serverAppConfigs/AGENTS.md

AGENTS.md — serverAppConfigs Agent Rules

Overview

This repository is the single source of truth for the two homelab servers and the FrankBook-Pro Mac mini. It tracks the Podman container stacks and the LLM inference fleet. One repo, one checkout per host, one deploy script. All services run from /home/dank/serverAppConfigs on each host (the 2026-08-22 merge absorbed the former vllmScripts repo, history preserved).

• Dankserver (192.168.0.235) — media/library stack (RomM, Copyparty, Calibre-Web, Jellyfin, Plex) plus the frank-primary LLM service.
• Dankclaw (192.168.0.143) — AI/search/dev/observability stack (SearXNG via Gluetun, ComfyUI, OpenWebUI, Hermes, Gitea, Grafana, Prometheus, Loki, Alloy, Pi-hole) plus TWO LLM services: frank-large (halogen container on the gfx1151 iGPU) and frank-small (llama.cpp CUDA on the RTX 3090). Caddy on Dankclaw reverse-proxies both hosts.
• FrankBook-Pro (192.168.0.112) — macOS mini running the *arr stack (Prowlarr, Sonarr, Radarr, Lidarr, Readarr, qBittorrent) as quadlets inside its podman machine; all egress rides the Gluetun WireGuard netns. Deployed via the frankbook-pro handler.

Fleet topology (ACTIVE)

Target	Service	Runtime	Model	Port
Dankserver	frank-primary	tcclaviger RDNA4 vLLM fork (docker.io/tcclaviger/vllm@sha256:da4029666…, digest-pinned, plain podman run, root systemd unit)	orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8 (~172.8 GiB, 131 shards), TP=2, expert offload automatic (fork plans it), prefetch OFF, fp8 KV (--kv-cache-dtype fp8), no speculative decode, MNBT 2048 (arena-free: the registry prices overhead flat up to its nbt 4096 anchor), MoE gather 32×64, --expert-precise-memory (fitted 5.25 GiB overhead, not the 7.60 GiB default), VLLM_CUDAGRAPH_MEM_BASE_MB=0 / _PER_GRAPH_MB=12 (the plan's CUDA-graph term priced at the measured capture — 0.31 GiB then a NEGATIVE −0.06 GiB delta, so the 256 MiB analytic base was fiction → graph term 0.07 GiB), GPU_MAX_HW_QUEUES=2, mounted cache dir at /root/.cache (the image ships no /root/.cache at all; unmounted, every boot re-pays torch.compile took 181.20 s), 3 slots × 200704 ctx at KV_BYTES 3,166,500,000/rank (~400,700 KV tokens at the engine's 7,891.8 B/token here) — the pool, not the slot count, caps resident conversations (measured live: GPU KV cache size: 401,408 tokens, Maximum concurrency for 200,704 tokens per request: 2.00×; the 3.15e9 pool below it measured 398,658/1.99×), and shorter requests use the spare blocks. --max-num-seqs must stay 3: SEQS=2 cannot boot this model — torch.fx ConstraintViolationError on query_start_loc, measured at BOTH MNBT 1024 and 2048. util stays 0.982 BY LAW: at le=1 the worker init refuses budget == total with ValueError: Free memory on device cuda:N (31.29/31.86 GiB) on both ranks before a single weight loads — measured crash loop 2026-10-03. Arena at shape G: 17.77 GiB / 7,762 slots measured, KV 401,408 tokens / 2.00× concurrency, 0.89 GiB of real free VRAM per card held through a live 4,096-token generation (the ~3.5 GiB of driver-reported free VRAM at shape E was estimator slack — the registry's 9190 B/token KV reserve vs the real 7,892, and the 5.25 GiB overhead vs far less — NOT allocatable), and the client contract wired: `--reasoning-parser …
Dankclaw	frank-large	halogen container (ghcr.io/peonist-ai/halogen@sha256:dc0a39a0…, digest-pinned, plain podman run via scripts/serve-frank-large.sh, root systemd unit)	peonist-ai/halogen-qwen3.8-27b qwen3.8-27b-p1w4d-d2.hgn (35,865,565,184 B; the .hgn format loads in NO other engine — llama.cpp, vLLM and transformers all refuse it), DFlash2 spec decode (HALOGEN_DRAFTER=2), 1 slot × 200,700 ctx — HALOGEN_KV_SLOTS=1 is load-bearing: batching XOR speculation, slots>1 silently disables the drafter (measured: 8 slots at serial ~6 t/s/stream), d2 KV (2× fp8, 64 KiB/token → 12.2 GiB pool, engine sizes it and prints a boot ledger naming any clamp), TEXT-ONLY (the engine ignores the vision encoder — vision is frank-small's job), censored base (the ROCmFPX AGENT rollback is the refusal escape hatch), no server-side sampler: absent temperature is GREEDY, so the card's thinking recipe (temp 1.0 / top_p 0.95) rides the wire in models.yml. Publisher card 31.7 t/s mean across eight text workloads vs the retired ROCmFPX endpoint's measured 27.6–28.1. Prompt over the slot AND max_tokens over HALOGEN_MAX_TOKENS_CAP are both 400s naming the limit, never truncations. Serve passes NO --ulimit nofile (rootless runc cannot raise Dankclaw's 524288 hard limit — see Key Lessons). Cutover 2026-10-05	7770
Dankclaw	frank-small	llama.cpp master, CUDA sm_86 build (root systemd unit, relinked)	bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF Q6_K (9B dense qwen3_5 hybrid, 7.26 GiB GGUF, no MTP weights so no spec decode) + the repo's f16 mmproj for vision, on the RTX 3090, 2 × 200,704 (pool 401,408 — NOT the native window) with fully unquantized KV (f16 K + f16 V) under flash attention, batch 4096 / ubatch 2048 (doubled 2026-10-05: measured +4.6 % prefill at 8,928 tokens for 616 MiB), 21,846 used / 2,279 free of 24,576 MiB, 94.4 tok/s decode. Resized 2026-10-03 (3 × 180,224 + q8_0 V → 2 × 200,704 + f16 V → + Q6_K rung: the freed slots paid for exact values AND the higher quant; the rung costs a measured −10 % decode — weight-bandwidth-bound); rollback MODEL_FILE=…/MiMo-Q5KM/…Q5_K_M.gguf (buys the 10 % back) or CTX_SIZE=540672 PARALLEL=3 CACHE_TYPE_V=q8_0 (Q5 only)	7771

Retired endpoints (frank-npu, and every prior runtime) are in the runbook index. The full swap history lives in serverAppConfigs-docs/LLM/index.md — do not grow it here.

Repository Layout

• dankserver/ — compose.yaml + Caddyfile + .env.example for the media stack
• dankclaw/ — compose.yaml + Caddyfile + .env.example for the AI stack
• shared/ — homepage, Prometheus, Loki, Alloy, SearXNG configs
• scripts/ — deploy.sh / deploy-plan.py / deploy-remote.sh, LLM serve scripts, systemd units, env tunables, bench harnesses
• old/ — retired LLM scripts (reference only)
• serverAppConfigs-docs/ — knowledge base (Obsidian vault, OKF format)
• tests/ — bats suite + Playwright e2e

Git Workflow

All edits flow through git. Never edit files on a remote machine directly.

1. Source the live config from the target machine
2. Commit the fetched version here (redact secrets into .env)
3. Edit locally
4. Test — ./tests/run.sh must pass
5. Commit and push — origin is Gitea on Dankclaw
6. Pull on the target (deploy.sh does this)
7. Deploy — scripts/deploy.sh plans and dispatches only changed services

Deployment

./scripts/deploy.sh dankclaw
./scripts/deploy.sh dankserver
./scripts/deploy.sh dankclaw --service searxng
./scripts/deploy.sh dankserver --service frank-primary
./scripts/deploy.sh dankclaw --dry-run

The planner maps changed paths to services; the remote runner restarts only the selected service and records its revision outside the checkout. Documentation-only changes are a no-op. LLM services are selected only by their own files — every endpoint (frank-large/frank-small/frank-primary) by its unit/env/serve trio in scripts/ — so unrelated container changes cannot interrupt inference. Set FRANK_SUDO_PW for headless runs on a host without passwordless sudo; the value never enters the repo.

deploy.sh <target> with no --service is a fleet-wide DRAIN, not a sync. It also walks every per-service revision marker under the deploy-state dir and restarts each service whose marker is behind HEAD and whose paths changed since that marker — so a backlog accumulated by service-scoped deploys fires all at once, and one failing health gate aborts the run with every later service left pending. Name the service unless a drain is what you want (./scripts/deploy.sh dankserver --service frank-primary).

Deployment completeness (MANDATORY)

A deploy is not complete when deploy.sh exits green. It requires:

1. ./tests/run.sh — all seams green, and
2. the Playwright e2e suite (tests/e2e/run.sh) — runs automatically as the [3/3] gate at the end of deploy.sh against the LIVE fleet; a behavior-affecting deploy that fails it exits DEPLOY INCOMPLETE. Bypass only with --skip-e2e in an emergency.

graph LR
    A[Edit locally] --> B[Tests pass] --> C[Commit + push]
    C --> D[deploy.sh plan] --> E[Target pulls ff-only]
    E --> F[Dispatch selected service] --> G[Per-service health]
    G --> H[Playwright e2e gate] --> I{green?}
    I -->|yes| J[Deploy complete]
    I -->|no| K[DEPLOY INCOMPLETE]

Boot Persistence & Self-Healing (MANDATORY)

Every deployed, working Podman container MUST have a self-healing, boot-persistent Quadlet unit. A container that runs only from compose.yaml or a manual podman run dies silently at reboot — rootless Podman does not auto-start bare containers, and restart: unless-stopped does not survive the explicit stop that shutdown performs. Never report such a service as "up".

Unit requirements:

• Boot persistence — [Install] WantedBy=default.target, installed to ~/.config/containers/systemd/<name>.container, enabled in the user manager, with linger on for the deploy user (loginctl show-user dank -p Linger → Linger=yes).
• Self-healing — Restart=always (or on-failure) with a short RestartSec; quadlet health check where the service has one.
• Repo provenance — the .container file lives in this repo (<host>/<svc>/quadlet/), is installed by the deployer, and is covered by the quadlet test seam. The host-side unit is generated state; the repo file is the source of truth.

Before declaring a service deployed: systemctl --user is-enabled <name> → enabled, is-active → active, and the container answers its health check. A running container with no unit is a gap: write the unit, test it, deploy it.

This rule exists because reboots found gaps twice (2026-08-26: Caddy as a bare compose container never returned; 2026-09-11: five observability services had no units). The observability services now have units under dankclaw/observability/quadlet/ and their compose definitions were removed — one owner per container name. The stack is configured and delivering as of 2026-09-17: a Grafana datasource and the llm-fleet dashboard are provisioned from shared/grafana/provisioning/, shared/prometheus/rules/ llm-fleet.yml loads 5 alert rules + 1 recording rule, and firing rules are delivered: Prometheus → Alertmanager (dankclaw/observability/quadlet/ alertmanager.container, loopback :9093) → host-side Signal relay (scripts/signal-relay.py, user unit, 192.168.0.143:18093 — the host LAN IP, the one bind a container can reach as host.containers.internal; iptables admits loopback, the host and the podman bridge subnet only) → signal-cli account 318311. Prometheus targets read down for structural reasons (HTML-scraped "probes", Gitea metrics off); the remedy is a blackbox exporter. See serverAppConfigs-docs/Infrastructure/Observability.md.

Testing

Tests are mandatory before commit (bats required on the dev machine):

./tests/run.sh

Seams: compose validity, env consistency (no leaked secrets), Caddyfile validity, quadlet units, docs completeness including vault link integrity, script syntax (shell + python), LLM unit integrity, LLM secret hygiene, deployment flow, service-scoped planning, homepage config/pages/deploy/checker, and e2e_scaffold — which fails if a Playwright spec file's number prefix is in no project, because such a spec silently never runs. See serverAppConfigs-docs/Architecture/Testing Policy.md.

Secret Rules

• Never commit real secrets. The live Gitea DB password and sudo password are test-enforced leak checks.
• .env files are git-ignored; commit only .env.example with placeholders. Exception: scripts/*.env LLM tunables (paths and numbers, never credentials).
• Do not document passwords anywhere in the repo. SSH keys are the auth mechanism; FRANK_SUDO_PW covers headless sudo and never enters the repo.
• A token passed as curl -H "Authorization: Bearer …" is world-readable. It lands in /proc/<pid>/cmdline, so any ps/pgrep -a on the host prints it — that leaked an HF token into a session transcript on 2026-09-16, while merely checking whether a download was still running. Pass secrets out of band (curl -K <mode-600 file> carrying the header, or HF_TOKEN in the environment for hf/hf_transfer) and never pgrep -a a fetch command that carries an auth header.

External Exposure Policy (MANDATORY)

Every Caddy-proxied service is either behind the auth chain (Anubis :8923 → Authentik embedded outpost) or explicitly internal-only — no public DNS name, no vhost. New services default to internal-only until guarded.

When guarding a new vhost — provider first, then route:

1. Create the outpost proxy provider FIRST (external_host = exact vhost URL with trailing slash; internal_host = http://caddy:8017). If the vhost flips first, authenticated traffic hits the outpost error page.
2. Flip the vhost to reverse_proxy anubis:8923 with header_up X-Real-Ip {remote_host}, and add the post-auth @<svc> host <vhost> handler on the http://:8017 listener routing to the real upstream.
3. Anubis guard set, Caddy TLS edge, and outpost provider list change together in one deploy.
4. Service-to-service traffic targets internal addresses (loki:3100), never the public hostname — non-browser clients cannot solve challenges.

Exempt categories (document the reason in the vhost comment): the IdP itself (auth.), git smart-HTTP endpoints, services with their own complete auth.

Deletion & Trash Policy (MANDATORY)

Never rm. Move deletions into trash so they stay recoverable:

1. In-repo files → .AGENTTRASH/ beside the item's directory.
2. Elsewhere → dated staging dir ~/.AGENTTRASH/YYYYMMDD-topic.
3. List staged items and sizes; wait for explicit user approval to purge.

rm, rm -rf, and git clean -f are forbidden without explicit permission. git * --no-verify is forbidden. The trash is a temporary hold, not an archive.

Benchmarking Policy (MANDATORY)

STOP the production service BEFORE benchmarking — these servers serve real users, and a bench beside a live service starves the GPU and corrupts results. Use the runtime's own harness: llama-bench for the llama.cpp endpoints; a live llama.cpp server also reports honest per-request numbers in the /completion timings object (prefill/decode rates, draft_n, draft_n_accepted), which is how frank-large and frank-small were measured. Rolling your own bench requires asking first. Scratch ports are not used on LLM hosts — a test process cannot share the single GPU anyway; validate against the service itself.

(A gufo-era law; frank-large ran ROCmFPX llama.cpp 2026-10-02 → 10-05 and is the halogen container since, but the rollback tags keep it live.) gufo bench cannot rank speculative-decode settings, and will invert the ranking. Measured 2026-09-28 on frank-large at depth 32768: its corpus reports acceptance=0.933 ± 0.000 with an identical trace sha256 across draft widths, i.e. the same highly-predictable text every run, and 93 % of every proposal accepted. It reported 29–37 t/s decode where real traffic gives ~15 t/s, and it made --draft-tokens 16 look 6 % faster than 7 when real text made it 14 % slower. Use it for shape, not for A/B. For an A/B, replay real text through the live server and read the Server-Timing header — but note the unit rides in the name half (ttft;dur=402.485), so splitting on = alone keys the dict ttft;dur and every lookup silently returns nothing. Generate at least 128 tokens for the measurement: at 8 tokens, inter_token is first-token noise and the run-to-run spread swamps every real difference.

This covers any GPU experiment, not just benchmarks — including a bare 2-rank RCCL/torch.distributed probe. On 2026-09-16 one such probe was run against the live service "because it only needs 200 MB of VRAM"; the host went hard down ~20 minutes later (no ARP reply, 100% ICMP loss from two subnets, no recovery — a power cycle was required). A hung AQL dispatch wedging the GPU driver is the plausible mechanism, and this host has precedent for going unreachable until a power cycle. If an experiment needs GPUs, stop and mask first, or use a host that serves nobody. Telling an off machine from a wedged one: ip neigh show <ip> from a peer — FAILED means off/NIC dead, while a merely wedged kernel usually still answers ARP.

1. Stop and mask (self-healing would restart it):
   • Dankclaw frank-large (root unit, halogen podman on the iGPU): sudo systemctl stop frank-large && sudo systemctl mask frank-large
   • Dankclaw frank-small (root unit, llama.cpp on the 3090): sudo systemctl stop frank-small && sudo systemctl mask frank-small
   • Dankserver frank-primary (root unit, tcclaviger vLLM on the 2x R9700): sudo systemctl stop frank-primary && sudo systemctl mask frank-primary
2. Bench. NOTE: systemctl mask FAILS on a relinked unit (the unit file in /etc/systemd/system/ is itself a symlink into the checkout — mask refuses to replace it). Stop is enough for the bench window; the deploy restores the unit, so re-deploy after benching instead of unmasking.
3. Unmask and start; verify curl http://192.168.0.<host>:7770/health.

CRITICAL: never systemctl disable a system unit on these hosts — it deletes the unit file from /etc/systemd/system/, not just the symlink. mask symlinks to /dev/null and is safe/reversible.

Model Config (harness)

The harness reaches the fleet through .omp/agent/models.yml in the operator's home — /home/blank/.omp/agent/models.yml on the dev box, which is where the harness runs; there is no copy on the servers. Each provider is one HOST:PORT; each model id is the served alias: frank-primary → 192.168.0.235:7770, frank-large → 192.168.0.143:7770, frank-small → modelRoles.

Resolution rules: the default model resolves as a defaultProvider + defaultModel pair (the provider must serve the model); a subagent override uses a bare model id, which resolves only when unique across providers. Two ports on one host need two providers — the file cannot express "same host, different port".

When renaming a service, retiring an endpoint, or moving a model: update the provider block, model id, contextWindow, maxTokens, input in models.yml, then fix every modelRoles entry that named the old endpoint — a role pointing at a provider the file does not define fails at call time. As of 2026-10-05 two of the three endpoints accept images (input: [text, image]) — frank-large went TEXT-ONLY with the halogen swap (the engine ignores the vision encoder; the BF16-mmproj era ended 2026-10-05) — while frank-primary takes them through a BF16 mmproj loaded explicitly (this build never autoloads projectors), and frank-small through mmproj-MiMo-V2.6-Distill-Qwen-9B-f16.gguf passed as --mmproj — the MiMo GGUF itself carries no vision tensors. frank-small's contextWindow is 200704 (resliced from 180224 on 2026-10-03), the per-slot window, not the model's native 262144: the harness must compact before llama.cpp answers a longer prompt with a 400. Note also that the reasoning channel differs by runtime: llama.cpp AND SGLang (qwen3 parser, live-verified) emit reasoning_content, vLLM emits reasoning — models.yml sets reasoningContentField per endpoint to match. There is no audio input anywhere in the fleet: that capability left with the FLM NPU endpoint on 2026-09-11, returned 2026-09-17 with the Gemma4 projector's audio encoder, and left again with the Gemma4 retirement on 2026-10-01 (the MiMo tower is vision-only).

frank-large — halogen container, Qwen3.8-27B + DFlash2 (Dankclaw, ACTIVE)

The halogen engine (ghcr.io/peonist-ai/halogen@sha256:dc0a39a0…, digest-pinned, published tag 0.1.4) as a podman container driven by scripts/serve-frank-large.sh under a root systemd unit (scripts/frank-large.service, User=dank, Restart=always, ExecStop=podman stop -t 120 frank-large-halogen, the script exec podman wait so the restart policy is real) on port 7770 from the Strix Halo iGPU. Engine stdout goes to /home/dank/logs/frank_large.log — the boot ledger (KV pool, drafter state, any clamp) lives there, not in the journal. Serving peonist-ai/halogen-qwen3.8-27b: qwen3.8-27b-p1w4d-d2.hgn (35,865,565,184 B ≈ 33.4 GiB, sha256 274c3fc7…, 1352 tensors; p1 weights / w4 activations / d2 KV) + the flat tokenizer/ directory (5 files). Cutover 2026-10-05 from the ROCmFPX llama.cpp fork (frank-large-rocmfpx-q6agent-uncensored-mtp, measured 27.6–28.1 t/s prose), chosen for the publisher's 31.7 t/s DFlash2 mean.

The format is the lock. .hgn loads in nothing else — the runtime is chosen by the FILE, as ROCmFPX's tensor types were before it. This is NOT the 2026-09 halogen-flash-server image (different repo, Flash-Next w4b era); the halogen-server repo restarted versioning at 0.1.x.

Two product regressions, both accepted by Frank before deploy: the endpoint is text-only (README honest limits: the vision encoder is ignored — frank-small owns images now) and the checkpoint is a finetune of the censored Qwen/Qwen3.8-27B base (refusals can return; the rollback is one commit-revert away with the AGENT payload still on disk).

Engine laws that shape the config (all in scripts/frank-large.env with evidence): HALOGEN_KV_SLOTS=1 — batch XOR speculation, slots>1 turns the drafter off silently. HALOGEN_SLOT_CTX=200700 — the per-conversation HARD ceiling (400 naming the limit, never truncation); models.yml contextWindow: 200700 must match or the harness discovers the wall as a 400. HALOGEN_MAX_TOKENS_CAP=65536 + HALOGEN_QUEUE_TIMEOUT=7200 are a coupled pair; the harness maxTokens: 57344 stays under the cap (over-cap is a 400 naming the limit — README, never a 413). Thinking tokens COUNT against the budget: exhausted mid-thought is finish_reason: "length" with empty content and the reasoning in reasoning_content. HALOGEN_DRAFTER=2 = DFlash2. No server-side sampler: an absent temperature is greedy, so the thinking recipe exists only in models.yml extraBody. response_format is unsupported (no constrained decoding). Thinking lands in reasoning_content (harness field unchanged); the same Qwen3.8 template rejects reasoning_effort outside xhigh|medium|low, so the models.yml fold stays mandatory.

Fetch: scripts/download-frank-large-halogen.sh into /home/dank/models/halogen-qwen3.8-27b/ (-f + byte count + LFS sha256 + --verify-only, no token on the command line; resume goes against the RESOLVED CDN URL because curl drops Range across HF's 302). The deployer and serve script both refuse a missing tokenizer file or any symlink — an HF cache snapshot symlinks each entry into blobs/ and those links dangle inside the container mount.

Runbook: serverAppConfigs-docs/LLM/frank-large-halogen27b-dflash2.md. Rollback: revert the swap commit and redeploy — the ROCmFPX payload (/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/), the fork build and scripts/build-rocmfpx.sh (pin 36e9acd) all remain on disk, seam-tested. The older gufo tag frank-large-gufo-q4kxl-20261001 stays a re-download rollback (payloads removed 2026-10-03; scripts/build-gufo.sh stays).

frank-small — llama.cpp on the RTX 3090 (Dankclaw, ACTIVE)

llama.cpp master (CUDA sm_86 build /home/dank/GitPackages/llama-master/build-cuda/bin/llama-server) as a root systemd unit (scripts/frank-small.service, relink_united into /etc/systemd/system/), serving bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF at Q6_K (7,793,710,624 B, sha256 ef96d05a…) plus mmproj-MiMo-V2.6-Distill-Qwen-9B-f16.gguf (918,166,048 B, sha256 ff348f…), both /home/dank/models/MiMo-Q6K/ (the projector is a hardlink of the Q5KM copy — the dir is self-contained), on port 7771. The rung went Q5_K_M → Q6_K on 2026-10-03 the same morning as the reslice, funded by the freed slots. The runtime arrived 2026-10-01 night for one reason: asymmetric KV — --cache-type-k f16 + --cache-type-v q8_0 is the fleet's only per-tensor KV lever, and SGLang's --kv-cache-dtype moves K and V together. On 2026-10-03 Frank spent the headroom that lever bought: the reslice from three slots to two made room for fully unquantized V (CACHE_TYPE_V=f16); both flags stay explicit in the serve script. Vision came 2026-10-02 via --mmproj: the GGUF carries no v.* tensors, so the projector file is the entire vision path (and ~1.4 GiB of the card).

Architecture: qwen3_5 hybrid — 32 layers, 8 full-attention + 24 Gated DeltaNet, native 262,144 window. KV exists only in the 8 attention layers (4 KV heads × 256 dim × 2 B ≈ 2 KiB/token/layer, ~16 KiB/token for f16 K+V); the DeltaNet state is a fixed per-slot buffer, so slot count — not context — drives that term.

The shape (2026-10-03): --ctx-size 401408 --parallel 2 = TWO conversations of 200,704 tokens each (llama.cpp divides the pool: n_ctx_slot = CTX_SIZE / PARALLEL), --flash-attn on (the seam pins it), f16 K + f16 V, all layers on the GPU, --mmproj for vision, and since 2026-10-05 batch 4096 / ubatch 2048 (doubled; the boot-pair A/B is in the runbook). Measured trade at 8,928 prompt tokens: prefill 4,024 → 4,209 t/s, +4.6 %, decode flat (91.9–93.0 vs 93.8–94.3 — noise), and a 279-token prompt, which sits below either ubatch, did not move at all. The old "ubatch 2048 failed the scratch boot" note belongs to the Q5_K_M predecessor and does NOT generalize: this shape boots 2048 clean. Measured VRAM 21,846 used / 2,279 free of 24,576 MiB against 21,276 / 2,849 for the 1024 arm, reproduced byte-identically the same hour — the doubling costs 616 MiB and lands at the BOTTOM of the 2–3 GiB floor Frank asked for (4096/1536 is the middle rung if that margin is ever wanted; non-KV 9,302 MiB, up from 8,732 by exactly the compute buffers; f16 K+V is 32 KiB/token across the 8 attention layers; quote nvidia-smi's free field, never total − used — the driver holds ~450 MiB of reserve the subtraction hides). The 2026-10-02 3 × 180,224 + q8_0-V shape stays one env edit away (CTX_SIZE=540672 PARALLEL=3 CACHE_TYPE_V=q8_0, and it must pair with Q5_K_M — Q6_K breaks the floor there). Three native windows need 18.4 GiB of KV alone and were never on the table.

Verified at deploy (2026-10-03): boot logs n_slots = 2, n_ctx_slot = 200704 and loaded multimodal model; decode 94.44 tok/s at temp 0 on the probe that measured 104.84 on Q5_K_M — the Q6_K rung costs a measured −10 %: decode on this card is weight-bandwidth-bound and Q6_K streams ~10 % more bytes/token; health 200, e2e 33/33; native tool_calls; thinking in reasoning_content; a 512×512 solid-red PNG answers Red through an image_url content part (two traps: the legacy top-level images field is ignored, and a TINY probe image undersupplies image tokens — a 10×10 PNG made the model claim "it's a video"; probe with ≥256 px). A prompt over the slot is a 400 exceed_context_size_error, which is why models.yml sets contextWindow: 200704.

Runbook: serverAppConfigs-docs/LLM/frank-small-llama-mimo-q6k-2x200704.md (the boot-pair table that prices pool changes lives in the retired frank-small-llama-mimo-q5km). Rollback: MODEL_FILE=/home/dank/models/MiMo-Q5KM/MiMo-V2.6-Distill-Qwen-9B-Q5_K_M.gguf buys back 875 MiB and the −10 % decode rate (3,571 free, 104.84 tok/s, both measured); the older shape CTX_SIZE=540672 PARALLEL=3 CACHE_TYPE_V=q8_0 pairs with Q5 only; SGLang rollback in frank-small-sglang-mimo-v26 (venv on disk at /home/dank/GitPackages/sglang-cu128/, INT8 payload staged to ~dank/.AGENTTRASH/20261002-retired-llm-payloads/).

frank-primary — tcclaviger vLLM, Qwen3.8-Flash-Next FP8 (Dankserver, ACTIVE)

The tcclaviger RDNA4 vLLM fork (docker.io/tcclaviger/vllm@sha256:da4029666…, digest-pinned) serving orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8 (~172.8 GiB, 131 shards) on both Radeon AI PRO R9700s (gfx1201), TP=2, fp8 KV, no speculative decode, expert prefetch OFF, and since 2026-10-03 the shape-G geometry (shape E + MNBT 2048 + the CUDA-graph plan term at the measured floor): --max-model-len 200704 × 3 slots with --kv-cache-memory-bytes 3,166,500,000/rank (~400,700 KV tokens — the engine prices this pool at 7,891.8 B/token here), util 0.982 (the law — see the unbootable-1.0 lesson below), --expert-precise-memory, VLLM_CUDAGRAPH_MEM_BASE_MB=0 / _PER_GRAPH_MB=12, GPU_MAX_HW_QUEUES=2 and a mounted /root/.cache. models.yml carries contextWindow: 200704 for this endpoint — the served window, not the model's native 262144, so the harness compacts before vLLM answers a longer prompt with a 400.

--max-num-seqs 2 is impossible on this model. Measured 2026-10-02 by single-variable boot probe against the then-deployed shape: SEQS 2 fails at BOTH MNBT 1024 and 2048 with torch.fx.experimental.symbolic_shapes.Constraint ViolationError: Constraints violated (L['query_start_loc'].size()[0]) … specialized it to be a constant (3); SEQS 3 passes at both. model_runner.py sizes that input buffer max_num_reqs + 1 and _dummy_run sets num_reqs = min(max_num_batched_tokens, max_num_reqs), so the width is SEQS + 1 for any MNBT ≥ SEQS — MNBT only moves the compile RANGE ((1,1024) vs (1,2048)). --enforce-eager bypasses the path but drops CUDA graphs. The same failure class killed the 2026-10-01 bf16 2×262144 shape.

The expert arena has THREE reachable levers: max_model_len, the overhead term, and the CUDA-graph estimate. The fork's planner sizes expert cache as budget − non-expert − KV reserve − graphs − overhead, and the KV reserve is registry-priced at 9190 B/token × max_model_len: 262144 → 2.48 GiB reserve → arena 14.30 GiB / 6,248 slots; 200704 → 1.91 GiB → 14.87 GiB / 6,496 slots. Δreserve 0.57 GiB = Δarena 0.57 GiB. The 6.1e9 → 3.15e9 pool cut bought zero arena; it buys runtime headroom (the shape it replaced died with 0 bytes free on a real 995-token request). The second lever is --expert-precise-memory, ON since 2026-10-02: it swaps the registry's default overhead (7.60 GiB = fitted + compile/graph-capture headroom) for the fitted figure, and the 2.35 GiB difference lands in the arena — measured plan lines at 200704 × 3: overhead 5.25 GiB, expert cache 17.22 GiB / 7,522 slots (+1,026 / +16 %), swap region 13.7 GiB / 5,986 slots, boot healthy in 420 s. The pool stayed at 3.15e9 deliberately: because the pool is not an arena lever, cutting it would cost concurrency (a 2.2e9 probe reports 277,686 tokens / 1.38×) and buy no slots. The arena (2.344 MiB per slot) is the decode-critical region — cutting it for KV headroom collapses the offload hit rate.

Free VRAM on this endpoint is NOT arena headroom (2026-10-03 audit). rocm-smi reads ~3.5 GiB free per card after hours of real traffic, but the planner charges budget = total × util against terms that overstate reality: the KV reserve is registry-priced at 9190 B/token (1.91 GiB at 200704) while the real pool bills 2.93 GiB at 7,892 B/token, and the precise overhead (5.25 GiB) is far above the idle-measured figure. So the visible free space is the estimator gap, not allocatable arena. gpu_memory_utilization is a real arena lever (each +0.01 = +0.32 GiB) but 0.982 is the law, and 1.00 is UNBOOTABLE on this fork — measured 2026-10-03: the worker init demands free > budget, and at le=1 budget == device total, so both ranks raise ValueError: Free memory on device cuda:N (31.29/31.86 GiB ...) before a single weight loads, twice (2-restart crash loop, zero payload progress). The CLI field accepting le=1 is the trap; Frank's 0.982 is where it always ships. What IS reachable from flags: VLLM_CUDAGRAPH_MEM_BASE_MB / _PER_GRAPH_MB, read at call time by estimate_cudagraph_memory() — the analytic default (256 MiB + 52 MiB/graph) charged 0.55 GiB where the engine's own post-capture line reported 0.31 GiB, and a second boot (with the plan already charging 0.32 GiB) reported -0.06 GiB, so the 256 MiB base pool is pure over-charge: shape G charges 0 base + 12 MiB/graph = 0.07 GiB, which still prices per graph for a future 18-graph spec-decode shape (0.21 GiB). Verify every deploy by comparing the Estimated CUDA graph memory line with Graph capturing finished … took X GiB; if capture exceeds the estimate, raise _PER_GRAPH_MB back toward 52. MNBT 1024 → 2048 costs ZERO arena: the registry overhead is base + 0.22 GiB × max(0,(nbt−4096)/1024) anchored at nbt 4096, so every rung ≤4096 charges the same 5.25 GiB — and the fitted figure was fitted AT 4096.

A container's compile cache does NOT survive podman rm -f, and on this endpoint that is 181 s per boot. Measured 2026-10-02: the image ships no /root/.cache at all (fresh container: No such file or directory); after a boot it holds /root/.cache/vllm at 1,154,714,333 B in 4,665 files plus /root/.cache/comgr at 229 MB, all in the ephemeral layer, and the engine prints torch.compile took 181.20 s in total (173.19 s of it one graph, Compiling a graph for compile range (1, 1024)). Every deploy, restart and reboot re-pays it. The serve script now mounts FRANK_PRIMARY_CACHE_DIR at /root/.cache — the parent, so the vLLM AOT / torch_compile tree (VLLM_CACHE_ROOT, whose default is /root/.cache/vllm) and the comgr HIP-kernel JIT cache both persist, and nothing in the image is hidden because the image has no such directory. The compile RANGE is part of the cache key, so an MNBT change costs one recompilation, not a permanent tax. Same class of fix as the recipe's /cache/{triton,vllm,inductor} volumes. Proven live 2026-10-02 on consecutive boots: 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) for compile range (1, 1024) from the cache, took 2.143 s / torch.compile took 3.97 s in total / init engine (profile, create kv cache, warmup model) took 41.13 s (compilation: 3.98 s). Time to /health is dominated by the 172.8 GiB payload load, so the visible win is the ~184 s of inductor/AOT build, not a proportional boot cut.

Measured 2026-09-29 tuning campaign (fp8 + MTP4 era, since retired): decode 37.4-40.6 tok/s, acceptance ~54 %; the 256k ctx arm measured 34.1-36.5; 2-way concurrency sums BELOW parity (25.49 vs 26.8 single) — slots are a capacity feature, not a throughput one. V-only q8 KV is NOT possible via config: stock --kv-cache-dtype is K+V together; a V-only quant needs a fork patch.

Unit and self-healing: root systemd unit (scripts/frank-primary.service, User=dank, LimitMEMLOCK=infinity), ExecStop=podman stop -t 120 frank-primary-qwen. The serve script drives podman run directly, gates the digest-pinned image, the payload's config.json and its exact 131-shard count and the GPU device nodes, clears stale container identities, then blocks on podman wait. deploy_llm_primary gates the same, waits out :7770, relinks/enables/restarts the unit and health-gates with the MainPID check. Boot cost ~7 min warm (the 172.8 GiB payload load; the ~184 s inductor/AOT build inside it is what the mounted cache now skips — see the compile-cache lesson below); health returns 503 while loading.

Runbook: serverAppConfigs-docs/LLM/dankserver-qwen38-fp8-expert-offload-tuning.md. Retired GLM experiment: serverAppConfigs-docs/LLM/frank-primary-glm53-flash-w4a16.md.

Key Lessons Learned

Durable operational lessons from the fleet's history. Full detail in the linked runbooks.

PCIe / multi-GPU

• RCCL collectives need PCIe atomics on every participating GPU. RCCL's kernels carry a hidden hostcall-buffer parameter; ROCr refuses to dispatch them when the device has no atomics, and the failure surfaces as hipErrorIllegalState ("the operation cannot be performed in the present state") from hipLaunchKernel, not as anything that names atomics. The value is consumed at driver probe time, so writing the AtomicOp Requester bit with setpci at runtime changes nothing. Check the chain: lspci -vv -s <gpu> and every upstream bridge must show AtomicOpsCap: Routing+ (a Routing- bridge is a hardware wall), and the endpoint's Device Control 2 (setpci -s <gpu> CAP_EXP+28.w) must have bit 6 set. AMD desktop platforms put one card on CPU root ports (atomics work) and the rest behind the chipset (usually not) — that asymmetry is what blocked frank-medium's TP=2. A one-card collective proves nothing: NCCL short-circuits at world_size=1.
• That wall was physical, and it was fixed on 2026-09-20 by a board swap. Both paths that put two cards on CPU root ports were: an M.2→PCIe x4 riser into a CPU-attached M.2 slot, or a board whose CPU lanes bifurcate to x8/x8. dankserver took the second: ASUS ProArt B850-Creator WiFi NEO, PCIEX16_1 + PCIEX16_2 at x8/x8 Gen5 on root ports 00:01.1 / 00:01.3, every bridge now Routing+ and both endpoints ReqEn+ — the exact precondition the old test plan named. Bandwidth was never the constraint (a 27B TP=2 allreduce at batch 4 is ≈10 MB per token-step, ≈1.4 ms at Gen4 x4 and under 0.4 ms at Gen5 x8). Proven on the same day. vLLM's R4D one-shot allreduce runs TP=2 here and frank-medium serves on it; llama.cpp's RCCL path was built too (GGML_HIP_RCCL=ON) and its tensor mode measured — real collective, slower shape. Note the collective choice matters: TheRock's RCCL builds lack gfx1201 device code (invalid device function), which is why R4D is the path vLLM uses at all. Evidence matrix and the measured three-arm verdict: Dual-R9700-Cooperation.
• When a GPU collective dies, read the runtime, not the framework. The message that identified the cause came from AMD_LOG_LEVEL=3 (Pcie atomics not enabled, hostcall not supported → AQL dispatch failed!), not from vLLM or RCCL. AMD_LOG_LEVEL costs one env var and prints the failing HIP call with its return code.
• The software path that puts both cards on ONE model without a collective: llama.cpp's layer split. On the old board vLLM could not do it in any of its parallelisms — TP needs an allreduce, PP needs the same p2p primitive, and both died on the card that could not dispatch an RCCL kernel. --split-mode layer in llama.cpp needs no collective: layers are distributed in order and only the boundary activation crosses PCIe, once per forward pass. Measured here: 1991 t/s prefill / 18.8 tok/s decode layer-split, against 272 / 26.3 tensor (butterfly fallback). This was what frank-medium ran until 2026-09-20, when the board swap made collectives legal and radiance TP=2 beat it on every axis (95.8 vs 45.7 t/s decode, 0.21 vs 0.72 s TTFT, 259 vs 72 t/s burst-4) — see frank-medium-radiance-mxfp4; the llama shape is the rollback target frank-medium-llamacpp-dflash2.
• Layer split is also the cross-vendor answer — and -ot is what cannot be combined with it. On Dankclaw (2026-09-16) a CUDA 3090 + Vulkan iGPU layer split ran a real two-device graph: -dev CUDA0/Vulkan1 -sm layer -ts 0.15/0.85 held 14,994 MiB on the 3090 and gave 27.2 tok/s with 268.5 t/s prefill at 64k depth (iGPU alone: 25.8 / 231.8). What does NOT work is per-tensor placement across devices: --override-tensor <group>=<device> sets has_tensor_overrides(), 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 is what 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 (ggml-backend.cpp:941, reported with the leaf's own op name (NONE)). =CPU is the only override target that works, because the CPU backend is last in the list. Pure Vulkan changes nothing here — the gate is the override, not the driver mix. See x4-pair-experiment.
• A host↔3090 link at x4 measures 6.72 GB/s, and that is a design constraint, not a detail. pcie-link-test.cu on Dankclaw (2026-09-16, services stopped): pinned H2D 6.72 GB/s, D2H 6.60 GB/s — 85 % of the Gen4 x4 ceiling — with async no better than sync and pageable within 1 % of pinned, i.e. the copies are link-limited rather than staging-limited (the usual "pageable costs 2x" rule does not apply here). The card is x16-capable and the Strix Halo port beneath it is x8-capable (LnkCap: 16GT/s, Width x8 on 00:03.1), so the x4 lives in the adapter or cable; the link idles at 2.5 GT/s x4 and trains up under load. Consequence for any engine that streams weights across it: a 3 MiB slab costs 474 us, and a 111 GiB A3B model needs ~1.70 GiB of active experts per token ⇒ 3.7 tok/s (4.6 with a 20 % resident cache), against 20.46 tok/s for the same model with the experts computed in host RAM by llama.cpp (-ot exps=CPU, whence the link carries activations only) and 25–26.5 on the iGPU. Streaming over PCIe loses; computing in RAM does not care. FreeToken — expert LRU in host RAM, exactly this streaming shape — also loads only modelopt block-FP8 safetensors for qwen3_5_moe, so a Q4_K_M GGUF is unloadable there before bandwidth is even the question. See x4-pair-experiment.

ROCm / GPU

• Layer split was the deployed shape on both multi-device llama.cpp services, for two different reasons. On Dankserver it was the only way both R9700s served one model (no collective could be dispatched on card #2) until the 2026-09-20 board swap; frank-medium now runs radiance TP=2 instead. On a RCCL-less fork build it is also the fast one: tensor parallel falls back to a butterfly allreduce and collapses prefill ~7x (measured 271.8 vs 1991 t/s). Pin the spec-decode drafter to one device — an unpinned draft under tensor mode aborts the KV build, and on a layer split the natural home is the card that ends up holding the target's last layers.

• **Bundling a ROCm runtime next to a llama.cpp binary: cp -r SRC DST/ is a

  trap.** With a missing DST, cp creates DST as a copy of SRC and drops

  one directory level. rocBLAS and hipBLASLt look for their device code at a

  fixed depth (../lib/rocblas/library, ../lib/hipblaslt/library/<arch>),

  and the flattened copy fails at request time, not load time: the engine

  reads the weights, answers /health and dies on the first completion with

  `Cannot read .../TensileLibrary.dat: No such file or directory for GPU arch

  gfx1201`. Declare the destination, then copy the contents.

• mmap must stay on for the big-PLE models: --no-mmap forces the ~50 GiB PLE table into anonymous memory → silent cgroup kill, dmesg is the only evidence. mmap pages ARE page cache. Never pin the PLE tensor with -ot (23 → 13.4 t/s).
• HIP reads bare env vars TEMP, HOST, GPU_MAX_HEAP_SIZE. Serve scripts export env files with set -a, so a tunable named TEMP leaks into HIP (TEMP=1.0 sent the JIT cache to path 1.0 and segfaulted every load, fixed 2026-08-13 by renaming to TEMPERATURE). Never name a tunable these; check with strings /opt/rocm-*/lib/libamdhip64.so.* | grep -xE "TEMP|HOST|GPU_MAX_HEAP_SIZE".
• -DGGML_CUDA_FA_ALL_QUANTS=ON is required for q8_0 KV + flash attention on CUDA builds; without it prefill collapses ~12x (95 vs 1300+ t/s, measured 2026-08-17).
• Flash-Next MTP heads do not load on any current loader — they need full per-layer blk.<il>.hc_attn_norm.weight and ship only blk.48 (verified 3x). Don't re-attempt without new loader support.
• QSA flash-attn assert: GGML_ASSERT(a-type == GGML_TYPE_F32) on qwen4exp needs scripts/patches/qwen4exp-qsa-f32-mask.patch — but only on the agentionai FAST layer set; the kingjones STRIX set works unpatched.
• GPU budget on Dankclaw is the GTT aperture (~114 GiB), not the rocm-smi "dedicated" figure. The Radeon 8060S (gfx1151) reports 116,732 MiB (llama-server --list-devices; gtt_total=122402963456 in /sys/class/drm/card1/device/). The older "64 GiB, SVM-backed" lesson was already stale — the CIRU Orca predecessor held 92.6 GiB of weights on the iGPU. free/dmidecode show the CPU-side carve-out — intentional, do not change.

Load / crash diagnosis

• Model load is slow (minutes cold); health returns 503 while loading; do not kill the process. deploy.sh waits up to 20 minutes (LLM_WAIT). On halogen, the FIRST request after a (re)start can go silent ~45–60 s (health engine_unresponsive, 503s) and then self-recovers — do not restart over it.
• Crash-loop procedure: (1) journalctl -u <svc> -n 100 + kernel line via journalctl -k (segfault at address 8 in libamdhip64 = NULL unref in HIP); (2) run serve script, then raw binary — different outcome isolates the env file; (3) gdb -batch -ex run -ex "bt 25" --args bash scripts/serve-<svc>.sh — llama-server buffers stdout and loses the log tail on SIGSEGV; trust gdb, not the log; (4) amdgpu: SVM mapping failed spam is a symptom, not a root cause.
• Journal signatures: client abort logs W srv stop: cancel task, id_task = N; natural stop logs only slot release with truncated = 0. The server never preempts a running generation; excess tasks queue when slots are busy.

Host memory / swap

• Swap resize ordering: create and swapon the replacement BEFORE swapoff of the old one, or live pages must land in RAM that MemAvailable overstates. Dankclaw swap (64 GiB at /swap.img, fstab-driven) protects the co-tenants, not frank-large — pinned weights don't swap, and paging the KV pool is worse than the OOM it prevents.
• Host swap state is invisible to deploy.sh (no repo file) — check swapon --show before reasoning about headroom.
• Silent RAM corruption outranks every software theory. frank-primary's ~40k-token output corruption (2026-10-04/05, repetition/garbage in long agent conversations) was marginal DDR5 at 4800 MT/s on four DIMMs, fixed by 4600. Consumer non-ECC memory raises no MCE for a silent flip, and a short in-band pattern check (60 s, 8 GiB of 246 GiB) is not a stability proof — combined, those two instruments said "ruled out" while the RAM was the cause. "4800 is the JEDEC floor on 5600-rated DIMMs" is a rating, not a margin guarantee. When the fault is data-shaped (NaN states, garbage tokens) and a software revert doesn't move it, price in the RAM before the runtime — and remember a revert tested under the same marginal hardware is itself confounded. See the runbook's 2026-10-05 section.

Runtime-specific guards

• Rootless podman cannot RAISE rlimits, and --ulimit nofile=1048576 dies at container create, not at runtime. runc reports error setting rlimit type 7: operation not permitted (type 7 = RLIMIT_NOFILE) when the requested hard limit exceeds the host process's hard limit — on Dankclaw that ceiling is 524288. Inherited-equal limits are fine, so memlock=-1:-1 works iff the unit grants LimitMEMLOCK=infinity (systemd is privileged; the value flows through podman to runc without any raise). Upstream halogen's own run lines pass NO --ulimit at all — copy vendor run lines before adding flags (measured 2026-10-05 on the frank-large cutover).
• Kairic Edge (if ever re-deployed): the pinned CK commit needs third_party/composable_kernel/include/ck/config.h generated manually (untracked; git clean removes it), and its greedy-argmax fastpath rejects non-greedy requests and reasoning budgets. Re-test chat before production.
• FastMTP sidecar: build must be commit 4df29be4 + the HauhauCS patch; stock binaries fail the draft load with a tensor-shape mismatch.
• A GGUF's custom tensor type picks the runtime — read the histogram before promising a file to a loader. Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf is ftype 115 with 311 tensors of type 103 (GGML_TYPE_Q8_0_ROCMFPX); stock llama.cpp and gufo refuse it at load, so ROCmFPX was not a preference but the only door. Counting tensor types out of the GGUF KV section takes one python read and settles a runtime argument in minutes. The build identity matters as much as the fork: the 2026-08-25 ROCmFPX builds SIGSEGV in ggml_compute_forward_set_rows on this same file, so scripts/build-rocmfpx.sh pins commit 36e9acd by SHA.
• Speculative-decode flags move as a SET. Handing the ROCmFPX fork --model-draft with no --spec-type (the naive way to build a "no MTP" bench arm) segfaulted at load, core dumped, no diagnostic (2026-10-02). A baseline arm must strip --model-draft AND every --spec-draft-* AND --spec-type, or it measures a crash instead of the absence of drafting.
• An orphaned scratch server taxes the whole host. A bench child left alive by a client-side exception kept ~26 GiB of weights resident on a 121 GiB box, pushed available to 3 GiB, and every later rate was suspect. Record the child pid and kill it in a finally, then prove free came back before believing a number. Kill by pid — never pkill -f llama-server, which also hits frank-small.

llama.cpp serving

• MTP is a property of the FILE, not of the flags. The -mtp GGUF build carries the head as blk.<n>.nextn.* (qwen35.nextn_predict_layers = 1); --spec-type draft-mtp then sets load_mtp, and the draft context runs on the MAIN model's weights (shares_model), so there is no second copy of a 12 GiB model in VRAM and no -md argument. The plain build has no MTP tensors at all, and pointing --model-draft at the file would load it twice.
• What can be chained, what cannot, and what a chain costs on a hybrid. --spec-type a,b builds every enabled implementation and chains drafting through them, in a fixed internal order (ngram types first, then the drafters — CLI order is irrelevant). The gate is ctx_dft: draftless types (the ngram family) are added unconditionally, while draft-mtp, draft-dflash, draft-dspark and draft-eagle3 all require a draft CONTEXT. So a chain may hold one context-backed drafter plus any number of ngram types — draft-dflash,ngram-mod is legal and measured (2.6x the drafter alone on prompt-echo traffic, byte-identical to it on prose at the fleet's 48/64 window) — but draft-mtp,draft-dflash dies at load: "context type MTP requested but model doesn't contain MTP layers", because MTP builds its context against whatever model the draft context holds. Two context-backed types cannot share one process, and MTP's head only loads when MTP is the draft model. A LEGAL chain can still lose on a recurrent model: draft-mtp,ngram-mod on frank-large's qwen35 hybrid measured −15 % prose / −7 % block-copy (2026-10-03) — adding the ngram layer logs recurrent state rollback is not compatible with 'ngram-mod' - disabling rollback support, and MTP verification depends on that rollback, so acceptance drops on ALL traffic even where the matcher fires. Measure chains per architecture, never by analogy.
• A drafter's KV scales with the TARGET's context, and its window policy decides how much that costs. The draft context is built with cparams.n_ctx = llama_n_ctx(ctx_tgt) (common/speculative.cpp) — there is no flag to cap it — so a drafter that declares NO sliding window caches the whole window on whatever device --spec-draft-device names. Swapping DFlash2 (dflash.attention.sliding_window = 2048, ~40 MiB of draft KV) for DSpark (no window; 5 layers x 2 x 8 kv heads x 128 dim x 2 B = 20 KiB per position, so 5.0 GiB at 262144) put ROCm1 at 33.9 GB of 34.2 GB and killed the FIRST draft graph capture with ROCm error: out of memory — after the model had loaded and /health had answered. Read sliding_window out of the drafter's GGUF header before sizing any spec-decode deployment, and prefer a skewed -ts over a context cut when the full window matters. The pair was swapped DFlash2 -> DSpark -> DFlash2 in one day (2026-09-16) and the round trip moved the balance point FOUR layers (~4.3 GiB of headroom, 40:25 -> 36:29), which is the practical size of the effect. See frank-medium-llamacpp-dflash2.
• --ubatch-size has an OPTIMUM on the 2x R9700 layer split, not a plateau — and the curve is BACKEND-SPECIFIC, so never carry it across (next bullet). Its compute buffers land about TWICE as heavy on the card that holds the output layers and the drafter. Measured on frank-medium 2026-09-16 at the deployed shape (262144 over two slots, a 64k prefill measured at 64k depth, scripts/bench-batch-ubatch.sh): prefill runs 1046 tok/s at ubatch 256, 1068 at 384, 1193 at 512 (the peak), 1135 at 768, 1152 at 1024, 1151 at 1280, 1165 at 1536, 1148 at 1792 and 1078 at 2048. Both ends lose — below 512 the SMs are not saturated, and 2048 is 10 % off the peak — and 512 -> 1536 costs ROCm0 1.30 GiB and ROCm1 2.41 GiB to go 2.4 % SLOWER. The logical --batch-size does not matter: 2048 through 8192 land within 2 % of each other at a fixed ubatch and use identical VRAM. Decode never moves at all (15.88 tok/s at depth 65536 on every cell). Because the buffers are asymmetric, a split balanced at one batch is unbalanced at another: 512 -> 1536 shifted frank-medium's balance by a layer, so 36:29 became 35:30. Re-measure the balance after changing the batch. (8192/2048 failed to allocate its 2020 MiB pp buffer here under the DSpark drafter, when ROCm1 had 0.08 GiB free. That was a CONTIGUITY failure, not a total-bytes one, and no totals model — the sweep's or the seam's — can see it.)
• On the CUDA 3090 the same knob has NO interior optimum up to 2048, so frank-small ships 4096/2048. Measured 2026-10-05 by boot-pair A/B on the live service: at 8,928 prompt tokens ubatch 1024 gives 3,972 / 4,077 t/s and ubatch 2048 gives 4,207 / 4,212 — +4.6 % — for 616 MiB (free 2,849 → 2,279). Decode did not move (91.9–93.0 vs 93.8–94.3, i.e. noise), and a 279-token prompt, which is below either ubatch, came back identical — that is the mechanical proof the knob was the only variable, since a prompt shorter than the ubatch cannot see it. The logical --batch-size was again free. And the predecessor's "ubatch 2048 failed the scratch boot" was a shape observation, not a law: the same file boots 2048 clean at 2 × 200,704. Bench it by restarting the SERVICE through a transient FRANK_SMALL_ENV_FILE drop-in, never a scratch port — one card holds one server. See frank-small-llama-mimo-q6k-2x200704.
• One -ts step is one LAYER, and a hybrid stack's layers are not the same size. qwen35 is 16 x (3 Gated DeltaNet + 1 Gated Attention), so a step is worth ~0.43 GiB on a DeltaNet boundary and ~1.43 GiB on a full-attention one (0.45 GiB of weights plus ~1.0 GiB of its own f16 KV at 262144) — both measured on frank-medium. Balance by measurement, never by interpolating between two splits; an affine fit reproduces the deployed point and overstates the gaps in between. tests/frank_medium_unit.bats encodes that fit as a gate. See frank-medium-llamacpp-dflash2.
• ngram-mod prints no startup banner. The load log shows the drafter's adding speculative implementation 'draft-dflash' and nothing at all for the matcher, so a chain can look unengaged while it is working. Identify it behaviourally (draft_n > 0 only where text repeats) or by the per-task draft acceptance = ... mean len = ... line. It costs no VRAM (a 64 MiB host-RAM hash table) and its window decides which traffic it can see: too short and it proposes on ordinary prose fragments, which costs verification without acceptance.
• --parallel divides the pool, so MORE conversations cost WINDOW, and the per-slot number is the contract — not ctx-size. n_ctx_slot = CTX_SIZE / PARALLEL, and llama.cpp refuses a prompt longer than the SLOT (400 exceed_context_size_error: request (184001 tokens) exceeds the available context size (180224 tokens)) even when the pool has room. So frank-small's 2 → 3 reslice moved each conversation from the model's native 262,144 down to 180,224, and models.yml contextWindow had to follow or the harness would discover the ceiling as a 400 on a long conversation. Side effect worth knowing: the shorter slot made decode ~5 % FASTER (106.8 vs 100.4 tok/s) — less attention to scan and one fewer DeltaNet state resident. The 2026-10-03 reslice went the other way (3 → 2 slots, 180,224 → 200,704, pool 401,408) to fund unquantized f16 V, and contextWindow followed again; decode landed 104.8 tok/s — the ~2 % off 106.8 is the exact-V tax. The same day the freed space also bought the Q6_K rung (now 94.4 tok/s — the rung itself is a further measured −10 %, weight-bandwidth, not slot geometry). All moves measured; both slot directions cost or buy window, never pool. See frank-small-llama-mimo-q6k-2x200704.
• Price a VRAM budget from two boots, never from a formula — and an mmproj costs ~1.5x its file size. The frank-small 3090 fit was solved from two measurements: 2 × 262,144 with no projector = 20,190 MiB used (KV 12,800 ⇒ non-KV 7,390), 3 × 196,608 with the f16 projector = 22,972 used (KV 14,113 ⇒ non-KV 8,859). The difference is the vision tower: a 918 MB file that bills ~1.4 GiB, because the encoder's activation and compute buffers ride with it. Subtracting the two rows is exact; extrapolating weights+KV+ "a bit for buffers" is what leaves 1.1 GiB free when you meant 2.7. Then solve KV ≤ card − non-KV − headroom and round to a multiple of PARALLEL × 4096. Recorded in scripts/frank-small.env so the next resize starts from measurement.
• images and image_url are NOT interchangeable: only the content part reaches the model. --mmproj loads, usage.prompt_tokens rises, and the model still invents an answer if the client used the legacy top-level images field on /v1/chat/completions — frank-small answered "Brown." for a solid-red PNG that way and "Red." through the OpenAI image_url part. The failure is silent because a hallucinated color is a plausible answer. Probe vision with a solid color and expect an exact word.
• --jinja is what tools needs. On 2026-09-16 a subagent died with 500 tools param requires --jinja flag, from the ik_llama.cpp runtime that then served frank-large: every tool-using agent on that endpoint was broken and the failure read like a model fault. Mainline llama.cpp 0.4.1 defaults --jinja on and the serve script pins it; halogen also accepts tools and returns tool_calls. Check tool support before routing a role at an endpoint.
• Quantized KV on CUDA needs GGML_CUDA_FA_ALL_QUANTS=ON — the same option and the same ~12x prefill collapse described in the ROCm note above. f16 KV needs nothing; frank-small ran f16 for that reason until 2026-09-17, when build-cuda was rebuilt with the option ON so the endpoint could take q8_0 KV.
• llama-bench separates a LIST with / and SWEEP values with ,. Both -dev and -ts follow this: -dev CUDA0,Vulkan1 is two single-device configurations, never one two-device run, and -ts 0.85,0.15 is two sweep values each becoming [0.85, 0, …] / [0.15, 0, …], i.e. device 0 takes everything. Write -dev A/B -ts x/y. This artifact produced a wrong fleet conclusion on 2026-09-16 — "CUDA + Vulkan cannot serve one model", with two bogus OOMs that were really "all 111 GiB on the 24 GiB card" — and the bench table's own dev column was printing the truth the whole time. Any placement-dependent number needs an independent check: sample the other device's memory (nvidia-smi) while the run is loaded.
• A model's own template can REJECT the harness's thinking level. Qwen3.8's template accepts only xhigh, medium and low for reasoning_effort and raises on anything else — HTTP 500, not a clamp — so the harness's levels must be folded onto that set in models.yml (compat.reasoningEffortMap). A role whose level is unmapped fails on its FIRST request, and the failure names Jinja, not the model: every subagent on frank-small died with Unexpected reasoning effort high until high: xhigh was added on 2026-09-16. Halogen accepts the wider set, so the two entries legitimately differ.

Ops

• Never pkill -f a pattern that appears in your own command line — it kills your ssh session. Use pkill -x <exact-name>.
• Both hosts allow passwordless sudo for the deploy user. Use plain sudo in committed scripts (no echo pw | sudo -S); FRANK_SUDO_PW covers hosts without it.
• ComfyUI on Dankclaw runs with GPU access (/dev/kfd + /dev/dri via AddDevice, GroupAdd=keep-groups, plus the udev rule that makes those nodes world-writable — user-unit spawns carry no supplementary groups). It must NOT pass --cpu: the gfx1151 ROCm wheels exist and the unit's --cpu era ended 2026-09-16, when a stale device-less container was found restarted 9,146 times (RuntimeError: No HIP GPUs are available).
• SearXNG outbound traffic must stay on the Gluetun tunnel; do not move its ports off the gluetun service.
• amd_iommu=off on Dankclaw (bought 13–16 % prefill for frank-large, 2026-09-11) permanently retired the XDNA2 NPU: the driver's SVA/PASID bind needs an enabled IOMMU (/dev/accel/accel0 fails err=-19). Revert path in serverAppConfigs-docs/LLM/dankclaw-iommu-translation-off.md.
• Rootless Quadlet containers run with --replace --rm: podman logs covers only the CURRENT instance; history across recreations lives in journalctl --user -u <name>.service. Read rootless storage WITHOUT sudo (per-user storage — sudo podman logs finds nothing).
• A dead rootless podman pause process breaks ALL container management on the host, and podman system migrate panics in 5.7.0. Symptoms: every podman command fails with invalid internal status … could not find any running process, every container start fails binding podman's bridge addresses (Error starting server failed to bind udp listener on 10.89.x.1:53: Cannot assign requested address), and every Restart=always quadlet loops. Container storage is intact; what is gone is /run/user/<uid>/libpod/tmp/alive (a UUID token, not a pid) and the rootless netns holding the bridges — deleting the stale token is not enough, and podman system migrate segfaults in teardownNetworkBackend. Repair: reboot (/run/user clears and all 28 quadlets return on their own — check Linger=yes and WantedBy=default.target first). Seen 2026-09-16, hours into heavy GPU work on the same host — but the trigger is NOT established (journalctl -k recorded no OOM kill; the kernel had no other word on it). Treat the pause process as host state that can die for its own reasons, and reach for the reboot when it does.
• A container cannot reach a Network=host service at the host's LAN IP. netavark DNATs published ports only: from a bridge container 192.168.0.143:<published> answers while 192.168.0.143:<host bind> and the gateway 10.89.1.1 are both refused. Use host.containers.internal (169.254.1.2). This is what 502'd hermes.franksdank.cloud until the Caddyfile upstream was changed on 2026-09-16 (its vhost comment had claimed the LAN IP was open from that bridge — it never was).
• A *.gguf filter misses mradermacher's payloads: large quants ship as .partNofM fragments. mradermacher/Qwen3.8-Flash-Next-Uncensored-i1-GGUF carries Q4_K_M (37.00 + 37.00 + 36.97 = 110.97 GiB) and Q4_K_S (35.00 + 35.00
  • 34.09 = 104.09 GiB) as parts, while the smaller rungs (Q2_K, Q3_K_, IQ3_, IQ4_XS) are single files. Fetch the parts, verify each against the Hub's LFS sha256, concatenate in numeric order — that is how the deployed 111 GiB file was assembled. Listing the Hub tree and filtering endswith(".gguf") makes Q4_K_S look like it does not exist.
• A fetch script must -f and verify, or it reports success on an access error. curl -o FILE URL exits 0 on HTTP 401/403 and writes the error body as the "artifact": a gated repo yielded three 153-byte shards with identical sha256 values and a cheerful DOWNLOAD_DONE in 20 seconds for 104 GiB. Use -f/--fail-with-body, compare each file's byte count to the Hub's declared size, and check the LFS sha256 — scripts/download-frank-medium-gguf.sh already does all three; ad-hoc fetches tend not to. Gated repos (e.g. orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF, same 104.09 GiB Q4_K_S) return 403 even with the host's HF token until that account accepts the gate.

Runbook Index

Active:

Service	Runbook
frank-large (halogen container, Qwen3.8-27B .hgn + DFlash2, 1×200,700, text-only, Dankclaw)	serverAppConfigs-docs/LLM/frank-large-halogen27b-dflash2.md
frank-small (llama.cpp CUDA, MiMo-V2.6-Distill-Qwen-9B Q6_K, 2×200,704 f16 KV, Dankclaw)	serverAppConfigs-docs/LLM/frank-small-llama-mimo-q6k-2x200704.md
frank-primary (tcclaviger vLLM, Qwen3.8-Flash-Next FP8, Dankserver)	serverAppConfigs-docs/LLM/dankserver-qwen38-fp8-expert-offload-tuning.md
Dankclaw kernel posture (amd_iommu=off)	serverAppConfigs-docs/LLM/dankclaw-iommu-translation-off.md
Deployment flow + incidents	serverAppConfigs-docs/LLM/deployment.md

Retired (rollback targets and history):

Runtime	Host	Window	Runbook
ROCmFPX llama.cpp fork Qwen3.8-27B Uncensored Q6_0_ROCMFPX_AGENT + MTP (frank-large)	Dankclaw	2026-10-03 → 10-05 (rollback target, payload + fork build on disk)	frank-large-rocmfpx-q6agent-uncensored-mtp.md
ROCmFPX llama.cpp fork Qwen3.8-27B Q8_0_ROCMFPX_AGENT + MTP (frank-large)	Dankclaw	2026-10-02 → 10-03 (rollback target, payload on disk)	frank-large-rocmfpx-q8agent-mtp.md
gufo v0.1.1 Qwen3.8-27B UD-Q8_K_XL/UD-Q4_K_XL + DFlash2 (frank-large)	Dankclaw	2026-09-28 → 10-01, again 10-01 → 10-02 (payloads removed from disk 2026-10-03; rollback = re-download)	frank-large-gufo-qwen38-27b.md
vLLM TheRock MiMo-V2.6-Distill-Qwen-9B BF16 + cross-model MTP (frank-large)	Dankclaw	2026-10-01 (same-day rollback to gufo)	frank-large-vllm-mimo-v26-bf16.md
llama.cpp master MiMo-V2.6-Distill-Qwen-9B Q6_K (frank-small)	Dankclaw	2026-10-01 → 10-02 (the rung returned ACTIVE 2026-10-03 at the 2×200,704 shape — frank-small-llama-mimo-q6k-2x200704.md)	frank-small-llama-mimo-q6k.md
llama.cpp master MiMo-V2.6-Distill-Qwen-9B Q5_K_M 2×200,704 f16 KV (frank-small)	Dankclaw	2026-10-02 → 10-03 (rollback target — one MODEL_FILE line, buys back 875 MiB + the −10 % rung tax; owns the boot-pair table)	frank-small-llama-mimo-q5km.md
SGLang MiMo-V2.6 INT8-W8A16-AutoRound (frank-small)	Dankclaw	2026-10-01 (same-day swap to the GGUF path)	frank-small-sglang-mimo-v26.md
tcclaviger vLLM GLM-5.3-Flash W4A16 int4, +/- DFlash2 drafter (frank-primary)	Dankserver	2026-09-30 → 10-01	frank-primary-glm53-flash-w4a16.md
llama.cpp b11030 Qwen3.8-27B UD-Q8_K_XL + DFlash2 (frank-large)	Dankclaw	2026-09-21 → 09-28	frank-large-unsloth-qwen38-27b-q8xl.md
R9V radiance-vLLM Qwen3.8-Flash-Next-Uncensored IQ4_XS (frank-primary)	Dankserver	2026-09-27 (same-day swap to Q8_0)	frank-primary-r9v-radiance-qwen38-uncensored.md
llama.cpp b11030 Qwen3.8-Flash-Next Q8_0 + shared MTP (frank-primary)	Dankserver	2026-09-26 → 09-27	frank-primary-llamacpp-qwen38-flash-next-q8.md
ik_llama.cpp Qwen3.8-Flash-Next-Uncensored i1-Q4_K_M (frank-large)	Dankclaw	2026-09-15 → 09-16	frank-large-ik-uncensored.md
vllm-radiance TP=2 MXFP4 + DFlash2 (frank-medium)	Dankserver	2026-09-20 → 09-26	frank-medium-radiance-mxfp4.md
Lucebox DS4.1 DeepSeek-V4.1-Flash ROCmFP23 (frank-primary)	Dankserver	2026-09-26 (same-day swap)	frank-primary-lucebox-ds4.md
kingjones STRIX Qwen3.8-Flash-Next ROCmFP4	Dankclaw	2026-08-30 → 09-11	kingjones-strix-rocmfp4.md
halogen-flash-server Qwen3.8-Flash-Next w4b (frank-large)	Dankclaw	2026-09-16 → 09-19	halogen-flash-server.md
CIRU v4.4.1 Orca Qwen3.8-Flash (frank-large)	Dankclaw	2026-09-19 → 09-21	frank-large-ciru-orca.md
unsloth llama.cpp b11030 Qwen3.8-Flash-Next UD-IQ4_XS (frank-large)	Dankclaw	2026-09-21 (same-day swap)	frank-large-unsloth-qwen38-flash-next.md
Kairic Edge Qwen3.8-27B IU4 W4A4 (frank-large)	Dankclaw	2026-09-21 (same-day swap)	frank-large-kairic-edge-qwen38-27b.md
ISTA-DASLab Qwen3.8-27B-GSQ-RCO IQ3_S (frank-small)	Dankclaw	2026-09-16 → 09-17	frank-small-qwen38-gsq-rco.md
vLLM Gemma4-12B w4a16 (frank-small)	Dankclaw	2026-09-19 (same-day revert)	frank-small-vllm-gemma4.md
llama.cpp Gemma4-12B QAT Q4_K_M (frank-small)	Dankclaw	2026-09-17 → 10-01	frank-small-gemma4-12b.md
agentionai ROCmFP4-FAST imatrix	Dankclaw	2026-08-29 → 08-30	agentionai-fast-imatrix.md
eemin Qwen3.8-27B ROCmI4 W4A4 MTP	Dankclaw	2026-08-18 → 08-25	eemin-rocmi4-mtp.md
Kairic Edge IU4	Dankclaw	2026-08-22 (same-day revert)	kairic-edge.md
CIRU ActiveFPX v3 / PromptForge	Dankclaw	through 2026-08-22	ciru-activefpx-promptforge.md
DFlash2 Vulkan	Dankclaw	2026-08-21 (same-day revert)	dflash2-large-vulkan.md
FastFlowLM NPU (frank-npu)	Dankclaw	twice, retired 2026-09-11	frank-npu.md, FastFlowLM NPU Inference.md
FreeToken (official + Huihui FP8, frank-small)	Dankserver	2026-09-02 → 09-15	freetoken-frank-small.md
FastMTP patched llama.cpp	Dankserver	retired 2026-08-18	qwen3.8-spec-matrix.md
ik_llama.cpp Ornith CRACK	Dankserver	retired 2026-09-02	ikllama-ornith-dankserver.md
llama.cpp layer split + DFlash2 (frank-medium)	Dankserver	2026-09-16 → 09-20	frank-medium-llamacpp-dflash2.md

Retired runbooks are under serverAppConfigs-docs/LLM/. Retired scripts live in old/ and .AGENTTRASH/.

Why

Direct edits on the remote drift from version control, and future agents won't know what changed. Git is the single source of truth. The two workloads share two hosts, one Caddy, and one Gitea; one repo keeps them in one history.

Summary

Context	Action
Editing a compose/Caddyfile/serve script	Through git — fetch, commit, edit, test, push, pull, deploy. Never edit on the remote.
Config exists only on remote	Pull it into the repo first — commit the fetched version (redacted) before editing.
Unstaged local changes	Commit or discard — check git status before starting work; ask if unsure.
Deploying a committed change	scripts/deploy.sh <target> — pushes, pulls ff-only, plans, dispatches, health-checks.
Container without a Quadlet unit	Give it one — boot-persistent, self-healing .container via git.
Benchmarking	Stop and mask the service first, bench, unmask, restart, verify health.
Deleting a file	Move it to trash — .AGENTTRASH/ (in-repo) or ~/.AGENTTRASH/YYYYMMDD-topic. Never rm.
Tests exist	Run them — verify tests pass after every change. Do not commit failing tests.
Real secrets	Never commit — host .env files; .env.example holds placeholders only.