The fast char-rp-reasoning seat works: ~77 tok/s (vs GGUF ~59.5, base NVFP4 ~53),
MTP draft-acceptance 32-40%, mean acceptance length 2.19. Same Heretic2/NEO-CODE
model, NVFP4 + native qwen3_5_mtp spec-decode.
Full end-to-end recipe + the four landmines in docs/runbooks/heretic2-nvfp4-mtp-seat.md:
(1) load as AutoModelForImageTextToText not AutoModelForCausalLM (namespace/gibberish);
(2) modelopt format not compressed-tensors (compressed-tensors MTP = 0% accept);
(3) modelopt 0.45 <-> transformers 5.12.1 FusedMoE crash (guarded in quant_modelopt.py);
(4) vLLM 0.24.0 does NOT propagate modelopt exclude_modules to the spec-decode draft
model -> BF16 mtp head gets quantized -> shape crash; no checkpoint config fixes it
(is_layer_skipped is exact-membership not glob) -> fix is a mounted sitecustomize that
force-skips mtp.* in is_layer_skipped (upstream vLLM bug to report).
Scripts: quant_modelopt.py (FusedMoE guard + single-shard export + multimodal load),
finalize_modelopt_mtp.py (splice bf16 mtp), serve_modelopt_mtp.sh, run_quant_modelopt.sh,
sitecustomize-mtp-workaround.py.
Root-caused the NVFP4 gibberish to a quant-namespace bug: quant_nvfp4.py loaded
via AutoModelForCausalLM -> text-only Qwen3_5ForCausalLM -> flat model.layers.* keys,
but vLLM 0.24 serves only Qwen3_5ForConditionalGeneration (whose weight mapper needs
model.language_model.*). Fixed by loading as AutoModelForImageTextToText; NVFP4 now
serves coherent (validated greedy on ana-ml2 GPU0).
Base NVFP4 (compressed-tensors) measured ~53 tok/s (~= GGUF at batch-1, no single-stream
win) and its MTP is 0% acceptance (vLLM's Qwen3_5MTP drafter loads the bf16 mtp head only
off a modelopt main-model checkpoint). Added quant_modelopt.py (nvidia-modelopt PTQ,
matches AEON's NVFP4 W4A4 g16 + lm_head/linear_attn/visual exclusions) as the path to
working native MTP; graft + splice + serve otherwise unchanged.
Captures the full pipeline recipe (graft->quant->splice->config->serve) with every
gotcha found this session, the 3 gibberish suspects, and the diagnostic ladder
(validate native-config no-MTP coherence FIRST) for a fresh session to finish the
chase. Also stages the NVFP4 scripts + 512-row calib. Recent decisions: NEO-CODE
seat swap (R36), webhook ALLOWED_HOST_LIST fix. Lessons: validate-tracer-bullet-first,
mtp-graft-dropped-at-load, gitea-204-red-herring.
The ufw fix (prior commit) was necessary but insufficient. The DECISIVE blocker
was gitea webhook.ALLOWED_HOST_LIST = 'external, 10.100.0.0/16' (NH3 only) —
corviduo-dev is 10.250.50.152 (Anaheim), so gitea refused to deliver ('deny
10.250.50.152') and never opened the TCP connection. Fixed to 'external,
10.0.0.0/8' (whole fleet, matches the ufw choice) + gitea restart.
Listener now logs every delivery (source-IP/hmac_ok/ref/action) — the old
log_message=pass silence hid the whole failure. Proven end-to-end: real gitea
delivery -> hmac_ok=True, ref=main, 202 deploying -> green deploy.
render-verify caught it: Dvalin's calib tool_calls carry OpenAI wire-form JSON
string arguments, but the Qwen3.6 chat template does .items() on arguments (needs
a dict) → jinja TypeError. Parse string->dict in render_verify + the quant's
load_calib_chat. Confirmed: renders the exact qwen3_coder XML the seat emits
(prefixed bifrost.soong-lab.*, v0.3.13 generate_portrait, <think>, <tool_response>).
Extracted from the deployed backend (bifrost/tools.py) via the venv with a
capturing mock register_tool — the LIVE schema, not a stale copy. OpenAI-function
form for brokkr/Dvalin's ~128-row tool-call-XML calib slice (R36 #355 anchor).
The auto-deploy silently never worked: corviduo-dev's ufw is default-deny and
port 9010 was never allowed, so gitea's webhook deliveries timed out (DROP).
v0.3.6 was a manual deploy; v0.3.7-v0.3.13 never auto-deployed. The setup-time
'test-delivery 204' was gitea queuing, not the listener receiving. Fixed by
'ufw allow from 10.0.0.0/8' (operator-directed). Confirmed end-to-end.
R36 gate (2026-07-14) validated NEO-CODE ships on all axes: tool-calling 0.967
(attach_tool 1.00, 0 runaways — #355 eliminated), prose genre-artifact-fine
(less clichéd than gen), refusal uncensored-as-spec + CSAM-clean.
#355 root cause was MODEL-level, not the reasoning-budget-forcing bug: Deckard
emitted Qwen's native qwen3_coder XML tool format malformed -> llama.cpp leaked
the closing tags into the arg value -> Bifrost attach_tool schema error -> retry
-> reasoning runaway to max_tokens. NEO-CODE emits the same native format cleanly
on the same seat/parser -> no schema error -> no runaway. The fix was the model
swap; there was never a wrong parser (the XML is Qwen3.5/3.6-native).
- reasoning seat: Deckard-PKD (Qwen3.5) -> NEO-CODE=Heretic2-Thinking (Qwen3.6-27B) Q5
- samplers: card defaults (temp 1.0 / top_p 0.95 / top_k 20 / min_p 0.0), DRY dropped
- ctx: 256K max; custom llama.cpp kept (qwen3_coder parse + PR#25544 belt-and-suspenders)
- persistent-memory ACTIVE 1 marked resolved
soong-dev found the studio serving a stale web/ (52015 vs 55025 bytes — missing the
01-Role section, favicon, thinking-status): the deploy rsynced backend/ but never web/,
so SOONG_LAB_WEB_DIR stayed pinned to the initial manual copy while the backend updated.
Deploy now rsyncs BOTH backend/->studio AND web/->SOONG_LAB_WEB_DIR (read from the env)
on every green run. Verified: served frontend now 55025 bytes, current.
Per operator call (no gitea write token on the Worldtree-team VM): a 2-min systemd
--user timer on nh3-dev polls corviduo's last-deploy.json and pings soong-dev via
althing on a NEW red deploy (green stays silent). Delivers soong-dev's red-run
visibility without a credential on corviduo. Tested (red detect+format DRY, green quiet).
The char-rp-reasoning (Deckard) seat now runs llamacpp-charrp:custom-latest via a
new LLAMA_REASONING_IMAGE var (Magidonia char-rp stays on stock — no reasoning bug).
Fixes Worldtree #355 at the source (budget multi-terminator handles Qwen3.5's
<tool_call> reasoning end-tag). Live 2026-07-13: Deckard loads on Blackwell, serves
coherent, reasoning bounds at the 400 budget. Rollback via .env LLAMA_REASONING_IMAGE.
Durable record of the custom llama.cpp the char-rp-reasoning seat will run to
fix Worldtree #355 (reasoning-budget forcing broken in stock b8840 — single
end-tag </think> can't match Qwen3.5's <tool_call> reasoning terminator, so the
budget never force-closes and reasoning runs away to max_tokens). PR #25544
adds multiple terminating sequences; unmerged upstream, so we build it.
- build.sh: reproducible recipe (clone master@6eddde0 + merge PR #25544 +
resolve the 1 server-common.cpp conflict + CUDA build for Blackwell sm_120)
- README.md: why + acceptance test + rollback + REMOVE-WHEN-MERGED tracking
Image llamacpp-charrp:6eddde0-pr25544 BUILT + smoke-tested on ana-ml2; seat
swap pending. See also auto-memory reference_charrp_custom_llamacpp_pr25544.
Adds the rsync --link-dest hourly snapshot job (nh3-dev:~/development ->
nh3-nas, 48-snapshot retention, secrets/build-dirs excluded) that closes the
no-off-box-backup gap exposed by the 2026-07-12 working-dir clobber. Script
mirrors the live ~/.config/dev-backup/dev-backup.sh; runbook covers restore.
Point the in-page TTS at the zonos-gateway wrapper on irv-ml1:8890 (direct,
so streaming isn't buffered by LiteLLM): OpenAI-shape body (model: ext-tts,
input, voice), Cora default voice, float32@44.1kHz PCM decode. Quotes are
joined into a single stream call (prosody — no per-sentence chunking).
Migration regex rewrites stale saved endpoints (:8299/:8210, /tts[/stream],
:4000) to the new one.
New TTS service entry + reproducibility_audit row for the zonos-gateway
wrapper (irv-ml1:8890) — the ext-tts-aliased OpenAI facade over Zonos.
23 fields across Text&voice / Expression / Prosody / Quality / Sampling /
Output section groups; live voice dropdown from /v1/voices; response
format pcm|wav (audition UI forces wav). Distinct from the older down
zonos :8203 entry. jsonschema-validated.
The live gateway config has served char-rp-reasoning as deckard-pkd-27b (:8018)
since the 2026-07-08 A/B; the standalone doc had frozen on QwQ-RpR-v4. Corrects
seat 4 (backend + samplers + server-side DRY/reasoning-budget notes).
Also snapshots session state in persistent-memory.md: phantom-qwen verified
already-clean, ana-docker docker log-cap (logrotate copytruncate, no bounce),
and the granite→gen memory_extractor bind live on demo+personal.
Drop-in system prompt for an agent whose quoted output is voiced by mOrpheus: speak in
double quotes (only quoted text is voiced), phrase each quoted line as one coherent
utterance (per-quote prosody), and the sparse/boundary/no-stack tag rules (measured
stability on the early checkpoint). Lists the honored tag set.
Per-sentence chunking generated each sentence cold, flattening intonation/prosody that
spans the whole quoted line. Chunk by QUOTED SECTION instead — each contiguous quote is
generated whole (max_tokens 2400) so its prosody stays intact; multiple quotes in a reply
still play serially on the shared clock. extractQuotes already returns exactly these spans;
dropped splitSentences.
Split the quoted dialogue into sentences and stream each as its own short /tts/stream
request (max_tokens 900), queued back-to-back on one shared AudioContext clock (speechHead)
so playback is gapless and in order. First sentence starts fast; each chunk is short so it
generates cleanly (no ramble/cap risk); the next sentence generates while the current plays.
A newer reply supersedes via the ttsGen counter; 🔊 replays.
Browsers suspend the Web Audio AudioContext until a user gesture; speakQuotes fires on
reply-complete (no active gesture), so a suspended context played silently. Prime/resume
the context on any click or keydown (capture phase) so it's running before playback.
Server side was fine throughout (/tts + /tts/stream both 200 with valid audio).
- max_tokens default 2400->3500 (~42s) in wrapper + gateway-chat client, with a _cap()
clamp so prompt+gen never exceeds MAX_CTX (4096) — a cloning ref block is ~1100 tokens,
so an unclamped 3500 would overflow context on the clone path.
- Staged clone voices: /voices dir of <name>.wav + <name>.txt, each encoded to its Orpheus
reference block at startup; voice="<name>" zero-shot clones it. Beatrice (a chatterbox
reference) staged as the first normal-voice clone. GET /voices lists baddy + clones.
- compose: mount voices dir + pass MORPHEUS_MAX_LEN to the wrapper (clamp must match engine).
vLLM concurrency (measured, --max-num-seqs 8, 250-tok reqs): near-linear batching — 8
concurrent finish in the same ~2.8s as 1 (707 tok/s, 8.1x single, flat per-req latency).
Chunked-sentence production can fan out for ~8x throughput; CPU SNAC decode is the scale
bottleneck, not generation.
Cut-offs were the max_tokens=1200 ceiling (~14.6s of audio), not memory (~1250 tokens
<< 4096 context). Diagnosis: the repetition penalty is load-bearing for clean stops —
rep 1.0 => the model never emits end-of-speech and rambles to the cap; rep 1.1 (the
wrapper default) => clean natural stop. So normal lines already complete; only genuinely
long dialogue (>~14.6s, ~25+ words) hit the cap. Raised default + client max_tokens to
2400 (~29s), still within the 4096 context (no memory cost). Verified: a 49-word line
now finishes at 16.73s (was clipped at 14.6s).
Wrapper gains POST /tts/stream: reads the vLLM token stream, decodes SNAC in WINDOWED
CHUNKS (every 6 frames, decode [2 ctx | 6 | 2 ctx] and emit only the middle 6 — context
both sides => seamless), and streams raw PCM16 (24kHz mono) as it generates. Windowed
(not per-frame) because per-frame CPU decode's per-call overhead x ~60 frames serialized
to ~7s (RTF 2.2); windowed keeps up (RTF ~0.97). Whole-clip /tts kept for non-browser use.
gateway-chat plays the stream via the Web Audio API (fetch reader -> int16->float32 ->
scheduled AudioBufferSourceNodes on a running clock; a new reply supersedes the prior
stream via a generation counter; 🔊 replays). Measured: TTFA 0.80s (was ~4.5s whole-clip),
RTF 0.97, full-duration match. CORS already covers the new route.
Deployed: tts rebuilt on irv-ml1, page pushed to ana-docker.
Gateway-chat now auto-plays quoted text from each assistant reply through the mOrpheus
TTS endpoint. Sidebar gains a 🔊 toggle + endpoint/voice fields (persist in localStorage,
prefilled to irv-ml1:8299 / baddy). On reply-complete, straight and typographic double
quotes are extracted, joined, POSTed to /tts, and the returned WAV plays (click 🔊 to
replay; a new reply interrupts the prior clip).
Requires CORS on the wrapper (page served from ana-docker:8091 fetches irv-ml1:8299
cross-origin) — added CORSMiddleware(allow_origins=[*]) to the mOrpheus tts app (internal-
only endpoint). Verified end-to-end: preflight + POST return ACAO=*, valid 24kHz WAV.
Deployed: tts container rebuilt/recreated on irv-ml1; page pushed to ana-docker conf
(bind-mounted, live on next request).
The gen seat's vLLM served-name was still qwen3.6-27b-aeon, a stale skin
left over from the AEON-27B → 35B-A3B-heretic swap — it named neither the
right family (aeon) nor size (27b vs 35B-A3B). Renamed the served-name to
qwen3.6-35b-a3b-heretic (+ -thinking) on ana-ml2 :8015 via the stack .env,
and repointed litellm's gen / gen-reasoning / summarizer-large model refs +
comments to match, so /v1/models, the gateway config, and spend-logs all
reveal the actual model in the request path.
Verified end-to-end: gen -> 'PIPELINE OK', gen-reasoning -> content + reasoning
surfaced, all three aliases healthy. char-rp / char-rp-reasoning untouched.
dvalin confirmed the live A/B-proven set IS canonical for Deckard as a dark-RP reasoning seat:
temp 1.0/top_p 0.95/top_k 40/min_p 0.05, no presence/rep penalty, DRY 0.8 server-side. Endorsed
over the card's base-thinking (top_k 20/min_p 0/presence 1.5). No value change; comment + memory
record the confirmation + tuning ladder (flat->min_p 0.08, loops->DRY 0.9, over-damped->DRY 0.6/off).
Operator wanted a reasoning-RP model that tolerates DRY (RpR-v4 forbids rep/DRY -> a
1/30 loop tail). Ran the full A/B on brokkr's 30-prompt D1 suite (content-only, slop-scored):
- Deckard-PKD (Qwen3.5-27B, DavidAU creative tune) WON: 0/30 loops, 0/30 refusals, clean
managed reasoning (native Qwen3.5 <think>/enable_thinking), DRY-tolerant, ~57 tok/s,
runs on the base llama-swap b8840 image. -> now the char-rp-reasoning seat (:8018).
- RpR-v4: 0 refusals but 1/30 loop (no-DRY). Pantheon-27B: clean slop but 7/30 explicit
refusals + needs the newer ggml-org/llama.cpp image (Qwen3.6 won't load on b8840).
Snowdrop + Gembrain (Gemma-4): floored (llama.cpp can't manage their reasoning without
the vetoed template hacks). Losers kept on disk as alternates.
- char-rp (Magidonia) unchanged; gen unchanged. gateway char-rp-reasoning -> Deckard
sampler (temp 1.0/top_p 0.95/top_k 40/min_p 0.05; DRY server-side).
dvalin-smithy offered a follow-up sampler pass for char-rp / char-rp-reasoning after
they accumulate real Worldtree/SillyTavern character-role traffic. Parked as a future
option (thread 01KX1DS6...) — nothing to tune until there's live-session data.
char-rp -> TheDrummer Magidonia-24B-v4.3 Q6_K (Magistral prose, ~65 tok/s,
zero refusal, tight POV) via llama.cpp (:8016).
char-rp-reasoning -> ArliAI QwQ-32B-RpR-v4 Q5_K_M (abliterated managed reasoning,
~52 tok/s, reasoning surfaces in reasoning_content) via llama.cpp (:8018).
- New canonical stack stacks/char-rp-gguf/ (llama-server x2, GPU0-pinned, ~86/97G
co-resident with gen). GGUF sidesteps the vLLM-NVFP4 + Mistral-tokenizer traps that
killed the Angel serve. Never Ollama.
- Best-of-breed per seat: no single dense 24-32B is both an elite non-thinking prose
seat AND a clean managed-reasoning seat on llama.cpp (Magidonia [THINK] boundary is
loose; Cydonia-R1 <think> runs away; QwQ is template-managed). Pantheon-Reasoning-27B
stays rejected (re-censors in <think>; RpR-v4 abliterated reasoning is the fix).
- Gateway rewired: char-rp->:8016, char-rp-reasoning->:8018, Mistral/QwQ samplers,
dropped the Qwen enable_thinking kwarg. One-model Magidonia fallback documented.
- Retired the ms32-24b-angel stack.
comfy-dev's explicit-over-implicit call: arbo now sends train_id, so the
worker no longer derives the loras/trained/{train_id}/ namespace from
output_dir.parent (which coupled it to arbo's handoff layout). train_id is
optional + path-safe-validated; when present it wins, else the path
derivation remains as the fallback. Wired through TrainRequest ->
validate_request -> published_relative_path -> _publish_lora. 18 tests green.
On a train reaching succeeded, IN ADDITION to output/{name}.safetensors
(unchanged download source), COPY it into ComfyUI's loras search path at
/storetank/arbo/models/loras/trained/{train_id}/{name}.safetensors and
return published_lora_name (the ComfyUI-relative LoraLoader string) in the
terminal GET /train/{id} payload (arbo Phase 2 auto-registration, §4.1/§7).
- Copy not move; a publish failure NEVER fails the train (keeps succeeded,
omits published_lora_name, logs the reason to the tailable run log).
- INV-T7-safe: a copy to a fixed computed path, no new free-form args.
- train_id derived from the handoff layout (output_dir.parent.name).
- Provisions loras/trained/ (arbotrain 2775, group-write per the Phase-1
lesson; world-readable/traversable for ComfyUI) via the deploy playbook.
- ComfyUI verified to resolve nested loras subfolders (no flat fallback).
- Pure path helper unit-tested; 16 tests green.
The first real arbo train 422'd: SDXL checkpoints live at
/storetank/arbo/models/checkpoints/ (the 2026-06-13 move to the 1.8TB
/storetank volume), which wasn't in ALLOWED_MODEL_ROOTS — the old roots
predated the move (/worktank/models is gone, /worktank/comfyui host path
is empty; ComfyUI mounts /storetank/arbo/models -> /basedir/models inside
its container). Allowlist /storetank/arbo/models (llmuser-readable,
world-readable tree), drop the two stale roots. Regression test added (15 green).
comfy-dev cross-check: Sindra v1/v2 used alpha=dim/2 (0.5 LoRA scaling),
which produced the validated likeness; the initial alpha=dim (1.0) was a
stronger, unvalidated default. Align the default to the proven value
(operator/per-request can still override). Tests updated (14 green).
Host service (runs as llmuser, owns /opt/fluxgym + GPU access) that runs
sd-scripts SDXL LoRA training on demand for arbo — the infra-ops half of the
in-arbo LoRA training Phase 1 ownership split (vh/arbo
docs/contracts/in-arbo-lora-training-phase1.contract.md §4.1/§2).
- Fixed-invocation only (INV-T7): bounded params -> one sd-scripts command
shape; every param range/allowlist/path-containment checked before spawn;
bad request = 422, never a silent downgrade. 14 unit tests green.
- Thin supervisor: never imports torch; subprocesses the fluxgym venv's
accelerate. 1-job-at-a-time (arbo lease is the serializer, 409 is backstop).
Durable job records + boot reconciliation (§4.3).
- API: POST /train, GET /train/{id}[/log], POST /train/{id}/cancel,
GET /gpu-status (per-device VRAM + tts_on_3090 co-OOM signal), GET /healthz.
- Wire-shape (§7 resolved with comfy-dev): shared /worktank/arbo/train handoff
(group arbotrain, setgid 2770); worker binds 0.0.0.0:8203, arbo reaches via
host.docker.internal:host-gateway (reachability proven on 172.20.0.1:8203);
device-aware TTS steering via /gpu-status.
Deployed to irv-ml1 via playbooks/deploy-lora-training-worker.yaml (elway,
idempotent); systemd unit active; /healthz + /gpu-status verified live.
vh