A torch-profiler kernel-time window (496 decode tokens, rank0 trace parsed for
device-activity time) on the deployed GLM-5.3-Flash shape falsifies this page's
own "the expert gather is the bound" explanation:
46.18 s of kernel time / 496 tokens = ~93 ms per token (actual step ~104 ms)
34.8 % vectorized_elementwise_kernel 708,500 launches (1,428/token)
10.0 % r4d_gemm_moe_w4a16_nt_b16_m64 (the MoE GEMM)
6.7 % elementwise_kernel_manual_unroll
6.4 % wvSplitK_hf_sml_ (split-K GEMM)
8.9 % router: sbtopk + warpMergeSortTopK + bitonicSortKVInPlace
6.5 % r4d_mhc_* (hyper-connections)
1.05 % r4d_moe_lru_gather_kernel <-- the retracted suspect
The earlier arithmetic was circular (it assumed the gather consumed the whole
step). The real bound is unfused pointwise glue: the fork compiles its own Qwen
trunk (support_torch_compile in models/qwen4_exp/amd/model.py) and not its GLM
trunk, so the engine boots with "provides neither a compiled submodule nor
breakable CUDA graphs. Overriding to FULL_DECODE_ONLY." That also explains why
MTP4 measured flat — pointwise cost scales with the token count, so spec decode
amortizes nothing here — and it retires arena/hit-rate tuning (the gather is 1 %).
Staged as documented patch artifacts (not wired into any serve script; the
engine does not yet boot with them) — scripts/patches/glm53-compile-0{1,2,3}*.patch:
-01 models/glm5next/common/model.py decorate Glm5NextModel (R4D_GLM_COMPILE)
-02 model_executor/kernels/r4d_lib.py freeze 10 capability probes that do
os.path.exists/.so lookups inside the traced forward. functools.lru_cache
does NOT protect them: dynamo re-traces the wrapped body even when warm
(verified in-image); a plain module-global read compiles.
-03 compilation/wrapper.py make fullgraph=True opt-out, because
suppress_errors does NOT cover graph breaks and this trunk can never
satisfy fullgraph (indexer metadata syncs to host — .item() at
glm5next/amd/sparse_indexer.py:263,419).
With all three the trunk traces (Dynamo bytecode transform time 1.08 s) and dies
on KeyError: in_proj_qkvbfg_a in the mHC/KDA path — a latent bug the compile
exposes, and the remaining blocker.
Also records two structural constraints: no new HIP kernels are possible (r4d.so
ships without sources), which binary-blocks the fused router at GLM's 288
experts; and DFlash2 for GLM is a TARGET-side port, not a drafter port
(incoai/GLM-5.3-Flash-DFlash2 is architectures ["DFlash2DraftModel"] +
model_type qwen3, already mapped at registry.py:636; glm5next lacks only the
aux-hidden-state plumbing deepseek_v2.py demonstrates).
Runbook frontmatter, intro, mermaid and wall section, AGENTS.md and the work log
corrected to match the measurement. tests/run.sh: 362 seams pass.
Payload: Intel/GLM-5.3-Flash-W4A16-AutoRound (AutoRound symmetric int4 group
128, auto_round:auto_gptq, 34 shards + 2 extras, 113,074 tensors, ~170 GiB).
It is the only W4A16 checkpoint this fork can serve with expert offload: its
quant_method "auto-round" is overridden to "inc", which routes the MoE to
libr4d's grouped W4A16 GEMM (R4dW4a16MoEMethod). A compressed-tensors W4A16
build of the same model resolves to TritonWNA16Experts, which has no offload
recipe, and refuses to serve at all (170 GiB of experts cannot fit 64 GiB of
VRAM, so offload is mandatory).
Shape: tcclaviger RDNA4 vLLM fork (same digest-pinned image as the 2026-09-29
qwen cutover), TP=2, 131,072 ctx x 2 seqs, bf16 KV (fp8 is ILLEGAL here — it
disables the sparse MLA backend's rope-free Triton path, the only MLA decode
gfx12 has), util 0.98, NBATCH 2048, --enable-expert-offload
--expert-precise-memory, prefix caching, glm47 tool + reasoning parsers,
MTP and expert prefetch absent by measurement.
Measured: decode 9.9-10.0 tok/s, prefill 918 tok/s (warm), arena 14.51 GiB =
2,383 slots/rank over 12,096 host-backed experts (73.64 GiB pinned/rank).
THE 35 tok/s TARGET IS NOT MET and the shortfall is measured: live telemetry
(R4D_LRU_TELEMETRY=1) gives a 0.53-0.66 per-layer hit rate, a step moves
0.82 GiB of misses in ~100 ms = 8.2 GB/s effective against a link measured at
22.74 GB/s — the same structural gather-path gap the qwen campaign found.
Growing the arena 13.08 -> 14.51 GiB changed decode by nothing.
Rejected by measurement: MTP4 (head loads, acceptance 3.76-4.90, but decode
unchanged and prefill collapses 918 -> 255 — a verification forward gathers the
union of the batch's experts), NBATCH 8192 (identical prefill).
Also: scripts/patches/gpu_worker-kv-cap.patch (fork-local fix for a KV-reserve
model that under-counts GLM's per-sequence KDA state; not needed by this shape),
runbook, AGENTS.md row + section, both indexes, work log.
Rollback: 2339fae (qwen FP8 + MTP4).
Asked for iterative decode/prefill improvement at a 32k working point
without touching the model or the quantization. There isn't one: gufo's
served settings are already at this engine's tuned point, so no config
change was made and none is proposed.
Measured on the repo's own prose and shell sources at 32,762 prompt
tokens, salted per run against a warm prefix cache, 3-5 repetitions:
* --prefill-chunk 512 -> 8192 moves prefill 410.7-412.4 t/s. A 16x change
in chunk size, 0.4% out. The lever is dead at this depth.
* --draft-tokens 5 -> 10 gives 13.79-15.86 t/s decode, all inside the
run-to-run noise. The 5-rep confirmation REVERSED the 3-rep leaders:
dt9+chunk4096 looked best and came last (12.86), the deployed config
came first (14.71).
* --draft-policy fixed vs adaptive: no effect. --min-draft-tokens:
rejected by the engine ("DFlash2 requires ... 1"). MTP: unsupported for
Q8 targets (token_embd.weight is Q8_0).
* --context 262144 -> 40960, tested as a diagnostic: no gain, and it
would break models.yml's declared contextWindow. Not shipped.
What does move the number is speculation itself — 6.65 t/s off vs ~15 t/s
on, 5.3x — and it is already enabled.
The two traps that cost most of the wall time are now in AGENTS.md's
benchmarking policy, because both produced FALSE results:
1. gufo bench cannot rank speculative settings and inverted the ranking.
Its corpus reports acceptance=0.933 +/- 0.000 with an identical trace
sha256 across draft widths — the same highly-predictable text every
run. It reported 29-37 t/s where real traffic gives ~15, and made
dt16 look 6% faster than dt7 when real text made it 14% slower.
2. A live A/B needs >=128 generated tokens. At 8, inter_token is
first-token noise; the first pass produced a nonsense ordering that
the 128-token pass reversed.
Plus the Server-Timing header is `ttft;dur=402.485` — the unit rides in
the NAME half, so splitting on "=" keys the dict "ttft;dur" and every
lookup silently returns nothing. Four configs reported "no Server-Timing"
while the header was present on all of them.
The disk was at 93% with 137G free. Nine retired payloads are now moved
to ~/.AGENTTRASH/20260928-retired-models/ with a MANIFEST.md giving a
restore command per item — moved, not deleted, and nothing is freed until
the purge is approved. Every one maps to a runtime already retired in the
runbook index.
Only the .gguf was taken from Qwen3.8-27B-UD-Q8_K_XL/: the live
frank-large DFlash2 drafter lives in that directory and stays put. The
manifest warns against restoring the directory wholesale.
Kept: checkpoints/ (23G) — unreferenced by the repo, but ComfyUI is live
and its workflows were not searched, so it is not confirmed dead. Also
logs/, because x4e-run.log is a SPARSE file (3.3G apparent, 4K real) and
frees nothing.
The reference check caught a live bug: serve-frank-small.sh's fallback
MODEL_DIR still pointed at the retired GSQ-RCO payload frank-small served
on 2026-09-16/17. frank-small.env has always overridden it, so the drift
was invisible — but a missing or unreadable env file would have turned a
config problem into a "missing asset" naming a model that is no longer
there. Repointed at Gemma4-12B-QAT/, which the env file already sets.
Not taken, by request: podman images (279G reclaimable, 95%) and the
uv+pip caches (73G). Together 352G of regenerable garbage, larger than
what was reclaimed here.
The 29.3G payload rollback target went with the rest, so the payload-only
rollback is now a re-fetch first; the runbook and AGENTS.md say so.
The ad-hoc status commands kept producing wrong readings, most seriously
reporting "MemAvailable 0.0 GiB" on a host that actually had 240 GiB free.
Cause: /proc/meminfo separates key from value with a TAB, so
grep MemAvailable /proc/meminfo | tr -s " " | cut -d" " -f2
returns an empty field and awk prints 0.0. A false "host is out of memory"
reading is worse than no reading, because it invites an unnecessary abort.
dankstatus.sh parses on a colon split (tab-agnostic), cross-checks
MemAvailable against MemTotal before printing either, and refuses to print
a zero value. It also pins the rocm-smi CSV column order (device, total,
used) which was previously read swapped, and takes only the first line of
systemctl is-active/is-enabled, which emit more than one line here.
Incidental finding while validating: rocm-smi reports card2 (the 7700X iGPU)
as 0.50 GiB total, not the 123.0 GiB torch.cuda reports. torch was exposing
the GTT aperture. The device is smaller than documented, not larger.
Tests: 346/346 pass (no test surface touched).
gufo is a vertical inference engine built for exactly this hardware
(Ryzen AI MAX+ 395 / Radeon 8060S, gfx1151), and Dankclaw's frank-large
has swapped runtimes seven times in search of a better one. This one
ships no prebuilt Linux binary, so the runtime is BUILT FROM SOURCE on
the host and the payload, port, alias and client contract stay exactly
where they are.
Runtime: gufo v0.1.1 (b0f8673) at /home/dank/GitPackages/gufo-v0.1.1,
built by the new scripts/build-gufo.sh. It links the host ROCm 7.2.4
through its RPATH — the opposite of the unsloth prebuilt it replaces,
which carried its own — so two host workarounds are encoded in that
script: Ubuntu's librocwmma-dev installs a truncated rocWMMA header set
that shadows AMD's and cannot be overridden with -I in HIP mode, and the
ROCm clang prefers Ubuntu 26.04's GCC 16 headers over the GCC 15 whose
libstdc++ is actually installed.
Shape is unchanged: one session, full 262144 ctx, the same UD-Q8_K_XL
target and the same DFlash2 drafter, greedy, text-only. Speculation
moves from llama.cpp's fixed depth 4 to gufo's adaptive block policy
capped at 7 proposals.
Client-contract deltas, all documented: the chat template is compiled
into the binary (no --chat-template-file, so the Qwen-Sharp 73k-token
failure mode is gone), reasoning_effort now takes the full OpenAI set
(so models.yml's reasoningEffortMap fold is no longer required — kept
anyway, so this commit does not also move every high role from xhigh to
high effort), and there is no KV-quant or prompt-cache-RAM flag.
The test seam now runs the serve script against a stub engine and
asserts on the argv it actually execs, instead of grepping the script's
text. The deploy handler gates on the runtime binary and returns before
it stops the serving unit.
Measured trade, from gufo's own Q8_K_XL tables against llama.cpp
b11069: prefill +37% to +80%, cold load 4.5x faster, decode +7% at zero
depth — but -14% at 65,536 tokens of depth and -24% at 131,072. On the
8-bit target the crossover is ~32k, and frank-large runs one 262,144
conversation. The Q4_K_XL target shows no such regression.
Discovery, not guesswork: dump the GGUF v3 headers of the fetched parts
(scripts/dump-gguf.py, added — a dependency-free reader that never
touches tensor data, so a 50 GiB part scans in milliseconds) before
the first load.
What the dump says (parts 2-3, 2026-09-26):
- The 51B-param PLE n-gram table is ONE tensor: per_layer_token_embd.
weight, dims [160, 320001536], Q8_0, 54.4 GiB — exactly the whole of
part 3. It is the same tensor the shared MTP head borrows, so the
=CPU override is required for placement AND for the draft load
(GPU-resident borrow: "invalid vector subscript").
- The model is qwen4exp, 512 experts, top-10 routing
(qwen4exp.ple.ngram_size=3, heads_per_ngram=8).
- Part 1 is metadata-only (ntensors=0); parts 2-6 carry tensors.
The placeholder pattern ngram_embeddings matched nothing — dropped from
OT_OVERRIDES (env, serve-script default, seam, AGENTS.md, runbook,
LLM index) before it could masquerade as a working override at load
time. Deployed set: exps=CPU,per_layer_token_embd.weight=CPU.
Suite: 346 green.
The 2026-09-26 same-day cutover that ended the Lucebox DS4.1 detour:
Dankserver's :7770 endpoint is now unsloth llama.cpp (the b11030-mix
prebuilt, ROCm gfx120X, bundled runtime) serving
unsloth/Qwen3.8-Flash-Next-GGUF at Q8_0 — 175.30 GiB across six sha256-
verified split parts — on both Radeon AI PRO R9700s, as a ROOT unit (the
frank-large shape: no container, no generated unit).
The deployed contract:
- Placement law: layer split -ts 1,1 with -ngl 999; the routed experts
and the 51B-param n-gram (PLE) table are COMPUTED in host RAM via
-ot exps=CPU,ngram_embeddings=CPU (the x4-pair measurement: streaming
over PCIe loses at 3.7 t/s, computing in RAM does not care).
per_layer_token_embd.weight=CPU is required by the shared MTP head's
borrow path (a GPU-resident borrow crashes the draft load).
mmap stays on; --cache-ram 16384.
- Spec chain draft-mtp,ngram-mod: the shared Q8_0 MTP head
(mtp-...-shared-Q8_0.gguf, 2,786,568,256 B) at depth 2, with the
draftless ngram-mod matcher chained behind it — the only legal way to
layer two spec implementations in one process.
- Vision explicit: --mmproj names the BF16 projector (this build never
autoloads; a server without it comes up healthy and text-only).
- Reasoning: --reasoning-format deepseek → reasoning_content; the
Qwen3.8 template's {xhigh, medium, low} reasoning_effort gate folded
by reasoningEffortMap in models.yml.
- One stream, the model's full 262,144 context; greedy server default
with Qwen-recommended sampling riding compat.extraBody.
The deployer's deploy_llm_primary is rewritten for the root unit: it
gates every artifact at its declared byte count (stat -c %s, never wc -c
— which read the whole 377 GiB lucebox file for minutes on 2026-09-26),
retires the lucebox user quadlet to dated home trash, waits out port
7770, relinks/enables/restarts the root unit and waits up to 1800 s for
/health. The quadlet-era tree under dankserver/frank-primary/ is
archived (dankserver/.AGENTTRASH/20260926-llamacpp-cutover/ on this
box; git history everywhere else) and its planner mapping replaced by
the scripts/ trio.
Observability flips back to the llamacpp namespace: a frank-primary
Prometheus job at 192.168.0.235:7770 (cross-host LAN — the netavark
published-port quirk that forces host.containers.internal for the two
Dankclaw endpoints is same-host only), the LlmEndpointDown runbook
text, and the dashboard description for three llama.cpp endpoints.
Docs: the new runbook (frank-primary-llamacpp-qwen38-flash-next-q8.md),
RETIRED banners on the lucebox and radiance runbooks and their index
entries, the AGENTS.md fleet section rewritten for the root-unit shape,
and the homepage frank-medium card renamed to frank-primary with the
Q8_0 Flash-Next spec (docs.html #models and the modality note updated:
images on frank-primary and frank-small, audio on frank-small only).
models.yml: the dankserver provider is "llamacpp (dankserver)" serving
frank-primary, contextWindow 262144, maxTokens 65536 (compaction-gap
law: 183,501 + 65,536 <= 262,144), input [text, image].
Suite: 346 green (frank_primary_unit.bats rewritten for the root unit —
16 tests pinning the placement law, the shared-head requirement, the
explicit projector, HIP env hygiene (BIND_HOST, never bare HOST), and
the deployer's byte-count gates; the quadlet-era planner and
changed-tests mappings rewritten for the trio; the observability scrape
test asserts three llamacpp jobs again).
Rollback: the lucebox image (localhost/lucebox-rocm:gfx1201) and the
378 GiB payload stay on the host; the Qwen payload stays fetched, so a
re-swap in either direction is a redeploy, not a refetch.
The serve script now passes --chat-template-file pointing at the
Qwen-Sharp v22.4.0 template. Without it the binary falls back to the
built-in peg-native template, which produces a ~73,600-token prompt for
a ~250-token request. With the sharp template wired in, the same request
produces a 62-175-token prompt.
Knowledge-base follow-up to a45f396. The runbook, both index one-liners, and
the AGENTS.md topology row all still declared FP8 active; flipped to
just1moremodel/Qwen3.8-27B-Uncensored-MXFP4-awq with unquantized bf16 KV (pool
486,464 tokens, full 262,144 window), and recorded the causal nuance: fp8 KV
was on both arms, so the FP8-vs-MXFP4 gap was weight precision, not KV - clean
KV is a quality hypothesis for W4A8, not a throughput one. No verdict yet.
The quadlet's "TEXT ONLY ... 400 here" comment was wrong: vision works on the
MXFP4 checkpoint (Qwen3_5ForConditionalGeneration carries the vision tower
through the Gemma4 merge path, unquantized) - verified by naming a solid red
48x48 PNG "Red". Comment corrected; harness models.yml input stays [text,image].
Docs-only; no re-deploy (the endpoint already serves vision).
The first cutover crash-looped the unit with 'invalid split file name':
the fetch script concatenated the three split parts into a single GGUF,
but part 1's header carries ggml.file_count=3, so the byte-correct
concatenation still advertises itself as a split and the loader demands
a -NNNNN-of-NNNNN filename.
- serve-frank-large.sh: -m points at part 1; llama.cpp loads the split
natively (parts 2-3 auto-discovered in the same directory)
- deploy-remote.sh: LLM_LARGE_MODEL gate path follows
- download-frank-large-unsloth.sh: assembly removed; the parts ARE the
canonical payload, verified against live Hub LFS oids
- docs: runbook model section + troubleshooting note, AGENTS.md, log
- the redundant 87 GiB assembled file is staged on dankclaw at
~/.AGENTTRASH/20260921-frank-large-assembled/
New runbook frank-large-unsloth-qwen38-flash-next.md (runtime, payload,
deployed config, the explicit -md, the 114 GiB GTT memory model, client
contract, rollback). frank-large-ciru-orca.md marked RETIRED (rollback
target, payload and runtime intact on disk). AGENTS.md topology row,
frank-large section, harness note, runbook index, and the GPU-budget
lesson (GTT aperture ~114 GiB, not the stale 64 GiB rocm-smi figure —
the CIRU Orca held 92.6 GiB on the same device). LLM index header,
mermaid, active list, and the vault root index updated. Work-log entry.
Runbook deployed-flags table and AGENTS.md now name the template
(froggeric v22.4.0, repo-tracked, terse by default) and the rollback
target; the stale qwen-fixed-v22.3.jinja row and the long-gone
MM_MERGE_ENGINE note are gone.
DFlash2 (100.0 t/s decode, 287.1 t/s burst-4) beats MTP self-speculation
(85.9 t/s, 179.7 t/s) despite a lower accept length (4.24 vs 5.66): the
small separate drafter steps faster than the target's own MTP layer applied
autoregressively. DFlash2 is the deployed shape; both swap-script paths now
default to dflash/7. Docs updated.
The radiance TP=2 cutover changed the metric namespace frank-medium exports
(llamacpp:* -> vllm:*) and left two scrape targets addressing host processes
by an address a container cannot reach.
- shared/prometheus/rules/llm-fleet.yml: two namespaces side by side. Added
LlmRequestsWaiting, LlmKvCacheSaturated (vllm-only gauges llama.cpp never
had), the job:llm_vllm_prefix_cache_reuse:rate5m and
job:llm_spec_accept_length:rate5m recording rules, and re-cut
LlmPrefixCacheCollapse / LlmSpecDecodeOff to OR both families with
per-family traffic floors so an idle endpoint cannot fire them. Every
expression validated against live Prometheus.
- shared/prometheus/prometheus.yml: frank-small -> host.containers.internal:7771
and frank-large scraped at host.containers.internal:7770. Both are HOST
processes on Dankclaw and netavark DNATs published container ports only, so
the LAN-IP targets were red by construction; frank-large's CIRU llama.cpp
build does export the llamacpp:* family, contra the old comment.
- tests/observability_quadlet.bats: the two seams now encode the new contract
(per-endpoint reachability by process type; every named series must be on the
measured list for its engine; engine-independent rules must name both).
- homepage docs.html/services.json: frank-medium sees pictures, frank-large is
CIRU Orca with a single seat (was halogen w4b, two seats, 200k).
- docs: radiance runbook is ACTIVE (deployed flags, three-arm battery, verified
capabilities, rollback, replica-era sections labelled); llama.cpp runbook
carries the SUPERSEDED banner; Dual-R9700-Cooperation states the measured
verdict; Observability records the namespace split and the DNAT lesson;
AGENTS.md topology/frank-medium/Model Config/Benchmarking Policy/runbook
index flipped; both indexes updated; work-log entry added.
Also repaired agent-pass damage: prose prepended above frontmatter openers,
duplicated runbook bodies, a literal tool-call token pasted into a runbook, and
an index destroyed by diff artifacts (restored from HEAD, flips re-applied).
Vault link check: 128 pages, 0 unresolved. ./tests/run.sh: 325 pass.
Reanalysis after the 2026-09-20 swap to an ASUS ProArt B850-Creator WiFi
NEO: both R9700s now sit on CPU GPP root ports (x8/x8 Gen5 bifurcation),
every bridge reads AtomicOpsCap Routing+, and both endpoints read
AtomicOpsCtl ReqEn+ (CAP_EXP+28.w = 0x1440) — the exact precondition
Dual-R9700-Cooperation's test plan named. TP=2 is therefore unblocked,
not proven: no collective has been dispatched since the swap,
GGML_HIP_RCCL is still OFF, and the 272 t/s "tensor" number on record
measured a butterfly fallback.
- Dual-R9700-Cooperation: short answer rewritten, re-measured 2026-09-20
topology table, rebuild-hypothesis door marked moot, probe plan is now
standing work.
- Dankserver System Specs: Motherboard section; GPU BDFs 0b:00.0 ->
07:00.0, iGPU 10:00.0 -> 11:00.0, kernel 7.0.0-31, SMT on, and the new
eno2 (192.168.0.247) second 5GbE port holding the low-metric default
route while eno1 keeps the fleet's static .235.
- frank-medium-llamacpp-dflash2: rationale and hardware row split past
from present; dated work-log sections stay as history.
- AGENTS.md + both indexes: no longer assert a wall that no longer exists.
Docs only; frank-medium stayed active and healthy throughout (verified
/health). ./tests/run.sh 331/331.
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.
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
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.
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.
The 393216/34:31/MTP deploy measured 30.78 GiB on ROCm1 against the
fit's 28.32 — the MTP draft buffers (one nextn layer's KV, 4096 B/pos,
plus draft buffers) came in 2.46 GiB above the estimate. Recalibrate
the seam's MTP constant to the measured 4.20 GiB total charge on card
1 and drop the floor 1.5 -> 1.0 GiB for this shape (ROCm1's 1.08 GiB
headroom is real; 35:30/36:29 fail card 0's floor at this pool, so
34:31 is the only split that fits 2x196k).
AGENTS.md and the runbook move to the MTP drafter: 34:31, 393216 pool
(196608 per conversation), the DFlash2 sidecar documented as a
one-line revert, and a dated part-2 section with the fit/measured
table and the recalibration.
AGENTS.md topology, Active Services, the CUDA-quant-KV lesson (f16 is no
longer why), and the runbook index. New frank-small-gemma4-12b runbook:
payload digests, the SWA/q8 arithmetic behind --fit's 262k window, the
MTP-sidecar + ngram-mod chain, the always-accepted reasoning_effort, and
the fit-probe error that looks like a fault at boot. The Qwen runbook and
both indexes carry the retirement; work log updated.
AGENTS.md topology row and section, and the runbook: the pool is 327680
positions (163840 per conversation over two slots), the split is 36:29 -- one
layer past the 34:31 balance at this context, on purpose, for headroom on the
drafter card (fit 29.19/26.04, measured 28.40/26.66 GiB at load, 3.46/5.20
headroom) -- and vision is on via the explicit --mmproj flag, because this
build has no autoload (the first deploy came up text-only and 500'd image
requests). The rope caveat is written down where the cap lives: the GGUF has
no scaling keys, so the top 65k of the pool extrapolates on the base rope.
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.
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.
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.
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.
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.
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.
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.
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.
Asked whether Vulkan on both devices unlocks the split. It does not: the
per-tensor design aborts byte-identically, because the gate is not the driver
mix but `--override-tensor` — `has_tensor_overrides()` is a required term of
the scheduler's `pipeline_parallel` flag (llama-context.cpp:433, the only place
in the tree that reads it), and that flag sets `n_copies` in
`ggml_backend_sched_new`. With `n_copies == 1` a pre-allocated weight on a
non-primary device has no backend and the graph aborts.
Chasing it exposed a harness artifact that invalidated the earlier conclusion:
llama-bench splits a device LIST with `/` and SWEEP values with `,`, so
`-dev CUDA0,Vulkan1` was two single-device runs all along (and `-ts 0.85,0.15`
two sweep values, i.e. device 0 takes everything — which is where the bogus OOM
came from). With the correct syntax two-device splits work in both flavours:
pure Vulkan -dev Vulkan1/Vulkan0 -ts 0.85/0.15 -> 27.55 tok/s @64k (12.8 GiB on the 3090)
cross-vendor -dev CUDA0/Vulkan1 -ts 0.15/0.85 -> 27.23 tok/s @64k, pp512 268.5 (15.0 GiB)
iGPU alone 25.77 tok/s @64k, pp512 231.8
+7 % decode, +16 % prefill — real but not a regime change, and it costs the
3090 ~15 GiB, which this fleet spends on its only vision endpoint. The design's
premise still fails: whole layers carry their experts (1.62 GiB each), the 12
attention layers are spread every 4th, and the 32.8 GiB PLE must stay on the
main device, so KV cannot be consolidated on the 3090.
Also records the evening's incident (a dead rootless podman pause process took
all container management down; `podman system migrate` panics in 5.7.0; reboot
repaired it) and the netavark host-address lesson from the hermes 502.
The 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.
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.
halogen-flash-server is ACTIVE again with its 2026-09-16 measured ledger (budget
33.5 GiB against a need of 27.2 — no clamp — and 14.7 GiB left for everything
else), ik_llama.cpp is retired with its measurements kept as the record of why,
and the new frank-small runbook covers the -mtp file, the 64 KiB/token KV
arithmetic, the MTP contract, and the reasoning_effort trap that cost me every
subagent on the endpoint.
AGENTS.md: topology, Active Services, runbook index, and three new llama.cpp
lessons (MTP is a property of the file; --jinja is what `tools` needs; a model's
template can reject the harness's thinking level). Three pre-existing defects
found while editing it: a truncated sentence mid-file, a harness path that does
not exist on either server, and the benchmarking stop/mask recipe still calling
frank-large a root unit.
Truthfulness notes worth keeping: the co-tenant stall NOTE prints again, so the
claim that it "no longer prints" is gone and the three observed block counts are
recorded instead of a guessed threshold; and the Observability page says out
loud why frank-large has no scrape job rather than leaving a silent absence.
New page: LLM/Dual-R9700-Cooperation.md — registered in the vault index, with
the measured evidence rather than the theory.
**The answer.** Card #2 sits behind the B650 chipset, whose bridges advertise
`AtomicOpsCap: Routing-` (read-only), so ROCr refuses to fill the hostcall
buffer RCCL's kernels declare and no RCCL op can dispatch on it. Two paths fix
it: an M.2→PCIe x4 riser into the free, CPU-attached M.2_2 slot (~$30; per
ASUS's spec page M.2_1 and M.2_2 are CPU-attached, PCIEX16_2 is chipset in x1
mode, and there is no x8/x8 bifurcation on this board), or a board whose CPU x16
bifurcates to x8/x8. Bandwidth is not the constraint: ≈10 MB per token-step ≈
1.4 ms at Gen4 x4.
**New measured results.** A bare 2-rank torch.distributed run: rank 0 (CPU-lane
card) passes broadcast/all_reduce/barrier; rank 1 (chipset card) fails all
three, and `send`/`recv` fails too — which kills pipeline parallelism as an
escape, since vLLM's PP handoff is exactly that primitive.
**Correction.** I previously said no software path exists. That overshot. One is
open and unverified: recompiling RCCL so its kernels stop declaring a hostcall
buffer (`-mprintf-kind=buffered` is the claim). Counter-evidence recorded: RCCL's
device source contains no live printf/assert/__assertfail, so the trigger is
unexplained, and the claim traces to search summaries, not a primary source. The
page carries the two-command test that would settle it. Also corrected: 0.9.3 is
radiance's version, not vLLM's — the base is vLLM 0.27.1 / torch 2.11.0 / ROCm
7.14, and its shipped RCCL exposes no `ncclDevComm*`, so torch's symmetric-memory
path cannot run.
**Rule added to the benchmarking policy**: it covers ANY GPU experiment, not just
benchmarks — including a bare 2-rank probe. Running one against the live service
on 2026-09-16 preceded the host going hard down (no ARP, no ICMP, power cycle
required); a hung AQL dispatch wedging the driver is the plausible mechanism.
The policy also now says how to tell off from wedged: `ip neigh` from a peer.
Retiring dankclaw/compose.yaml and adding alert rules invalidated claims across
the repo, including in the rule file agents read first. Clean cutover includes
the prose.
**Tests that had gone hollow.** Four assertions grepped dankclaw/compose.yaml
for a service name. With the file gone they passed by *absence* — a gate that
can no longer fail. Each now asserts the real invariant ("no dankclaw compose
file competes with the quadlet units") and keeps a live check for a re-added
entry, so it still fails when one appears: verified by recreating a compose file
with a comfyui service and watching three tests go red, then removing it.
**AGENTS.md, which was wrong about ComfyUI.** It said ComfyUI runs CPU-only
("no ROCm wheels for gfx1151, CUDA images wrong"). That was true in the
compose era and is false now: the restored unit runs with /dev/kfd + /dev/dri and
must NOT pass --cpu. An agent trusting the rule file would have re-broken it.
Also corrected: the observability paragraph ("RUNNING BUT UNCONFIGURED, no
datasource, no rules" — all three have been configured since 2026-09-16), the
Active runbook row that still named halogen for frank-large, halogen's absence
from the Retired table, the benchmark policy naming halogen's harness, and the
seam list.
**Unit and page claims.** prometheus.container said "no alerting rules anywhere
in this stack"; node-exporter.container said Dankserver runs from its compose
stack; Prometheus.md documented three jobs as working (they scrape HTML: they
can never report up) and said no dashboard existed; index.md and the three
authentik service pages pointed at the retired file; Authentik.md's upgrade
instruction told you to bump an image tag in it.
One bad edit of my own on the way (a stale line number clobbered a See Also
entry in Prometheus.md) — caught by reading back, fixed, and the vault link
checker now proves the cross-references hold.
The two replicas behind :7770 were on plain round-robin, which throws away
prefix-cache hits: an agent loop re-sends its system prompt and file context
every turn, and alternating replicas means every other turn pays a cold
prefill. Measured with 90k-token contexts, warm-turn TTFT was 2676 ms.
Pinning everything to one replica is not the answer either — a cold batch
(the shape a benchmark or a burst of parallel subagents produces) then loses a
card's worth of prefill: 16 requests took 30.1 s pinned against 18.6 s spread.
So the policy routes by request shape. A continuation (more than one message —
the agent-loop signature) is pinned to the replica that served its prefix;
anything else spreads across the cards by counter. Measured: 19.97 s for the
burst and 1095 ms warm-turn TTFT — the best of both, without a regression on
either workload.
The policy lives in proxy/routing.py, separate from the HTTP plumbing in
proxy/affinity-proxy.py, so it can be tested without aiohttp, a network or a
GPU. tests/frank_medium_proxy.bats covers it, and it earned its keep: the
first version hashed fresh prefixes to pick a card, which sends every member of
a same-prefix burst to one card (29.8 s, measured), and a stale index in the
pin path returned 500 on every continuation. Both are now assertions.
ROUTING=roundrobin and ROUTING=affinity stay available to re-measure the
choice on this host; the default is adaptive.
The radiance MXFP4 stack cannot run TP=2 on this host. RCCL's collective
kernels carry a hidden hostcall-buffer parameter, and ROCr refuses to
dispatch them on a device without PCIe atomics: the AQL dispatch fails with
hipErrorIllegalState ("the operation cannot be performed in the present
state"), which is what every two-rank collective hit. The second R9700 sits
behind the chipset bridges 05:00.0/06:09.0, which advertise
`AtomicOpsCap: Routing-` — a read-only capability, so no software can enable
atomic routing on that path and the driver's
pci_enable_atomic_ops_to_root() fails for that card. The first card, on CPU
root port 00:01.1, has atomics enabled.
Ruled out: kernel version (7.0.0-28/-30/-31 identical, after a full upgrade),
iommu=pt, P2P hardware (45.3 GB/s between the cards), the shipped librccl (a
from-source rebuild reproduces the failure at the same call path), vLLM and
pynccl (a bare two-rank torch.distributed allreduce fails the same way), all
NCCL transport knobs, and single-rank tests (NCCL short-circuits at
world_size=1, so one-card success proves nothing).
Deployment: three rootless Quadlets — frank-medium-a (:7771, card 1),
frank-medium-b (:7772, card 2), both TP=1 at 114688 context (262144 and
131072 do not fit one card's KV) with vLLM-profiled KV — and
frank-medium-proxy (:7770), a round-robin that preserves the fleet contract
of one endpoint and one served model name. frank-medium.container stays as
the TP=2 target for a host whose second card is atomic-capable; the deploy
handler does not install it and trashes any stale copy.
Measured through the proxy (vllm bench serve, random dataset, --ignore-eos):
conc 1 42.8 tok/s, conc 4 147.6 tok/s, conc 8 230.8 tok/s, 32k prefill at
~2.8k tok/s per card, TPOT p50 23.6-28.8 ms.
Also: healthcheck reads PORT so one probe serves all three units, the hermes
example and the harness models.yml drop to the real 114688 context, and the
runbook records the full evidence chain.
Comrade Chief Designer
Frank
