The vLLM detour (a2b39b1, same day) is retired: the Q4_K_M llama.cpp setup
is the requested shape and the vLLM KV-quant options on sm_86 all measured
dead (fp8 hardware-impossible, int8 page-size unification, turboquant
mm-prefix). Root unit restored, vLLM quadlet/Dockerfile/fetch script moved
to .AGENTTRASH, runbook listed as same-day-reverted. Rollback target:
frank-small-vllm-gemma4.
int8_per_token_head: page-size unification fails (scale tensor breaks
the 256/512 head-dim ratio). turboquant_k8v4: backend validation fails
(no mm-prefix attention, which Gemma4's vision/audio path requires).
fp8: hardware-dead on sm_86. bf16 it is; history documented in
frank-small.env, seam pins auto + no backend pin.
The VLLM_ATTENTION_BACKEND env var does not exist in this vLLM dev
build (warns 'Unknown vLLM environment variable' and is ignored), so
the pin landed nowhere and Gemma4's hook force-selected TRITON_ATTN
again. Pass --attention-backend TURBOQUANT as a CLI flag instead.
Gemma4's config hook force-selects TRITON_ATTN for the mixed 256/512
head dims whenever no backend is selected, and TRITON_ATTN rejects the
turboquant KV dtype (measured: 'kv_cache_dtype not supported', crash
loop). The TURBOQUANT backend accepts any head size and the
turboquant_* dtypes, and the Gemma4 MTP speculator carries the
target's selected backend through to the draft layers.
int8_per_token_head keeps a separate fp32 scale tensor per (token,
head); Gemma4's hybrid head dims (SWA 256 / full 512) make the
per-token page bytes 4160 vs 8256, not an integer ratio, so
unify_kv_cache_spec_page_size raises NotImplementedError — nine
crash-loop restarts measured. TurboQuant k8v4 packs its block scales
inside the tensor, keeps the 2:1 page ratio, and has a dedicated
backend with an explicit Ampere path. 289,647 -> ~772k-token pool.
fp8 KV is hardware-dead on the 3090 (sm_86, no cvt.e4m3x2), so int8 is
the cheapest quant the TRITON_ATTN backend can run: pure ALU dequant,
same backend, no fp8 tensor cores needed. 289,647 -> ~579k-token pool,
two -> four full-context sessions. Quality verified against the bf16
baseline (needle-in-haystack at 213k + short battery) before the swap.
Runtime swap on the RTX 3090 (port 7771):
- vLLM gemma-x86_64-cu129 dev build, digest-pinned, audio extra pip-installed
into localhost/vllm/vllm-openai:frank-small (Dockerfile, no COPY lines)
- google/gemma-4-12B-it-qat-w4a16-ct (10.26 GiB int4) + Google's bf16 MTP
assistant (845 MB) — the w4a16-ct assistant cannot draft in vLLM (drafter
path skips Marlin; acceptance 1.00), bf16 measures 2.80-3.56
- root systemd unit -> rootless Quadlet (dankclaw/frank-small/quadlet/);
manual CUDA device+lib mounts (no nvidia-container-toolkit on host)
- bf16 KV (fp8 hardware-dead on sm_86), 289,647-token pool, 2 concurrent
full-context sessions, CUDA graphs on
- native text+image+audio: the fleet's multimodal endpoint
- measured: pp@67k 1293 t/s, pp@213k 608 t/s, tg@213k ~13.6 t/s,
tg@85k ~17.6 t/s
- host driver 610.57.04 (Ubuntu 26.04); sm_86 is correct, no arch override
- planner/scoping: frank-small now selected by dankclaw/frank-small/ +
env/serve/download trio; old unit retired to scripts/.AGENTTRASH/
- docs: new runbook frank-small-vllm-gemma4.md; AGENTS.md, LLM index,
top docs index updated; old runbook moved to retired
--cache-ram caps the total bytes of saved prompt states in HOST RAM, and
alloc() refuses any single state larger than the whole cap. That refusal was
firing in production: 137 times in the two days from 2026-09-16, at 11.0-11.4
GiB per state, so every conversation past ~170k tokens was uncacheable and
re-prefilled from scratch on its next turn.
The state law at this shape is 68 KiB/token (16 full-attention layers x 4 kv
heads x 256 dim x 2 B x 2 for K/V, plus the MTP draft context's 4096 B), so
8192 MiB held ~120k tokens and 16384 holds ~240k: a FULL 196608-token slot
state (~12.8 GiB) now fits with ~3 GiB spare.
Host RAM only, allocated on demand and never pinned, so the VRAM fit is
untouched. New seam: the cap must clear one full slot's state and stay under
half the host's RAM (it fails on the old 8192 and passes on 16384). Also
corrects the adjacent claim that 0 is the build's default for
--ctx-checkpoints -- this binary's --help says 32.
Isolated-port (:7790) 2k-token decode with AVG rocm-smi util over the
request (peak util lies: a layer split touches both cards every token):
35:30 draft ROCm1 31.52 t/s util0 41% util1 98% <- deployed
36:29 draft ROCm1 31.49 t/s util0 43% util1 98%
35:30 draft ROCm0 31.52 t/s util0 42% util1 98%
36:29 draft ROCm0 31.48 t/s util0 43% util1 98%
The x1 card (ROCm1) stays the 98% bottleneck in every configuration.
Moving the MTP drafter to the x16 card does not flip the balance: the
per-step cost is the verify phase (both cards run the full target) plus
the logits, which are pinned to the card holding the last layers. A
target-layer move (36:29) buys 2 points on the x16 card and erodes its
VRAM margin (fit free 0.45 GiB) for no throughput change. The x1 card is
compute-bound, not PCIe-bound (~5.2 MB activation per ubatch on the link),
so saturating it costs no tokens. 35:30 is the balance; no split change.
A pending entry for frank-default-sightsound (FLM NPU endpoint retired
2026-09-11) survived in the dankclaw state store and made every --pending
run exit 2 at the dispatcher before the Playwright e2e gate could
execute. The pending pass now drops entries the target no longer allows
into ~/.AGENTTRASH and keeps going. Behavioral seam test added.
- services.json: frank-large/medium/small move from admin-only to the user
page with current descriptions (halogen w4b iGPU; 35:30 layer-split MTP
27B on the R9700s; Gemma4-12B QAT vision path on the 3090) and docs
links to /docs#models
- docs.html: static user documentation (access chain, model guide, media
how-tos, troubleshooting); served at /docs via the post-auth homepage
listener; shipped by deploy_homepage; footer + index link to it
- tests: pages seam asserts the docs contract (page, link, Caddy route,
deployer copy)
- vault: Homepage.md synced (docs landing shipped, cards reclassified)
The two R9700s sit on unequal PCIe paths: ROCm0 (03:00.0) is CPU-root-port
Gen4 x16 (~31.5 GB/s); ROCm1 (0b:00.0) rides the B650 chipset through
bridges negotiating Gen4 x1 (~1.97 GB/s). With the MTP drafter, output
layers and heavier compute buffers pinned to the last-layers card, live
rocm-smi read 98% busy on the x1 card vs 40% on the x16 card. Moving one
DeltaNet layer (34) to ROCm0 shifts ~0.43 GiB off the bottlenecked card;
the seam fit lands 30.59/30.01 with gap 0.58 GiB, inside floor and slack.
36:29 fails both gates, so this lever is spent.
signal-cli 0.14.7 parses an all-digit -a value as a phone number (the
numeric account name never resolves), and the message is a -m option,
not a positional (a positional is parsed as a second recipient).
Verified live: corrected invocation returns type SUCCESS.
Every webhook died on 'cannot access local variable _last_sent' because
the prune comprehension rebound the module global inside the function.
Prune in place instead.
The bridge-subnet-only pair dropped the host's own health probe
(source 192.168.0.143/127.0.0.1) and failed a healthy deploy. The
deployer now installs the full set: ACCEPT 127.0.0.1, ACCEPT the host
IP, ACCEPT 10.89.0.0/16, then DROP.
169.254.1.2 is not a local address on this host (netavark runs its
bridges in aardvark-dns' netns), so the relay cannot bind it — and a
loopback bind refuses container traffic. Probe-socket proof: a host
socket on 192.168.0.143:18094 answers the container's 169.254.1.2:18094
dial while the LAN IP refuses from the container. The relay binds
192.168.0.143:18093 and the deployer installs an iptables INPUT pair
(ACCEPT 10.89.0.0/16, DROP rest) so the unauthenticated endpoint stays
off the LAN.
A socket bound to 127.0.0.1 refuses container traffic that arrives via
the podman bridge: the first live webhook failed with 'dial tcp
169.254.1.2:18093: connection refused'. The relay now binds the
host-gateway endpoint — link-local to the podman bridge, off the LAN —
and the deployer's health probe follows it there.
- runbook: identity now 4x262144, audio verified, throughput table with
the OOM and negative-result rows, deployed flags updated to reality.
- AGENTS.md frank-small paragraph matches the deployed shape.
- log.md: the sweep entry.
- scripts/frank-small.env: my earlier line-range edit had clobbered
NGL=999 and FLASH_ATTN=on out of the file; the seams caught it. The
live service was unaffected because the script's defaults match the
declared values exactly, but the file contract is restored.
Suite: ALL TESTS PASSED (321), vault 125 pages 0 unresolved.
Swept on the 3090 at the deployed 4x262144 geometry, salted API probes:
- ubatch 512: pp@254k ~918 t/s, tg@250k 90-204 (noisy)
- ubatch 1024: pp@254k 962 t/s (+4.8%), tg@250k 257-302
- ubatch 2048: OOM — the MTP draft context needs ~3.3 GiB of pp compute
buffer against 2.2 GiB headroom; do not try.
- draft depth 4 + poll 100 decoded WORSE at 250k (230-257) than 3/50:
the longer verification round re-reads more long-ctx attention than the
extra accepted token earns. Kept 3/50, now with the evidence attached.
Idle VRAM 23,300 MiB; all ubatch-1024 variants idled identically
(compute buffers reserved at max graph size), so 8192 buys the deepest
prefill window for free over 4096.
Both were implicit build defaults (3 and 50); the throughput sweep needs
to move them without editing the exec line each time. Defaults preserve
current behavior exactly. Additive only.
DFlash2's sidecar (1.92 GiB + buffers on ROCm1) kept the 393216 pool off the table; the MTP head lives in the target file, so its cost is one nextn layer's KV (4096 B/pos) plus small buffers. Fit: ROCm0 29.77 / ROCm1 28.32 GiB, over the floor, 34:31 the balance point at this pool. n_max 3 is the measured MTP winner. The seam's fit is now drafter-aware, the slack is the deployed integer gap (1483 MiB), the ctx cap 393216 with the rope caveat (top 131k extrapolates on the base rope).
CTX_SIZE=1048576 / PARALLEL=4: every slot gets the model's full trained
window. Measured fit arithmetic on the 3090: 1 full-context slot idles
at 11,996 MiB, 4 at 22,346 -> 3,450 MiB per slot, ~8,546 MiB fixed.
A 5th slot projects 25.8 GiB past the 24,576 MiB card, and the KV pool
allocates upfront, so 4 is the ceiling. Explicit CTX_SIZE bypasses --fit
on purpose: fit maximises the TOTAL pool and would divide a smaller
window across slots at parallel>1. The slot-grid seam now pins both the
full-window division and the budget inequality so a future capacity
bump fails pre-deploy, not in the boot allocation.
Audio verified while measuring: the mmproj's audio encoder transcribes
input_audio parts verbatim (JFK WAV sample). Header comments updated.
The Qwen3.8-27B IQ3_S file carried its own MTP head; Gemma4's ships as a
separate assistant file (mtp-gemma-4-12B-it.gguf, arch gemma4-assistant,
4 nextn layers) that loads via --model-draft and shares the target's last
layers' KV. The draftless ngram-mod matcher chains behind it in the same
--spec-type value — two context-backed drafters cannot share one process,
so this chain is the only shape that carries both.
KV is q8_0 (16 KiB/token/layer vs 32 at f16) and CTX_SIZE is unset so
--fit sizes the pool against what the 7.4 GiB of payloads leave on the
24 GiB card. build-cuda was rebuilt with -DGGML_CUDA_FA_ALL_QUANTS=ON,
which quantized KV on CUDA requires. The three seams that pinned the Qwen
contract (MTP-in-file, the Qwen mmproj name, the f16 KV arithmetic) now
pin the Gemma4 contract instead, and the deploy handler's presence check
points at the new payload.
The 36:29/327680/vision deploy came up healthy and text-only: image requests
answered 500 'image input is not supported - hint: ... provide the mmproj'
with no mmproj line in the load log, so llama-server fb27a525 does not
autoload mmproj*.gguf from the model directory. Name the file, the way
frank-small's serve script already does, and pin the flag in the seam.
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.
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.
`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.
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 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.
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.
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 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.
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.
'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.
Frank