Record the live numbers and the UTIL 1.0 verdict (unbootable: worker init
demands free > budget; ValueError on both ranks before a single weight loads).
Driver used 30.97 of 31.86 GiB; the 0.89 GiB/card margin held through a live
4,096-token generation at ~38-39 tok/s.
The arena is budget - non-expert - KV reserve - graphs - overhead. With util
capped at 0.982 (never 1.00) the graph estimate is the only plan term this repo
can shrink: the fork charges an analytic 256 MiB + 52 MiB/graph = 0.55 GiB where
the engine's own post-capture line reports 0.31 GiB, so pricing it at
256 + 12/graph returns ~0.24 GiB to the expert arena. MNBT 1024 -> 2048 is
arena-free: the registry overhead is base + 0.22 GiB per 1024 ABOVE its nbt 4096
anchor, and the fitted 5.25 GiB was fitted AT 4096.
Live after the 1ac4d48 deploy: 'expert_precise_memory': True, runtime overhead
5.25 GiB, expert cache space 17.22 GiB / 7522 slots, swap region 13.7 GiB /
5986 slots, GPU KV cache 398,658 tokens / 1.99x, pool 3,150,000,000, MNBT 1024
-- the probe's projection is now production.
Compile cache proven on consecutive boots of that shape: cold 'Compiling a
graph for compile range (1, 1024) takes 179.24 s' / 'torch.compile took
188.26 s'; warm 'Directly load the compiled graph(s) ... from the cache, took
2.143 s' / 'torch.compile took 3.97 s in total' / 'init engine ... took 41.13 s
(compilation: 3.98 s)'. Host dir 1.4 GB / 4,707 files. Time to /health stays
payload-bound (490 s warm), so the claim is ~184 s of inductor/AOT build per
restart, not a proportional boot cut.
Tests: 367/367.
Measured by boot probe on the shape-D base (production stopped, candidate
booted AS the service), plan numbers read off the engine's own log:
runtime overhead 7.60 -> 5.25 GiB (precise gfx1201 TP2 registry)
expert cache 14.87 -> 17.22 GiB
slots 6,496 -> 7,522 (+1,026, +16%)
swap region 13.7 GiB / 5,986 slots
boot healthy in 420 s, no OOM
'--expert-precise-memory': True also appears in non-default args, so a silent
argparse drop is visible. The pool is NOT cut (option A) because the pool is
not an arena lever: 3.15e9/rank keeps shape D's measured 1.99x concurrency,
and a 2.2e9 arm probed in the same window (277,686 tokens, 1.38x) is recorded
as the conservative fallback.
Boot-cost bug found while diffing Frank's 3xR9700 recipe: the image ships no
/root/.cache at all, our boot writes 1.1 GB of vLLM AOT/torch_compile
artifacts (4,665 files) plus 229 MB of comgr JIT kernels into the ephemeral
layer, and podman rm -f discards them — so every deploy/restart/reboot re-pays
'torch.compile took 181.20 s in total' (173.19 s of it one graph). Mounting
FRANK_PRIMARY_CACHE_DIR at /root/.cache persists both and hides nothing.
GPU_MAX_HW_QUEUES=2 added at Frank's direction (real HSA knob); not A/B
measured. Retracted my own recipe-diff claim that cudagraph_capture_sizes
needed pinning: capture is 3 PIECEWISE + 2 FULL graphs, 3 s, 0.31 GiB.
Tests: 367/367.
`deploy.sh <target>` with no `--service` also walks the per-service revision
markers and restarts everything behind HEAD whose paths changed — a drain, not
a sync. That is how the latent plex gate bug surfaced: the backlog fired all
at once and plex's http:// readiness probe, which can never succeed against a
PMS that drops plain HTTP off-loopback, burned its full 600 s and aborted the
run with every later service left pending.
Both facts recorded: AGENTS.md gains the drain law under Deployment, log.md
gains the plex entry with the evidence chain.
index.md intro and mermaid nodes still described 4 x 262144 fp8/bf16 and
frank-small at 2 x 262144; the active-deployment list had no frank-primary
entry at all. Fixed both, plus a malformed GLM list item.
log.md gains the shape-D entry (the SEQS=2 constraint, the arena-lever
measurement, the 2x2 probe table, the measured 1.99x), and the frank-small
entry's stale consequence bullet now points at the resolution.
frank-small runbook: the compaction-fit note is corrected - the 200704 window
moved frank-primary's trigger to 140,492, which fits a 180,224 slot again.
Comments/docs only - no functional change. The live boot of c769ff2 reported:
GPU KV cache size: 398,658 tokens,
Maximum concurrency for 200,704 tokens per request: 1.99x
expert cache space 14.87 GiB / 6,496 slots of 2.344 MiB
The arena prediction (6,496) landed exactly; the concurrency prediction (2.04x)
did not, because the per-token KV price is SHAPE-DEPENDENT: 7,701 B/token at
MAXLEN 262144 (6.1e9 -> 792,069 tokens) but 7,892 at 200704 (3.146e9 ->
398,658). Never extrapolate a concurrency figure across geometries.
Also smoke-verified: three concurrent 67-token requests answered in parallel,
~4.1 s each - the 3 slots are real, and the pool admits 2 full-length.
Re-applies the desired shape incrementally on the proven-good fp8 3x262144
base, after bisecting why the first attempt (30904ba) crash-looped.
Root cause: `--max-num-seqs 2` cannot compile on this model. AOT
build_guards(strict_error=True) inside profile_run -> _dummy_run specializes
`query_start_loc`'s width to SEQS + 1 (3 at SEQS 2, 4 at SEQS 3) and the
ConstraintViolationError kills engine init ~7 min in. MNBT is orthogonal:
SEQS 2 failed at MNBT 1024 AND 2048; SEQS 3 passed at both. KV dtype was a
red herring. Measured by four single-variable boot probes.
Arena lever measured: the expert arena is sized from `max_model_len` via the
registry-priced KV reserve (9190 B/token x max_model_len), NOT from
--kv-cache-memory-bytes. 262144 -> 14.30 GiB / 6,248 slots; 200704 ->
14.87 GiB / 6,496 slots. The 6.1e9 -> 3.15e9 pool cut bought zero arena and
is a runtime-headroom lever.
Shape: MAXLEN 200704, SEQS 3 (unchanged - the only slot count that compiles),
KV_BYTES 3.15e9/rank = 2.04x full-length concurrency at 7,701 B/token fp8,
MNBT 1024, gather 32x64, no spec decode, no prefetch, no precise-memory.
The 2026-10-01 bf16 shape (2 x 262144, --expert-precise-memory)
crash-looped at graph capture — the arena is DEVICE-resident and that
shape had 0.13 GiB to spare. The plain bf16 2 x 262144 also did not stay
up: 0 bytes free at runtime, and a real client request (995-token prompt,
max_tokens=261149) killed the worker with .
The correction is a pool cut, not a downgrade. FP8 KV (7,701 B/token),
two 200704 windows (401,408 tokens -> 3.091e9 B floor, pinned 3.15e9),
MNBT 1024. The fork's expert planner sizes the arena AFTER the KV
override, so the ~2.95 GiB/rank reclaimed becomes arena: ~7,500 slots
vs 6,248. Two contexts at 200704 is a capacity contract, not a
throughput one (2-way concurrency summed BELOW parity at 4064 ctx), so
the slot buys the VRAM headroom long generations need.
Refs [[dankserver-qwen38-fp8-expert-offload-tuning]].
- scripts/frank-primary.env: MAXLEN 200704, SEQS 2, KV_BYTES 3.15e9,
MNBT 1024, rewritten shape comment
- scripts/serve-frank-primary.sh: header comment updated
- tests/frank_primary_unit.bats: assertions updated, new seam pins the
arena-floor pool law (the pool must move with a context change or the
reduction buys nothing)
- docs + AGENTS.md: fleet topology and deployed shape corrected
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
Staged 59 GiB of retired LLM payloads on Dankclaw to
~dank/.AGENTTRASH/20261002-retired-llm-payloads/ (vLLM MiMo BF16, Qwen3.5-9B
MTP source, both DFlash dirs, Gemma4-12B-QAT, SGLang MiMo INT8) and corrected
every doc claim about what is on disk: gufo's rollback pair (Q4_K_XL weights +
the Q8_K_XL DFlash2 drafter) and frank-small's Q6_K stay in place, so the
rollback paths in the tags' env files still resolve. The vLLM-era frank-large
fetch trio moves to old/ with its runtime; runbook links updated.
frank-large: new runbook frank-large-rocmfpx-q8agent-mtp.md — the ftype-115
tensor-type lock-in that forced the fork, the pinned build (set_rows history),
the measured MTP A/B (17.0 vs 7.6 t/s decode, acceptance 0.462), the live
client contract (reasoning_content, effort fold, vision restored) and rollback.
frank-small: new runbook frank-small-llama-mimo-q5km.md; q6k marked retired
(rollback target). gufo runbook marked retired (rollback target). Both indexes,
the fleet topology table, the vision/models.yml notes, the benchmarking law and
three new Key Lessons updated.
New runbook carries the roof math, the measured three-arm spec table, the
greedy-text numerics trap, the reasoning-field contract change and the
rollback tag. LLM index and AGENTS.md fleet table, bench policy and runbook
index follow; the frank-small index entries catch up to the SGLang cutover
(entry was missed on 2026-10-01). Work log records the cutover.
Frank's call. Payload: ajgazin's abliterated Swift-1.5 UD-Q8_K_XL (29.3 GiB,
SWIFT OPEN LICENSE) -> unsloth/Qwen3.8-27B-GGUF UD-Q4_K_XL (16.35 GiB,
sha256 3f227079...b01e, vanilla/Apache-2.0). NOTE: the endpoint is no longer
uncensored — that was one of the reasons for the Swift payload on 2026-09-28.
New fetcher scripts/download-frank-large-q4.sh (no auth header: repo is public
and ungated; size + sha256 verified after transfer). The speculative drafter
and its fetcher are unchanged.
Shape: --sessions 4 --context 163840 with unquantized KV. gufo exposes no KV
quantization lever (execution_policy.hpp pins kv_cache_storage to kFp16), so
f16 is inherent.
KV arithmetic: f16 KV = 64 KiB/token (16 full-attention layers x 4 KV heads x
256 head_dim x 2 x 2B; the other 48 layers are linear-attention with a
per-sequence fp32 state). 4 x 163840 = 655,360 tokens = 40.0 GiB KV, each
session exactly 10 GiB. Budget ~62 GiB of the 114 GiB GTT.
Seam retargeted (FETCH -> the q4 script, TARGET_STUB, context/sessions pins).
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.
Both of Frank's changes at once.
KV is bf16 (--kv-cache-dtype auto) at two full-size windows (2 x 262144).
bf16 measures 13,839 B/token (bf16 3x241664 boot: 11,047,010,304 B -> 798,280
tokens), decomposing as 12,276 attention + 1,563 mamba/side caches vs fp8's
6,138 + 1,563 = 7,701. 524,288 tokens needs 7.256e9 B; pinned at 7.3e9
(~527,500 tokens, 3,206 margin). Pool cost: 5.68 -> 6.76 GiB per rank.
--expert-precise-memory selects the fitted overhead (5.25 GiB) over the default
(7.60 GiB = fitted + compile/graph-capture headroom). It moves only the plan's
term, so the expert arena goes 6,248 -> ~7,278 slots (~1,025 more) at the
cost of boot-time capture slack. Overhead adds 0.22 GiB/1024 nbt above 4096, so
MNBT 2048 pays nothing.
Seam + docs + AGENTS row + log updated.
Frank's call 2026-10-01: MAXLEN 262144 x SEQS 3 = 786,432 tokens of contract.
Measured fp8 KV costs 7,698 B/token (4-slot boot 8,311,013,376 B -> 1,079,582
tokens; 5-slot boot 10,388,766,720 B -> 1,350,182 tokens), so the pool floor
for the contract is 6.054e9 B. Pinned at 6,100,000,000 (~792k tokens, ~5.9k
margin) so the remaining ~2.2 GiB versus the 4-slot shape goes to the expert
ARENA - the fork sizes its decode-critical expert cache from the remainder
after weights, pool and workspace.
MNBT 2048 (Frank): the arena sweep measured 6,499 slots at every rung <= 4096,
so the larger batch costs the arena nothing.
Seam retargeted (3 seqs / 6.1e9 / MNBT 2048); docs + AGENTS row + log updated,
including the closed DCP investigation.
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.
Five 262144-token contexts passed every boot gate (KV 1,350,182 / 5.15x) but
left 0 bytes free at runtime: a real client request (995-token prompt,
max_tokens=261149, sampling params straight from models.yml) killed Worker_TP1
with torch.OutOfMemoryError 'Tried to allocate 50.00 MiB ... 0 bytes free'.
Boot is not proof of runtime survivability.
Four slots x 262144 fp8 (KV_BYTES 8,311,013,376/rank) returns ~1.93 GiB per
rank for activations/graph capture.
Unit: Restart=always instead of on-failure — a runtime worker fault makes
vLLM's API server exit clean, so ExecStart returned 0 and on-failure never
fired, leaving the endpoint down until a human noticed. Also
StartLimitIntervalSec=0 and OOMScoreAdjust=-500.
Frank's call after the night ladder: five contexts at the model's full native
262,144 window beat six shortened ones. Six full-length contexts do NOT fit at
any KV dtype (6 x 241664 fp8 asked 10.70 GiB against 9.18 GiB free — the
6-seq graph/activation footprint leaves too little for the pool); 5 x 262144
at fp8 PASSED on the first rung: KV_BYTES 10,388,766,720/rank, measured
'GPU KV cache size: 1,350,182 tokens, Maximum concurrency for 262,144 tokens
per request: 5.15x' (39,462 tokens of margin).
Also documented: the SEQS=3 bf16 max (241,664, KV 732,359 / 3.03x) measured
earlier the same night, and the root-cause read that the 7.2 GiB KV ceiling
law was the MTP4 workspace, not VRAM. Seam and harness (contextWindow 262144)
re-pinned.
Night ladder result (Frank's subtract-16k method): 262144 and 245760 both OOM
at CUDA graph capture ('Tried to allocate 76.00 MiB ... 0 bytes free'), 237568
passed (KV 718,606 tokens), and the +4096 bonus step PASSED -> 241664 is max
possible and DEPLOYED at util 0.98 (KV 732,359 tokens / 3.03x concurrency at
241,664-token requests). Util stays 0.98: lowering it shrinks the engine's own
budget instead of buying slack.
Client contract restored (the GLM rollback d91379d dropped all three and the
qwen rollback inherited the gap): --reasoning-parser qwen3 fixes the malformed
thinking block (thinking was landing inline in content with a dangling
</think> and a null reasoning field), --enable-auto-tool-choice with
--tool-call-parser qwen3_xml makes tools work, and
--default-chat-template-kwargs enables enable_thinking + preserve_thinking.
Harness: contextWindow 241664 (compaction gap recomputed), preserve_thinking
moved into chat_template_kwargs.
- AGENTS.md: fleet table + frank-primary section back to qwen FP8 + MTP4
(37.4-40.6 tok/s, 54% acceptance); GLM runbook moved to the retired index.
- runbook: RETIRED verdict — DFlash2 drafter served and drafted but
acceptance pinned 1.00-1.38 on every text class including verbatim
repetition (mis-wired aux capture, not a weak drafter); 5x slower than
baseline at ~1.2 acceptance.
- log.md: 2026-10-01 rollback entry.
- deploy-remote.sh: LLM_PRIMARY_CONTAINER back to frank-primary-qwen;
glm-probe joins the legacy sweep list.
- frank_primary_unit.bats: re-pinned to the qwen shape — KV qualification
law (7.2 GiB/rank, 3.65x concurrency; 9/16 GiB OOM), fp8 q8/q8 KV
PRESENT, MTP4 PRESENT (3.1-3.4x over the 12.11 no-spec baseline), no tool
dialect, automatic offload. The old 'MTP buys zero' law carried the
prefetch tax and is retired.
The frank-primary section still carried the retracted framing ('a latent bug the compile
exposes' in the mHC/KDA path) and listed only three compile patches. It now records:
six apply-verified artifacts across two WIP changelists (compile 01-05 + the DFlash2
enabling patch), the measured facts (lru_cache does not protect a probe; suppress_errors
does not cover breaks; the trunk can never satisfy fullgraph because its indexer
metadata syncs to host), the single remaining blocker
(KeyError: 'in_proj_qkvbfg_a', raised in a callee dynamo inlines into
Glm5NextDecoderLayer.forward, not at any of the four nn_modules lookups and not a literal
string lookup -> currently a dynamo-side name-resolution failure, not a demonstrated fork
bug), and the two operational gotchas this session paid for (the launcher silently
dropping VLLM_COMPILE_FULLGRAPH=0, and the PYTHONPATH sitecustomize fork bomb). The
DFlash2 paragraph now says the aux-hidden-state contract is written, and that the
engine's refusal was literally 'Model does not support EAGLE3 interface'.
tests/run.sh: 362 seams pass.
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.
Frank
Comrade Chief Designer
Comrade Chief Designer
