Runbook, both indexes, AGENTS.md (topology row, endpoint section, three new
Key Lessons, one stale SGLang label in the bench list) and the bundle log now
describe the deployed shape: pool 540,672, 180,224 tokens per conversation,
f16 K + q8_0 V, mmproj vision, 21,684 used / 2,441 free of 24,576 MiB, 106.8
tok/s decode, 3,814 t/s prefill.
The claims that mattered:
- three contexts buy 180,224 tokens per conversation, not the native 262,144,
and a longer prompt is a hard 400 - so models.yml contextWindow is 180224
and frank-primary's 183,501-token compaction request no longer fits here
(forward pointer added to the retired SGLang runbook, which asserted it)
- the vision tower bills ~1.4 GiB for a 918 MB file, established by subtracting
two measured boots, not by a formula
- /v1/chat/completions ignores the legacy top-level 'images' field: the model
invents a color; only the image_url content part reaches it
Measured after the first 3-slot deploy: 3 x 196,608 booted (n_slots=3,
n_ctx_slot=196608, multimodal loaded) but used 22,972/24,576 MiB - 1,153 MiB
free, outside the 2-3 GiB band. The two measurements now recorded in
frank-small.env price the vision tower at ~1.4 GiB (non-KV 7,390 -> 8,859
MiB), so the pool drops 49,152 tokens (16,384/slot, 1,176 MiB) to land ~2.7
GiB free. The bats budget gate is recalibrated to the measured 9 GiB non-KV
cost, which now rejects the 3 x 196,608 shape instead of blessing it.
CTX_SIZE 589824 / PARALLEL 3 -> three independent 196,608-token conversations
(a mamba hybrid cannot share one KV cache, so per-slot window = pool/slots;
3 x 262,144 would need 18.4 GiB of KV alone and busts the card).
Vision comes back via bartowski's mmproj-MiMo-V2.6-Distill-Qwen-9B-f16.gguf
(918,166,048 B, sha256 ff348f31..., fetched + verified by
scripts/download-frank-small.sh): the Q5_K_M GGUF ships no v.* tensors, so
--mmproj IS the image path. MMPROJ_FILE without a colon default so an explicit
empty value is a real text-only opt-out, and the deployer now gates on the
projector exactly as the serve script does.
Budget is measured, not guessed: the 2 x 262,144 shape used 20,190/24,576 MiB,
which prices non-KV cost at 7,390 MiB. Three slots + projector land near
22.5 GiB, i.e. ~2 GiB spare - the bottom of the requested band, to be
confirmed by measurement and laddered if it lands outside 2-3 GiB.
One quant step down from Q6_K on the same CUDA build and the same shape:
2 x 262,144 with K f16 + V q8_0. The 0.86 GiB freed from the weights is
headroom, not a new window — the native 262,144 cap is already reached,
and f16 V would still cost 3.96 GiB more than the pool can carry.
The KV pool stays pinned at 606,208 tokens (card is full: 531 MiB free at
the old shape; autofit's 687,388 would leave 0.53 GB and OOM on the first
long request). 2 x 262,144 = 524,288 leaves ~82k tokens of pool slack.
Motivation: harness compaction. frank-primary's contextWindow 262,144
compacts at 70% = 183,501 tokens, and the compaction request carries a
32,768-token summary budget -> 216,269 must fit ONE frank-small request.
The 151,552 slot window could not hold it; the native window does.
No spec-decode interaction: checkpoint ships zero mtp.* tensors (config
lies), seam keeps every --speculative-* flag away from the engine.
Replaces llama.cpp Gemma4-12B-QAT (payload + binary retained as rollback).
SGLang in a self-contained python3.12 venv with an extracted CUDA 13.0
nvcc toolchain for the sgl-kernel/flashinfer JIT.
Shape: fp8 KV, pinned 606,208-token pool / 4 x 151,552 slots, bf16 mamba
states pinned at 32 slots (the engine autofills the ssm pool to a fixed
~5.1 GB regardless of dtype), eager prefill (no headroom for graph
capture), no spec decode (the checkpoint ships no mtp.* tensors despite
mtp_num_hidden_layers=1 — NEXTN drafts with an uninitialised 4.24 GiB
ghost at 0% acceptance).
Verified: 83.9 tok/s decode, 170k single prompt, 4 concurrent 54k
requests, vision (in-checkpoint tower), reasoning_content field.
The build default is 8192 MiB. dankclaw is memory-tight (frank-large's
70 GiB GTT plus the rest of the stack) and Gemma4 runs fully on the 3090
(-ngl 999), so the CPU-side cache is not the hot path. Cap it at 2048 to
give back ~6 GiB of LPDDR. Gemma4 is not a DeltaNet model, so the
>12.8 GiB CACHE_RAM floor does not apply.
Wired through frank-small.env (CACHE_RAM) -> serve-frank-small.sh
(--cache-ram), with a wiring seam so the cap cannot be silently unwired.
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
- 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.
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.
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.
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.
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.
Frank