03 frank-large/frank-small/frank-npu
rename cutover; hermes-webui resolved those ids to nonexistent models and
requests failed. Live config fixed in place (backup kept on-host as
config.yaml.bak-20260905); this backup now mirrors the live shape: named
providers for aux routes, title_generation/triage/kanban routed to their
actual models, moa + mcp sections, signal disabled.
Wget-style (no-JS) clients must never receive application bytes from guarded
services: copyparty serves authenticated traffic with anonymous read perms, so
the guard chain is the only boundary between an unauthenticated client and the
file bytes. Assertions target leakage markers in app code (Copyparty UI title,
homepage panel grid, services.json JSON shape), not gate internals, so any
guard-route regression fails the suite while any gate implementation passes.
Also route 15-*.spec.ts into the guest project (empty storage state).
Flip the homepage Sonarr card from the user category to admin and repoint
its link at the LAN-direct port (192.168.0.112:8989). Retire the unguarded
public vhost sonarr.franksdank.cloud (direct reverse_proxy, no auth) and its
post-auth @sonarr handler from the Caddyfile; drop the Pi-hole addn-hosts A
record with it. No public Porkbun record ever existed (NXDOMAIN at the zone),
so after this there is no WAN path to Sonarr: users request TV via Seerr
(guarded), admins reach Sonarr's own UI from the LAN only.
Fixes the seerr/sonarr mix-up from 19b2589/e0bade6.
FrankBook-Pro becomes the only *arr/qBittorrent deployment. Every fleet
container (qB + five arrs) now shares gluetun's netns, so all egress —
torrent, indexer APIs, metadata lookups — exits the WireGuard tunnel by
construction (AllowedIPs=0.0.0.0/0); the arrs' own ports publish on the
gluetun unit, none on the arr units themselves.
- dankserver/compose.yaml: remove the six arr services (the Aug 20 compose
copy — never healthy, superseded; config dirs trashed on the host)
- dankclaw/Caddyfile + shared/homepage/services.json: arr vhosts and
probes repoint 192.168.0.235 -> 192.168.0.112 (FrankBook-Pro)
- frankbook-pro units: five arrs join gluetun netns (Wants/After/Network);
gluetun publishes 9696/8989/7878/8686/8787 alongside 8080/6881
- tests/frankbook_quadlet.bats: new VPN egress seam — every arr+qB unit
must share gluetun's netns and publish no port; gluetun must publish the
whole fleet's ports
- deploy-plan.py: gluetun changes carry the whole netns family (a gluetun
recreation kills all six joiners); dankserver compose set loses the arrs
- AGENTS.md: deployment is complete only when unit tests AND the Playwright
E2E suite both pass
Rootless podman on dankserver already maps container root (PUID=0) to
the dank user (uid 1000) by default — /proc/<pid>/uid_map reads
0->1000, 1..->100000+. Verified across plex, tautulli, jellyfin,
calibre-web, copyparty, romm. userns_mode: host is unnecessary, and
podman-compose 1.5.0 cannot apply it anyway (compose services run in
pods; --userns and --pod cannot be set together). The real fix for the
never-booted arr stack was chowning the root-owned /opt/<svc>/config
dirs to 1000:1000.
The arr services were committed (b9de886) with PUID=0 but no userns_mode.
Under rootless podman's default subuid mapping, container-root cannot write
host-root-owned bind mounts, so every app died in logger init
(UnauthorizedAccessException on /config/Sentry) before binding its port —
the stack 502'd since it was first brought up (2026-09-04 00:33).
Per the fleet convention (AGENTS.md, plex/tautulli migration): PUID=0 +
UserNS=host maps container-root to dank (1000), which owns /opt data roots
and /mnt/entropy. Config dirs chowned 1000:1000 at bring-up.
Also drops a duplicate 'networks: [appnet]' key on prowlarr.
The arr stack (prowlarr, radarr, sonarr, lidarr, readarr, qbittorrent)
had A records in addn-hosts but no Caddyfile blocks. This adds:
- Six public vhost blocks in the dankserver section, each proxying to
192.168.0.235:<port> (exempt by construction: own auth, no Anubis)
- Six @arr named handlers inside :8017 for authenticated traffic routing
All 233 bats tests pass.
Boot tag flips from qwen3.8-distilled-heretic:9b to nuextract3:4b (verified
live: per-request routing already serves every registered tag; the boot tag
only decides which model is resident at start, so the first request after a
swap pays the load cost). FLM_CTX_LEN=-1 drops the global --ctx-len override
so each pack loads with its own trained context (nuextract3 32768, gemma4
131072).
Live run against the fleet showed the outpost redirects direct navigation to
auth.franksdank.cloud with the guarded host preserved as redirect_uri, rather
than rendering the flow UI in place. The security property under test is that
an unauthenticated visitor never sees app content; assert that shape (uidField
visible + no leak markers + gate is either in-place or a redirect carrying the
return path) instead of hostname preservation.
Mechanical identity rename across the knowledge base, AGENTS.md, README, the
hermes config example, and homepage services.json. No behavior change; served
aliases on the LLM endpoints change with the deploy of commit 69f87c8.
14-guard.spec.ts: every authentik-guarded vhost (www, search, romm, files,
seerr, grafana, prometheus, loki) must challenge an unauthenticated direct
navigation; plex is asserted exempt (PMS answers directly, no IdP form).
11-admin.spec.ts: admin direct-nav to grafana/prometheus/loki passes through
with host preserved and no uidField challenge.
frank-default-large -> frank-large (Dankclaw :7770), frank-default-small ->
frank-small (Dankserver :7770), frank-default-sightsound -> frank-npu
(Dankclaw :7771). Units, env files, serve scripts, deploy-plan mappings,
deploy-remote dispatch (stop/disable old units + relink + enable new), and the
test seams move together. Retired frank-default-npu.* artifacts moved to
scripts/.AGENTTRASH/ (tracked-in-history; ignored on disk).
Served aliases become frank-large / frank-small via --alias; the NPU endpoint
serves the whole FLM registry, so its rename changes no served tags.
Post-change doc reconciliation for the hermes vhosts + DNS deletion batch:
- Hermes-Dashboard.md: access is now the hermesadmin.franksdank.cloud vhost
(direct Caddy proxy to 192.168.0.143:9119, no guard chain — exempt because the
dashboard enforces its own basic-auth login), plus the raw LAN port. The old
text claimed an Authentik-guarded public vhost, which was decommissioned.
- Hermes-WebUI.md: description updated — the webui is now also reachable via the
unguarded hermes.franksdank.cloud vhost (its own password auth is the gate).
- LAN-DNS.md: "add a name in two places" corrected to a two-or-three-step
decision; internal-only names (prometheus/loki/grafana/seerr) deliberately get
no public A record and resolve on the LAN only, and DNS-01 issuance is
unaffected because the challenge is a TXT record in the same zone.
- Seerr.md: DNS row corrected — addn-hosts only since 2026-09-02; NXDOMAIN on WAN.
- hermes-dashboard.container: unit comment no longer claims Authentik termination.
No behavior change in this commit; tests 233/233 green (docs-completeness seam included).
Both Hermes services get direct (unguarded) vhosts: each authenticates itself
(webui password; dashboard basic_auth in /data/config.yaml), so no Anubis
challenge and no Authentik proxy provider is required. Upstreams are the host
LAN IP because Caddy itself sits on dankclaw_default, where the gateway
address 10.89.1.1 has nothing serving 8787. Internal-only by DNS absence:
hermes.franksdank.cloud already resolves via Pi-hole addn-hosts; hermesadmin
is added in this same commit so the caddyfile/addn-hosts cross-check (test 178)
stays green.
New page Architecture/Container Network Namespaces.md: which local services are
addressable from which Podman network namespace on Dankclaw (rootless podman +
netavark). Verified live 2026-09-02: gateway/webui/CDP run in the host netns and
dial loopback; hermes-dashboard is the only Hermes container on a bridge network
(dankclaw_default, 10.89.1.0/24 gw 10.89.1.1) and can dial published ports only —
unpublished host-netns listeners (signal-cli :8080, webui :8787) are refused from
that namespace even though both bind 0.0.0.0.
Corrected while verifying the above:
- Hermes.md: data root is /home/dank/Containers/hermes/data (not the retired
/home/dank/hermes-data); gateway dials the Signal daemon on loopback; the
daemon itself is a host user unit (signal-cli-daemon.service) bound to
0.0.0.0:8080, so the Signal HTTP API is LAN-reachable.
- Hermes-Dashboard.md: it is a Quadlet unit (Network=dankclaw_default), not a
compose service; data root + /opt/data anonymous volume corrected.
- Hermes-WebUI.md: binds 0.0.0.0:8787 in the host netns (verified: unauthenticated
GET /health from another LAN machine returned 200) — the SSH tunnel is an access
path, not a containment boundary; HealthCmd and data-path corrections.
- Network Architecture.md: no hermes vhost exists in either the tracked or the live
Caddyfile (hermes.franksdank.cloud resolves but routes nowhere); SearXNG exits the
tunnel via the shared named network searxng.network, not service:gluetun.
Docs only — deploy is a no-op for this change.
PMS 1.43 serves TLS-only on 32400: a plain-HTTP GET gets a TLS alert and
curl reports "empty reply from server" (exit 52, code 000), so the card read
down while Plex answered 200 at /web/index.html over HTTPS. TCP connect is the
honest liveness signal for this port; the public URL stays HTTP because Caddy
terminates TLS and re-keys the path.
The Hermes card probed 127.0.0.1:8090, a port retired when the hermes
stack moved from compose (8090 web dashboard) to quadlets: the dashboard
now publishes 9119 and the webui listens on 0.0.0.0:8787 in the host netns.
Nothing has listened on 8090 since, so the card reported "down" while the
gateway, dashboard, and webui were all running.
Probe 127.0.0.1:8787 instead — the webui runs HERMES_HOME=/data, the same
agent state as the gateway, and /health answers only when that server is
alive (uptime, sessions, active runs). The prober runs on Dankclaw, so the
loopback probe hits the host netns where the webui binds.
The C5 probe failed for the wrong reason: plex.container lost its Network=
key in the same edit churn that ate tautulli's ContainerName. Plex ran on
the default podman network — published 32400 still worked (clients fine)
but tautulli could not resolve the plex alias, breaking the seeded
PMS_Web_Host=plex. Restore appnet membership; structure now parallel to
tautulli.container. Lesson recorded: after multi-hunk line edits on these
unit files, diff the whole file — key-order churn has eaten three keys
already.
Without ContainerName= the v1 generator names the container
systemd-tautulli; the fleet references containers by their plain names
(podman inspect/exec, deployer health checks). Container was healthy but
renamed — caught by post-cutover verification, not by the seam; the seam
only asserts file keys, so pin the name here and keep plex/tautulli
structures parallel.
- tautulli.container: HealthCmd switches to the repo's proven curl pattern
(jellyfin unit) — quadlet's systemd-style value parsing mangles embedded
double quotes; PublishPort=8181:8181 restored (a mis-aimed line edit
replaced it; the PublishPort seam test caught the regression)
- deploy_plex readiness gate: 36x5s -> 120x5s; the 2026-09 cutover showed
the first container boot of a years-old library runs ~10 min while PMS
pre-binds 32400 and empty-replies; 180s failed the gate twice with a
healthy, still-booting service
Evidence from the failed first cutover: podman top shows container root
maps to dank (1000) under UserNS=host rootless (subuid range 100000:65536
cannot express 997), so PUID=997 fell back to root and PMS crash-looped on
write access to its 997:984-owned data root. Fleet model instead: PUID=0/
PGID=0 (container root = dank, same as jellyfin/copyparty data roots) and
deploy_plex chowns /var/lib/plexmediaserver to 1000:1000 inside the
first-cutover block. Zero bytes moved; rollback regains one chown -R
997:984 step, documented in Plex.md. Seam test pins PUID=0/PGID=0 and the
absence of any userns key.
podman 5.7 podman-user-generator on dankserver hard-fails the whole unit on
the unknown key ('unsupported key UsernsMode in group Container'), so the
first cutover attempt masked the native unit and then found no plex.service
at all. Swap both units to UserNS=host (same --userns=host flag), update the
ownership-contract seam test to pin the exact key, and correct both docs.
Pin the seam: the v1 generator rejects unknown Container keys silently at
deploy distance — the test now pins UserNS=, not the misspelled alias.
- runtime rows: podman quadlet under the systemd user generator, pinned
image digests, in-place bind mounts, identity model, appnet aliases
- cutover/rollback section on Plex.md (mask-native-unit lever, decommission
gate); drop the false 'updates are apt-only' and 'snap refresh' gotchas
- update path is now deploy.sh dankserver --service plex|tautulli
- plex: lscr.io/linuxserver/plex:1.43.3.10896-cb3ebc72d-ls322 (verified ==
ghcr latest digest; plan's ls9 pin superseded by the re-verify clause),
/var/lib/plexmediaserver bind-mounted in place as /config, /mnt/entropy
mounted ro at the same path, PUID=997/PGID=984 + UsernsMode=host so the
container's file identity is byte-compatible with the native unit
- tautulli: lscr.io/linuxserver/tautulli:v2.18.1-ls242 (verified == ghcr
latest digest), fresh /home/dank/Containers/tautulli config root,
PUID=1000/PGID=1000, python3 /home healthcheck
- both on dankserver_appnet with aliases; 32400/8181 published as before
- deploy_plex(): install unit, stop+mask native plexmediaserver.service on
first cutover (mask, never disable), gate on /identity for 180s
- deploy_tautulli(): install unit, stop+disable snap on first cutover,
seed config.ini with PMS_Web_Host=plex only if absent
- allow-lists (deploy.sh valid_service, deploy-remote.sh service_is_allowed),
planner routing (SERVICE_ORDER, DANKSERVER_SERVICES, quadlet_services),
test-changed cutover case, quadlet seam enumeration + new
userns/PUID=997/PGID=984 ownership-contract test
The loki/grafana/seerr proxy providers were script-provisioned with no
flows and no Application: null flows made the embedded outpost's Rust
deserializer reject the whole config (all guarded vhosts down after any
outpost reload), and the missing apps bounced post-login to authentik's
user UI instead of the app. Backfilled both live; documented the two
provisioning rules and updated the provider inventory. SSO e2e now
asserts the app landing (pathname leaves /if/, no IdP markers in body)
instead of only the hostname.
- qB 5.x's 204 login breaks Readarr's retired client at runtime; 4.6.7 answers
legacy 200+Ok. Pin recorded in the unit and documented.
- Prowlarr download client created via schema template; all five clients test
200 through the podman gateway.
- PBKDF2 gotcha: qB 5.x conf rewrite left a hash 4.6.7 rejects; fresh rewrite
of the same line restored login.
- E2E: torrent through the Panama tunnel at ~17 MB/s, completed into Finished
Torrents, deleted with data.
podman 5.7.0 refuses ghcr.io/goauthentik/ldap:2026.8.0's nested manifest
list ('Unexpectedly received a manifest list instead of a manifest for a
single image'); pin the linux/amd64 single-platform digest instead.
Document the Mac *arr/qBittorrent stack (VM placement, gluetun credential
gate, socat LaunchAgents, PBKDF2 qB seed) and log the day: dankserver
recovery + full quadlet cutover of the six media services, the frankbook
skip-propagation/PBKDF2 handler fix, the 7-service frankbook deploy, and
the jellyfin GPU-parity finding.
The frankbook handler ran port verification even for services the
credential gate skipped, so a parked qbittorrent failed the deploy and
aborted the five arrs. Gate now lives Mac-side (test -s gluetun/env
inside the VM) before restart+verify.
The qB first-boot seed wrote a plaintext WebUI\Password, which qB >= 4.2
ignores (PBKDF2 @ByteArray only), and omitted HostHeaderValidation=false,
which would 401 the LAN forwarder path. Seed now generates the PBKDF2
hash in the VM (python3 hashlib) and sets ServerDomains=* +
HostHeaderValidation=false. Arr categories (lidarr/radarr/tv-sonarr)
pre-created with save_path=/Volumes/COMPLETE/Finished Torrents so any
grab lands in the review dir regardless of wiring details.
Seven units installed INSIDE the Mac VM (/home/core/.config/containers/
systemd) by a new deploy_frankbook_stack handler: gluetun (ProtonVPN
WireGuard, env-file credentials, publishes qbit's 8080/6881), qbittorrent
(Network=gluetun.container, verbatim /Volumes data roots, first-boot
config seed with fresh machine-local WebUI password and the Finished
Torrents category), and prowlarr/sonarr/radarr/lidarr/readarr (grab-only,
config migrated from dankserver at provisioning, readarr pinned to
amd64-1.0.980). deploy.sh routes the new target through the ssh-config
alias with REMOTE_SHELL=/usr/local/bin/bash (macOS bash 3.2 lacks
associative arrays); the planner learns FRANKBOOK_SERVICES and the
gluetun->qbittorrent netns pairing; new seam test pins the contract.
Two live-verified loader failures on Dankserver (2026-09-01), each a
clean exit at load (no swap-thrash, 37 GB fits RAM):
1. ornith-ai/Ornith-1.5-35B-A3B-FP8 (compressed-tensors PER-CHANNEL fp8):
ValueError: Missing MoE expert source layers: {'gate_up': [0, 1, ...]}
— nonstandard expert tensor naming, provider finds no expert sources.
2. lamianlbe/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-fp8
(compressed-tensors BLOCK-fp8 128x128, standard per-expert naming):
RuntimeError: The size of tensor a (2048) must match the size of
tensor b (...) in _dequant_fp8_weight — the dequant path multiplies
weight * weight_scale directly and cannot expand compressed-tensors
block scales.
Empirical matrix for FreeToken v0.1.2 on qwen3_5_moe (RTX 3090):
modelopt block-fp8 (quant_method 'fp8' + weight_scale_inv) -> LOADS
compressed-tensors per-channel FP8 -> fails
compressed-tensors block-fp8 128x128 -> fails
batsclamp/Huihui-...-abliterated-FP8 is the proven uncensored checkpoint;
restoring it until a modelopt/NVFP4 requant of a wanted model exists or
FreeToken learns compressed-tensors.
Phase 1/2 of the podman standardization plan: service pages updated to the
quadlet reality (units, Containers data roots, env-file secret convention,
the v1-generator findings each caught live), Storage Policy documents the
two rootless ownership regimes and the merge-move migration, Deployment
Flow's service-manager table no longer claims dankserver media is bare
inventory-only. Bundle change log entry added.
Ornith-1.5-35B-A3B-FP8 failed to load on FreeToken v0.1.2: its
compressed-tensors checkpoint uses a nonstandard expert tensor layout
(FreeToken: 'Missing MoE expert source layers: gate_up'). Retarget to
the HauhauCS Aggressive FP8 finetune of the SAME Qwen3.6-35B-A3B arch
Huihui used, which carries compressed-tensors BLOCK-fp8 (128x128) with
the standard per-expert naming (mlp.experts.N.gate_proj/up_proj/down_proj
+ .weight_scale) that FreeToken's MoE loader reads.
- MODEL_DIR -> /home/dank/models/hauhaucs-qwen36-35b-a3b-aggressive-fp8
- use the checkpoint's own chat_template.jinja (drop the Sharp override;
forcing a template the finetune wasn't trained on risks format drift)
- pool plan unchanged (same ~35 GB FP8 mass / qwen3_5_moe arch)
The cutover unit joined dankserver_appnet but dropped the compose-era
0.0.0.0:3923 mapping (live inspect: bridge, 3923 published), so
copyparty listened inside the netns while the host refused. New seam
asserts every host-facing port matches the pre-cutover PortBindings:
romm 8085->8080, jellyfin 8096, calibre-web 8083, copyparty 3923.
The podman 5.x v1 quadlet generator rejected the whole jellyfin unit on
the unsupported Device= key (generator log: unsupported key 'Device' in
group 'Container'), so no jellyfin.service was generated and the cutover
restart failed with 'Unit not found' after the data root had already
migrated cleanly. gluetun proves AddDevice= is the house key. New seam
test denies Device= across all six dankserver units and pins the
AddDevice=/dev/dri replacement.
rename() of a directory requires write on the directory itself; mariadb/
redis data roots are owned by the containers' mapped subuids (100998),
so the deploy user cannot move them. Use run_privileged mv (numeric
ownership preserved = rootless mapping contract) and skip the chown pass
for romm via the new keep mode — chowning mapped subuid data to the
deploy uid breaks the in-container mysql/redis users. Collisions with
rescued anonymous-volume stubs now merge one level deep instead of
stranding real data under /opt. Reverse the podman rm order so the
dependent romm container is removed before its backends.
dankserver's ~/.config/containers/systemd did not exist (no quadlet units
had ever been installed there); the dankserver media handlers cp'd into it
without mkdir -p, failing before touching any container.
- MODEL_DIR -> /home/dank/models/ornith-1.5-35b-a3b-fp8 (37 GB,
compressed-tensors per-channel FP8, same qwen3_5_moe arch/mass as
Huihui so the pool plan carries unchanged)
- CHAT_TEMPLATE_FILE -> scripts/chat-template.jinja (Qwen-Sharp, shared
with frank-default-large); the checkpoint's stock template
mis-serializes reasoning per its own model card
- env/serve/unit headers updated; loader-risk + BF16-does-not-fit note
mv of a data root onto a non-empty directory fails (Directory not empty).
The romm/copyparty cutover creates Containers/<svc>/ before migrating:
the seed env file lands there first, and the romm anonymous-volume rescue
extracts into it before the /opt move. Merge the old root's entries in
without overwriting existing state (mv -n), warn and leave residue if a
collision kept the new side, and keep the plain rename when the
destination is absent or empty.
The first deploy after a unit-file edit restarted with the stale generated
unit (generator had not seen the new Volume= paths), crash-looping hermes on
statfs ENOENT for the just-renamed data root. Reload between cp and restart.
Repoint Volume=/EnvironmentFile= in the 8 affected quadlet units to the
caddy-established Containers/<svc> layout: gitea (data, postgres, env),
authentik (data, postgres), hermes (data, webui checkout). Data-root mv on the
target happens at deploy time before each group's unit restart; same-fs rename,
instantly reversible. Running containers keep their inode-bound binds until the
unit restart flips them to the new path.
Cutover of the six surviving dankserver services from compose-era bare
containers to boot-persistent Quadlet units (WantedBy=default.target; podman
5.x v1 generator honors it at user-manager start).
Units are faithful to live inspect ground truth (image pins = running tags,
not compose inventory):
- romm trio shares one data root: /home/dank/Containers/romm mounted whole at
/romm; mariadb datadir and redis data dir live inside it. Aliases mariadb/
redis preserved on dankserver_appnet via Network=...:alias=. Secrets stay in
the host-local env file (EnvironmentFile + ${VAR} expansion in the generated
ExecStart, gitea precedent). copyparty passes its admin account via
Exec=-a ${COPAY_ACCOUNT}:${COPAY_PASSWORD} (Exec= not Command=: the v1
generator hard-fails on Command=); entropy stays rw only there.
- jellyfin keeps every creation-time env as Environment= lines (zero drift);
Device=/dev/dri + GroupAdd=keep-groups preserve hardware transcoding.
- calibre-web pins the exact running image tag (id 986277563e76 =
0.6.27-ls397), pasta network preserved.
Deployer: first-cutover handlers retire the old container, rescue unshadowed
anonymous-volume content (podman volume export + tar --skip-old-files), mv
data roots on the same fs, then start units in dependency order with health
gates. Re-deploys restart only the selected unit. deploy-plan.py maps compose
keys mariadb/redis to the new unit names; allowlists updated. New seam test
tests/dankserver_quadlet.bats pins boot-persistence + v1-generator contract;
test-changed.sh maps the new dirs to it. Full suite green (209 tests).
-v /w::A granted anonymous full rwmda on the LAN (verified: anon PUT -> 201,
PROPFIND -> 207). New perms ::r:ra,dank: browse/download for anonymous
(required by the guarded files.franksdank.cloud chain), admin for dank,
no w/m/d for anyone. Runbook corrected with the --help-accounts semantics.
Adds a digest-pinned headless-shell Chrome DevTools Protocol daemon as a
boot-persistent quadlet unit (chrome-devtools-protocol.container). The
entrypoint override drops the image's socat front so the CDP endpoint binds
loopback-only on 127.0.0.1:9222 in the host network namespace; only the two
host-network consumers (Signal gateway, webui server) can drive it.
hermes-webui flips Network=dankclaw_default -> Network=host so its agent
process shares the daemon's namespace (0.0.0.0:8787 bind is LAN-facing as
before; no exposure change). deploy_hermes installs the new unit and starts
it if inactive without restarting in-flight browser sessions.
DNS at Porkbun was renamed (chat->ai, jellyfin->jelly) and the qbittorrent,
*arr-stack, comfyui, hermes, dankclaw, and dankserver public records were
deleted. This brings the Caddy config in lockstep: ai/jelly vhosts now match
the new names; comfyui (guarded), hermes (guarded), and the *arr/qb vhosts
are removed as their DNS is gone. The legacy dankclaw/dankserver redirect
vhosts stay: Pi-hole still answers those names on the LAN.
LAN clients now connect directly to 192.168.0.143:8787 (password-gated via
HERMES_WEBUI_PASSWORD); the SSH tunnel via port 22 remains the sole
off-network ingress. Recreation also re-reads the rotated password from
dankclaw/.env via the [Container] EnvironmentFile.
Verified live 2026-08-31: guarded path serves the Anubis JS challenge at
HTTP 200 to non-browser API clients (status-only consumers silently get
HTML); direct published port 8081 from pass_ip sources returns real JSON.
Decision: no second agent-only instance — same engines and same
rate-sensitive VPN exit IP would see double request volume for zero gain.
Server-side implementation landed on Dankclaw: dedicated 'tunnel' account
(nologin shell, password-locked, key-only auth), sshd drop-in
99-tunnel.conf scoping AllowTcpForwarding to that user with PermitTTY off,
and permitopen="localhost:8787" key-option guidance so a leaked tunnel key
can only forward to the webui port. WebUI stays loopback-only published.
Token exchange chain verified end-to-end in-process on the live provider
(pk 9): authorize -> ORM-minted code -> form-encoded token exchange with
Basic client auth -> userinfo. Key findings documented:
- Application row + grant_types must exist before authorize can complete
- token endpoint consumes form-encoded bodies only (JSON body => no
grant_type parsed -> unsupported_grant_type); client auth is Basic-only
- PropertyMappingEvaluator.evaluate(*args,**kwargs) forwards kwargs to
BaseEvaluator.evaluate(expression_source), so context kwargs become the
expression source and every evaluation raises a swallowed exception;
userinfo therefore answers sub-only claims for every provider in this
fork. Scope granting filter (provider= M2M descriptor) is always empty.
- Grafana identity = sub hash until the evaluator bug is fixed upstream.
The embedded outpost carries no provider token (token=None), so the
auth_proxy identity header was never set and no user could log in. RP
mode is topology-independent: challenge gate stays in front, login runs
inside Grafana via the OAuth2 code flow against auth.franksdank.cloud.
Provider 'grafana' (pk 9, confidential) created via ak shell; secret
written straight into the live .env (never in git). Auth URL uses the
global /application/o/authorize/ path — this fork resolves the provider
by client_id query param; per-slug authorize does not exist.
Caddyfile comment corrected: no identity header is forwarded to Grafana.
Grafana leaves the exempt-by-construction set: grafana.franksdank.cloud
now terminates at the Anubis guard (anubis:8923) and authenticated
traffic returns at the internal :8017 listener, where an @grafana Host
handler proxies through to grafana:3000 with the outpost-proxy trust
token (GRAFANA_PROXY_TOKEN = client_id_client_secret of the new
'grafana-proxy' ProxyProvider on the embedded outpost; created via
ak shell before this route flip — provider-first ordering per the
External Exposure Policy).
The Grafana admin account is intentionally kept: auth.proxy in
auto-upsert mode only creates non-admin users, so the local admin
remains the only way to administer Grafana. The old edge-basicauth
wrapper is retired (shared family password, no per-user audit trail).
- Caddyfile: grafana vhost flipped to the guarded chain; @grafana
handler added at the :8017 return listener.
- compose.yaml: GF_AUTH_PROXY_ENABLED/HEADERS/TOKEN on the grafana
service (auth_proxy trusted-proxy CIDRs left at the image default —
the proxying Caddy container is not statically addressable on
10.89.0.0/24, so Grafana trusts any in-network source and relies on
the token as the capability).
- .env.example: GRAFANA_PROXY_TOKEN placeholder so test env_consistency
keeps guarding the compose<->env contract.
Every Caddy-proxied service is now either behind the auth chain or
explicitly internal-only, with the provider-first flip procedure codified.
The comfyui GPU-access seam test now asserts the deployed contract
(group_add keep-groups + udev MODE 0666 on kfd/renderD*) instead of the
pre-fix 'video' group assumption.
Loki's public vhost had zero auth — any host that can resolve DNS could
query all logs via LogQL. The vhost now terminates at anubis:8923 and
authenticated traffic returns at the internal :8017 listener, where a new
@loki host handler routes to loki:3100. The outpost proxy provider
'loki-proxy' (external_host https://loki.franksdank.cloud/, internal_host
http://caddy:8017) was created before this deploy — provider-first
ordering so authenticated traffic never hits the outpost error page.
Ingestion is unaffected: Alloy pushes to loki:3100 on the compose network.
The flip hunk's first comment line was written with a literal ':' prefix
(edit-tool auto-prefix artifact). Caddyfile parses ':# ...' as a matcher token,
which made the following site block an unrecognized directive and failed
config validation — that is why the earlier deploy validated, failed, and
skipped the reload, leaving the live config pre-flip (comfyui still direct).
The first revision guarded the kfd line with SUBSYSTEM=="misc", which never
matches (misc-class devices carry no subsystem for udev matching), so the rule
was read but never fired and /dev/kfd stayed 0660 root:render. Name-only match
applies 0666 on creation and on trigger.
Also fixes test-changed.sh to exit 0 when changed paths map to zero tests
(udev rules are mapped-but-untestable) instead of invoking bats with an empty
selection.
The comfyui container is spawned by a systemd user unit whose spawn context
carries no supplementary groups, so render-group access to /dev/kfd and
renderD128 was structurally unreachable (keep-groups preserved an empty set;
named group_add 'render' fails name lookup inside the userns). World-mode on
the compute nodes is the only spawn-context-independent grant; display cards
keep their modes. Apply: install to /etc/udev/rules.d, udevadm control
--reload-rules, then trigger by subsystem to re-apply modes to live nodes.
Also maps dankclaw/udev/* in the focused-test selector (not bats-testable).
Rootless podman cannot resolve the literal group name 'render' inside the
container user namespace and fails the start with 'unable to find group
render'. keep-groups preserves the launching user's real supplementary
groups (dank is in render on this host), which is the canonical rootless
ROCm device-access recipe.
/dev/kfd and /dev/dri/renderD128 are root:render 0660 on this host. The
container passed the device nodes but only joined group video(44), so torch's
HIP init failed with 'No HIP GPUs are available' and main.py never bound
:8188 — every proxy hop to the app returned 502. Group membership is a
create-time property, so this deploy recreates the container.
comfyui.franksdank.cloud was exempt (direct reverse_proxy to comfyui:8188).
The comfyui-proxy proxy provider was created on the embedded outpost first
(internal_host http://caddy:8017, flows copied from hermes-proxy), then the
vhost flips onto the guard chain and authenticated traffic routes back via a
@comfy host match. ComfyUI has no auth of its own; this closes the last
unguarded admin surface on the internal hosts.
The quadlet generator space-splits HealthCmd into the healthcheck argv, so the
quoted /bin/sh -c form arrived mangled and failed with exit 1 forever (FailingStreak
569) while the DB itself accepted connections. Bare argv survives both exec forms.
Env file is the source of truth (unit ExecStart runs the serve script which
sources it); serve-script default and the deployer health-gate URL flipped in
lockstep so the post-restart health gate polls the new port. Docs, fleet table,
and both index bullets updated to match.
Prior docs claimed unknown tags silently serve llama3.2:1b. Live probes on the
live endpoint disprove it: /v1/chat/completions answers HTTP 200 with a flat
{"error":"Failed to load <tag> model!"} body and no choices; /v1/completions
answers a nested OpenAI-shaped server_error (code 500) under HTTP 200. Valid
tags route per-request to the named model (response echoes the requested tag);
no restart needed to switch between registry tags. Corrected in the runbook,
the AGENTS.md runtime section, the serve-script header comment, and the
historical FastFlowLM incident note.
frank-default-sightsound (FLM v1.0.1, XDNA2 NPU, port 52626) is live on
Dankclaw. Documents the first-deploy crash-loop root cause: iommu=off on the
kernel cmdline breaks the XDNA2 SVA bind (amdxdna_drm_open ret -19), so every
/dev/accel/accel0 open fails with ENODEV; fixed with iommu=pt + reboot.
Whole-registry single-port serving with per-request routing by exact FLM tag;
unknown tags fall back to llama3.2:1b (no 404). Health = GET /v1/version.
Vision + audio towers proven live (image_url color probe; audio control-pair
token-count probe). Updates AGENTS.md fleet table + runtime section, both
index pages, and the change log.
The v1 quadlet generator skips the whole unit when [Container] holds an
unsupported key; the stale key was a leftover of a draft edit. Network=host is
the only host-networking key this generator version accepts.
The v1 generator on this host skips a unit whose [Container] section holds an
unknown key, so the deploy-time regeneration dropped hermes.service entirely
while the container kept running unmanaged. Network=host generates
--network host verbatim (v1 passes extension-less values through).
The gateway unit ran on dankclaw_default and could not reach the host
daemon: from that network only DNAT'd published ports accept connections to
the LAN address. The compose era ran the gateway with network_mode: host, so
the daemon was reachable; the unit silently changed topology at takeover and
broke the adapter. Host networking + a loopback dial removes the dependency
entirely — the gateway is an outbound-only client and publishes nothing.
The trio ran as bare compose containers and died silently at the Aug 26
reboot — nothing restarts compose containers at boot. Each now has a
boot-persistent unit (WantedBy=default.target; the v1 quadlet generator on
this host cannot enable generated units). The deploy handler retires the
compose-era containers only while the units are not yet active, then
(re)starts each unit.
Units run the image under a localhost/hermes-agent:latest alias tag (the
quadlet create path on this host cannot resolve registry-prefixed refs from
local storage); the deploy handler tags first. The gateway's :8090 publish
is dropped — nothing consumes it; the public vhost now routes to the
dashboard's login-gated UI on 9119.
The gateway's SIGNAL_HTTP_URL pointed at the host LAN IP. From container
space on netavark networks only DNAT'd published ports accept connections
to that address; :8080 has no forward, so every connection was refused
and the Signal adapter stayed disconnected since the Aug 26 reboot.
host.containers.internal is the verified working path (probed live from a
probe container on dankclaw_default).
The public hostname also routed to hermes:8090, which serves nothing —
the gateway has no HTTP listener. It now routes to the dashboard's login-
gated admin UI on :9119, behind the existing Anubis+Authentik guard.
The first deploy attempt pulled c3ed3a6 on the target but the remote runner
rejected the explicit --service selection because service_is_allowed() keeps
its own per-target allowlist; add the new LLM service to it.
New Dankclaw service frank-default-sightsound: FLM 1.0.1 serving the
Gemma4-E4B-IT-NPU2 pack (vision+audio+reasoning, 131072 ctx) on the XDNA2
NPU at port 52626. Model choice rationale: gemma4-it is the only NPU2 model
family with BOTH image and audio towers; the uncensored candidate qwythos:9b
is text-only, so sight+sound wins over uncensored here.
Deployer wiring: planner maps env/unit/serve files to the new service name;
deploy-remote gains deploy_llm_sightsound (relink + daemon-reload + enable +
restart + health gate on /v1/version — FLM has no /health route). Unlike the
retired frank-default-npu relink-only path, this unit is boot-enabled.
- kingjones-strix-rocmfp4 reactivated as the ACTIVE large runbook:
runtime build recipe (WEBUI=OFF gotcha), deployed serving profile,
fa-on-works-unpatched evidence, MTP heads documented unloadable,
swap procedure + 2026-08-30 validation record
- agentionai-fast-imatrix marked RETIRED (rollback record kept)
- freetoken-frank-default-small: model/lineage/template sections now
Huihui abliterated FP8; new operational quirk: chat API 422 without
explicit model field
- LLM index + AGENTS fleet table/runtime sections current; both
services' history chains fixed (small was still Tiel in AGENTS.md)
- worklog entry added; models.json contextWindow for large corrected
to the real per-slot 131072 (llama clamps n_ctx_train/parallel)
- same qwen3_5_moe arch and native block-fp8 format (e4m3, 128x128
weight_block_size, un-fused per-expert layout) as the official
Qwen FP8 checkpoint: identical FreeToken loader path, pool numbers
unchanged (2048 MoE slots / 453005 KV tokens / 4 requests)
- checkpoint downloaded to /home/dank/models/huihui-qwen36-35b-a3b-abliterated-fp8
(26 shards, ~35G); official checkpoint kept for rollback
- ships froggeric fixed Qwen3.6 chat template, read from checkpoint
- also refresh both .service Descriptions (large still said STRIX_LEAN)
- runtime: charlie build-qwen38 -> kingjones30/ROCmFPX (card recipe
build, LLAMA_BUILD_WEBUI=OFF since npm/HF-bucket provisioning fails;
all 35 of our serving flags verified present in the tree)
- model: agentionai FAST imatrix -> kingjones777 STRIX single GGUF
(on-host copy sha256 == HF LFS oid c6770d74...49fa verified)
- args unchanged from our validated profile (401408x2, b4096/ub512,
t16/tb16, f16 KV, fa on, fit off); mmap stays default per card
(PLE streams via page cache; --no-mmap = silent cgroup OOM)
Meta-header routing and forward_auth removed from the described design — both
disproved live. Documents: Host-based provider routing with per-guarded-host
proxy providers (lockstep rule), the null-flows wedge that starves the
embedded outpost config fetch, the Anubis /outpost.goauthentik.io/* transport
exemption, the single-file bind-mount stale-policy inode gotcha, Invalid next
URL as a dead design, and the public-DNS record requirement per guarded host.
Login/PoW/POST traffic to the outpost URL space cannot solve a fresh PoW
challenge per request; exempt the transport prefix (authentik's own session
gate still protects it) so the guarded-host sign-in journey completes.
The redirect-to-flow design with an absolute cross-host next= was rejected by
the flow's open-redirect guard at completion ("Request has been denied:
Invalid next URL"), and meta-header routing does not exist in this build
(providers route on the ORIGINAL Host). :8016 now proxies every request whole
to authentik-server:9000 preserving Host; the embedded outpost serves login
in place and forwards authenticated traffic back to :8017. Requires one proxy
provider per guarded host (provisioned live: search/hermes/prometheus/romm/
files + www, flows backfilled). Tests flipped from the meta-header claim to
the upheld contract: Host preserved, no absolute next= redirects.
fa on aborted graph build: build_attn_qsa called ggml_fill on the
F16-converted kq mask (GGML_ASSERT F32 at ggml.c:5371). Patch mirrors the
MLA sparse path in llm_graph_context::build_attn (F32 base mask for the
top-k surgery, F16 cast only at the attention call, gated on
cparams.flash_attn). Vendored at scripts/patches/qwen4exp-qsa-f32-mask.patch
(sha256 f2f4a486...); the build-qwen38 binary on Dankclaw is rebuilt with
it and validated live at CTX_SIZE=32768 (health 200, real completion).
- Main model (v2 joined PLE layout, sha256 1e4160ca…) loads and serves on
build-qwen38 only with --flash-attn off: the QSA indexer path calls
ggml_fill on the F16-cast kq mask and aborts graph build (GGML_ASSERT
F32 at ggml.c:5371). Verified live: model loaded, health 200, real
completion.
- MTP draft head (agentionai FAST, 2.28 GiB) does not load on this binary:
qwen4exp loader requires full per-layer blk.<il>.hc_attn_norm.weight,
head carries only blk.48. Third confirmation (EasiiX Q8_0 + agentionai
FAST vs kingjones STRIX_LEAN previously, now own-matched pair). Spec
decoding runs ngram-mod only; head retained for a future compatible build.
- FLASH_ATTN=off and SPEC_TYPE=ngram-mod wired through env + serve script;
runbook and AGENTS.md updated with the verified constraints.
Model: agentionai/Qwen3.8-Flash-Next-ROCmFP4-FAST-imatrix-GGUF (87.06 GiB,
Q4_0_ROCMFP4_FAST imatrix-calibrated, per-head PLE table inside the GGUF —
fully VRAM-resident). Draft: agentionai/Qwen3.8-Flash-Next-MTP-ROCmFP4-FAST
(2.28 GiB, matched ROCmFP4-FAST quant, ~66% acceptance at n-max 3).
Vision tower (f16) loaded via mmproj/.
Runtime: charlie12345/ROCmFPX build-qwen38 (HIP, ROCm 7.2.4, gfx1151) —
the only binary with the ROCmFP4 FAST types. Replaces the jcbtc CIRU v1.1
deployment (ciru-ai runtime).
Profile: card sampling table (temp 0.7, top_p 0.8, top_k 20, min_p 0.0,
presence 1.5) mixed with the validated build-qwen38 serving profile:
262144 ctx x 1 slot, b 4096 / ub 4096, t 16 / tb 16, f16 KV, fa on,
cont-batching, prompt cache + 8 GiB cache-ram + idle slots, 16 ctx
checkpoints (no --checkpoint-min-step — unsupported on this binary),
--fit off (auto-fit segfaults on qwen4exp hyper-connection config),
--spec-type draft-mtp,ngram-mod with n-max 3 (adaptive p-split measured
worse per author card), p-min 0.75. Qwen-Sharp chat template kept;
reasoning stays in DeepSeek format.
The 2026-08-28 failure of this draft head on the kingjones STRIX_LEAN
model was a cross-vendor pairing problem: the head carries
blk.48.hc_attn_norm.weight (verified in-GGUF, same shape as the working
CIRU draft). Docs: new runbook agentionai-fast-imatrix, index updates.
Runtime: ciru-ai/Qwen3.8-Flash-CIRU-STRIX-IU4 llama.cpp at tag v1.1, built
with the vendor build-linux-amd.sh flag set plus GGML_CUDA_FA +
GGML_CUDA_FA_ALL_QUANTS (the Q4_1_MOE_FORCE_J path is live in the hipified
mmq kernels). Model: jcbtc/Qwen3.8-Flash-CIRU-STRIX-IU4 (Q4_1 IU4 experts,
MTP Q8_0 draft, CIRUPLE1 PLE sidecar) with checksums verified at download.
Vendor production profile verbatim on first deployment: 262144 ctx x 1 slot,
b 2048 / ub 512, t 8 / tb 8, f16 KV, fa on, cont-batching, prompt cache +
8 GiB cache-ram + idle slots, 32 ctx checkpoints (min step 8192), MTP depth
3 with p-split 0.10, PLE pager env knobs (Q41_MOE_FORCE_J=32,
PLE_WORKERS=16, PLE_STRICT_SHA=0, HIPBLASLT=1). Qwen-Sharp chat template
kept deliberately; reasoning stays in DeepSeek format.
This 2026.x build does not mount /outpost.goauthentik.io/* for proxy-provider
meta-header routing (probed live: ping answers 204, app paths with full meta
headers 404). Guarded vhosts now redirect to /if/flow/default-authentication-flow/
with next= back to the app — unauthenticated sees the login page in place,
authenticated sessions complete instantly and bounce straight to the app.
The proxy provider matches external_host exactly, so Host=www matched zero
providers and Authentik answered its error page while the naked domain
served the homepage. Flip the topology: www.franksdank.cloud is the one
guarded vhost carrying the homepage through Anubis -> Authentik -> :8017;
franksdank.cloud (plus legacy names and the bare-IP catch-all) 308-redirects
onto www before any challenge. Tests flipped to assert the new contract.
All four arr roots live: /tv, /movies via API; /music, /books created after
v1 validator layers peeled back (quality/metadata profile ids must reference
existing profiles; v1 has no /api/v1/profile endpoint). Auth pitfall: the
literal AuthenticationRequired=Disabled string throws DryIoc.ScopedItemException
in v1 AuthorizationMiddleware on every request — the valid no-auth state is
ABSENCE of the key. Fixed live on four containers, verified boot clean.
Sonarr/Radarr root folders created via API (201). Lidarr/Readarr deferred to
first-run wizard (v1 validators demand quality-profile ids that don't exist
yet). Root cause of the write failures: LSIO PUID=1000 under rootless podman
maps to host uid 101000 — a phantom user owning nothing; PUID=0 maps to the
host user that owns every media dir.
Under rootless podman, LSIO's PUID=1000 maps to host uid 101000 — a phantom
user owning nothing: Sonarr's own validator rejected /tv because the process
could not write it. Container-root under a single-user userns maps to the host
user (dank), which owns every /mnt/entropy media dir and /opt config dir.
Identity is now mapping-independent instead of depending on subuid offsets.
calibre-web untouched (working; churn avoided).
Verified live: indexers->Prowlarr->per-arr QBittorrent download clients with
category routing. Readarr's client parked (qB proxy auth defect in abandoned
upstream). EZTV/1337x parked with reason (CloudFlare-blocked at this egress);
Nyaa/sukebei/RuTor/Byrutor/YTS passed live tests and are enabled. Prowlarr v2
Applications rows now carry real arr baseUrl+apiKeys; before this every sync
attempt threw NullConfig in GetIndexerMappings.
Verified live: AuthenticationMethod=None + AuthenticationRequired=Enabled on
all five arrs; unauthenticated API probes return 401 both locally and through
the vhosts. Also records the as-built divergences (entropy mounts, LA timezone,
readarr tag pin, prowlarr mount fix) that supersede the design compose.
Kestrel rejects the h2c upgrade preface; Caddy surfaced that as a 502 per
request while the backend answered 200 on plain HTTP/1.1 from Caddy's network.
Contrast recorded: Plex PMS needs the TLS transport, Jellyfin needs none.
Jellyfin's Kestrel backend rejects the h2c upgrade preface, so Caddy turned
every request into a 502 while the same port answers 200 over plain HTTP/1.1
from Caddy's network. Default transport (HTTP/1.1) restores the vhost.
The page claimed an env-var interface (COPAY_PASS/COPAY_MODE) and a /data
volume convention that this image does not implement: the ac edition consumes
accounts and volumes as argv flags (-a USER:PASS, -v SRC:DST:PERMS), serves
the LUKS array at /w, and answers HTTP 500 with its own error page when zero
volumes are configured. Image ref corrected to copyparty-ac (bare latest tag
does not exist on ghcr).
Ground truth from the image itself: python3 -m copyparty --help-accounts shows
-a takes USER:PASS and -v takes SRC:DST:PERMS:FLAGS — the COPAY_PASS/COPAY_MODE
envs this compose carried are not consumed by this image at all (zero COPAY_
references in the package). With no accounts and no volumes configured, the ac
edition serves its cwd read/write for everyone: the stack was running an
unauthenticated world-writable share of the 91T array. Now: admin account dank
with the A flag on volume /w; anonymous gets nothing.
The ac edition's entrypoint (cpp.sh -> python3 -m copyparty) consumes config
via /z/initcfg, not argv; -v is the volume flag and ate the next arg,
crash-looping argparse. The image mounts everything under /cfg/copyparty as
volumes by marker-file convention — plant the marker on the host instead.
The image entrypoint is /bin/ash /z/cpp.sh -c /z/initcfg: copyparty flags are
passed through command, and the ac edition mounts volumes with -m, not via a
COPAY_VOL env. Without any volume mounted every request resolved zero volumes
and the app answered HTTP 500 with its own error page. Verbose flag included
for the bring-up; drop it once stable.
The ac edition maps volumes from COPAY_* envs; without a volume root every
request resolves zero volumes and the app answers HTTP 500 with its own error
page. Point it at /w where the LUKS array is mounted.
ghcr.io/9001/copyparty:latest does not exist — ghcr answers 403 on the token
request for nonexistent repos, which is why the bring-up failed at pull. The
upstream publishes ghcr.io/9001/copyparty-ac (ac = full edition) and serves
whatever is mounted at /w, so the compose mounts the LUKS array there; config
persists in /opt/copyparty/cfg per upstream layout.
Jellyfin ran as user 1000:1000 inside a rootless user namespace, where host
uid 1000 maps to container uid 0 — the pinned uid is unmapped, so /config/log
creation EACCES'd and the service SIGSEGV'd on boot. Container root (= host
dank) is correct under rootless podman; group_add 992 (render) still maps via
the userns group map, so GPU access is preserved.
Readarr: lscr.io/linuxserver/readarr:latest does not exist — LSIO ships only
arch-prefixed nightly/version tags for readarr (verified against the ghcr tag
API). Pin amd64-1.0.980, the newest stable tag.
The compose stack mounted /mnt/media/{movies,shows,tv,music,books,downloads}
— a prefix that has never existed on this host. The 91T LUKS array mounts at
/mnt/entropy and its libraries are case-exact (Movies, TVShows, Music,
eBooks). Also restores prowlarr's config mount (eaten by an off-by-one in the
prior edit) and fixes TZ New_York->Los_Angeles on the arr stack.
Live-verified 2026-08-29: plexmediaserver.service is an enabled native
systemd unit on Dankserver (not a quadlet; updates via apt, not deploy.sh),
and tautulli runs as a snap with no auth wall on 0.0.0.0:8181. The hunt list
named both; neither had a page. Index linked per the docs-completeness seam.
The policy must ALLOW Anubis's own endpoints before the challenge-all rule
or the challenge challenges itself: browsers POST-solve via make-challenge
then GET-navigate pass-challenge, and a challenged answer renders as the
browser's own fetch-failure error page. Non-JS clients never reach the
verify step, which is why every curl probe looked green while every human
hit a wall. Also records that v1.13.0's valid action verbs are
ALLOW/CHALLENGE/DENY (ACCEPT crash-looped the unit on first deploy).
The pinned binary's policy parser rejects unknown actions and exits 1 on
startup (verified against the image binary: the valid action literals are
ALLOW, CHALLENGE, DENY). The first cut of the own-endpoints accept rule
used ACCEPT and crash-looped the unit on deploy.
The catch-all path_regex "." policy challenged everything, including
Anubis's own API namespace /.within.website/x/cmd/anubis/*
(make-challenge, pass-challenge, static assets). In whole-request proxy
mode a JS-capable browser POSTs its challenge solution to
pass-challenge and received another challenge page (HTML) instead of the
verification verdict (JSON): the challenge app's fetch failed and the
browser rendered its own error — 'the request filters did not return an
alternative response'. Non-JS clients never POST the solution, so every
curl probe looked green while every human hit a wall.
Accept rule ordered above the catch-all; policy is volume-mounted ro from
this checkout, so the deploy updates it in place.
The forward_auth target /outpost.goauthentik.io/auth/caddy does not exist
in authentik 2026.8 (verified live: 404 for every header combination; the
URL tree has no outpost auth prefix at all — the outpost worker dials back
to core over ws/outpost/<uuid>/ instead). Unauthenticated subrequests got
404, Caddy treated >=400 as 'not authenticated' and redirected browsers to
the dead path — users saw authentik's own Not-Found page.
Fix: guarded vhosts now proxy the whole request through authentik-server:9000
with X-authentik-meta-outpost/-provider/-app headers (the exact header set
this build's Traefik controller emits, adapted for Caddy). Unauthenticated
clients get the login flow served in place; once a session exists,
authentik proxies on to the provider's internal_host — now
http://caddy:8017, a new post-auth internal listener that routes straight to
the app. The chain terminates; nothing dials back into authentik.
Live DB changes applied alongside (postgres is the source of truth):
provider pk=1 authentication_flow=default-authentication-flow (was NULL —
the actual cause of the 404s) and internal_host=http://caddy:8017 (was
self-referential http://authentik-server:9000).
The quadlet-dir allowlist silently dropped every change under
dankclaw/authentik/ — the new outpost unit (and its siblings) now route
by exact unit name, mirroring the searxng/gitea pattern. Adds
authentik-outpost to the service inventories; it is a quadlet unit, not
a compose service, so it stays out of the compose set.
In authentik 2026.8.x the proxy-provider URL space (/outpost.goauthentik.io/*)
is served by the Rust worker binary, not the server — verified live: with a
tenant-matching Host header, /if/flow/... answers 200 on the server while
/auth/caddy still 404s. The outpost instance existed in the DB but no process
ever served it, so Caddy's forward_auth subrequests hit a 404 and browsers got
authentik's own Not-Found page instead of the authentication flow.
The new unit runs the same image as server/worker with Exec=worker and flips
into outpost mode via AUTHENTIK_HOST/AUTHENTIK_INSECURE/AUTHENTIK_TOKEN (the
token is read from the DB into the host .env at deploy time; it never enters
the repo). It dials back to core over websocket and serves forward-auth on its
own 9000. Deploy handler restarts only this unit — the rest of the stack is
untouched.
Replaces the pre-Anubis direct-routing diagram with the live chain:
Caddy TLS -> Anubis :8923 verify -> accepted traffic back to Caddy's
internal listener :8016, where per-vhost routing resumes and Authentik's
embedded outpost gates wrapped apps via forward_auth. Guarded set
(homepage, search, hermes, prometheus, romm, files) and the
exempt-by-construction set (auth, plex, gitea git paths, grafana) are
listed explicitly. Access Policy bullet updated: basic auth was replaced
by the challenge-then-forward_auth layering.
Three compounding deploy-handler bugs found by live deployment:
- podman image exists/pull used a mixed-case ghcr.io ref that podman's
reference parser rejects before any network I/O (repository name must be
lowercase); all refs now lowercase — ghcr resolves case-insensitively.
- the unit runs the image under a localhost/ alias tag: this host's podman
5.7 quadlet path fails container create with 'image not known' for
registry-prefixed refs even when present in storage (verified live).
- the health probe was shell-form and the distroless ko-build image has no
shell; HealthCmd is now a JSON exec-form array, which passes through the
quadlet generator's argv assembly verbatim.
Tests updated to assert the alias-tag reality plus the pull guard.
The image has no shell (ko-build distroless; entrypoint /ko-app/anubis),
and this host's quadlet generator joins a bare HealthCmd into one
\x20-escaped argv element that podman then runs shell-form — exec of a
missing shell exits 1 and the container sits unhealthy. A JSON array in
HealthCmd passes through the generator verbatim and executes the binary
directly.
The v1.13.0 image is distroless: no shell, no HEALTHCHECK in the image
config (verified live: Healthcheck config is null on this host's image),
so quadlet inherits nothing. Declare the probe explicitly in exec form;
'anubis --healthcheck' exits 0 when the challenge server answers.
On this host's podman 5.7 quadlet path, a registry-prefixed Image ref
fails to resolve from local storage at container create time even when
the image is present ('image not known'); the identical argv works from
a shell or transient unit. Verified live: an identical unit with
Image=localhost/techarohq/anubis:v1.13.0 starts immediately. The deploy
handler now tags the alias after the guarded pull; upgrades re-tag
after pulling the new tag.
image inspect on an absent reference can exit non-zero for reasons
unrelated to existence in a non-tty deploy context; image exists is
the boolean primitive and cannot misreport. Absent-image behavior
(pull from ghcr) is unchanged.
podman's reference parser rejects mixed-case repository refs before any
network I/O ('repository name must be uppercase'); upstream's own install
docs pull lowercase (ghcr resolves case-insensitively). The first deploy
of this chain exit-2'd on that parse error — the image was not yet
pulled, so the guarded vhosts 502 until this deploy lands.
The first deploy run dispatched caddy (hot reload, no restart) but exit-2'd
on the inventory guard before deploy_anubis could run: deploy.sh's
valid_service was extended but the remote runner's per-target inventory
was not. Guarded vhosts 502 until this lands.
Bots and AI scrapers hit the SSO-fronted vhosts and eat Authentik auth
work for 403s. Anubis verifies clients first: DENY for known AI scraper
UAs, JS challenge for everyone else; verified traffic resumes the normal
Authentik forward_auth chain on Caddy's internal listener (:8016).
Chain: client -> Caddy (TLS) -> Anubis :8923 (verify) -> Caddy :8016
(auth + routing resumes; Host survives both hops) -> app. Reverse-proxy
through Anubis, not forward_auth: at the pinned tag a CHALLENGE answers
HTTP 200 + HTML, and forward_auth treats 200 as "authenticated" — the
page would pass through unchallenged. Exempt by construction (never
routed through Anubis): plex (plain HTTP legacy), auth.* (IdP login
must be reachable ungated), gitea (git smart-HTTP clients cannot solve
JS challenges), and every app with its own login.
Pinned ghcr.io/TecharoHQ/anubis:v1.13.0 — the tag was source-verified
before pinning: at this tag the policy parser is JSON-only with a
minimal bots[] schema (no imports, no CEL expressions), flags are the
config surface via flagenv (TARGET/POLICY_FNAME map to --target/
--policy-fname), and the image's exec-form HEALTHCHECK is inherited by
the quadlet unit (dodges the v1 generator's HealthCmd space-split quirk).
Rollback nuance: stopping anubis 502s everything behind the chain (the
public listeners proxy through it); a true rollback is reverting the
Caddyfile and redeploying, not just systemctl --user stop anubis.
Verified live on dankclaw: the podman 5.x v1 quadlet generator splits
HealthCmd values on spaces into the container healthcheck argv array, so
quoted probes (sh -c "pg_isready ...") arrive mangled and every probe
exits 127 — postgres read 'unhealthy' with the database actually up.
Probe forms now use only space-free argv tokens: pg_isready passes its
args directly; the server probe runs manage.py check --database default
(verified exit-0 standalone in the image).
Probed the live generator on dankclaw: podman 5.x v1 hard-fails unit
generation on dash-prefixed options (--shm-size killed authentik-server's
unit) and rejects Command= entirely; Exec= is the accepted role selector.
Drop --shm-size (uwsgi uses unix sockets under /run, not /dev/shm; default
64m shm is fine) and drop systemctl --user enable — generated units are
transient and cannot be enabled; WantedBy=default.target in each unit file
is what makes them boot-persistent on this generator.
The compose-era authentik trio died at every boot (rootless podman does not
auto-start bare containers). Convert the trio to Quadlet units installed by
deploy-remote.sh into ~/.config/containers/systemd/, enabled + WantedBy=
default.target:
- 3 unit files under dankclaw/authentik/quadlet/ (postgres -> server -> worker
ordering via Wants=/After=; secrets stay in the git-ignored host .env,
expanded by systemd at start time exactly like the gitea units)
- deploy_authentik() handler: install units, daemon-reload, retire stale
compose-era containers only while units are inactive, enable once, then
start in dependency order with podman-health gates between stages
- compose.yaml: authentik service definitions retired to a comment pointing at
the quadlet dir (same pattern as searxng/open-webui/gitea/caddy)
- tests/authentik.bats rewritten from compose assertions to unit-file
contract assertions (images, Exec roles, ordering, loopback-only publish,
pg_isready probe, secret hygiene, dispatch routing)
All in-repo path references (AGENTS.md, README, old scripts, report artifacts,
test code paths) rewritten in the same commit; focused-test map extended for
.gitignore, old/scripts/*, dankclaw/comfyui/*, and report/* per this repo's
own convention ('add one before committing'). Obsidian wikilinks are
title-based and unaffected.
Both the EasiiX Q8_0 and agentionai ROCmFP4-FAST MTP draft heads fail
with 'missing tensor blk.0.hc_attn_norm.weight' against the STRIX_LEAN
Uncensored model. The qwen4exp binary expects this hyper-connection
attn norm tensor which neither draft GGUF carries. Using ngram-mod
only until a compatible MTP head is found.
The EasiiX Q8_0 draft was incompatible with STRIX_LEAN (missing
blk.0.hc_attn_norm.weight). Switched to agentionai/Qwen3.8-Flash-Next-
MTP-ROCmFP4-FAST-GGUF (2.28 GiB, ROCmFP4-FAST quant matched to target).
n_max 3 per the HF card (n_max 4 measured worse). SPEC_TYPE=draft-mtp,ngram-mod.
The script runs ON Dankclaw but was calling ssh() to SSH back to
dankclaw for service management. Nested SSH fails (no key auth for
localhost). Replaced ssh() with local() using subprocess.run(shell=True)
for systemctl and sed commands.
The env file had the old STRIX paths (directory doesn't exist on disk).
Updated to the actual running STRIX_LEAN (Uncensored) variant:
- MODEL_DIR → kingjones-Qwen3.8-Flash-Next-Uncensored-ROCmFP4-STRIX_LEAN
- MODEL_FILE → 3-shard Q4_0 (00001-of-00003)
- MMPROJ_PATH → mmproj-Qwen3.8-Flash-Next-Uncensored-BF16.gguf
- serve-frank-large.sh: add --mmproj $MMPROJ_PATH
Reconciles the uncommitted local changes that were on Dankclaw.
serve-frank-large.sh: only pass --model-draft when SPEC_TYPE contains
'draft-mtp'. ngram-only configs (ngram-mod, ngram-simple, etc.) do not
need a draft model file — loading one wastes VRAM and may error.
bench-spec-decode.py: new benchmark script. Sweeps 4 configs (none,
ngram-mod, draft-mtp, draft-mtp+ngram-mod), generates 16k tokens per
config at temp 0 from a minimal prompt. Uses HTTP API (llama-bench
cannot target --spec-type). Follows benchmarking policy: stop+mask
the production service, bench, unmask+restart, health check.
podman's quadlet generator creates the default.target.wants symlink from
the unit's [Install] section at every user-manager start (is-enabled
reports 'generated' for all healthy quadlets, pihole included). An
explicit 'systemctl --user enable caddy.service' fails with 'transient or
generated' and aborted the deploy after the unit was already installed.
systemd-sysctl applies /etc/sysctl.conf AFTER all of sysctl.d/, so the
stale ip_unprivileged_port_start=80 line there beat every drop-in and kept
the runtime value at 80 after the first deploy pass. deploy_pihole now
comments out an active line in /etc/sysctl.conf (idempotent) and generalizes
the sysctl.d conflict scan to any file setting a value above 53.
Caddy left compose.yaml for a quadlet unit (caddy.service): joins
dankclaw_default for container-name upstreams, Pull=never for the local-only
image, Porkbun credentials via /home/dank/Containers/caddy/env written by
deploy from .env, bare-IP-catch-all health check. deploy_caddy installs,
enables, and starts the unit after retiring the bare container (state is in
host volumes).
Pi-hole: the repo now owns the sysctl drop-in
(dankclaw/pihole/sysctl/99-dnsmasq.conf, ip_unprivileged_port_start=53);
deploy_pihole installs it to /etc/sysctl.d/, stages conflicting later-sorted
drop-ins in the home trash, and applies 'sysctl --system' (the boot code
path). This fixes the 2026-08-26 crash loop (port-53 bind denied, restart
counter 1645) caused by a second drop-in resetting the value to 80.
Tests: new caddy_quadlet.bats seam; caddyfile/deployment_flow assertions
repointed from the removed compose service to the quadlet; pihole.bats pins
the sysctl drop-in and its deployment; test-changed.sh routes dankclaw/caddy/
and the pihole subtree to their seams. Docs: Caddy, Pi-hole, Dankclaw Stack,
work log.
A bare compose container dies silently at reboot (rootless Podman has no
boot-time auto-start; 'restart: unless-stopped' does not survive shutdown's
explicit stop). Rule established by the 2026-08-26 Dankclaw reboot, which
left Caddy down and Pi-hole crash-looping on a sysctl conflict.
--ctx-size is the total pool divided among --parallel slots, so with
PARALLEL=3 the 160k target must be a 491520 pool (163840/slot) to
actually deliver 160k context per slot. Env, serve default, bats updated.
Reconfigure to 3 parallel 160000-ctx slots (was 1 slot). KV cache is
now asymmetric: K at q8_0, V at q5_0, both with Hadamard. The 3 slots
triple the KV, so expert blocks 27-34 are pinned back to CPU (freed
~8.7 GiB) leaving output-adjacent 35-40 on CUDA. Q8_0 model,
split-mode none, spec disabled (ik limits spec to 1 slot). AGENTS.md:
drop the scratch-port rule (test/bench processes cannot share the
single GPU with the running service anyway). Env, serve script, bats,
docs updated.
Switch to CRACK Q8_0 model (37.8 GB, mmproj f16). Servt as a single
160000-ctx slot (PARALLEL=1); speculative decoding stays disabled. KV
cache is q6_0 + Hadamard. iq5_k is not a valid KV-cache type in ik_llama
(cache whitelist excludes iq5_*); q8_KV reproduces the ggml.c:5427 view
assert even at PARALLEL=1/no-spec, so q6_0 is the verified working 6-bit
choice. Env, serve script, tests, and KB updated. Deployment was left
down after the isolation test; this commit redeploys it with the
verified Q8_0/160k/q6_0 config.
Fused MoE (fused_moe / fused_up_gate / fused_mmad) is on by default in
ik_llama - no flag needed; only -no-fmoe disables it. Slide two more
expert blocks (28,29) from CPU onto the RTX 3090 (~1.26 GiB) for the
fused-MoE run, leaving ~1.6 GiB headroom. Update override regex, test,
and KB (including a note that the headroom is now tighter than the
2-3 GiB ideal).
--jinja: ik_llama only renders a custom --chat-template-file as real Jinja
when --jinja is passed (inputs.use_jinja ? apply_jinja : non-jinja
fallback). Without it the Qwen-Sharp template's thinking/image/tool
formatting is silently dropped. Add USE_JINJA=1 -> --jinja.
Experts: the RTX 3090 had ~5.4 GiB headroom; pull expert blocks 30,32,33,34
back to CUDA (each block ~630 MiB) from CPU for lower latency. Update the
OVERRIDE_EXPERT regex, its test, and the KB with the headroom-check note.
ik_llama limits speculative decoding to a single server slot (-np 1) and
hard-exits at load if a spec stage is configured with --parallel > 1:
"Speculative decoding is currently limited to a single server slot
(-np 1)." This service keeps PARALLEL=3 (one slot per user), so set
SPEC_TYPE=none to disable speculation. It is a low-parameter MoE (3B
active of 35B), so throughput is fine without it. MTP stays disabled.
KB and tests updated.
The ik_llama fast q8_KV cache type crashes at model init on this
QWEN3MOE/MTP + multimodal arch at commit b166e269 with a ggml_view size
assert (ggml.c:5427) right after the output buffer is created, looping
the service. Empirically isolated by toggling cache type: q8_0 and f16
load and serve (smoke-tested correct output), q8_KV asserts. Switch to
q8_0 + Hadamard (K/V), which still gives the 8-bit cache + Hadamard
quality win. KB updated with the crash lesson.
--split-mode graph requires >= 2 devices in ik_llama.cpp and errors hard
at model load on a single RTX 3090 ("It is not possible to use split mode
'graph' with less than 2 devices"), looping the service. Set SPLIT_MODE
=none (the default), which works with the --override-tensor CPU expert
offload. KB updated with the two-device requirement so we do not assume
graph degrades gracefully on one GPU.
Switch frank-default-small on Dankserver from upstream llama.cpp (b10121)
to ikawrakow/ik_llama.cpp (b166e269, built with the staged pip cu13
toolkit). The model (Ornith-1.5-35B-A3B CRACK Q6_K) is unchanged and the
old binary is kept for rollback.
Runtime changes:
- KV cache quantized to q8_KV with --k-cache-hadamard/--v-cache-hadamard
- --split-mode graph (single-GPU no-op, retained per operator pref)
- migrate ngram speculation to canonical
--spec-type ngram-mod:n_min=48,n_max=64,ngram_size_n=24
(legacy --spec-ngram-mod-* are rejected in ik main)
- N_GPU_LAYERS 999 (ik -ngl rejects the literal "all"); experts stay
pinned to CPU via --override-tensor
Build: GGML_CUDA_FA_ALL_QUANTS=ON + GGML_IQK_FA_ALL_QUANTS=ON mandatory
for the quantized KV cache; recipe + reasoning baked into
docs/LLM/ikllama-ornith.md. Tests and both indices updated.
Caddy v2.11.4's caddyfile adapter rejects 'tls { insecure_skip_verify }'
inside transport http (unrecognized subdirective '{') and also rejects
the handler-level form. The working form, verified live against PMS with
a throwaway caddy (200 OK), is the bare token:
reverse_proxy 192.168.0.235:32400 {
transport http {
tls_insecure_skip_verify
}
}
PMS 1.43.x on :32400 accepts, reads, and silently closes every non-TLS
connection from a non-loopback source (loopback exempt). The plain-HTTP
reverse_proxy therefore got EOF and Caddy answered an empty-body 502.
Both Plex vhosts now dial the upstream over TLS with
insecure_skip_verify (PMS self-signed cert, pinned LAN hop only;
client-facing Let's Encrypt TLS unchanged).
dankware carries two +DefaultRoute links (enp47s0 eth, wlan0 wifi) that
both receive the same DHCP DNS list. systemd-resolved picked a different
'current' resolver per link, so flat-name resolution flapped between the
LAN IP (192.168.0.143 via wlan0/Pi-hole) and the WAN IP (Comcast v6 via
enp47s0).
Pin 192.168.0.143 as the static IPv4 resolver on both NetworkManager
connections (ipv4.method=auto keeps the DHCP address/gateway/route,
ignore-auto-dns=yes uses only Pi-hole, ignore-auto-routes=no keeps the
DHCP routes), then restart systemd-resolved. resolved.conf's own DNS= is
a per-link fallback, so the link must be pinned.
Flat names now resolve to 192.168.0.143 on every query; apex/unknown
subdomains still fall through to Porkbun (71.227.227.172).
The v6 restore commits dropped --chat-template-file and
--chat-template-kwargs while the artifact check still validated the
template file, so the Sharp template was checked but never applied.
Re-add both flags (with --jinja) so reasoning is emitted in DeepSeek
format and preserve_thinking/preserve_reasoning are honoured.
Reformat the script one-flag-per-line with long flag aliases
(--gpu-layers, --ctx-size, --ubatch-size, --cache-type-k/v) like
serve-frank-small, and update the 2 frank_large_config assertions
that grep the old short-flag forms (-ctk/-ctv, -ub).
- scripts/serve-frank-large.sh
- tests/frank_large_config.bats
The pre-commit hook runs scripts/test-changed.sh, which requires a
focused-test mapping for every staged path. The quadlets were moved into
dankclaw/<svc>/quadlet/ but the mapping still only matched top-level
dankclaw/*.container, so any nested quadlet or the pihole addn-hosts/env
data files fell through to the catch-all 'die' and blocked commit.
Add the nested-quadlet mappings (pihole-specific first, then the
dankclaw/*/quadlet/* fallback) plus the pihole data-file paths.
- Place ik's Trellis (IQ*_KT), IQK (IQ*_K), MXFP4, and dedicated fast KV type
Q8_KV in a single high->low bits-per-value table with ROCmFPX TurboQuant KV.
- Mark mainline-provided types (*), KV-native vs weight-family-reused-for-KV.
- Hadamard annotated as a transform on top of a base quant, not a row.
- bpw read from exact ggml block structs (QK_K=256, fp16 scale): type_size x 8 / QK,
verified against the struct comment "6.5625 bits per weight" and the trait table.
Notes where legacy nominal figures (4.0/4.25) diverge.
- AGENTS.md: add benchmark directive (use llama-bench / vllm bench; ask before
rolling your own).
Fan out subagents to compare mainline llama.cpp, charlie12345/ROCmFPX,
and ik_llama.cpp, then add their findings to the project knowledge base:
- docs/Research/Llama Cpp Runtimes.md (positioning + quantization + server fit)
- docs/Research/Llama Cpp Code Diff.md (code-level quant diff, table, recommendations)
- docs/Research/SGLang.md (what SGLang is, comparison, our-boxes fit)
- docs/Research/index.md (Research progressive-disclosure index)
Cross-linked from docs/index.md (root) and docs/LLM/index.md.
AGENTS.md:
- Add Documentation Authoring section: docs/ is an OKF bundle; point doc authors
at the canonical spec in the system-config knowledge base
(~/.pi/agent/docs/references/okf-spec.md), not the orphaned
docs/references/okf-spec.md; consolidate read-index/one-concept-per-file/
progressive-disclosure/citation rules.
- Map AGENTS.md to docs_completeness.bats in the focused pre-commit hook (it was
last committed before the hook existed and had no mapping).
frank-default-large now serves the kingjones Qwen3.8-27B-ROCmFPX-Q8_0-AGENT
model with a standalone MTP draft head (Q4_0), f16 K + TurboQuant V KV cache,
and the Sharp chat template. Context is maxed to 262144 (one slot); reasoning
is on with preserve_thinking + preserve_reasoning in the chat kwargs.
- frank-default-large.env: Q8_0 AGENT model paths, CTX_SIZE=262144, REASONING=on,
REASONING_FORMAT=none, CHAT_TEMPLATE_KWARGS {enable_thinking, preserve_thinking,
preserve_reasoning}
- serve-frank-large.sh: standalone draft-mtp profile (already switched off the
Kairic/PromptForge path)
- download-q8agent.sh: reproducible HF fetch of the three payload files
- docs/LLM/kingjones-q8agent-rocmfpx.md: Q8_0 AGENT runbook
- AGENTS.md: topology + ACTIVE runtime section + TTM/GTT load note
- qwen3.8-strix-optimizations, Dankclaw Improvements: TTM pages_limit now 114 GiB
(applied 2026-08-25), full 262k KV + Q8 payload fit in one slot
- frank_large_config.bats: assert 262144 context, reasoning + preserve_* enabled
The TTM/GTT was raised to 114 GiB (29,883,536 pages) to hold the full 262k KV.
SearXNG queries all ~50 upstream engines from one Gluetun VPN exit IP, so
rate-sensitive engines (Google, DuckDuckGo) answer with reCAPTCHA/429 and
SearXNG suspends the offending engine for hours-to-days, showing up as
"unavailable". Add docs/Dankclaw/SearXNG Providers.md covering: the
suspension model, the two failure modes (CAPTCHA vs selector drift), engine
config (settings.yml engines: list — duplicate google, timeout,
retry_on_http_error, JSON feed), request tuning (outgoing block, UA rotation,
suspended_times — correcting the old rate_limit/useragents blog myths),
routing/proxies (per-engine proxies, dedicated 2nd Gluetun, dedicated IP), and
verification (/json, /config, check.searx.space). Crosslink into SearXNG.md
and index.md.
| Dankserver | `frank-medium` | mainline llama.cpp master fb27a525 (root systemd unit, gfx1201 HIP build) | ajgazin Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP UD-Q8_K_XL + z-lab Qwen3.8-27B-DFlash2 Q8_0 drafter, 2x Radeon AI PRO R9700 in ONE layer-split process (35:30), 262144 ctx over 2 slots, batch 6144x512, text only | 7770 |
| Dankclaw | `frank-large` | halogen-flash-server 0.5.8 (rootless Quadlet) | peonist-ai Qwen3.8-Flash-Next w4b `.hgn` on the gfx1151 iGPU (Strix Halo LPDDR), 200192 ctx per conversation over one 400384-position KV pool / 2 slots, text only | 7770 |
The GPU budget on Dankclaw is the rocm-smi figure: 64 GiB VRAM reported by
rocm-smi (SVM-backed). The docs' older 96 GiB figure was the addressable
SVM pool. The GPU node reports `local_mem_size 0`, so all allocations are
GTT/SVM-backed. Trust rocm-smi and the runtime. Do not use `free` or
`dmidecode` for the GPU budget. They show the CPU-side carve-out. This is
intentional. Do not change it.
Retired runbooks are under `serverAppConfigs-docs/LLM/`. Retired scripts live
in `old/` and `.AGENTTRASH/`.
## Why
Direct edits on the remote drift from version control. Future agents will not
know what changed. Git is the single source of truth. The two workloads share
two hosts, one Caddy, and one Gitea. One repo keeps them in one history.
> User request → Fetch from remote → Commit here → Edit → Test → Commit → Push → Pull on remote → Deploy + restart
Direct edits on the remote drift from version control, and future agents
won't know what changed. Git is the single source of truth. The two workloads
share two hosts, one Caddy, and one Gitea; one repo keeps them in one
history.
## Summary
| Context | Action |
|---------|--------|
| Editing a compose/Caddyfile/serve script | **Work through git** — fetch, commit, edit, test, push, pull, deploy. Never edit on the remote directly. |
| Config exists only on remote | **Pull it into the repo first** — commit the fetched version (redacted) before making changes. |
| Unstaged local changes | **Commit or discard** — check `git status` before starting work. Ask the user if unsure. |
| Deploying a committed change | **Run `scripts/deploy.sh <target>`** — it pushes, pulls on the target, restarts the stack and the LLM service, and checks both health endpoints. See [docs/LLM/deployment.md](docs/LLM/deployment.md). |
| Benchmarking | **Stop the service first** — scratch port, restart, verify health. See the Benchmarking Policy above. |
| Editing a compose/Caddyfile/serve script | **Through git** — fetch, commit, edit, test, push, pull, deploy. Never edit on the remote. |
| Config exists only on remote | **Pull it into the repo first** — commit the fetched version (redacted) before editing. |
| Unstaged local changes | **Commit or discard** — check `git status` before starting work; ask if unsure. |
Git is the single source of truth. A `deploy.sh` script moves committed
changes to the target hosts over SSH. One run per target covers both
workloads on that host: the podman stack and the LLM service.
## Workflow
```mermaid
graph LR
A[Edit in this repo] --> B[Commit]
B --> C[Push to origin]
C --> D[deploy.sh: pull on target]
D --> E[restart stack]
E --> F[health check]
```
1. Edit files locally in this repository.
2. Commit and push.
3. Run `./scripts/deploy.sh <dankclaw|dankserver>`.
4. The script pulls the repo on the target and restarts the podman stack
and the LLM service.
5. It verifies service health after restart (homepage and LLM endpoint).
## Deploy Script
```
./scripts/deploy.sh dankclaw # pull + restart on Dankclaw
./scripts/deploy.sh dankserver # pull + restart on Dankserver
./scripts/deploy.sh dankclaw --dry-run
```
SSH uses the existing key pair (`dank@dankclaw`, `dank@dankserver`). No passwords are stored or documented in this repo. If sudo is needed, use the interactive sudo on the target — never commit a password.
The deploy script polls the health endpoint of each service after restart. Service healthchecks are defined in each compose file. A service that fails its healthcheck is reported, not hidden.
## Tests
Run the test suite before pushing:
```
./tests/run.sh
```
The suite validates compose YAML, env consistency, Caddyfiles, docs completeness, and script syntax. See [[Testing Policy]].
## See Also
- [[Storage Policy]]
- [[Testing Policy]]
- [LLM deployment docs](../LLM/deployment.md) — the LLM half of the same flow
One Caddy fronts both hosts. It runs as a container on Dankclaw (192.168.0.143) and terminates TLS for every service on the LAN. Pi-hole on Dankclaw answers the service names for LAN clients; Porkbun answers them for the WAN. The G36 router firewall governs external access. SearXNG outbound traffic exits through the Gluetun VPN container.
## Topology
```mermaid
graph TB
subgraph WAN["G36 router 192.168.0.1"]
EXT[Internet]
end
EXT -->|port forwards| C
subgraph Dankclaw[192.168.0.143]
C[Caddy :80 :443]
D[Pi-hole :53]
C -->|comfyui:8188| F[ComfyUI]
C -->|192.168.0.143:3000| G[Gitea quadlet]
C -->|192.168.0.143:8081| SE[SearXNG quadlet]
C -->|192.168.0.143:3005| O[OpenWebUI quadlet]
SE -->|network_mode service| GL[Gluetun VPN]
GL -->|tunnel| EXT
end
subgraph Dankserver[192.168.0.235]
C -->|192.168.0.235:8085| R[RomM]
C -->|192.168.0.235:3923| P[Copyparty]
C -->|192.168.0.235:8083| W[Calibre-Web]
C -->|192.168.0.235:32400| PL[Plex host service]
end
CLIENT[LAN clients] -->|DHCP DNS| D
```
Dankserver runs no proxy of its own. Its services expose published LAN ports, and the Dankclaw Caddy routes the names to those ports.
## Hostnames
| Scope | Convention | Resolved by |
|-------|------------|-------------|
| LAN | `romm.franksdank.cloud` → 192.168.0.143 | Pi-hole (`addn-hosts`) |
| WAN | `romm.franksdank.cloud` → 71.227.227.172 | Porkbun A record |
Users type a name and nothing else. Caddy maps the name to the upstream. The split between the two resolvers removes the need for router NAT loopback. See [[LAN-DNS]] for the flow and the router hand-off.
## TLS
Caddy holds one Let's Encrypt certificate per vhost, issued and renewed through the DNS-01 challenge via the Porkbun module. No inbound port 80/443 is required. Details in [[Caddy]].
## Gluetun Path (SearXNG only)
SearXNG attaches to the Gluetun network namespace (`network_mode: service:gluetun`). Only SearXNG shares the tunnel. Everything else on Dankclaw uses the normal LAN route. This matches KB Pattern 2 in [[Gluetun]].
## Access Policy
- The **G36 router firewall** governs what the WAN can reach. Port forwards are manual router steps; the repo tracks no router state.
- Caddy adds HTTP basic auth on the homepage and on sensitive services.
- LAN subnet: `192.168.0.0/24`.
## See Also
- [[LAN-DNS]] — name resolution, Pi-hole, router hand-off
All persistent data uses **bind mounts**: a host directory mapped directly into the container. No named volumes. This matches the preference of this repo and the existing Gitea stack on Dankclaw.
## Why Bind Mounts
- The exact host location of every byte is visible and controllable.
- Backups are a simple `tar` of known host directories.
- The data survives container recreation and `podman-compose down`.
- Ownership is explicit: everything runs as UID 1000 (`dank`) so host tools can read it.
This repository tests its configuration at defined seams. The tests run with `bats`; run them with `./tests/run.sh` before every push.
## Seams Under Test
| Seam | File | What it asserts |
|------|------|-----------------|
| Compose validity | `tests/compose_validity.bats` | YAML parses, services have images, no named volumes |
| Env consistency | `tests/env_consistency.bats` | every `${VAR}` is defined in `.env.example`; no real secrets; the live gitea password never appears |
Tests are written first. A failing test means the seam is not satisfied; the config change that makes it pass is the only acceptable fix. No `--no-verify`, no deleting tests without approval.
The last two rows are bare-metal systemd services, not containers. This
repo owns their scripts, units, and knowledge base (see
[docs/LLM/index.md](../LLM/index.md)). Current mapping (2026-08-22):
`frank-default-large` runs the CIRU ROCmFPX llama.cpp on Dankclaw;
`frank-default-small` runs the upstream llama.cpp on Dankserver.
## Boot Persistence
The Gitea trio, the SearXNG trio (gluetun, searxng, valkey), and open-webui run as quadlet units, so they start at boot. User linger is on. ComfyUI runs as a systemd unit around a bare process. The remaining services in the table are defined in `compose.yaml` and are not yet boot-persistent.
Deploy the SearXNG trio with `scripts/deploy-searxng-quadlet.sh dankclaw` and open-webui with `scripts/deploy-open-webui-quadlet.sh dankclaw`. The Gitea units install to `~/.config/containers/systemd/` (see [[Gitea]]).
## Routing Rules
- SearXNG outbound traffic exits through the [[Gluetun]] VPN container.
- Every other service uses the normal LAN route.
- The homepage lists all services with status.
## See Also
- [[index]] — repository index
- [[Network Architecture]] — routing and proxy
- [[Observability]] — Grafana, Prometheus, logs
- [LLM fleet knowledge base](../LLM/index.md) — the model-serving endpoints on this host
Gitea is a self-hosted Git service with issues, pull requests, and a web UI. On Dankclaw it runs as three Quadlet-managed containers in the rootless podman session of user `dank`.
| `gitea-caddy` | `gitea-caddy.service` | `docker.io/library/caddy:latest` | Reverse proxy for the public domain |
The containers share the private bridge network `systemd-gitea`. The quadlet network file `gitea.network` enables DNS, so the containers resolve each other by name. The automatic `podman` default network has DNS disabled. Do not run the Gitea containers on it.
User-level linger is enabled for `dank` (`loginctl enable-linger dank`). The units start at host boot without a login session. Each unit carries `[Install] WantedBy=default.target`. The podman generator applies it at boot; no symlink is needed.
The containers use the quadlet fresh-start model. A stop removes the container (`podman run --rm --replace`). The next start recreates it. The bind-mounted data survives.
## Secrets
The database password lives in `/opt/gitea/.env` on the host (mode 0600). The quadlet files reference it with `EnvironmentFile=`. The password never enters this repository. Placeholders live in [`dankclaw/gitea/.env.example`](../../dankclaw/gitea/.env.example).
Do not change the password in place. Change the PostgreSQL role first, then update `/opt/gitea/.env` and the `[database] PASSWD` line in the app.ini (see Gotchas), then restart the stack.
## Configuration files
| File | Location | Repo copy |
|------|----------|-----------|
| Quadlet units | `~/.config/containers/systemd/gitea-{postgres,server,caddy}.container`, `gitea.network` | [`dankclaw/gitea/quadlet/`](../../dankclaw/gitea/quadlet/) |
| Caddyfile | `/opt/gitea/data/caddy/Caddyfile` | host-only (not in repo) |
| `/var/lib/postgresql/data` | `/opt/gitea/data/gitea-postgres` | PostgreSQL data |
| `/etc/caddy/Caddyfile`, `/root/.local/share/caddy` | `/opt/gitea/data/caddy/...` | Caddy config and TLS state |
Back up `/opt/gitea/data/` (tar) or `pg_dump` the database. See the knowledge base page [[Gitea]].
## Gotchas
- **app.ini wins over env vars.** Gitea writes `/data/gitea/conf/app.ini` (host: `/opt/gitea/data/gitea-gitea/gitea/conf/app.ini`) on first start. The `[database] HOST` and `PASSWD` lines persist and override the `DATABASE_HOST`/`DATABASE_PASSWD` env vars. When you rename the database container, edit the `HOST` line.
- **The files in the data directory are owned by host UID 100999**, not `dank` (rootless ID mapping: container uid 1000 → host 100999). Edit them inside the container with `podman exec --user 1000 gitea-server ...`.
- **The host auto-reboots.** `unattended-upgrades` reboots Dankclaw when a package needs it. The quadlet units make Gitea survive that. See the 2026-08-16 incident below.
- **Port 3000 vs 3001.** The live host port is 3000. The full-stack [`dankclaw/compose.yaml`](../../dankclaw/compose.yaml) maps 3001. Align the two before deploying the full stack.
- **Legacy compose retired.** `/opt/gitea/podman-compose.yml` is retired to `/opt/gitea/.deprecated/`. Do not run it. Its containers collide with the quadlet units on ports 3000, 2222, 80, and 443.
- **Rootless cannot bind ports below 1024 by default.** `gitea-caddy` publishes 80 and 443. Without a host sysctl, `rootlessport` fails with `permission denied` and the unit crash-loops under `Restart=always`. The fix is a one-time sysctl: `net.ipv4.ip_unprivileged_port_start=80`, persisted in `/etc/sysctl.d/99-rootless-ports.conf`. Applied 2026-08-18 after a 13,000-restart crash loop.
## Incident: 2026-08-16 — down after unattended reboot
`unattended-upgrades` rebooted Dankclaw at 15:01 on 2026-08-16. The Gitea containers ran through podman-compose with `restart: unless-stopped` and no systemd units. Rootless podman does not auto-start containers at boot without units. Gitea stayed down until a manual start at 16:02. The fix: migrate the stack to quadlet units with linger (this page). After the fix, verify on the next reboot.
The Hermes web dashboard is a browser UI for managing the agent. It shows config, API keys, and sessions. It is a management tool. It is not a chat window. Chat happens over Signal or the CLI. See [[Hermes]].
| Data | shares `/home/dank/hermes-data` with the gateway |
## Why a separate container
The gateway container runs `hermes gateway run` (the Signal platform). That entrypoint does not start the dashboard. The dashboard is a separate process. So it runs in its own compose service (`hermes-dashboard`). Both services mount the same data volume. Both restart at boot (`restart: unless-stopped`).
## Authentication
The dashboard refuses to bind to a non-loopback address without an auth provider. There is no unauthenticated public bind. This repo uses basic auth:
-`dashboard.basic_auth.username` and `dashboard.basic_auth.password_hash` live in the git-ignored `/home/dank/hermes-data/config.yaml`.
- The stored value is a one-way scrypt hash. The plaintext password is never written to a file or the repo.
- The hash is generated with `plugins.dashboard_auth.basic.hash_password`.
## Access
Open `http://192.168.0.143:9119` in a browser on the LAN. Log in with the configured username and password.
## Gotchas
- The dashboard and the gateway share one data volume. A config change made in the dashboard takes effect when the gateway restarts.
- If the shared password is reused elsewhere, treat it as exposed. Rotate it and regenerate the hash.
-`--skip-build` uses the prebuilt web assets in the image. Do not remove it unless the assets are rebuilt.
Hermes is the self-improving AI agent from Nous Research. It runs in a terminal, Telegram, Discord, Slack, WhatsApp, or Signal, and keeps a persistent learning loop: it creates skills from experience, searches past conversations, and builds a model of the user across sessions.
The *arr suite is a collection of automated media management tools for Usenet and BitTorrent. For Dankserver — which runs Copyparty, Calibre-Web, RomM, and Jellyfin — the *arr stack provides automatic downloading and library management that Copyparty alone does not cover: quality profile enforcement, series/movie discovery, metadata enrichment, and automated file organization.
The core chain is: **Prowlarr** (indexer API aggregator) → **Sonarr/Radarr/Lidarr/Readarr** (media-specific managers) → **qBittorrent** (download client) → **Jellyfin** (media server scan → auto-refresh).
Use Copyparty for curated file collections. Use *arr for automatic downloads. Use both together for maximum functionality — point *arr download directories at Copyparty's watched folders.
Copyparty is a lightweight file server and media streamer that runs as a single Python binary. It handles large payloads, media streaming, resumable uploads, directory browsing, and WebDAV from one process.
A simple HTTP check on port 3923 works. The root path returns HTTP 200 with the directory listing.
## WOPI Integration
Copyparty v1.20.9+ supports WOPI (Web Application Open Platform Interface) for editing Office documents in the web UI via Collabora Online. Enable with the `--wopi-accs` flag to limit WOPI access to specific paths.
Jellyfin is a free, open-source media server and suite of applications for organizing, managing, and streaming personal media collections (movies, TV shows, music, photos) to networked devices. It is a community fork of Emby, written in C# (.NET), and uses FFmpeg for transcoding.
The database must be on local storage (SSD/NVMe preferred). Media libraries can be on NFS/mergerfs. If media is on NFS, set `DOTNET_SYSTEM_IO_DISABLEFILELOCKING=true`.
## Authentication
- **Built-in auth (default):** Per-user libraries, parental controls, max bitrate limits. Manage via Dashboard → Users.
- **Caddy basic auth:** Single login layer. Does not integrate with per-user Jellyfin permissions.
- **LDAP/OIDC:** Available via plugins (LDAP-Auth, Jellyfin SSO). Integrates with Authentik, Keycloak, etc.
For Dankserver, built-in auth is recommended. Use Caddy basic auth on top for external access.
## Metadata Scraping
| Provider | Content | Type |
|----------|---------|------|
| TheMovieDb (TMDb) | Movies, TV shows, posters, fan art, overviews | Default |
| OMDb API | Movies only, English metadata | Default |
| TheTVDB | TV shows, episodes | Plugin |
| fanart.tv | Extra artwork, banners, thumbnails | Plugin |
| AniDB | Anime metadata | Plugin |
| MusicBrainz | Music metadata | Plugin |
| Open Subtitles | Subtitle downloads | Plugin |
| Intro Skipper | TV show intro detection | Plugin |
Run Jellyfin alongside Plex for an incremental migration. Point Jellyfin at the same media folders as Plex. Watch history and metadata will not transfer cleanly.
7. Configure Caddy reverse proxy. Set Known Proxies in Jellyfin.
8. Add media libraries in Dashboard.
9. Install plugins: Open Subtitles, fanart.tv, Intro Skipper.
10. Test playback on Android/Chrome.
## Gotchas
- **GPU passthrough silently failing:** `/dev/dri` must be passed and the correct render node used. Verify with `vainfo` inside the container.
- **Wrong render group GID:** Common GID is 992, but varies per host. Run `getent group render` to find it.
- **Subtitle burn-in triggers full transcode:** PGS (image-based) subtitles force full video transcoding. SRT/ASS (text-based) play fine.
- **Database on NFS:** SQLite on network storage causes slow startup and corruption risk. Keep `/config` on local storage.
- **Client bitrate limit too low:** Many clients default to 20 Mbps. Set Max Streaming Bitrate to Original or Maximum.
- **Naming convention mismatches:** Non-standard folder/file names cause missing metadata. Use `Movie Title (Year)` for movies and `Show Name/Season 01/Show Name S01E01.mkv` for TV.
## See Also
- [[Dankserver Stack]]
- [[ARR-Stack]] — automated download into Jellyfin libraries
- [[Copyparty]] — alternative file viewer for existing collections
Lidarr manages music albums in the *arr stack on Dankserver. It automatically searches, downloads, and organizes albums from indexers into the Jellyfin music library.
Prowlarr is the indexer management layer for the *arr stack. It aggregates indexers (torrent and Usenet) and exposes a unified API that Sonarr, Radarr, Lidarr, and Readarr consume.
qBittorrent is the torrent download client for the *arr stack on Dankserver. It receives download requests from Sonarr, Radarr, Lidarr, and Readarr via category tagging.
Radarr manages movies in the *arr stack on Dankserver. It automatically searches, downloads, and organizes movies from indexers into the Jellyfin media library.
> **RETIRED (May 2026).** The Servarr Team retired Readarr: its metadata became unusable and the community migration to Open Library stalled. There is no upstream development. Alternatives: the Bookshelf community fork, or the Calibre content server. See [Readarr (Retired)](https://wiki.servarr.com/readarr).
Readarr manages ebooks and audiobooks in the *arr stack on Dankserver. It automatically searches, downloads, and organizes books into the Calibre library.
Sonarr manages TV shows in the *arr stack on Dankserver. It automatically searches, downloads, and organizes episodes from indexers into the Jellyfin media library.
Alloy is Grafana's log and telemetry shipper. It replaces Promtail (EOL 2026-03-02). One Alloy instance runs per host and forwards container and journal logs to Loki, and optionally metrics to Prometheus.
**Hosts:** Dankclaw + Dankserver (one per host) — **Docs:** https://grafana.com/docs/alloy/latest
## Facts (from research)
| Item | Value |
|------|-------|
| Image | `franksdank/alloy:latest` (local build with a static busybox healthcheck) |
| Port | none published (push-only) |
| Config | River syntax, `config.alloy` |
| Successor to | Promtail (EOL 2026-03-02) |
| Inputs | journald socket, container log files |
| Output | Loki push endpoint |
## Why Alloy and not Promtail
Promtail reached end-of-life on 2026-03-02. Grafana Labs directs new installs to Alloy. This stack uses Alloy from day one.
## Configuration
River config at `shared/alloy/config.alloy`:
- Read journald logs from the host's `/run/systemd/journal/socket`.
- Read container logs from `/var/lib/containers/...` (podman) or the docker path.
- Forward to `http://loki:3100/loki/api/v1/push` with `host` and `job` labels.
- The journald socket mount is required for journal scraping.
- Container log paths differ between docker (`/var/lib/docker/containers`) and podman (`/var/lib/containers/storage/...`); point the config at the right one per host.
- One Alloy per host; do not run a single instance that tries to reach both hosts' logs.
Grafana visualizes the metrics that Prometheus collects from every service. It runs on Dankclaw and is reachable at `grafana.franksdank.cloud` (via Caddy, with basicauth).
Point Grafana at the Prometheus service by name (`http://prometheus:9090`) on the shared podman network.
## Healthcheck
Grafana responds on `/api/health` with 200 and `"database":"ok"`.
## Gotchas
- Change the default admin credentials at first boot (env or UI).
- Dashboards provisioned as YAML survive container recreation; UI-created ones live in the data volume — back it up.
## See Also
- [[Dankclaw Stack]]
- [[Prometheus]] — the datasource
- [[Observability]] — the whole stack
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Frank