frank-primary: 2 x 200704 fp8, pool cut for the expert arena, MNBT 1024

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
autoresearch: harness setup
2026-10-02 09:54:08 -07:00 · 2026-10-02 07:50:50 -07:00 · 2026-10-02 03:46:50 -07:00 · 2026-10-02 03:17:35 -07:00 · 2026-10-02 03:08:00 -07:00 · 2026-10-02 02:31:39 -07:00
239 changed files with 16882 additions and 4829 deletions
@@ -37,3 +37,4 @@ __pycache__/
 # Editor/OS noise
 *.swp
 .DS_Store
+third_party/
@@ -9,11 +9,11 @@ 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-medium` LLM service.
+  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-flash-server on
-  the gfx1151 iGPU) and `frank-small` (llama.cpp on the RTX 3090). Caddy on
+  Alloy, Pi-hole) plus TWO LLM services: `frank-large` (ROCmFPX llama.cpp
+  fork 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
@@ -24,9 +24,9 @@ merge absorbed the former `vllmScripts` repo, history preserved).

 | Target | Service | Runtime | Model | Port |
 |--------|---------|---------|-------|------|
-| Dankserver | `frank-medium` | mainline llama.cpp master fb27a525 (root systemd unit, gfx1201 HIP build) | ajgazin Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP UD-Q8_K_XL + z-lab Qwen3.8-27B-DFlash2 Q8_0 drafter, 2x Radeon AI PRO R9700 in ONE layer-split process (35:30), 262144 ctx over 2 slots, batch 6144x512, text only | 7770 |
-| Dankclaw | `frank-large` | halogen-flash-server 0.5.8 (rootless Quadlet) | peonist-ai Qwen3.8-Flash-Next w4b `.hgn` on the gfx1151 iGPU (Strix Halo LPDDR), 200192 ctx per conversation over one 400384-position KV pool / 2 slots, text only | 7770 |
-| Dankclaw | `frank-small` | mainline llama.cpp master 37b53fd (root systemd unit) | ISTA-DASLab Qwen3.8-27B-GSQ-RCO IQ3_S (`-mtp` + BF16 mmproj) on the RTX 3090, 131072 ctx / 1 slot, MTP speculative decoding, the fleet's vision endpoint | 7771 |
+| 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 1024, MoE gather 32×64, **2 seqs × 200704 ctx** at KV_BYTES 3,150,000,000/rank (fp8 measures 7,701 B/token), pool deliberately small so the expert arena keeps what it reclaims (~7,500 slots vs 6,248 at the 6.1e9 shape) — util 0.98, and the client contract wired: `--reasoning-parser …
+| Dankclaw | `frank-large` | **ROCmFPX llama.cpp fork** (`kingjones30/ROCmFPX` @ `36e9acd`, gfx1151, built by `scripts/build-rocmfpx.sh`; bare `llama-server`, root systemd unit) | `kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF` **`Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf`** (26.28 GiB, ftype 115 ≈ 8.39 bpw: 353 F32 + 187 Q8_0 + 311 `GGML_TYPE_Q8_0_ROCMFPX` tensors — only this fork can read them), **MTP spec decode from the repo's own `mtp-Qwen3.8-27B-Q4_0.gguf` (n-max 4, pinned to ROCm0)**, **1 session × 262,144 (native window)** with **f16 KV**, the bundled BF16 mmproj for vision, thinking on with the card's recipe (temp 1.0 / top_p 0.95 / top_k 20). Measured 2026-10-02: decode **17.0–17.1 t/s** with MTP vs **7.6 without (+123 %)**, draft acceptance 0.462, prefill 325–351 t/s, prompt-cache replay 0.5 s, boot→healthy 16 s, GTT 64.5 GiB. Cut over 2026-10-02 from gufo; rollback tag `frank-large-gufo-q4kxl-20261001` | 7770 |
+| Dankclaw | `frank-small` | llama.cpp master, CUDA sm_86 build (root systemd unit, relinked) | bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF **Q5_K_M** (9B dense qwen3_5 hybrid, 6.4 GiB GGUF, no MTP weights so no spec decode) + the repo's f16 mmproj for **vision**, on the RTX 3090, **3 × 180,224** (pool 540,672 — NOT the native window; three contexts cost 81,920 tokens per conversation) with **asymmetric KV f16 K + q8_0 V** under flash attention, 21,684 used / **2,441 free** of 24,576 MiB, 106.8 tok/s decode, 3,814 t/s prefill. Resized twice 2026-10-02 (Q6_K → Q5_K_M → 3 slots + vision → this pool); rollback `models/MiMo-Q6K/` or `PARALLEL=2 CTX_SIZE=524288` | 7771 |

 Retired endpoints (`frank-npu`, and every prior runtime) are in
 the [runbook index](#runbook-index). The full swap history lives in
@@ -61,15 +61,16 @@ All edits flow through git. Never edit files on a remote machine directly.
 ./scripts/deploy.sh dankclaw
 ./scripts/deploy.sh dankserver
 ./scripts/deploy.sh dankclaw --service searxng
-./scripts/deploy.sh dankserver --service frank-medium
+./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 — frank-large by `dankclaw/frank-large/quadlet/`, frank-small
-and frank-medium by their unit/env/serve trio in `scripts/` — so unrelated
+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.

@@ -122,14 +123,18 @@ 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 as of 2026-09-16**: a
-Grafana datasource and the `llm-fleet` dashboard are provisioned from
-`shared/grafana/provisioning/`, and `shared/prometheus/rules/llm-fleet.yml`
-loads 5 alert rules + 1 recording rule. Nothing is *delivered* — there is no
-Alertmanager, so a firing alert reaches nobody (an operator decision, not an
-oversight). Five 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`.
+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

@@ -203,14 +208,27 @@ archive.

 **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` (llama.cpp), `vllm
-bench` (vLLM), and halogen-flash-server's own `bench`/`sweep` modes for
-frank-large, which is the active runtime there again. Rolling your own bench
+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 is ROCmFPX llama.cpp since 2026-10-02, but the rollback tag keeps 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
@@ -223,16 +241,17 @@ 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` (rootless Quadlet again since 2026-09-16):
-     `systemctl --user stop frank-large`, then move its
-     `~/.config/containers/systemd/frank-large.container` aside and
-     `systemctl --user daemon-reload` — the quadlet generator re-creates that
-     unit from the file on every reload, so the source is the lever, not `mask`.
+   - Dankclaw `frank-large` (root unit, vLLM 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-medium` (system unit):
-     `sudo systemctl stop frank-medium && sudo systemctl mask frank-medium`
+   - 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
@@ -244,9 +263,8 @@ deletes the unit file from `/etc/systemd/system/`, not just the symlink.
 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-medium` →
+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` →
-`192.168.0.143:7771`. Per-role routing lives beside it in `config.yml` under
 `modelRoles`.

 Resolution rules: the default model resolves as a `defaultProvider` +
@@ -259,102 +277,181 @@ 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 2026-09-16 the image path is `frank-small` (`input: [text, image]`), and
-`frank-large` is text-only because its halogen deployment loads no vision
-tower. There is **no audio** input anywhere in the fleet: that capability left
-with the FLM NPU endpoint on 2026-09-11.
+of 2026-10-02 **all three endpoints accept images** (`input: [text, image]`):
+`frank-large` through the BF16 mmproj that ships inside its GGUF payload,
+`frank-primary` 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 Q5_K_M
+GGUF itself carries no vision tensors. `frank-small`'s `contextWindow` is
+**180224**, 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-flash-server (Dankclaw, ACTIVE)
+### frank-large — ROCmFPX llama.cpp, Qwen3.8-27B Q8_0_ROCMFPX_AGENT + MTP (Dankclaw, ACTIVE)

-halogen-flash-server 0.5.8 as a **rootless Quadlet again since 2026-09-16**
-(`dankclaw/frank-large/quadlet/frank-large.container`), serving
-`peonist-ai/halogen-qwen3.8-flash-next` (115.55 GiB w4b `.hgn` checkpoint + a
-quality overlay that re-fetched to 2.40 GiB on 2026-09-16) on port 7770 from
-the gfx1151 iGPU. Two conversations of 200,192 tokens resident (pool 400,384,
-2 slots), greedy default, reasoning extracted to `reasoning_content`, and **no
-vision tower** — that moved to frank-small. It replaced ik_llama.cpp, which had
-replaced it for two days.
+The **kingjones30/ROCmFPX llama.cpp fork** at commit
+`36e9acd40e10a87cd3c3ef8ec734668757dc8520`, built for gfx1151 by
+`scripts/build-rocmfpx.sh` into `/home/dank/GitPackages/kingjones-ROCmFPX/build`,
+as a **root systemd unit** (`scripts/frank-large.service`, `User=dank`,
+`Restart=always`) running a bare `llama-server` (no container) on port 7770
+from the Strix Halo iGPU. Boot to healthy: **16 s**.

-The engine **fetches its own weights on first start** (`HALOGEN_DOWNLOAD`, into
-the `/home/dank/models/halogen` bind mount): clearing that directory and
-restarting the unit is the supported way to pick up upstream updates, and it
-needs the directory to exist **empty** — a missing directory fails the
-container with `Error: statfs /home/dank/models/halogen: no such file or
-directory` and the unit loops on `Restart=on-failure` until `mkdir -p` brings
-it back. The sidecars are versioned in place, so a stale copy is invisible from
-the file name (both overlays were ~95 MB larger upstream than the copies this
-host had been serving).
+**The runtime is not interchangeable — it is chosen by the FILE.**
+`kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF`'s best rung,
+`Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf` (28,214,425,952 B, sha256 `f3ef2bda…`,
+ftype 115 ≈ 8.39 bpw), stores **311 tensors of type 103,
+`GGML_TYPE_Q8_0_ROCMFPX`** — a quantization that exists in no other loader:
+stock llama.cpp and gufo both refuse it at load (`unknown tensor type`). The
+file is 353 F32 + 187 plain-Q8_0 + 311 ROCmFPX; the AGENT routing variant
+keeps the output-head tensors at plain Q8_0 (187 vs 2 in the non-AGENT file),
+which is why the card reports MTP acceptance 0.953 vs 0.911 — highest quality
+AND the better drafter base. Vision comes from the bundled
+`mmproj-Qwen3.8-27B-BF16.gguf` (931,145,888 B).

-Measured boot ledger (2026-09-16, `HALOGEN_VERBOSE=1` ships in the unit):
-device budget `121 − 67.7 − 20 = 33.5 GiB` against a need of `~27.2 GiB`, so
-the configured pool was **not clamped**; 68.0 GiB of weights locked + 11.0 GiB
-of KV pool + 16.5 GiB of working memory = 95.4 GiB in all, leaving 14.7 GiB
-(475 contiguous 2 MiB blocks) for everything else: frank-small on the 3090 and
-~28 co-tenant containers. The engine printed its co-tenant stall NOTE and
-logged 4,350 compaction stalls (285 failed) while reserving the pool, so read
-that as "tight and fragmented", not "out of memory".
+**The pin is load-bearing.** The 2026-08-25 ROCmFPX builds segfaulted in
+`ggml_compute_forward_set_rows` on exactly this file; `build-rocmfpx.sh`
+checks the commit out by SHA, so a fork-side force-push cannot silently
+re-break the endpoint.

-Runbook: `serverAppConfigs-docs/LLM/halogen-flash-server.md`.
-Rollback: [[frank-large-ik-uncensored]].
+**Spec decode is the performance feature, and it is measured.** The draft is
+the repo's own `mtp-Qwen3.8-27B-Q4_0.gguf` (1,680,271,648 B, sha256 `051a1764…`,
+`nextn_predict_layers = 1`), pinned to the same card:
+`--spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0
+--spec-draft-p-min 0.0 --spec-draft-device ROCm0 --spec-draft-ngl 99`.
+Measured 2026-10-02 in one session, temp 0, 2,202-token prompt / 512 tokens,
+`cache_prompt: false`, production service stopped for each arm:
+**MTP on 17.00–17.14 t/s decode** (draft acceptance 0.462, mean acceptance
+length ≈ 1.86) against **7.60–7.66 t/s with the draft removed — +123 %**;
+prefill pays 8–15 % (381 → 325–351 t/s). Acceptance is far from the card's
+0.95 (that number is the card's own text at a shallow prompt); the gain is
+real and n-max 4 is where it was measured, never swept. This replaces the
+gufo Q4_K_XL arm at ~15 t/s, so the endpoint got higher quality AND more
+speed at once.
+
+**Shape and budget:** ONE session at the native 262,144 window, unquantized
+f16 KV (~64 KiB/token → ~16 GiB, affordable inside the iGPU's ~114 GiB GTT
+aperture), `--fit off --flash-attn on --direct-io --ctx-checkpoints 16
+--cache-ram 8192 --cache-idle-slots`. Measured GTT 64.5 GiB with model +
+draft + projector resident. Prompt cache verified: a repeat pass over the
+same 6k prompt replays in 0.5 s.
+
+**Client contract (llama.cpp, live-verified 2026-10-02):** `--jinja` is on
+(without it `chat_template_kwargs` are silently ignored), tools are native
+(`get_weather` → `{"city":"Seattle"}`), thinking lands in
+**`reasoning_content`**, `reasoning_effort` accepts ONLY
+`xhigh|medium|low` (anything else is HTTP 500 — `models.yml` folds `high` and
+`max` to `xhigh`), and thinking is enabled server-side with the card's
+recipe (temp 1.0 / top_p 0.95 / top_k 20 / min_p 0) plus
+`--reasoning-format deepseek --reasoning-budget -1`. Two traps: a request
+whose whole budget is consumed by reasoning returns `content: null` (bench
+clients must fall back to `reasoning_content` or they crash *after*
+measuring), and the fork logs
+`MTP boundary missing for seq_id=… ; resyncing` under concurrent draft
+resets — self-recovering, not a fault.
+
+Fetch: `scripts/download-frank-large-kingjones.sh` (`-f` + byte count + LFS
+sha256 verified, `--verify-only`, no token on the command line).
+
+Runbook: `serverAppConfigs-docs/LLM/frank-large-rocmfpx-q8agent-mtp.md`.
+Rollback: git tag **`frank-large-gufo-q4kxl-20261001`** — gufo v0.1.1
+(`scripts/build-gufo.sh` stays on disk for it), the Q4_K_XL Qwen3.8-27B
+payload and the DFlash2 drafter all remain on the host; restore the tag's
+unit/env/serve trio and redeploy.

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

-Mainline llama.cpp master `37b53fd` as a **root systemd unit** (repo
-`scripts/frank-small.service`, `relink_unit`ed into `/etc/systemd/system/`),
-serving `ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF` at IQ3_S — the `-mtp` build, so
-the model's own Multi-Token Prediction head drives speculative decoding — with
-the BF16 mmproj, on port 7771, 131,072 ctx / 1 slot, f16 KV, greedy default.
-Deployed flags are in `scripts/frank-small.env`; the serve script refuses to
-start if the weights, the projector or the binary are missing.
+**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_unit`ed into
+`/etc/systemd/system/`), serving
+`bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF` at **Q5_K_M** (6,876,124,704 B,
+sha256 `04cf4f…`) **plus `mmproj-MiMo-V2.6-Distill-Qwen-9B-f16.gguf`**
+(918,166,048 B, sha256 `ff348f…`), both `/home/dank/models/MiMo-Q5KM/`, on
+port 7771. The 2026-10-02 swap was a **quant rung down from Q6_K**, not a
+runtime change — the llama.cpp path 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; SGLang's `--kv-cache-dtype` moves K and V together.
+Vision came the same day via `--mmproj`: the GGUF carries no `v.*` tensors, so
+the projector file is the entire vision path (and ~1.4 GiB of the card).

-Verified at deploy (2026-09-16): `/v1/models` reports `frank-small`; a chat
-completion returns content plus `reasoning_content`; a tool call returns
-`tool_calls`; the slot timing lines show `draft acceptance = 0.66–0.94, mean
-len ≈ 3` (MTP engaged); an image request is answered from the projector while
-the halogen side refuses images with a named reason; 22,362 MiB of the card's
-24,576 MiB is in use; `n_slots = 1, n_ctx_slot = 131072`.
+**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.

-### frank-medium — llama.cpp layer split, DFlash2 + ngram-mod (Dankserver, ACTIVE)
+**The shape:** `--ctx-size 540672 --parallel 3` = **THREE conversations of
+180,224 tokens each** (llama.cpp divides the pool:
+`n_ctx_slot = CTX_SIZE / PARALLEL`), `--flash-attn on` (the q8_0-V path is
+only cheap under flash attention — the seam pins it), f16 K + q8_0 V, all
+layers on the GPU, `--mmproj` for vision. Measured VRAM **21,684 used / 2,441
+free** of 24,576 MiB — the 2–3 GiB headroom Frank asked for, landed by
+solving the pool against two measured boots: non-KV cost is 7,390 MiB without
+a projector (2 × 262,144 boot) and 8,859 MiB with one (3 × 196,608 boot), so
+vision costs ~1.4 GiB, not the 0.9 GiB the file size suggests. Three native
+windows (786,432 tokens) need 18.4 GiB of KV alone and were never on the
+table. Full f16 V would add ~4 GiB more and bust the card — that is the
+justification for the asymmetric contract.

-Mainline llama.cpp at **`fb27a525`** (a gfx1201 HIP build, `Root=`-less root
-systemd unit `scripts/frank-medium.service`) serving
-`ajgazin/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF` at **UD-Q8_K_XL**
-(29.29 GiB) with the **z-lab DFlash2** `Q8_0` drafter chained to the
-draftless **ngram-mod** matcher, in **one process across both R9700s**,
-`--split-mode layer` at a **35:30 tensor split** (the balance point for the
-deployed batch: the drafter rides ROCm1, so it takes fewer target layers),
-262144 ctx over **2 slots** of 131072, batch 6144x512, text only.  DSpark was
-put up in the drafter's place the same day and reverted: it declares no sliding
-window, so its draft KV spanned the whole context (5.0 GiB) and cost ROCm1 four
-layers of the split.
+Verified at deploy (2026-10-02): `/v1/models` reports `frank-small`; boot logs
+`n_slots = 3, n_ctx_slot = 180224` and `loaded multimodal model`; decode
+**106.8 tok/s** at temp 0 (prefill 3,814 t/s) — ~5 % faster than the
+2 × 262,144 shape, because a shorter slot means a shorter attention scan;
+three concurrent requests answered in parallel and a fourth queued behind a
+released slot; native `tool_calls`; thinking in `reasoning_content`; a
+solid-red PNG answers "Red." through an `image_url` content part (the legacy
+top-level `images` field is ignored by `/v1/chat/completions` — the model then
+invents a color). A prompt over the slot is a **400
+`exceed_context_size_error`**, which is why `models.yml` sets
+`contextWindow: 180224`.

-**Layer split is why both cards now serve ONE model.** Tensor parallelism is
-impossible (see the PCIe lesson below), and so is vLLM's pipeline parallelism —
-its stage handoff is an RCCL `send`/`recv` of the hidden state, the exact
-primitive that cannot be dispatched on card #2. llama.cpp's layer split needs
-no collective at all: layers are distributed in order and the boundary
-activation is a per-backend copy, once per forward pass. Measured on this host
-(the predecessor runtime): 1991 t/s prefill layer-split vs 272 tensor.
+Runbook: `serverAppConfigs-docs/LLM/frank-small-llama-mimo-q5km.md`.
+Rollback: [[frank-small-llama-mimo-q6k]] (Q6_K file on disk at
+`/home/dank/models/MiMo-Q6K/`); before that [[frank-small-sglang-mimo-v26]]
+(venv on disk at `/home/dank/GitPackages/sglang-cu128/`; its INT8 payload was
+staged to `~dank/.AGENTTRASH/20261002-retired-llm-payloads/` on 2026-10-02).

-This is the box's **only** LLM runtime: the vllm-radiance replica pair and its
-routing proxy (`frank-medium-a/-b/-proxy`) were retired the same day they went
-live, with their units moved to `~/.AGENTTRASH/20260916-vllm-radiance-retire/`.
+### frank-primary — tcclaviger vLLM, Qwen3.8-Flash-Next FP8 + MTP4 (Dankserver, ACTIVE)

-Files: `scripts/build-llama-rdna4.sh` (builds the runtime, gfx1201 only, private
-ROCm 7.14 toolchain bundled beside the binary), `scripts/download-frank-medium-gguf.sh`
-(resumable, byte-verified payload), `scripts/frank-medium.env`,
-`scripts/serve-frank-medium.sh`, `scripts/frank-medium.service`. The deploy
-handler refuses to start a unit whose runtime or payload is missing, retires
-the vLLM quadlets, waits for 7770 to be released (the retired proxy held it) and
-gates on `/health`.
+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**,
+**MTP4** speculative decoding, **expert prefetch OFF**, 200,000 ctx x 3 seqs,
+**fp8 KV**. Deployed 2026-10-01 by rollback from commit `2339fae` after the
+GLM/DFlash2 experiment was declared failed (same day).

-The ref is pinned by two features and must not float: **DFlash2**
-(ggml-org/llama.cpp#27342) and **qwen35 MTP**. The ROCmFPX fork the endpoint
-used earlier that day has neither.
+Measured 2026-09-29 tuning campaign: **decode 37.4-40.6 tok/s, acceptance
+~54 %** (MTP4 + prefetch off — the fleet's only endpoint that meets the
+35 tok/s target). 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.

-Runbook: `serverAppConfigs-docs/LLM/frank-medium-llamacpp-dflash2.md`.
-Retired runtimes: [[frank-medium-radiance-mxfp4]] (vLLM replicas, same day),
-[[frank-medium-rocmfpx]] (llama.cpp fork, earlier that day).
+The engine auto-sets the attention block size (mamba-page floor); the
+expert plan's registry prices KV per token; the expert arena (6,574 slots
+x 2.344 MiB = 15.05 GiB/rank) is the decode-critical region — cutting it
+for KV headroom collapses the offload hit rate.
+
+Unit and self-healing: root systemd unit (`scripts/frank-primary.service`,
+`User=dank`, `LimitMEMLOCK=infinity`), `ExecStop=podman stop -t 120
+frank-primary-glm`. 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; 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
@@ -376,29 +473,40 @@ linked runbooks.
  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`.
- **The fix is physical — and cheap on this board.** Two paths put both cards
-  on CPU root ports, which restores the atomics routing and therefore TP=2: an
-  M.2→PCIe x4 riser into a *CPU-attached* M.2 slot (on dankserver's TUF
-  GAMING B650-E WIFI that is M.2_2, which is free — per ASUS's spec page;
-  `PCIEX16_2`, where card #2 sits today, is chipset and x1-mode), or a board
-  whose CPU x16 bifurcates to x8/x8. Bandwidth is not the constraint: a 27B
-  TP=2 allreduce at batch 4 is ≈10 MB per token-step, ≈1.4 ms at Gen4 x4.
-  RCCL's P2P `send`/`recv` fails on such a card too, so pipeline parallelism is
-  not an escape either. Full evidence matrix, the open rebuild hypothesis, and
-  the test plan: [[Dual-R9700-Cooperation]].
+- **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 one software path that lets both cards serve ONE model: llama.cpp's
-  layer split.** vLLM cannot do it here in any of its parallelisms — TP needs
-  an allreduce, PP needs the same p2p primitive, and both die on the card that
-  cannot dispatch an RCCL kernel. `--split-mode layer` in llama.cpp needs no
+- **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). That is what `frank-medium` runs — see
+  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
@@ -437,9 +545,10 @@ linked runbooks.

 **ROCm / GPU**

- **Layer split is the deployed shape on both multi-device llama.cpp services,
-  for two different reasons.** On Dankserver it is the *only* way both R9700s
-  serve one model (no collective can be dispatched on card #2). On a
+- **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
@@ -472,9 +581,13 @@ linked runbooks.
 - **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 rocm-smi figure** (64 GiB, SVM-backed).
-  `free`/`dmidecode` show the CPU-side carve-out — intentional, do not
-  change.
+- **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**

@@ -513,6 +626,26 @@ linked runbooks.
  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**

@@ -585,6 +718,35 @@ linked runbooks.
  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.
+  Measured 2026-10-02. See [[frank-small-llama-mimo-q5km]].
+- **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
@@ -593,7 +755,9 @@ linked runbooks.
  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, which is why frank-small runs f16.
+  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
@@ -683,9 +847,9 @@ Active:

 | Service | Runbook |
 |---------|---------|
-| frank-large (halogen-flash-server, Dankclaw) | `serverAppConfigs-docs/LLM/halogen-flash-server.md` |
-| frank-small (llama.cpp Qwen3.8-27B GSQ-RCO, Dankclaw) | `serverAppConfigs-docs/LLM/frank-small-qwen38-gsq-rco.md` |
-| frank-medium (llama.cpp + DFlash2, Dankserver) | `serverAppConfigs-docs/LLM/frank-medium-llamacpp-dflash2.md` |
+| frank-large (ROCmFPX llama.cpp fork, Qwen3.8-27B Q8_0_ROCMFPX_AGENT + MTP4 + vision, Dankclaw) | `serverAppConfigs-docs/LLM/frank-large-rocmfpx-q8agent-mtp.md` |
+| frank-small (llama.cpp CUDA, MiMo-V2.6-Distill-Qwen-9B Q5_K_M, Dankclaw) | `serverAppConfigs-docs/LLM/frank-small-llama-mimo-q5km.md` |
+| frank-primary (tcclaviger vLLM, Qwen3.8-Flash-Next FP8 + MTP4, 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` |

@@ -693,8 +857,25 @@ Retired (rollback targets and history):

 | Runtime | Host | Window | Runbook |
 |---------|------|--------|---------|
-| kingjones STRIX Qwen3.8-Flash-Next ROCmFP4 | Dankclaw | 2026-08-30 → 09-11 | `kingjones-strix-rocmfp4.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 (rollback target) | `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 (rollback target) | `frank-small-llama-mimo-q6k.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` |
@@ -704,6 +885,7 @@ Retired (rollback targets and history):
 | 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/`.
@@ -13,7 +13,7 @@ one deploy script.
 | Host | User | IP | Podman stack | LLM service | LLM port |
 |------|------|----|--------------|-------------|----------|
 | dankserver | dank | 192.168.0.235 | media and library | frank-medium (ROCmFPX llama.cpp, Qwen3.8-27B AGENT) | 7770 |
-| dankclaw | dank | 192.168.0.143 | AI, search, development, observability | frank-large (CIRU ROCmFPX, Qwen3.8-27B) | 7770 |
+| dankclaw | dank | 192.168.0.143 | AI, search, development, observability | frank-large (gufo, Qwen3.8-27B UD-Q8_K_XL) | 7770 |

 ## Active Services

@@ -0,0 +1,104 @@
+#!/usr/bin/env bash
+# autoresearch.sh — canonical benchmark entrypoint: speed of frank-primary.
+#
+# Runs on the operator box and drives one bench window on DANKSERVER:
+#   1. fingerprint the candidate trio (env + serve script + unit)
+#   2. memo fast-path: an identical fingerprint that already produced a FULLY
+#      successful window (bench + verified production restore) returns the
+#      cached metrics without touching the fleet
+#   3. stage the candidate trio plus the measure client into
+#      /home/dank/.autoresearch-frank-primary/ (ephemeral; the repo checkout
+#      on the host stays the production source of truth)
+#   4. run the window (scripts/autoresearch/frank-primary-window.sh):
+#      stop + mask the production unit, boot the staged config AS THE SERVICE
+#      (same serve script, same container, same port 7770), wait for health,
+#      run the in-container usage-counter measure, then restore the production
+#      unit and verify its health — AGENTS.md Benchmarking Policy
+#   5. print METRIC lines (parse by run_experiment)
+#
+# Primary metric: decode_tps (HIGHER is better) — median across the three
+# prompt depths (115/1056/4096) of per-depth medians of
+# completion_tokens / (wall - ttft), 3 runs each after a warm pass,
+# temperature 0, the campaign's deterministic word-seed workload.
+#
+# Exit 0 with METRIC lines on success; non-zero on any failure (the window
+# script's EXIT trap restores production on every path).
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+HOST="${FRANK_PRIMARY_AR_HOST:-dankserver}"
+STAGE="/home/dank/.autoresearch-frank-primary"
+
+ENV_FILE="$SCRIPT_DIR/scripts/frank-primary.env"
+SERVE="$SCRIPT_DIR/scripts/serve-frank-primary.sh"
+UNIT="$SCRIPT_DIR/scripts/frank-primary.service"
+MEASURE="$SCRIPT_DIR/scripts/autoresearch/frank-primary-measure.py"
+WINDOW="$SCRIPT_DIR/scripts/autoresearch/frank-primary-window.sh"
+
+for f in "$ENV_FILE" "$SERVE" "$UNIT" "$MEASURE" "$WINDOW"; do
+    [ -f "$f" ] || { echo "[autoresearch] missing harness file: $f" >&2; exit 1; }
+done
+command -v python3 >/dev/null 2>&1 || { echo "[autoresearch] python3 required on operator box" >&2; exit 1; }
+
+# Fingerprint of everything that can change the engine's argv or environment.
+HASH="$(cat "$ENV_FILE" "$SERVE" "$UNIT" | sha256sum | cut -d' ' -f1)"
+echo "[autoresearch] candidate fingerprint $HASH" >&2
+
+ssh_run() { ssh -o BatchMode=yes -o ConnectTimeout=15 "$HOST" "$@"; }
+
+emit_metrics() { # $1 = run.json
+    python3 -c '
+import json, sys
+run = json.loads(sys.argv[1])
+m = run["measure"]
+print("METRIC decode_tps=%s" % m["decode_tps"])
+for k in ("decode_tps_115", "decode_tps_1056", "decode_tps_4096"):
+    print("METRIC %s=%s" % (k, m[k]))
+print("METRIC ttft_ms_4096=%s" % m["ttft_ms_4096"])
+print("METRIC kv_tokens=%s" % (run.get("kv_tokens") or 0))
+print("METRIC boot_seconds=%s" % (run.get("boot_seconds") or -1))
+print("METRIC completion_tokens_4096=%s" % m.get("completion_tokens_4096", 0))
+' "$1"
+}
+
+# --- memo fast-path -------------------------------------------------------
+cached_hash="$(ssh_run "cat '$STAGE/last/hash' 2>/dev/null" 2>/dev/null || true)"
+if [ -n "$cached_hash" ] && [ "$cached_hash" = "$HASH" ]; then
+    cached_run="$(ssh_run "cat '$STAGE/last/run.json' 2>/dev/null" || true)"
+    if [ -n "$cached_run" ]; then
+        echo "[autoresearch] memo hit — cached metrics for $HASH, fleet untouched" >&2
+        emit_metrics "$cached_run"
+        exit 0
+    fi
+fi
+
+# --- stage the candidate ----------------------------------------------------
+echo "[autoresearch] staging candidate on $HOST:$STAGE" >&2
+ssh_run "mkdir -p '$STAGE'"
+scp -q "$ENV_FILE" "$HOST:$STAGE/frank-primary.env"
+scp -q "$SERVE" "$HOST:$STAGE/serve.sh"
+scp -q "$MEASURE" "$HOST:$STAGE/measure.py"
+scp -q "$WINDOW" "$HOST:$STAGE/window.sh"
+ssh_run "printf '%s\n' '$HASH' > '$STAGE/hash'"
+
+# --- the window -------------------------------------------------------------
+# Budgets: bench boot up to 20 min (warm boots measured ~7 min), measure up
+# to 30 min, restore boot up to 20 min — call run_experiment with
+# timeout_seconds >= 3600.
+echo "[autoresearch] opening bench window (stop+mask -> boot -> measure -> restore)" >&2
+set +e
+ssh_run "bash '$STAGE/window.sh'"
+rc=$?
+set -e
+if [ "$rc" -ne 0 ]; then
+    echo "[autoresearch] window failed (rc=$rc); host-side evidence:" >&2
+    ssh_run "tail -60 '$STAGE/fail.log' 2>/dev/null || tail -40 '$STAGE/serve.out' 2>/dev/null || tail -40 '$STAGE/measure.out' 2>/dev/null || true" >&2 || true
+    exit 1
+fi
+
+run_json="$(ssh_run "cat '$STAGE/run.json'")" || {
+    echo "[autoresearch] could not read $STAGE/run.json back" >&2; exit 1; }
+echo "[autoresearch] window result:" >&2
+printf '%s\n' "$run_json" >&2
+emit_metrics "$run_json"
+exit 0
@@ -4,6 +4,11 @@
 # SearXNG
 SEARXNG_SECRET=changeme
 SEARXNG_BASE_URL=http://search.franksdank.cloud/
+# Kagi search API key (https://help.kagi.com/kagi/api/). Powers the paid
+# `kagi` engine in shared/searxng/settings.yml; deploy-remote.sh bakes it into
+# the live settings.yml via scripts/searxng-render-settings.py. The quadlet
+# does NOT pass it into the container environment.
+KAGI_API_KEY=changeme

 # Gluetun VPN (see https://github.com/qdm12/gluetun-wiki)
 # Single-peer WireGuard model. Fill from your provider's WireGuard config.
@@ -309,6 +309,10 @@ http://:8017 {
 	@home host www.franksdank.cloud
 	handle @home {
 		root * /srv/homepage
+		# Deploys overwrite /srv/homepage in place; without revalidation a
+		# browser keeps serving the pre-deploy CSS/JS/HTML. no-cache forces an
+		# ETag round-trip on every load (a cheap 304 when unchanged).
+		header Cache-Control "no-cache"
 		# RBAC (2026-09-01): /admin requires membership in the homepage-admins
 		# authentik group. In proxy-provider mode the outpost injects
 		# X-Authentik-Groups on its upstream request to this listener and strips
@@ -323,6 +327,16 @@ http://:8017 {
 			}
 			respond "Forbidden: homepage admins only" 403
 		}
+		# User docs page (static). Anchors (#models etc.) are client-side,
+		# so both /docs and /docs/ map onto the single docs.html.
+		handle /docs {
+			try_files /docs.html
+			file_server
+		}
+		handle /docs/ {
+			try_files /docs.html
+			file_server
+		}
 		@data path /services.json /status.json /app.js /style.css /deployed.js
 		handle @data {
 			file_server
@@ -21,4 +21,20 @@ RUN mkdir -p /opt/ComfyUI/models /opt/ComfyUI/input /opt/ComfyUI/output /opt/Com
 RUN grep -viE '^(torch|torchvision|torchaudio)([^a-z]|$)' requirements.txt > /tmp/req-notorch.txt \
    && pip install --no-cache-dir -r /tmp/req-notorch.txt

+# The city96/ComfyUI-GGUF custom node (loaded from the host's custom_nodes
+# volume) needs the gguf library to decode GGUF-quantized diffusion models.
+# sentencepiece and protobuf are its optional tokenizer deps. Baking them into
+# the image keeps the dependency across container recreations — the node itself
+# lives in the host volume, which persists, but the Python env does not.
+RUN pip install --no-cache-dir "gguf>=0.13.0" sentencepiece protobuf
+
+# Bake the qwen_image21 patch script into the image. It is run by the
+# startup wrapper before ComfyUI starts, so the patch survives container
+COPY dankclaw/apply_qwen21_patch.py /opt/ComfyUI/apply_qwen21_patch.py
+COPY dankclaw/comfyui-start.sh /opt/ComfyUI/comfyui-start.sh
+RUN chmod +x /opt/ComfyUI/comfyui-start.sh
+
+CMD ["/opt/ComfyUI/comfyui-start.sh"]
 EXPOSE 8188
+# v2: VAE 4D-to-5D latent expansion for video VAEs (qwen_image21)
+# v3: fix in_channels to use shape[0] (latent input dim) not shape[1] (hidden dim)
@@ -0,0 +1,52 @@
+import os
+
+# Apply qwen_image21 detection patch
+f = "/opt/ComfyUI/comfy/model_detection.py"
+with open(f) as fh:
+    code = fh.read()
+if "qwen_image21" not in code:
+    qwen21_check = (
+        "    if '{}txt_in.text_norm.weight'.format(key_prefix) in state_dict_keys:  # Qwen Image 2.1\n"
+        "        dit_config = {}\n"
+        '        dit_config["image_model"] = "qwen_image21"\n'
+        "        dit_config[\"in_channels\"] = state_dict['{}img_in.weight'.format(key_prefix)].shape[0]\n"
+        "        dit_config[\"num_layers\"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')\n"
+        "        return dit_config\n"
+    )
+    anchor = "    if '{}txt_norm.weight'.format(key_prefix) in state_dict_keys:  # Qwen Image"
+    code = code.replace(anchor, qwen21_check + "\n" + anchor)
+    with open(f, "w") as fh:
+        fh.write(code)
+    print("Applied qwen_image21 detection patch")
+
+# Add qwen_image21 to GGUF node arch list
+loader = "/opt/ComfyUI/custom_nodes/ComfyUI-GGUF/loader.py"
+if os.path.exists(loader):
+    with open(loader) as fh:
+        lcode = fh.read()
+    if '"qwen_image21"' not in lcode:
+        old = '"lumina2", "qwen_image"}'
+        new = '"lumina2", "qwen_image", "qwen_image21"}'
+        if old in lcode:
+            lcode = lcode.replace(old, new)
+            with open(loader, "w") as fh:
+                fh.write(lcode)
+            print("Applied qwen_image21 arch list patch")
+
+# Patch VAE decode to expand 4D latents to 5D for video VAEs (latent_dim=3)
+sd = "/opt/ComfyUI/comfy/sd.py"
+with open(sd) as fh:
+    sdc = fh.read()
+if "self.latent_dim == 3 and samples_in.ndim == 4" not in sdc:
+    anchor = "        if self.latent_dim == 2 and samples_in.ndim == 5:\n            samples_in = samples_in[:, :, 0]"
+    replacement = (
+        "        if self.latent_dim == 2 and samples_in.ndim == 5:\n"
+        "            samples_in = samples_in[:, :, 0]\n"
+        "        elif self.latent_dim == 3 and samples_in.ndim == 4:\n"
+        "            samples_in = samples_in.unsqueeze(2)"
+    )
+    if anchor in sdc:
+        sdc = sdc.replace(anchor, replacement)
+        with open(sd, "w") as fh:
+            fh.write(sdc)
+        print("Applied VAE 4D-to-5D latent expansion patch")
@@ -20,10 +20,10 @@ Wants=authentik-server.service
 [Container]
 ContainerName=authentik-outpost-ldap
 Network=dankclaw_default
-# Image=ghcr.io/goauthentik/ldap:2026.8.0 (amd64 digest; podman 5.7.0 refuses
+# Image=ghcr.io/goauthentik/ldap:2026.8.3 (amd64 digest; podman 5.7.0 refuses
 # the tag's nested manifest list, so pin the single-platform digest).
 EnvironmentFile=/home/dank/serverAppConfigs/dankclaw/.env
-Image=ghcr.io/goauthentik/ldap@sha256:fa33bc8f522113830e1662f4c034352f9eb08e0a7d3da2ddadad9a4b33b4cbf4
+Image=ghcr.io/goauthentik/ldap@sha256:2e7b259eb8c37faac916946ddf48ec681ee9d3a9f724fe8521bf161973f13cf9
 Environment=AUTHENTIK_HOST=https://auth.franksdank.cloud
 Environment=AUTHENTIK_INSECURE=false
 Environment=AUTHENTIK_TOKEN=${AUTHENTIK_LDAP_OUTPOST_TOKEN}
@@ -20,7 +20,7 @@ Wants=authentik-server.service

 [Container]
 ContainerName=authentik-outpost
-Image=ghcr.io/goauthentik/server:2026.8.0
+Image=ghcr.io/goauthentik/server:2026.8.3
 Exec=worker
 Network=dankclaw_default
 EnvironmentFile=/home/dank/serverAppConfigs/dankclaw/.env
@@ -19,7 +19,7 @@ Wants=authentik-postgres.service

 [Container]
 ContainerName=authentik-server
-Image=ghcr.io/goauthentik/server:2026.8.0
+Image=ghcr.io/goauthentik/server:2026.8.3
 Exec=server
 Network=dankclaw_default
 EnvironmentFile=/home/dank/serverAppConfigs/dankclaw/.env
@@ -16,7 +16,7 @@ Wants=authentik-server.service

 [Container]
 ContainerName=authentik-worker
-Image=ghcr.io/goauthentik/server:2026.8.0
+Image=ghcr.io/goauthentik/server:2026.8.3
 Exec=worker
 Network=dankclaw_default
 EnvironmentFile=/home/dank/serverAppConfigs/dankclaw/.env
@@ -0,0 +1,4 @@
+#!/bin/bash
+# Apply qwen_image21 patches before starting ComfyUI
+python3 /opt/ComfyUI/apply_qwen21_patch.py 2>/dev/null
+exec python main.py --listen 0.0.0.0 --port 8188
@@ -21,11 +21,13 @@
 # user-unit spawn — which carries no supplementary groups — can still open them.
 # Both halves are asserted by tests/comfyui_rocm.bats.
 #
-# EXPOSURE. Loopback only. ComfyUI has no authentication and executes arbitrary
-# node graphs, and it has no Caddy vhost, so the repository's rule applies:
-# an unguarded service is internal-only. The homepage status card probes
-# 127.0.0.1:8188, which this satisfies. To reach the UI from another machine,
-# guard it behind the auth chain and give it a vhost — do not widen the bind.
+# EXPOSURE. LAN-reachable, no auth. ComfyUI has no authentication and executes
+# arbitrary node graphs. It is internal-only (no public vhost, no guard chain);
+# it is reached on the LAN via the Pi-hole DNS record
+# (comfyui.franksdank.cloud -> 192.168.0.143, in dankclaw/pihole/addn-hosts),
+# which resolves to this host's IP — so the bind is widened from loopback to
+# 0.0.0.0:8188. The homepage admin card links to that URL and probes
+# 127.0.0.1:8188.
 #
 # DATA. Models, input, output and custom_nodes live under
 # /home/dank/comfyui-data (host state, not in the repo).
@@ -39,7 +41,7 @@ ContainerName=comfyui
 Image=localhost/comfyui-rocm:gfx1151
 Pull=never
 Network=dankclaw_default
-PublishPort=127.0.0.1:8188:8188
+PublishPort=0.0.0.0:8188:8188
 AddDevice=/dev/kfd
 AddDevice=/dev/dri
 GroupAdd=keep-groups
@@ -48,7 +50,7 @@ Volume=/home/dank/comfyui-data/input:/opt/ComfyUI/input
 Volume=/home/dank/comfyui-data/output:/opt/ComfyUI/output
 Volume=/home/dank/comfyui-data/custom_nodes:/opt/ComfyUI/custom_nodes
 Volume=/home/dank/serverAppConfigs/dankclaw/comfyui/comfyui-healthcheck.py:/usr/local/bin/comfyui-healthcheck.py:ro
-Exec=python main.py --listen 0.0.0.0 --port 8188
+Exec=/opt/ComfyUI/comfyui-start.sh
 # The probe lives in a file because the podman 5.x v1 generator splits
 # HealthCmd on spaces into an argv array; an inline `python3 -c "..."` becomes
 # a puzzle of quoting that fails silently.
@@ -1,219 +0,0 @@
-# Quadlet unit: frank-large (rootless, boot-persistent).
-#
-# Install path on dankclaw: ~/.config/containers/systemd/frank-large.container
-# Unit name: frank-large.service (container name: frank-large)
-#
-# The large inference service. The halogen-flash-server runtime (since
-# 2026-09-11, restored 2026-09-16 after a two-day ik_llama.cpp hybrid
-# interlude) serving peonist-ai/halogen-qwen3.8-flash-next (Qwen3.8-Flash-Next,
-# w4b .hgn checkpoint + quality overlay sidecar). It replaced the llama.cpp /
-# kingjones30-ROCmFPX STRIX deployment because the engine is ~4x faster on
-# prefill (the dominant cost of any real prompt) on this exact SKU. The RTX
-# 3090 it briefly shared with frank-large is now frank-small's card.
-#
-# SHAPE. The image runs both halves in one container: the engine on container
-# loopback (:8730, UNAUTHENTICATED — never publish it) and the OpenAI
-# front-end on :8731, published here as host 7770 so the fleet's clients,
-# services.json probe and docs keep the port they have always used. Same for
-# HALOGEN_MODEL_ID=frank-large: it is the id at /v1/models, in /health and in
-# every response, so no client had to be renamed. Halogen never rejects a
-# request for naming a different model, it just serves it.
-#
-# DEVICES. gfx1151 through the amdgpu/KFD stack: /dev/kfd plus the DRM render
-# node. Both are 0666 on this host, so the rootless user namespace can open
-# them (id dank is also in render/video). seccomp=unconfined and --ipc=host are
-# the image's documented requirements; memlock must be unlimited.
-#
-# DO NOT name an environment variable TEMP, HOST or GPU_MAX_HEAP_SIZE: the HIP
-# runtime reads bare TEMP for its JIT cache path (2026-08-13 incident).
-#
-# MEMORY. Weights are pinned (HALOGEN_FLASH_PIN_TRUNK default 1) — that is the
-# configuration the published decode numbers were measured with, and `free`,
-# MemAvailable and every monitor built on them will overstate free memory by
-# the size of the model (~68 GiB) because the kernel counts the locked weights
-# as reclaimable page cache. Halogen prints the corrected figure at startup.
-# The pinned trunk means this container holds most of the host; a co-tenant
-# that then takes the remainder can stall both for minutes at full CPU with no
-# disk activity and no output. That is documented behaviour, not a crash, and
-# the escape hatch is HALOGEN_FLASH_PIN_TRUNK=0 (at several times the decode
-# cost). Sizing is chosen for a SHARED 121 GiB host, and HALOGEN_KV_POOL_FIT
-# stays on as the safety net: if the machine refuses the configured pool the
-# engine lowers it itself and prints what it settled on. Read that ledger on
-# every restart — it is the ground truth for sizing, not this comment.
-#
-# POOL RETUNED 2026-09-16: 400,384 positions -- exactly two 200,192-token
-# conversations -- beside a per-conversation ceiling of 200,192 and two slots.
-# It replaces the 401,408 the 2026-09-13 trim left behind, which held one full
-# native-262k conversation beside a ~139k one, or three at 131k.
-#
-# MEASURED AT 401,408 (2026-09-13 deploy ledger): 68.0 weights + 11.0 KV +
-# 21.0 working = 100.0 GiB held, 9.1 GiB / 1396 contiguous 2 MiB blocks left,
-# engine need printed ~31.5 against the modelled ~33.4, and NO stall-warning
-# NOTE. Fit-check budget here: 121.2 - 67.7 - 20 = 33.5. That boot is the
-# baseline the 09-16 retune is measured against: the vision tower (1.97 GiB)
-# and roughly 4.2 GiB of arena come back to the host, which is what pays for
-# the co-tenant below.
-#
-# CO-TENANT since 2026-09-16: this host also serves frank-small on the RTX
-# 3090 (llama.cpp, ~12.2 GiB of weights that load through the page cache and
-# then live in VRAM, plus a CUDA context). It is the first real co-tenant this
-# service has had, which is why the arena was halved and the reserve raised.
-#
-# BOOT NOTE: on the first request after a (re)start the engine can go silent
-# for ~45-60 s ("has not answered PING") while it first-touches the freshly
-# reserved pool; /health reads engine_unresponsive during it and recovers by
-# itself. The quadlet health check's HealthStartPeriod=300s covers it; do not
-# restart the unit over it.
-#
-# WEIGHTS. HALOGEN_DOWNLOAD fetches the whole weights repo on FIRST START ONLY
-# (the entrypoint checks whether the checkpoint is absent) into the bind mount
-# below, which must therefore be read-WRITE. huggingface_hub resumes partial
-# files natively, and the quadlet runs the container with --replace --rm, so a
-# restart mid-fetch continues rather than starting over. Roughly 127 GB: the
-# 115.55 GiB checkpoint, the 2.31 GiB quality overlay, the 0.84 GiB vision
-# tower and the tokenizer. The repo is public and ungated, so no token is
-# needed; later starts open no outbound connections at all.
-#
-# VISION. Deliberately NOT set since 2026-09-16. The tower is a 1.97 GiB
-# charge against the same device budget as the pool and the arena, and the
-# fleet's image path moved to frank-small (llama.cpp on the RTX 3090, loaded
-# with --mmproj mmproj-Qwen3.8-27B-BF16). This server refuses image content
-# parts, and models.yml declares frank-large text-only to match.
-#
-# CLIENT CONTRACT (see serverAppConfigs-docs/LLM/halogen-flash-server.md).
-# There are no server-side sampler defaults: a request that sends no
-# temperature decodes GREEDY. The per-request budget defaults to 8,192 tokens
-# and REASONING COUNTS AGAINST IT, so a budget that runs out mid-thought
-# returns finish_reason "length" with empty content and the partial reasoning
-# in reasoning_content. The cap is 65,536; above it a request is a 400, never a
-# silent truncation.
-#
-# IMAGE. The unit runs the localhost/-aliased tag: this host's podman quadlet
-# path cannot resolve registry-prefixed refs from local storage at create
-# time. deploy-remote.sh pulls the ghcr.io ref and re-tags it before the
-# restart. Pull=never keeps boots off the registry.
-[Unit]
-Description=frank-large — halogen-flash-server, Qwen3.8-Flash-Next (w4b) on gfx1151
-After=network-online.target
-
-[Container]
-ContainerName=frank-large
-Image=localhost/peonist-ai/halogen-flash-server:0.5.8
-# Published port stays 7770: Hermes, open-webui, the harness and the homepage
-# probe all point here. 8730 (the engine) is deliberately not published.
-PublishPort=7770:8731
-AddDevice=/dev/kfd
-AddDevice=/dev/dri
-# Read-write: the first start downloads into it. Contains only weights.
-Volume=/home/dank/models/halogen:/models
-# Fetch the weights on first start only (absent checkpoint gates it).
-Environment=HALOGEN_DOWNLOAD=peonist-ai/halogen-qwen3.8-flash-next
-# Fleet identity: the id at /v1/models, /health and every response body.
-Environment=HALOGEN_MODEL_ID=frank-large
-# Per-conversation ceiling: exactly half the pool, so two conversations of the
-# full window are always resident. 200,192 rather than a round 200,000 is
-# pool/2. The model's native window is 262,144; a conversation past this
-# ceiling is refused, never silently truncated. Clients declare 200192 (see
-# the frank-large contextWindow in models.yml).
-Environment=HALOGEN_CTX=200192
-# The KV pool: 400,384 positions = 2 x 200,192 (the 2026-09-16 retune, down
-# from 401,408). ~28 KiB per position including block scratch, so ~11.0 GiB
-# held -- the 401,408 ledger entry measured exactly 11.0. The two slots still
-# OVERSUBSCRIBE the pool for any real request, because admission reserves
-# prompt + max_tokens: two full-window conversations each holding an output
-# budget are pool-bound, and the second waits in arrival order for the first
-# to release its region. Read the `kv pool:` lines on every boot;
-# HALOGEN_KV_POOL_FIT lowers the pool itself if the budget refuses the
-# configured one, and a green health check does not mean the pool you asked
-# for is the pool you got.
-Environment=HALOGEN_KV_POOL_POSITIONS=400384
-# Concurrency cap: two, one per half of the pool. Each slot is 111.0 MiB of
-# O(1) state, so this hands ~111 MiB back against the three set on 2026-09-12
-# for the fleet size of the time.
-Environment=HALOGEN_KV_SLOTS=2
-# Sizing is budget arithmetic, all of it. ONE budget covers the KV pool, the
-# per-call prefill arena and the vision tower together:
-#
-#   budget = 121.2 host - 67.7 resident weights - reserve
-#   need   = KV (~28 KiB/position) + arena + vision
-#   arena  = 3.80 GiB + 8.40 GiB x (HALOGEN_MAX_TOK / 16384)
-#
-# The arena term is linear in max_tok, and it reproduces five live restarts
-# AND upstream's three published "on the device" figures to within 0.1 GiB —
-# note those published figures are pool+arena totals, not pool alone, which is
-# the misreading that first made a doubling look affordable here. Corners this
-# box offered at a reserve of 24, at the old 262,144-position pool:
-#
-#   16384 -> need 28.6 vs 29.5  FITS, 0.9 spare   (doubling, 8-9% prefill cost)
-#   32768 -> need 37.0 vs 29.5  POOL CLAMPED to 262144, deploy still green
-#   65536 -> need 46.5 vs 29.5  OOM, will not boot; unit still reads ACTIVE
-#
-# LOWERED 32768 -> 24576 on 2026-09-16 for the frank-small co-tenant on this
-# same host. The arena's variable term goes 8.40 x 2 -> 8.40 x 1.5, handing
-# ~4.2 GiB back to the host at roughly a 4-5 % prefill cost -- the same lever
-# at 16,384 costs 8-9 %. The vendor's own notes name 24,576 as the largest
-# arena that allocated with headroom to spare, so this stays inside the range
-# it has tested.
-#
-# The arena is a PER-CALL PREFILL CHUNK, not a prompt ceiling: the
-# 200,192-token conversation is prefilled in arena-sized pieces, so a smaller
-# arena means more pieces and a slower prefill, not a smaller conversation.
-#
-# THE POOL, ARENA AND RESERVE ARE A MATCHED SET, NOT independent tunables --
-# the fit check's printed ledger on each boot is the arbiter, not these
-# comments.
-Environment=HALOGEN_MAX_TOK=24576
-# RAM the fit check refuses to count on, watched over by the engine's own
-# n-gram lookup table (47.7 GiB, read through the page cache and never held)
-# and, since 2026-09-16, by frank-small on the same host.
-#
-# RAISED 15 -> 20 with the 2026-09-16 retune. This knob feeds the fit-check
-# budget and is NOT an allocation: raising it cannot hand memory back to the
-# host, because the actual footprint is weights + KV + arena + vision, none of
-# which it touches — it can only take the pool away. What it does is keep the
-# fit check honest at the smaller arena. With the arena at 24,576 and no
-# vision tower, the engine's printed need falls to roughly 25 GiB against a
-# budget of 121.2 - 67.7 - 20 = 33.5, leaving ~8 GiB of margin, so the pool is
-# not clamped.
-#
-# The 2026-09-13 profile for comparison: 103.5 GiB held and "7.1 GiB left for
-# everything else" (310 contiguous 2 MiB blocks) at the doubled 524,288 pool,
-# which is why the pool came down; at 401,408 the same ledger read 9.1 GiB /
-# 1396 blocks. The 09-16 retune gives back the vision tower (1.97) and ~4.2 of
-# arena on top of that for the co-tenant, so the expected figure at this boot
-# is ~15 GiB left. Read the ledger; do not trust this arithmetic.
-#
-# To free host memory, change the POOL or the ARENA, and the reserve then
-# becomes free to rise: HALOGEN_KV_POOL_POSITIONS=262144 hands back 3.6 GiB,
-# HALOGEN_MAX_TOK=16384 hands back another 4.2. Both are one-line edits the
-# fit check confirms on the next boot.
-Environment=HALOGEN_HOST_RESERVE_GIB=20
-# Verbose engine logging: the vendor's first-line evidence for load,
-# allocation and admission questions. Shipping it on means an incident report
-# already has the full startup account instead of needing a restart-into-debug.
-# HALOGEN_DMALLOC_LOG is deliberately NOT set — it logs every allocation
-# over 64 MB, which is steady-state noise here; raise it only for a live
-# OOM hunt.
-Environment=HALOGEN_VERBOSE=1
-# Required by the image: the engine maps the checkpoint and registers it with
-# the GPU in place, which needs host IPC and an unconfined seccomp profile.
-PodmanArgs=--ipc=host --security-opt seccomp=unconfined
-Ulimit=memlock=-1:-1
-# The image ships its own HEALTHCHECK; this pins it. /health asks the ENGINE
-# for a PONG before answering, so 200 means "ready to serve", not "process
-# alive". During the first-start download it stays unhealthy by design.
-HealthCmd=halogen-healthcheck api
-HealthInterval=30s
-HealthTimeout=10s
-HealthRetries=3
-HealthStartPeriod=300s
-Pull=never
-
-[Service]
-# The first start reads ~68 GiB off disk and pins it before serving.
-TimeoutStartSec=900
-Restart=on-failure
-RestartSec=10
-
-[Install]
-WantedBy=default.target
@@ -0,0 +1,54 @@
+# Quadlet unit: alertmanager (rootless, boot-persistent).
+#
+# Install path on dankclaw: ~/.config/containers/systemd/alertmanager.container
+# Unit name: alertmanager.service (container name: alertmanager)
+#
+# The delivery half of the LLM alert rules. Until 2026-09-17 the rules in
+# shared/prometheus/rules/llm-fleet.yml evaluated and reached nobody — a
+# firing alert was visible in Prometheus and on the Grafana dashboard and
+# nowhere else, and an undelivered alert looks identical to a quiet one.
+# This unit closes that gap: Prometheus pushes to alertmanager:9093 over the
+# dankclaw_default network, and the single receiver webhooks the host-side
+# signal-relay user unit, which hands the text to signal-cli under Frank's
+# existing Signal account.
+#
+# PUBLISHING. 127.0.0.1 only, like every other observability unit: the
+# unauthenticated Alertmanager API (lifecycle endpoints are always on in
+# this binary — there is no flag to gate them) has no business on a host
+# interface, and no off-host consumer needs it. Caddy gains no vhost for
+# it; the homepage timer probes loopback.
+#
+# CONFIG. /opt/alertmanager/config/alertmanager.yml is installed from
+# shared/prometheus/alertmanager.yml by deploy-remote.sh on every deploy, so
+# the repo copy is the source of truth. STATE lives in /opt/alertmanager/data.
+#
+# Rootless: container root maps to the invoking user (dank) in this user
+# namespace and dank owns /opt/alertmanager/data. The image's default user
+# (nobody) maps to a subuid that cannot write it — same User=0:0 as prometheus.
+
+[Unit]
+Description=Alertmanager alert delivery (Dankclaw)
+After=network-online.target
+Wants=network-online.target
+
+[Container]
+ContainerName=alertmanager
+Image=docker.io/prom/alertmanager:latest
+Network=dankclaw_default
+User=0:0
+Exec=--config.file=/etc/alertmanager/alertmanager.yml --storage.path=/alertmanager
+PublishPort=127.0.0.1:9093:9093
+Volume=/opt/alertmanager/config:/etc/alertmanager
+Volume=/opt/alertmanager/data:/alertmanager
+HealthCmd=wget --spider -q http://localhost:9093/-/healthy
+HealthInterval=30s
+HealthTimeout=10s
+HealthRetries=3
+HealthStartPeriod=30s
+
+[Service]
+Restart=always
+RestartSec=10
+
+[Install]
+WantedBy=default.target
@@ -27,10 +27,11 @@
 #
 # RULES. Five LLM alert rules and one recording rule load from
 # /etc/prometheus/rules/ (installed from shared/prometheus/rules/ by the
-# deployer). There is still **no Alertmanager**, so nothing is delivered: a
-# firing alert is visible in Prometheus and on the Grafana dashboard's ALERTS
-# panel and reaches nobody. Delivery is an operator decision, not an oversight;
-# the fragility is that an undelivered alert looks identical to a quiet one.
+# deployer). Delivery went live 2026-09-17: the alerting block in
+# prometheus.yml points at alertmanager:9093, which webhooks every firing
+# group to the host-side Signal relay (192.168.0.143:18093, reached from
+# the container as host.containers.internal), which sends through signal-cli account 318311.
+# The Grafana dashboard's ALERTS panel remains the second surface.
 #
 # The config volume is also why rules belong in git: editing them in the UI is
 # lost on the next deploy, since the deployer overwrites the whole directory.
@@ -32,6 +32,52 @@ PublishPort=3005:8080
 Environment=OPENAI_API_BASE_URL=http://192.168.0.143:7770
 Environment=ANONYMIZED_TELEMETRY=False
 Environment=DO_NOT_TRACK=True
+
+# SSO via authentik (OIDC authorization-code flow, 2026-09-24). The OAuth
+# client (application slug "open-webui") lives in authentik; the client
+# SECRET stays off-repo in the git-ignored host env file (mode 600). The
+# container reaches the IdP over the public URL: Caddy publishes 0.0.0.0:443
+# and Pi-hole resolves auth.franksdank.cloud to the host LAN IP, so the
+# server-side code exchange works from this bridge. Roles come from the
+# "Open WebUI roles" scope mapping (entitlements "Open WebUI Users" /
+# "Open WebUI Admins"); OAUTH_MERGE_ACCOUNTS_BY_EMAIL folds SSO logins into
+# the existing local account (frank8465@protonmail.com). The OAuth vars are
+# read directly from the environment at runtime (the per-key config DB holds
+# only Admin-Panel overrides, so nothing is seeded on the live data dir).
+# NOTE: the OAUTH_SCOPES value MUST stay quoted — systemd splits unquoted
+# Environment= values on whitespace (unquoted, only "openid" survived).
+EnvironmentFile=/home/dank/Containers/open-webui/env
+Environment=OAUTH_CLIENT_ID=KmtzKc0bjWjaIFeM0puYudI1CLXR2p0jnPCSf4N0
+Environment=OAUTH_PROVIDER_NAME=authentik
+Environment=OPENID_PROVIDER_URL=https://auth.franksdank.cloud/application/o/open-webui/.well-known/openid-configuration
+Environment=OPENID_REDIRECT_URI=https://ai.franksdank.cloud/oauth/oidc/callback
+Environment=WEBUI_URL=https://ai.franksdank.cloud
+Environment=ENABLE_OAUTH_SIGNUP=true
+Environment=OAUTH_MERGE_ACCOUNTS_BY_EMAIL=true
+Environment="OAUTH_SCOPES=openid email profile roles"
+Environment=ENABLE_OAUTH_ROLE_MANAGEMENT=true
+# SSO-only (2026-09-24): the OIDC round-trip is verified, so the local
+# username/password path is retired. ENABLE_LOGIN_FORM hides the form;
+# ENABLE_PASSWORD_AUTH refuses password logins entirely, so the only way in
+# is the authentik OIDC flow (which folds into the local account by email).
+# The per-key config DB override (ui.enable_login_form) is flipped to false
+# in the same change — the DB wins over env, so both must agree.
+Environment=ENABLE_LOGIN_FORM=false
+Environment=ENABLE_PASSWORD_AUTH=false
+
+# Web search -> SearXNG, reached over the host-published port (NOT the
+# search.franksdank.cloud vhost): the Anubis->Authentik guard chain is a
+# browser mechanism and a plain HTTP client cannot solve it. The internal
+# port bypasses Caddy entirely. These env vars SEED the per-key config DB on
+# first run (fill-missing only); once the keys exist in /app/backend/data the
+# DB is authoritative, so this block documents the intent and makes a fresh
+# data dir come up wired. SearXNG's limiter exempts this source (the bridge
+# egress 10.89.x is in pass_ip), so the 4-req/IP/hour JSON cap never applies.
+Environment=ENABLE_WEB_SEARCH=True
+Environment=WEB_SEARCH_ENGINE=searxng
+Environment=SEARXNG_QUERY_URL=http://192.168.0.143:8081/search
+Environment=SEARXNG_LANGUAGE=all
+Environment=WEB_SEARCH_RESULT_COUNT=6
 Volume=/home/dank/Containers/open-webui/data:/app/backend/data

 # App liveness (/health) AND network liveness (a non-loopback interface must
@@ -1,21 +0,0 @@
-#!/usr/bin/env python3
-"""frank-medium container healthcheck: the port's /health must answer 200.
-
-Runs INSIDE the container. The vLLM /health endpoint only answers once the
-engine is up (weights loaded, cudagraphs captured), so a healthy container is
-a serving container, not merely a live process. PORT selects the endpoint so
-the single-GPU replicas (:7771/:7772) and the round-robin proxy (:7770) can
-share this probe. The HealthCmd in the quadlet splits on spaces into an argv
-array (podman 5.x v1 generator), which is why this lives in a file rather
-than an inline python one-liner.
-"""
-import os
-import sys
-import urllib.request
-
-port = os.environ.get("PORT", "7770")
-try:
-    with urllib.request.urlopen(f"http://127.0.0.1:{port}/health", timeout=5) as r:
-        sys.exit(0 if r.status == 200 else 1)
-except Exception:
-    sys.exit(1)
@@ -1,147 +0,0 @@
-# Quadlet unit: frank-medium (rootless, boot-persistent).
-#
-# Install path on dankserver: ~/.config/containers/systemd/frank-medium.container
-# Unit name: frank-medium.service (container name: frank-medium)
-#
-# The dankserver LLM endpoint (:7770, alias frank-medium), since 2026-09-15
-# running the vllm-radiance MXFP4 stack (stilldeadcode/vllm-radiance:0.9.3
-# image + the ggz14/radiance-vllm-mxfp4 patches) serving
-# just1moremodel/Qwen3.8-27B-Uncensored-MXFP4-awq on the 2x Radeon AI PRO
-# R9700 (gfx1201) cards. It replaced the charlie12345/ROCmFPX llama.cpp
-# deployment, which served the same alias from the Q8_0_ROCMFPX_AGENT GGUF;
-# the MXFP4 stack is tensor-parallel (TP=2, R4D one-shot all-reduce) with the
-# DFlash2-FP8 block-diffusion drafter (SPEC=7) — the measured production
-# configuration of the upstream repo, on our exact hardware signature
-# (2x7551-32624, see the KV pin below).
-#
-# SHAPE. The launcher the upstream repo ships binds the server with
-# --network=host (vLLM takes --host 0.0.0.0 --port 7770 directly), which is
-# what this unit does too — no PublishPort. The fleet contract is unchanged:
-# :7770, served model name frank-medium (SERVED_NAMES), OpenAI /v1.
-#
-# NOT INSTALLED ON THIS HOST. This is the target shape: one tensor-parallel
-# process across both R9700s. It cannot launch a collective here — RCCL's
-# kernels need PCIe atomics and the second card sits behind chipset bridges
-# advertising `AtomicOpsCap: Routing-` — so the active deployment is the
-# two-replica trio in this directory (frank-medium-a/b/proxy) and
-# deploy-remote.sh deliberately does not install this unit. Keep it for a
-# host whose second card is on an atomic-capable root port; the full evidence
-# is in serverAppConfigs-docs/LLM/frank-medium-radiance-mxfp4.md.
-#
-# DEVICES. /dev/kfd + /dev/dri through the amdgpu/KFD stack; HIP/ROCR device
-# indices 0,1 (the two R9700s). The 2 GiB Raphael iGPU enumerates as index 2
-# and is excluded both by the index pin and by the upstream MIN_GPU_MIB=8192
-# logic this unit replaces. --privileged --ipc=host --security-opt
-# seccomp=unconfined --cap-add SYS_PTRACE is the upstream launcher's exact
-# podman posture.
-#
-# MOUNTS. /models (rw checkpoints), the HF cache (tokenizer files),
-# /cache (Triton/inductor/aiter/vLLM compile cache — first start compiles for
-# several minutes, later starts reuse), /patches (the pinned radiance clone —
-# its patches are applied at EVERY container start by the entrypoint), /r4d
-# (the patched libr4d build; the image's shipped r4d.so NaNs the
-# gated-delta-net and must never be used), and the entrypoint + healthcheck
-# scripts from this checkout (mounted read-only straight out of
-# /home/dank/serverAppConfigs — deploy.sh git-pulls the checkout before
-# dispatching).
-#
-# IMAGE. The unit runs the localhost/-aliased tag: this host's podman quadlet
-# path cannot resolve registry-prefixed refs from local storage at create
-# time (same as frank-large on dankclaw). deploy-remote.sh pulls the
-# docker.io ref and re-tags it before the restart. Pull=never keeps boots off
-# the registry.
-#
-# KEEP IN SYNC: the radiance clone is pinned (97353293...); if it is updated,
-# re-diff dankserver/frank-medium/radiance-entrypoint.sh against the new
-# serve-mxfp4.sh container command and bump the pin deliberately.
-#
-# DO NOT name an environment variable TEMP, HOST or GPU_MAX_HEAP_SIZE: the
-# HIP runtime reads bare TEMP for its JIT cache path (2026-08-13 incident).
-[Unit]
-Description=frank-medium — vllm-radiance MXFP4 Qwen3.8-27B-Uncensored on 2x R9700 (gfx1201)
-After=network-online.target
-
-[Container]
-ContainerName=frank-medium
-Image=localhost/radiance-vllm:0.9.3
-# vLLM binds 0.0.0.0:7770 on the host network directly; no PublishPort.
-AddDevice=/dev/kfd
-AddDevice=/dev/dri
-GroupAdd=keep-groups
-# The upstream launcher's exact podman posture (rootless; podman-only flags
-# like --replace are not needed under systemd management).
-PodmanArgs=--privileged --ipc=host --security-opt seccomp=unconfined --cap-add SYS_PTRACE --network=host
-# The two R9700s. The Raphael iGPU (index 2, 2 GiB) stays out: shared system
-# memory is not a compute target.
-Environment=ROCR_VISIBLE_DEVICES=0,1
-Environment=HIP_VISIBLE_DEVICES=0,1
-# The uncensored MXFP4 checkpoint ships its fp8 MTP head already, so no
-# fp8_mtp.py rewrite step exists or is needed (unlike amd/Qwen3.8-27B-Quark-AWQ-MXFP4).
-Environment=SNAP=/models/Qwen3.8-27B-Uncensored-MXFP4-awq
-Environment=DRAFTER=/models/Qwen3.8-27B-DFlash2-FP8
-Environment=CHAT_TEMPLATE=/patches/qwen-fixed-v22.3.jinja
-# Fleet identity: the id at /v1/models and in every response body. Every
-# client (Hermes, the harness, services.json) points at frank-medium.
-Environment=SERVED_NAMES=frank-medium
-Environment=PORT=7770
-# Fixed 2-card tensor parallel: the tested configuration of the upstream
-# repo; the R4D one-shot all-reduce kernel serves it natively.
-Environment=TP=2
-# Measured production batch shape (upstream README): MAXSEQS 8 peaks
-# aggregate throughput at 573 t/s; MAXLEN is the model's native context.
-Environment=MAXSEQS=8
-Environment=CHUNK=8192
-Environment=MAXLEN=262144
-Environment=GPU_UTIL=0.98
-# Speculative decoding: the DFlash2 drafter at depth 7 is the configuration
-# every published number (224 t/s single-stream decode, 4749-4810 t/s
-# prefill) was measured with.
-Environment=SPEC_METHOD=dflash
-Environment=SPEC=7
-# KV cache pin, bytes. From the upstream repo's shipped kv-profiles.tsv row
-#   sig=2x7551-32624 maxseqs=8 chunk=8192 maxlen=262144 spec=dflash
-# which is EXACTLY this host (2x R9700 DID 0x7551, 32624 MiB each). Pinned
-# explicitly because this unit does not run the launcher's lookup; valid
-# only at this batch shape (change MAXSEQS/CHUNK/SPEC_METHOD => recalibrate
-# with calibrate-kv.sh or set KV_MEM=0 to let vLLM profile). Worth
-# 943,581 KV tokens vs ~892,799 profiled (~5.7%).
-Environment=KV_MEM=18563072000
-# Checkpoints (rw: nothing downloads at runtime, but vLLM mmaps the shards;
-# keep parity with the launcher's rw mount).
-Volume=/home/dank/models:/models
-# Tokenizer files ride the HF cache; HF_HUB_OFFLINE=1 is set in the
-# entrypoint.
-Volume=/home/dank/.cache/huggingface:/root/.cache/huggingface
-# Compile cache. Keyed on model + stack version + every graph-changing knob;
-# never share it across configurations (upstream rule).
-Volume=/home/dank/.radiance-cache-w4a8-093-gdnm-nqft-fp8s-gnq:/cache
-# The pinned radiance clone: patches applied at every container start.
-Volume=/home/dank/GitPackages/radiance-vllm-mxfp4:/patches:z
-# The patched libr4d (pin b9e42ab + repo extras, key b9e42ab-rx6). The
-# image's shipped r4d.so predates the GDN overflow fix and NaNs this model.
-Volume=/home/dank/.cache/radiance-libr4d/b9e42ab-rx6:/r4d:z
-# Entrypoint + healthcheck from this checkout, mounted read-only.
-Volume=/home/dank/serverAppConfigs/dankserver/frank-medium/radiance-entrypoint.sh:/usr/local/bin/frank-medium-entrypoint.sh:ro
-Volume=/home/dank/serverAppConfigs/dankserver/frank-medium/frank-medium-healthcheck.py:/usr/local/bin/frank-medium-healthcheck.py:ro
-# The entrypoint must be the container command; the image's own entrypoint
-# (which the patched script execs at the end) must not run unpatched.
-PodmanArgs=--entrypoint /usr/local/bin/frank-medium-entrypoint.sh
-# 200 only once the engine serves; during first-start compilation the
-# container is legitimately unhealthy. HealthCmd is split on spaces into an
-# argv array by the podman 5.x v1 generator, hence the file-based probe.
-HealthCmd=python3 /usr/local/bin/frank-medium-healthcheck.py
-HealthInterval=60s
-HealthTimeout=30s
-HealthRetries=3
-HealthStartPeriod=1800s
-Pull=never
-
-[Service]
-# First start compiles Triton/inductor kernels (several minutes) before the
-# engine loads ~9.5 GiB of weights per card.
-TimeoutStartSec=1800
-Restart=on-failure
-RestartSec=10
-
-[Install]
-WantedBy=default.target
@@ -1,229 +0,0 @@
-#!/usr/bin/env bash
-# frank-medium container entrypoint: vllm-radiance MXFP4 on 2x Radeon AI PRO
-# R9700 (gfx1201), serving just1moremodel/Qwen3.8-27B-Uncensored-MXFP4-awq.
-#
-# This is a fork of the patch-and-serve block inside radiance-vllm-mxfp4's
-# serve-mxfp4.sh (clone at /home/dank/GitPackages/radiance-vllm-mxfp4, pinned
-# commit 9735329348ae1c9319f01cbe5f7004ec5f4dba63). The published image alone
-# is NOT sufficient for this model: the shipped libr4d predates the
-# gated-delta-net overflow fix and NaNs the output, and the MXFP4 gates need
-# relaxing, so the repo's patches are applied at every container start from
-# the /patches mount. It runs INSIDE the container (executed as the image
-# command by the frank-medium quadlet); the launcher's host-side GPU/TP/KV
-# detection is not needed here because this fleet is fixed 2x gfx1201 and the
-# KV pin is pinned explicitly in the quadlet.
-#
-# KEEP IN SYNC: if the radiance clone is updated, re-diff this file against
-# the new serve-mxfp4.sh container command and against the DRY_RUN output.
-#
-# Env contract: every knob below is ${VAR:-default}; the quadlet sets the
-# fleet-specific ones (SNAP, PORT, SERVED_NAMES, TP, KV_MEM, ...).
-set -euo pipefail
-
-# ---------------------------------------------------------------- serving shape
-: "${SNAP:=/models/Qwen3.8-27B-Uncensored-MXFP4-awq}"
-: "${DRAFTER:=/models/Qwen3.8-27B-DFlash2-FP8}"
-: "${CHAT_TEMPLATE:=/patches/qwen-fixed-v22.3.jinja}"
-: "${PORT:=7770}"
-: "${SERVED_NAMES:=frank-medium}"
-: "${MAXLEN:=262144}"
-: "${MAXSEQS:=8}"
-# KV_MEM=0 is passed THROUGH as --kv-cache-memory 0, which the image's
-# patched KV sizing reads as "profile the shape" (the launcher's rule).
-# Blanking the variable instead omits the flag and makes vLLM size the cache
-# from --gpu-memory-utilization alone: measured on the single-GPU replicas
-# that overshoots and the allocation dies with torch.OutOfMemoryError right
-# after the weights load (2.16 GiB requested, 116 MiB free). Do not "fix"
-# this into an empty expansion.
-: "${KV_MEM:=18563072000}"   # shipped kv-profiles.tsv row 2x7551-32624 @ 8/8192/262144/dflash
-: "${SPEC_METHOD:=dflash}"
-if [ "$SPEC_METHOD" = dflash ]; then : "${SPEC:=7}"; else : "${SPEC:=4}"; fi
-: "${MIN_M:=0}"              # 0 = the W4A8 kernel serves every M (MIN_M=1 still routes M=1 to aiter)
-: "${R4D_ATTN:=1}"
-: "${ASYNC:=0}"
-
-if [ "$R4D_ATTN" = "1" ]; then ATTN=R4D; else ATTN=ROCM_AITER_UNIFIED_ATTN; fi
-if [ "$ASYNC" = 1 ]; then ASYNC_FLAG="--async-scheduling"; UNPAD=false; else
-  ASYNC_FLAG="--no-async-scheduling"; UNPAD=true; fi
-
-# Drafting config. dflash: the DFlash2-FP8 block-diffusion drafter (its
-# attention backend must support full cuda graphs -> TRITON_ATTN). The
-# disable_padded_drafter_batch unpad is the single-stream lever (~+50%).
-if [ "$SPEC_METHOD" = dflash ]; then
-  if [ ! -f "$DRAFTER/config.json" ]; then
-    echo "frank-medium: no dflash drafter at $DRAFTER" >&2
-    exit 1
-  fi
-  DRAFT_SAMPLE=${DRAFT_SAMPLE:-greedy}
-  SPEC_CFG="{\"method\":\"dflash\",\"model\":\"$DRAFTER\",\"num_speculative_tokens\":$SPEC,\"attention_backend\":\"TRITON_ATTN\",\"disable_padded_drafter_batch\":$UNPAD,\"draft_sample_method\":\"$DRAFT_SAMPLE\"}"
-  RADIANCE_DRAFT_RERANK=${RADIANCE_DRAFT_RERANK:-80}
-  RADIANCE_VERIFY_HEAD=${RADIANCE_VERIFY_HEAD:-1}
-else
-  SPEC_CFG="{\"method\":\"mtp\",\"num_speculative_tokens\":$SPEC,\"attention_backend\":\"$ATTN\",\"disable_padded_drafter_batch\":$UNPAD}"
-  RADIANCE_DRAFT_RERANK=${RADIANCE_DRAFT_RERANK:-32}
-  RADIANCE_VERIFY_HEAD=${RADIANCE_VERIFY_HEAD:-0}
-fi
-
-# The AR size gate compares raw bf16 bytes: CHUNK x hidden(5120) x 2. Derived
-# so changing CHUNK cannot silently drop prefill back onto RCCL.
-export RADIANCE_AR_MAX_KB=$(( (CHUNK * 5120 * 2) / 1024 + 4096 ))
-
-# ---------------------------------------------------------------- runtime env
-export HF_HUB_OFFLINE=1
-export VLLM_LOGGING_LEVEL=${VLLM_LOGGING_LEVEL:-INFO}
-export VLLM_NO_USAGE_STATS=${VLLM_NO_USAGE_STATS:-1}
-export VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=1
-export VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER_MLA=0 VLLM_ROCM_USE_AITER_MOE=0
-export VLLM_ROCM_USE_AITER_LINEAR=0 VLLM_ROCM_USE_AITER_FP8BMM=0
-export VLLM_ROCM_USE_AITER_FP4BMM=0 VLLM_ROCM_USE_AITER_RMSNORM=0
-export NCCL_PROTO=Simple
-export RADIANCE_USE_R4D=${RADIANCE_USE_R4D:-1} RADIANCE_USE_R4D_AR=${RADIANCE_USE_R4D_AR:-1}
-export RADIANCE_DFLASH_SELECTOR_TOPK="${RADIANCE_DFLASH_SELECTOR_TOPK:-}"
-export RADIANCE_R4D_REPORT=1
-export RADIANCE_PRESHUFFLE=${RADIANCE_PRESHUFFLE:-1}
-export RADIANCE_MXFP4_SANITIZE="${RADIANCE_MXFP4_SANITIZE:-0}"
-export RADIANCE_MXFP4=1 RADIANCE_MXFP4_W4A8=1 RADIANCE_MXFP4_W4A8_MIN_M="$MIN_M"
-export RADIANCE_FAST_DRAFT=${RADIANCE_FAST_DRAFT:-1}
-export RADIANCE_DRAFT_TAU=${RADIANCE_DRAFT_TAU:-0.20}
-export RADIANCE_DRAFT_RERANK
-export RADIANCE_VERIFY_HEAD
-export RADIANCE_VERIFY_HEAD_MAX_M=${RADIANCE_VERIFY_HEAD_MAX_M:-32}
-export RADIANCE_MXFP4_DEBUG=${RADIANCE_MXFP4_DEBUG:-0}
-export RADIANCE_MXFP4_PUREQUANT=${RADIANCE_MXFP4_PUREQUANT:-0}
-export RADIANCE_MXFP4_SYNC=${RADIANCE_MXFP4_SYNC:-0}
-export RADIANCE_MXFP4_CLONE=${RADIANCE_MXFP4_CLONE:-0}
-export RADIANCE_MXFP4_CHECKX=${RADIANCE_MXFP4_CHECKX:-0}
-export RADIANCE_MXFP4_PADOUT=${RADIANCE_MXFP4_PADOUT:-0}
-export RADIANCE_MXFP4_TN4_MIN_M=${RADIANCE_MXFP4_TN4_MIN_M:-2048}
-# The decode-kernel band must cover MAXSEQS x (SPEC+1) rows: 64 covers 8x8.
-if [ "${MAXSEQS}" -gt 8 ]; then
-  export RADIANCE_MXFP4_DECODE_MAX_M=${RADIANCE_MXFP4_DECODE_MAX_M:-128}
-fi
-export RADIANCE_MXFP4_DECODE_NT=${RADIANCE_MXFP4_DECODE_NT:-1}
-export RADIANCE_MXFP4_A_TILED_MIN_M=${RADIANCE_MXFP4_A_TILED_MIN_M:-513}
-export RADIANCE_MXFP4_WPERM=${RADIANCE_MXFP4_WPERM:-1}
-export RADIANCE_TP_PAD=${RADIANCE_TP_PAD:-0}   # TP=2 here; 3 is the only non-zero value upstream supports
-export RADIANCE_TP_PAD_DRAFTER=${RADIANCE_TP_PAD_DRAFTER:-1}
-export RADIANCE_TP_PAD_STRICT=${RADIANCE_TP_PAD_STRICT:-1}
-export RADIANCE_GDN_MERGE_INPROJ=${RADIANCE_GDN_MERGE_INPROJ:-1}
-export RADIANCE_GDN_NORM_QUANT=${RADIANCE_GDN_NORM_QUANT:-1}
-export RADIANCE_GDN_STRIDED_GATES=${RADIANCE_GDN_STRIDED_GATES:-0}
-export RADIANCE_GDN_EMPTY_OUT=${RADIANCE_GDN_EMPTY_OUT:-0}
-export R4D_ATTN_FP8=${R4D_ATTN_FP8:-3}
-export RADIANCE_DFLASH_CALIB="${RADIANCE_DFLASH_CALIB:-}"
-export RADIANCE_DFLASH_CALIB_TOKENS="${RADIANCE_DFLASH_CALIB_TOKENS:-200000}"
-export RADIANCE_MXFP4_SHADOW="${RADIANCE_MXFP4_SHADOW:-}"
-export RADIANCE_MXFP4_KERNEL_N="${RADIANCE_MXFP4_KERNEL_N:-}"
-export RADIANCE_MXFP4_KERNEL_NK="${RADIANCE_MXFP4_KERNEL_NK:-}"
-export RADIANCE_MXFP4_CHECKALL="${RADIANCE_MXFP4_CHECKALL:-}"
-export RADIANCE_MXFP4_DECODE_KS="${RADIANCE_MXFP4_DECODE_KS:-}"
-export RADIANCE_MXFP4_DECODE_BK="${RADIANCE_MXFP4_DECODE_BK:-}"
-export RADIANCE_MXFP4_PERBLOCK_NK="${RADIANCE_MXFP4_PERBLOCK_NK:-}"
-export RADIANCE_MXFP4_REFLINEAR="${RADIANCE_MXFP4_REFLINEAR:-0}"
-export PYTORCH_CUDA_ALLOC_CONF="${PYTORCH_CUDA_ALLOC_CONF:-}"
-export RADIANCE_MXFP4_EPIFAST="${RADIANCE_MXFP4_EPIFAST:-1}"
-export RADIANCE_AR_OVERLAP=${RADIANCE_AR_OVERLAP:-0}
-export RADIANCE_GDN_FUSED_UPDATE=${RADIANCE_GDN_FUSED_UPDATE:-1}
-export RADIANCE_DYNAMIC_WIDTH=${RADIANCE_DYNAMIC_WIDTH:-1}
-export RADIANCE_DYNW_ALPHA=${RADIANCE_DYNW_ALPHA:-0.35}
-export RADIANCE_DYNW_MARGIN=${RADIANCE_DYNW_MARGIN:-2}
-export RADIANCE_DYNW_MIN=${RADIANCE_DYNW_MIN:-2}
-export RADIANCE_DYNW_MIN_BATCH=${RADIANCE_DYNW_MIN_BATCH:-3}
-export RADIANCE_AR_QNB=${RADIANCE_AR_QNB:-96}
-export RADIANCE_AR_QNT=${RADIANCE_AR_QNT:-1024}
-export MXFP4_CUMODE=${MXFP4_CUMODE:-0}
-export RADIANCE_STEP_TRACE=${RADIANCE_STEP_TRACE:-0}
-export RADIANCE_AR_OVERLAP_MIN_M=${RADIANCE_AR_OVERLAP_MIN_M:-2048}
-export RADIANCE_AR_OVERLAP_SLICES=${RADIANCE_AR_OVERLAP_SLICES:-4}
-export RADIANCE_MXFP4_EPIFAST=${RADIANCE_MXFP4_EPIFAST:-1}
-export RADIANCE_MXFP4_R4D_DECODE_MAX_M=${RADIANCE_MXFP4_R4D_DECODE_MAX_M:-0}
-export RADIANCE_TOPK_TRITON_MIN_ROWS=${RADIANCE_TOPK_TRITON_MIN_ROWS:-1}
-export RADIANCE_SKINNY_GEMM=${RADIANCE_SKINNY_GEMM:-1}
-export RADIANCE_MXFP4_HOIST_QUANT=${RADIANCE_MXFP4_HOIST_QUANT:-1}
-export RADIANCE_MXFP4_TRACED_QUANT=${RADIANCE_MXFP4_TRACED_QUANT:-1}
-export RADIANCE_FP8_STREAM=${RADIANCE_FP8_STREAM:-1}
-export RADIANCE_RMS_QUANT_FUSION=${RADIANCE_RMS_QUANT_FUSION:-1}
-export RADIANCE_NVFP4_MXFP4=${RADIANCE_NVFP4_MXFP4:-0}
-export RADIANCE_NVFP4_EXP=${RADIANCE_NVFP4_EXP:-mse}
-export RADIANCE_NVFP4_FP8_LAYERS=${RADIANCE_NVFP4_FP8_LAYERS:-mxfp4}
-export RADIANCE_NVFP4_BF16_LAYERS=${RADIANCE_NVFP4_BF16_LAYERS:-in_proj_ba}
-export RADIANCE_NVFP4_LMHEAD=${RADIANCE_NVFP4_LMHEAD:-bf16}
-export RADIANCE_GDN_PATHS=${RADIANCE_GDN_PATHS:-both}
-export RADIANCE_GDN_NANTRACE=${RADIANCE_GDN_NANTRACE:-0}
-export RADIANCE_MXFP4_CHECK_MAX_M=${RADIANCE_MXFP4_CHECK_MAX_M:-128}
-export RADIANCE_MXFP4_MHIST=${RADIANCE_MXFP4_MHIST:-0}
-export VLLM_CACHE_ROOT=/cache/vllm
-export TORCHINDUCTOR_CACHE_DIR=/cache/inductor
-export TRITON_CACHE_DIR=/cache/triton
-export AITER_ROOT_DIR=/cache/aiter
-export TRITON_CACHE_AUTOTUNING=1
-# The patched libr4d built by setup-mxfp4.sh, bind-mounted at /r4d.
-R4D_SO=${R4D_SO:-/r4d}
-
-# The norm+quant fusion turns on vLLM's passes (one --compilation-config
-# instance; argparse keeps the last, so entries accumulate into one flag).
-CC_ITEMS=""
-if [ "${RADIANCE_NORMQUANT_FUSION:-1}" != "0" ]; then
-  CC_ITEMS="\"pass_config\":{\"fuse_norm_quant\":true,\"fuse_act_quant\":true}"
-fi
-EXTRA=${EXTRA:-}
-if [ -n "$CC_ITEMS" ]; then
-  EXTRA="$EXTRA --compilation-config {$CC_ITEMS}"
-fi
-
-# ---------------------------------------------------------------- patch block
-# Byte-for-byte the launcher's container-start block. The patches come from
-# the /patches mount (the pinned radiance clone). SP must match the image's
-# venv layout; the image tag is pinned in the quadlet.
-SP=/opt/vllm/lib/python3.12/site-packages
-cd /patches
-python3 patch_quark_mxfp4.py
-python3 patch_nvfp4_mxfp4.py
-python3 patch_tp3_pad.py
-python3 patch_ar_maxbytes.py
-python3 patch_topk_triton_rows.py
-python3 patch_dflash_calib.py
-python3 patch_dflash_mxfp4_kv.py
-python3 patch_rmsquant_fusion.py
-python3 patch_verify_head.py
-python3 patch_kv_group_size.py
-python3 patch_topk_composite.py
-python3 patch_gdn_shared_build.py
-python3 patch_dflash_selector_topk.py
-python3 patch_gdn_merge_inproj.py
-python3 patch_dynwidth.py
-python3 patch_async_dynwidth.py
-python3 patch_step_trace.py
-python3 patch_ar_geometry.py
-python3 patch_ar_3rank.py
-python3 patch_gdn_glue.py
-# Non-fatal: fixes content=null on thinking-off requests; not required to serve.
-python3 patch_qwen3_thinkoff.py \
-  || echo "[radiance] WARNING: thinkoff patch did not apply; thinking-off requests will return empty content"
-cp mxfp4-configs/*.json "$SP"/aiter/ops/triton/configs/gemm/
-cp radiance_preamble.py /opt/radiance_preamble.py
-cp radiance_nvfp4.py radiance_mxfp4.py radiance_gdn.py radiance_rmsquant.py radiance_drafthead.py \
-   radiance_verifyhead.py radiance_gdnmerge.py radiance_aroverlap.py radiance_topk.py \
-   radiance_arnq.py radiance_tp3pad.py "$SP"/
-hipcc -O3 -w -std=c++17 -fPIC -shared --offload-arch=gfx1201 $([ "${MXFP4_CUMODE:-0}" = 1 ] && echo -mcumode) \
-  $(python3 -m pybind11 --includes) radiance_mxfp4_fp8.hip -o "$SP"/radiance_mxfp4_fp8.so
-if [ -n "${R4D_SO:-}" ] && [ -f /r4d/r4d.so ]; then
-  cp /r4d/r4d.so "$SP"/r4d.so
-  echo "[radiance] using patched r4d.so from $R4D_SO"
-fi
-# Leave /patches before exec: a stale radiance_mxfp4_fp8.so there would shadow
-# the one just compiled into site-packages (cwd precedes it on sys.path).
-cd /
-exec /opt/radiance_entrypoint.sh "$SNAP" --served-model-name $SERVED_NAMES \
-  --host 0.0.0.0 --port "$PORT" \
-  --kv-cache-dtype fp8 --tensor-parallel-size "$TP" \
-  --gpu-memory-utilization "$GPU_UTIL" \
-  ${KV_MEM:+--kv-cache-memory "$KV_MEM"} \
-  --max-model-len "$MAXLEN" --max-num-seqs "$MAXSEQS" --max-num-batched-tokens "$CHUNK" \
-  --attention-backend "$ATTN" \
-  --speculative-config "$SPEC_CFG" \
-  $ASYNC_FLAG $EXTRA \
-  --enable-prefix-caching --mamba-cache-mode align --enable-auto-tool-choice \
-  --tool-call-parser qwen3_coder --reasoning-parser qwen3 \
-  --override-generation-config '{"temperature":0.7,"top_p":0.95,"top_k":20}' \
-  --chat-template "$CHAT_TEMPLATE"
@@ -0,0 +1,63 @@
+# Quadlet unit: Calibre Content Server (rootless, boot-persistent).
+#
+# Install path on dankserver:
+# ~/.config/containers/systemd/calibre-server.container
+# Unit name: calibre-server.service (podman 5.x drops the container- prefix)
+#
+# The image is the CURRENT linuxserver/calibre (9.15.0), which is a desktop
+# image (Selkies web-streamed GUI on 8080/8181) with the headless
+# calibre-server available inside. The compose-era snippet in
+# serverAppConfigs-docs/Dankserver/Calibre.md predates the desktop revamp;
+# the dedicated headless images (ta264/docker-calibre, agentgreen) are
+# dormant (calibre 6.10 / 2017), so this unit runs calibre-server as the
+# image's s6-overlay main service via Exec=; the desktop also runs but its
+# ports are internal-only (see the port map below).
+#
+# Port map: Network=pasta is port-FORWARD mode on this host (the container
+# keeps its own netns; only PublishPort'd ports reach the host — verified
+# 2026-09-21: 8080/8090/8182 inside the container are not visible on the
+# host, only the 8084 forward). So the ONLY exposed surface is:
+#   8090 -> host 8084  calibre-server (the content server; 8080 inside the
+#                      container is the image's own nginx, 8083 on the host
+#                      is calibre-web)
+# The desktop GUI (nginx 8080/8181 + WS 8082, inside the container) runs but
+# is internal-only: no PublishPort. Exposing it is one PublishPort line away
+# if the browser GUI is ever wanted (the basic-auth credential is already in
+# the env file).
+# Auth: calibre-server 9.x keeps its users in a SQLite userdb (the pw column
+# is PLAINTEXT, so the file is mode 600). The single admin account is
+# bootstrapped once by the deploy handler (deploy_calibre_server) via
+# `calibre-server --manage-users -- add`, password piped on stdin with
+# printf (a trailing newline from echo lands in the stored password and
+# breaks every login). The desktop's basic auth (CUSTOM_USER/CUSTOM_PASSWORD)
+# reuses the same credential pair. Both live in the machine-local env file;
+# this unit holds no secrets.
+#
+# --auth-mode=basic is required: on a non-SSL port the server defaults to
+# Digest auth, which rejects plain Basic clients with 400 "Unsupported
+# authentication method".
+
+[Unit]
+Description=Calibre content server container
+After=network.target
+
+[Container]
+ContainerName=calibre-server
+Image=lscr.io/linuxserver/calibre:9.15.0
+Network=pasta
+EnvironmentFile=/home/dank/Containers/calibre-server/env
+Environment=PUID=1000
+Environment=PGID=1000
+Environment=TZ=America/Vancouver
+Volume=/home/dank/Containers/calibre-web/books:/books
+Volume=/home/dank/Containers/calibre-server/userdb.sqlite:/userdb.sqlite
+Exec=calibre-server --port 8090 --enable-auth --auth-mode=basic --userdb /userdb.sqlite /books
+PublishPort=8084:8090
+
+[Service]
+TimeoutStartSec=300
+Restart=always
+RestartSec=5
+
+[Install]
+WantedBy=default.target
@@ -16,7 +16,7 @@ After=network.target

 [Container]
 ContainerName=calibre-web
-Image=lscr.io/linuxserver/calibre-web:0.6.27-ls397
+Image=lscr.io/linuxserver/calibre-web:0.6.27-ls402
 Network=pasta
 Environment=DOCKER_MODS=linuxserver/mods:universal-calibre
 Environment=LANG=en_US.UTF-8
@@ -127,7 +127,7 @@ services:
  # Calibre-Web — e-book web frontend
  # ═══════════════════════════════════════════════
  calibre-web:
-    image: lscr.io/linuxserver/calibre-web:0.6.27-ls397
+    image: lscr.io/linuxserver/calibre-web:0.6.27-ls402
    container_name: calibre-web
    restart: unless-stopped
    environment:
@@ -4,7 +4,7 @@ After=network-online.target
 Wants=network-online.target

 [Container]
-Image=lscr.io/linuxserver/plex:1.43.3.10896-cb3ebc72d-ls322
+Image=lscr.io/linuxserver/plex:1.43.4.10903-e5521bd8c-ls325
 ContainerName=plex
 Network=dankserver_appnet:alias=plex
 Environment=PUID=0
@@ -4,7 +4,7 @@ After=network-online.target
 Wants=network-online.target

 [Container]
-Image=lscr.io/linuxserver/tautulli:v2.18.1-ls242
+Image=lscr.io/linuxserver/tautulli:v2.18.1-ls245
 ContainerName=tautulli
 Network=dankserver_appnet:alias=tautulli
 Environment=PUID=0
@@ -0,0 +1,76 @@
+#!/usr/bin/env bash
+# Download the z-lab Qwen3.5-9B-DFlash drafter for frank-large's DFlash
+# speculative-decode arm (vLLM --speculative-config method=dflash). Public
+# repo — no auth header on the command line.
+#
+# The drafter is a 6-layer qwen3 draft whose dflash_config pins
+# target_layer_ids [1,5,9,13,17,21,25,29] and block_size 16; its hidden_size
+# 4096 / vocab_size 248320 match the MiMo-V2.6-Distill-Qwen-9B target exactly
+# (both preserve the Qwen3.5-9B geometry), which is what lets vLLM load it as
+# the draft for that target.
+#
+# Usage:
+#   scripts/download-frank-large-dflash.sh [DEST_DIR] [--verify-only]
+set -euo pipefail
+
+REPO="z-lab/Qwen3.5-9B-DFlash"
+BASE="https://huggingface.co/${REPO}/resolve/main"
+
+DEST=""
+VERIFY_ONLY=0
+for arg in "$@"; do
+    case "$arg" in
+        --verify-only) VERIFY_ONLY=1 ;;
+        *) DEST="$arg" ;;
+    esac
+done
+DEST="${DEST:-/home/dank/models/Qwen3.5-9B-DFlash}"
+
+# name bytes sha256(- = non-LFS, size-only check)
+FILES=(
+  "model.safetensors 2583816465 0a42274b32554f48de1faa0d42824e9c2ceda649c30ae0a731cddf410dd698c7"
+  "config.json 1222 -"
+)
+
+mkdir -p "$DEST"
+status=0
+for entry in "${FILES[@]}"; do
+  read -r name size sha <<<"$entry"
+  f="$DEST/$name"
+  have=0
+  if [[ -f "$f" ]] && [[ "$(stat -c%s "$f")" == "$size" ]]; then
+    if [[ "$sha" == "-" ]] || sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+      have=1
+    fi
+  fi
+  if (( have )); then
+    echo "OK        $name"
+    continue
+  fi
+  if (( VERIFY_ONLY )); then
+    echo "MISSING   $name"
+    status=1
+    continue
+  fi
+  echo "FETCHING  $name"
+  tmp="$f.part"
+  curl -fsSL --fail-with-body -o "$tmp" "$BASE/$name"
+  mv "$tmp" "$f"
+  if [[ "$(stat -c%s "$f")" != "$size" ]]; then
+    echo "SIZE FAIL $name ($(stat -c%s "$f") != $size)"
+    status=1
+    continue
+  fi
+  if [[ "$sha" != "-" ]] && ! sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+    echo "SHA FAIL  $name"
+    status=1
+    continue
+  fi
+  echo "VERIFIED  $name"
+done
+if (( status == 0 )); then
+  echo "DOWNLOAD_DONE $DEST"
+else
+  echo "DOWNLOAD_FAILED $DEST" >&2
+fi
+exit "$status"
@@ -0,0 +1,90 @@
+#!/usr/bin/env bash
+# Download the BF16 MiMo-V2.6-Distill-Qwen-9B payload for frank-large
+# (vLLM/kyuz0 gfx1151, Strix Halo iGPU). Public repo — no auth header on the
+# command line (a Bearer token in argv is world-readable via /proc).
+#
+# Per-file FRESH downloads, no resume: the house rule (see
+# download-frank-small.sh); -f/--fail-with-body plus size+sha256 against the
+# Hub LFS oid is what makes an access error fail instead of "succeeding".
+#
+# Usage:
+#   scripts/download-frank-large-mimo.sh [DEST_DIR] [--verify-only]
+#
+# The verified set below is exactly what the 2026-10-01 bench window served:
+# 17.53 GiB of BF16 weights (4 shards, 18,819,720,848 B total with the rest).
+set -euo pipefail
+
+REPO="XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B"
+BASE="https://huggingface.co/${REPO}/resolve/main"
+
+DEST=""
+VERIFY_ONLY=0
+for arg in "$@"; do
+    case "$arg" in
+        --verify-only) VERIFY_ONLY=1 ;;
+        *) DEST="$arg" ;;
+    esac
+done
+DEST="${DEST:-/home/dank/models/MiMo-V2.6-Distill-Qwen-9B-BF16}"
+
+# name bytes sha256(- = non-LFS, size-only check)
+FILES=(
+  "model-00001-of-00004.safetensors 5276436216 aab052180118aee34abc3029b54eaa49096aac606b97d703866b420dceb703c3"
+  "model-00002-of-00004.safetensors 5033171280 7a0486565f06d25ac4628e9dba470dc3f604353471d240d5a0bf7128f64df396"
+  "model-00003-of-00004.safetensors 5234499504 6c73207563d1879bfd6c143a028cc70be68edff56280458c71a43df4240300f4"
+  "model-00004-of-00004.safetensors 3275613848 1379a7cf8c0b8555a39ab65a47e830e0eb45e776b46045734da3afd73c09eea2"
+  "model.safetensors.index.json 69217 -"
+  "config.json 2784 -"
+  "generation_config.json 126 -"
+  "chat_template.jinja 3916 -"
+  "tokenizer.json 19989325 06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523"
+  "tokenizer_config.json 1165 -"
+  "vocab.json 6722759 -"
+  "merges.txt 3353259 -"
+  "preprocessor_config.json 443 -"
+  "processor_config.json 1191 -"
+  "video_preprocessor_config.json 385 -"
+)
+
+mkdir -p "$DEST"
+status=0
+for entry in "${FILES[@]}"; do
+  read -r name size sha <<<"$entry"
+  f="$DEST/$name"
+  have=0
+  if [[ -f "$f" ]] && [[ "$(stat -c%s "$f")" == "$size" ]]; then
+    if [[ "$sha" == "-" ]] || sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+      have=1
+    fi
+  fi
+  if (( have )); then
+    echo "OK        $name"
+    continue
+  fi
+  if (( VERIFY_ONLY )); then
+    echo "MISSING   $name"
+    status=1
+    continue
+  fi
+  echo "FETCHING  $name"
+  tmp="$f.part"
+  curl -fsSL --fail-with-body -o "$tmp" "$BASE/$name"
+  mv "$tmp" "$f"
+  if [[ "$(stat -c%s "$f")" != "$size" ]]; then
+    echo "SIZE FAIL $name ($(stat -c%s "$f") != $size)"
+    status=1
+    continue
+  fi
+  if [[ "$sha" != "-" ]] && ! sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+    echo "SHA FAIL  $name"
+    status=1
+    continue
+  fi
+  echo "VERIFIED  $name"
+done
+if (( status == 0 )); then
+  echo "DOWNLOAD_DONE $DEST"
+else
+  echo "DOWNLOAD_FAILED $DEST" >&2
+fi
+exit "$status"
@@ -0,0 +1,85 @@
+#!/usr/bin/env bash
+# Download the source Qwen3.5-9B checkpoint — the ONLY place this model
+# family's MTP head (15 mtp.* tensors) exists. frank-large runs it as the
+# vLLM cross-model MTP draft (MiMo-V2.6-Distill-Qwen-9B itself ships zero
+# mtp tensors; that is why SGLang's NEXTN on frank-small draws 0.00
+# acceptance). Public repo — no auth header on the command line.
+#
+# NOTE the odd upstream shard names (model.safetensors-0000N-of-00004...).
+#
+# Usage:
+#   scripts/download-frank-large-qwen35.sh [DEST_DIR] [--verify-only]
+set -euo pipefail
+
+REPO="Qwen/Qwen3.5-9B"
+BASE="https://huggingface.co/${REPO}/resolve/main"
+
+DEST=""
+VERIFY_ONLY=0
+for arg in "$@"; do
+    case "$arg" in
+        --verify-only) VERIFY_ONLY=1 ;;
+        *) DEST="$arg" ;;
+    esac
+done
+DEST="${DEST:-/home/dank/models/Qwen3.5-9B}"
+
+# name bytes sha256(- = non-LFS, size-only check)
+FILES=(
+  "model.safetensors-00001-of-00004.safetensors 5276436216 db6f444b43d318c92f360a13a25561a6a65b10c0631b8ed305a426dbaa6c380e"
+  "model.safetensors-00002-of-00004.safetensors 5335161512 31c7d7e2dd5d207840b31cc59083c8f4c4718959149e0358c0364052bb9a0330"
+  "model.safetensors-00003-of-00004.safetensors 5368717440 7ec36ba3a4176a44c3c0876ad80c56a2f70c84bf008d82e9501df642f17dadec"
+  "model.safetensors-00004-of-00004.safetensors 3325995712 b62b0c4cd7e44edee103ee8f4fe225f246d5e768e07bfd5f25b63a8aa1fdd0c6"
+  "model.safetensors.index.json 79657 -"
+  "config.json 3126 -"
+  "chat_template.jinja 7756 -"
+  "tokenizer.json 12807982 5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42"
+  "tokenizer_config.json 16710 -"
+  "vocab.json 6722759 -"
+  "merges.txt 3353259 -"
+  "preprocessor_config.json 390 -"
+  "video_preprocessor_config.json 385 -"
+)
+
+mkdir -p "$DEST"
+status=0
+for entry in "${FILES[@]}"; do
+  read -r name size sha <<<"$entry"
+  f="$DEST/$name"
+  have=0
+  if [[ -f "$f" ]] && [[ "$(stat -c%s "$f")" == "$size" ]]; then
+    if [[ "$sha" == "-" ]] || sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+      have=1
+    fi
+  fi
+  if (( have )); then
+    echo "OK        $name"
+    continue
+  fi
+  if (( VERIFY_ONLY )); then
+    echo "MISSING   $name"
+    status=1
+    continue
+  fi
+  echo "FETCHING  $name"
+  tmp="$f.part"
+  curl -fsSL --fail-with-body -o "$tmp" "$BASE/$name"
+  mv "$tmp" "$f"
+  if [[ "$(stat -c%s "$f")" != "$size" ]]; then
+    echo "SIZE FAIL $name ($(stat -c%s "$f") != $size)"
+    status=1
+    continue
+  fi
+  if [[ "$sha" != "-" ]] && ! sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+    echo "SHA FAIL  $name"
+    status=1
+    continue
+  fi
+  echo "VERIFIED  $name"
+done
+if (( status == 0 )); then
+  echo "DOWNLOAD_DONE $DEST"
+else
+  echo "DOWNLOAD_FAILED $DEST" >&2
+fi
+exit "$status"
@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+# download-frank-primary-fp8.sh — fetch/verify the Qwen3.8-Flash-Next-FP8
+# payload (131 safetensors shards, 172.78 GiB total, plus aux files) for the
+# vLLM frank-primary deployment on dankserver.
+#
+#   bash scripts/download-frank-primary-fp8.sh [dest_dir]
+#
+# The manifest (scripts/manifests/qwen3.8-flash-next-fp8.manifest) carries
+# per-file sizes + LFS sha256s from the Hub; hf-fetch.sh is idempotent — an
+# interrupted re-run fetches only what is missing or corrupt, and a complete
+# on-disk payload is re-verified (full sha256 pass) with zero downloads.
+# The repo is public; no HF_TOKEN required.
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+dest="${1:-/home/dank/models/Qwen3.8-Flash-Next-FP8}"
+
+exec bash "$REPO_ROOT/scripts/hf-fetch.sh" "$dest" \
+    "$REPO_ROOT/scripts/manifests/qwen3.8-flash-next-fp8.manifest" \
+    Qwen/Qwen3.8-Flash-Next-FP8
@@ -0,0 +1,93 @@
+#!/usr/bin/env bash
+# Download MiMo-V2.6-Distill-Qwen-9B-INT8-W8A16-AutoRound for frank-small
+# (SGLang, RTX 3090). Public repo — no auth header on the command line.
+#
+# Per-file FRESH downloads, no resume: `curl -C -` once produced a 213 MB
+# oversized file on a 31 GiB payload; a fresh -f fetch plus size+sha256 is
+# the house rule.
+#
+# Usage:
+#   scripts/download-frank-small.sh [DEST_DIR] [--verify-only]
+#
+# LFS shards are verified against the Hub LFS sha256; non-LFS text files
+# against their declared byte size. Idempotent: complete files are skipped.
+set -euo pipefail
+
+REPO="letechlead/MiMo-V2.6-Distill-Qwen-9B-INT8-W8A16-AutoRound"
+BASE="https://huggingface.co/${REPO}/resolve/main"
+
+DEST=""
+VERIFY_ONLY=0
+for arg in "$@"; do
+    case "$arg" in
+        --verify-only) VERIFY_ONLY=1 ;;
+        *) DEST="$arg" ;;
+    esac
+done
+DEST="${DEST:-/home/dank/models/MiMo-V2.6-Distill-Qwen-9B-INT8-W8A16-AutoRound}"
+
+# name bytes sha256(- = non-LFS, size-only check)
+FILES=(
+  "model-00001-of-00008.safetensors 2946296312 ddb6117c04cb6feaf7c265e5d75efadf21949037d73253a8e4521d40e164366d"
+  "model-00002-of-00008.safetensors 1402134792 2dc78d55d1d33d7b3b21dd6f8b4b817b62c2fb1376e6448b322428798c10e257"
+  "model-00003-of-00008.safetensors 1427365904 5c797d14e345309d67c78306b9de4aa01a21dbd92dd98e40b0ad69b2900e9203"
+  "model-00004-of-00008.safetensors 1420041856 7402f852e2c7a24ebf4694b98a220e71dfb2c93ba3776272cf2acfc65acd8126"
+  "model-00005-of-00008.safetensors 1410655648 991cbf2eda2ff91fc3ffbb15591d7fcac340fb10d9c33bd3da20010e2f3f4090"
+  "model-00006-of-00008.safetensors 1427627856 efc42448054cfb5e3061e154ba4539204088ce95176d689a95b58de885ef7eeb"
+  "model-00007-of-00008.safetensors 2035442408 75a101fc71ae6f2e1f5c4816f711d62cf1d7a0e02ef6e5182361b8c11e72b626"
+  "model-00008-of-00008.safetensors 822153664 0fb90cb50e2d336fbad58bd4c01880eeddbf537c1dca1cf0ae291195d5efbb4b"
+  "tokenizer.json 19989325 06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523"
+  "model.safetensors.index.json 106278 -"
+  "config.json 4608 -"
+  "generation_config.json 161 -"
+  "quantization_config.json 1654 -"
+  "chat_template.jinja 3916 -"
+  "tokenizer_config.json 1168 -"
+  "preprocessor_config.json 443 -"
+  "processor_config.json 1191 -"
+  "README.md 8448 -"
+)
+
+mkdir -p "$DEST"
+status=0
+for entry in "${FILES[@]}"; do
+  read -r name size sha <<<"$entry"
+  f="$DEST/$name"
+  have=0
+  if [[ -f "$f" ]] && [[ "$(stat -c%s "$f")" == "$size" ]]; then
+    if [[ "$sha" == "-" ]] || sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+      have=1
+    fi
+  fi
+  if (( have )); then
+    echo "OK        $name"
+    continue
+  fi
+  if (( VERIFY_ONLY )); then
+    echo "MISSING   $name"
+    status=1
+    continue
+  fi
+  echo "FETCHING  $name"
+  tmp="$f.part"
+  # no -C resume; -f + --fail-with-body so an HTTP error is not a success
+  curl -fsSL --fail-with-body -o "$tmp" "$BASE/$name"
+  mv "$tmp" "$f"
+  if [[ "$(stat -c%s "$f")" != "$size" ]]; then
+    echo "SIZE FAIL $name ($(stat -c%s "$f") != $size)"
+    status=1
+    continue
+  fi
+  if [[ "$sha" != "-" ]] && ! sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+    echo "SHA FAIL  $name"
+    status=1
+    continue
+  fi
+  echo "VERIFIED  $name"
+done
+if (( status == 0 )); then
+  echo "DOWNLOAD_DONE $DEST"
+else
+  echo "DOWNLOAD_FAILED $DEST" >&2
+fi
+exit "$status"
@@ -0,0 +1,144 @@
+.gitattributes 1635 
+LICENSE 3235 
+README.md 65400 
+chat_template.jinja 8952 
+config.json 72514 
+generation_config.json 202 
+merges.txt 3353259 
+model-00001-of-00131.safetensors 1040155912 774f0ceeadb40d165f2b3ff397d5f3840e6ca8fcb8f3d39d8acb4fea9e52c941
+model-00002-of-00131.safetensors 1678209208 6841fe21fa8a8a7a693c585efe65cd2732889095b696da88bda0cb287366910b
+model-00003-of-00131.safetensors 993901136 974a2a2ab551f8f1405a4955ab32a8721c68c73dd85b382491d9f0e6a34ee752
+model-00004-of-00131.safetensors 1678209208 591f488ab5cd5f0bd4fe28266099523761b9b9339137800734c34ffd84595538
+model-00005-of-00131.safetensors 1717267690 c50bf465a4a0129f1a1196c0e75d881048238f7107b356fc734007eac0d3b123
+model-00006-of-00131.safetensors 1600008328 57f103367fe36c9dc355c648d521d026caabb4ce5ba51038c78bd935f6593736
+model-00007-of-00131.safetensors 1600008328 cde2a12854770c706a74a5c7b85ffc140e3d4dba93dd4bee1cd7aa07cc348c90
+model-00008-of-00131.safetensors 1600008328 98ae323b5a61abdcf58a0dc9db7eb1a5a748cebc1ef1932804a1411500c09d84
+model-00009-of-00131.safetensors 1600008328 70f4b66bca5c4c0b441cf4284d4ab311670e182bfc5ec51c7294a4054baeb20f
+model-00010-of-00131.safetensors 1600008328 6d5dfa1dfa9a6c482f100f6e75c1f5e8ac02391c62e844e53f4a53996baf2206
+model-00011-of-00131.safetensors 1600008328 de085adb563e530e7ad157220cd1a219be189a1db0fcc3f2507886a2d78bd4e1
+model-00012-of-00131.safetensors 1600008328 97dac7c366ba19858d78c01364054136449ccb3cf60768587461cbf0bb582c90
+model-00013-of-00131.safetensors 1600008328 467be10b473d25e0581f8fb1bc986916c8b9301272f72e7a05762484ffc080a9
+model-00014-of-00131.safetensors 1600008328 38def6e9bbe6f242c2f92858c3e9a2b356d74d0180e323677158016dbee32ae7
+model-00015-of-00131.safetensors 1600008328 90c6c665f1f819bc6350714652c8fb77064fec57832ef51b878a0cdf9f602236
+model-00016-of-00131.safetensors 1600008328 88185c2530641b37dcad80d37b505c5257035e87ad29fa86640aae4814903e4f
+model-00017-of-00131.safetensors 1600008328 42e3b96359116d7471a74cb7de08f6a4e9e173830d562f1f74ff09d1d253ad38
+model-00018-of-00131.safetensors 1600008328 040adc7270efca6498c6f6a07dd7da14925d20217fe6de67223d6856fb4f2ad7
+model-00019-of-00131.safetensors 1600008328 b8a831992deb1a694439e37b7136ff2e181457a8683bff65f1f4b99772cabe58
+model-00020-of-00131.safetensors 1600008328 9426112fbd49ef4f5e9dd38dc2f93b046602e6d92cff732d33b96413ff1b40dc
+model-00021-of-00131.safetensors 1600008328 fce8426665f749ef38c4fad33b612f892b728c0f74fe60cab0e43ece785b4e77
+model-00022-of-00131.safetensors 1600008328 a20dd415f280774321538e6b9e78fba9f8e218a9090eae52ea7172b85b9c5a6d
+model-00023-of-00131.safetensors 1600008328 4100591839471a8721d9cedb251554dcef5ebc1dfd96ea7dce8912dd1e9921b2
+model-00024-of-00131.safetensors 1600008328 bbd6d7a4b54b6a6032e15b7c2fc9d6526377f89992d61da8837ad059f3210b45
+model-00025-of-00131.safetensors 1600008328 f9fa7b04de155dd2be74b5693d5b36e5463dabf956b9be36f67d4aae6b8dd304
+model-00026-of-00131.safetensors 1600008328 193c6b9da2c6b69c34d8a14b65234871db272761e712c2e56bac3f13c2d0cbf7
+model-00027-of-00131.safetensors 1600008328 d7efa372552c07ee8783e27c4b8ebd12fc75333f92553861bec8c7c0155a03fe
+model-00028-of-00131.safetensors 1600008328 5fc13d084d62c06b928dfb388a1b36f5c27cd25bd3cbb68a0b21e1c4cfc68e6e
+model-00029-of-00131.safetensors 1600008328 a11f5a6360b349159575871db5c553c40c37be68382b82551a2c2372f15f2811
+model-00030-of-00131.safetensors 1600008336 cac780a675f87fc699435a5207cbecb30a887271b5514b60b7e45d9bebe0e1ec
+model-00031-of-00131.safetensors 1600008336 d390a872eb3928b9d7c0a720ced792cc111613812a163d9fd6c97fd35d633897
+model-00032-of-00131.safetensors 1600008336 3b36fe4e7215a12cdc6c7499aafde51cfbce51266785184b99d3838a6b37b2c8
+model-00033-of-00131.safetensors 1600008336 3fb0504f406c0c71ee40e9e158492973c5c35337889c81b974892b407620c64c
+model-00034-of-00131.safetensors 1600008336 6d989a812a8d8d9330c8c4d21d7be98a0114705ca3bd7834d43bb268bad6be08
+model-00035-of-00131.safetensors 1600008336 8e56ba0a714198bd07a2d8fe9d91fbf5c72601c1b98edf09b289a2538445cb75
+model-00036-of-00131.safetensors 1600008336 c23fd6f6cfa9acfa961ba26e7c9c0dd7b05782c5bbf8f55e2526119b5c516fd4
+model-00037-of-00131.safetensors 942343240 b824a11bf460bd78c068e3d301e914a0bec2f243a6eea4679be7200d179248e7
+model-00038-of-00131.safetensors 1678211256 716eda81a763ffdefd1ac5306b4c80eef123c4dee24ed3a565a22438497581cb
+model-00039-of-00131.safetensors 877983696 b91e3bfbcd765ce704db85b78f69055da59456ae733df635e3973ff2f966a148
+model-00040-of-00131.safetensors 1678211256 3311ed7821b68bddb1542f06ca349feec72e68288113fe4899b08a3b8156343f
+model-00041-of-00131.safetensors 1103350296 1d79949b61730a0fa1f6cd1af149fa50fedc051566cdea48283de8c7a32d972c
+model-00042-of-00131.safetensors 1678211256 3f61684338e203cb2ccff6e451eda5705e10043d386b7ed55d4732d23eb5c77b
+model-00043-of-00131.safetensors 1000455912 bab92cfefbb811ee63c4e32595024117c7df3537f5359606eb1de23fe4f115b2
+model-00044-of-00131.safetensors 1678211256 1d05f138848edca4138c3d1c6be4a432a0817eb29fa65f12d8611ecc7685c493
+model-00045-of-00131.safetensors 993902176 4610384e28f7a91b8bf2a1527178c82cbe7d50ecd9a33b61ee878fa8e8bced85
+model-00046-of-00131.safetensors 1678211256 5ee9b0e45810300cbd4fd6f2acfe85080e7032a9ce5d77038b86f357b6700c23
+model-00047-of-00131.safetensors 878065752 efa6c0cc0e527486dc5955b00fa75665151d99997b722d8e56bd32da575e414c
+model-00048-of-00131.safetensors 1678211256 57b148a266c55a896f171b5c98b350092bf394c1fb836658c58c589a0a317ccf
+model-00049-of-00131.safetensors 1096817160 801dd3b475b43413ade810f4e8d3ace69d6ef5ee4bac5cb98d15e369343a5569
+model-00050-of-00131.safetensors 1678211256 b4d00aca34ba7a9f8427983fd0829408fb5e15b2bb7268199e4010bf0c460a04
+model-00051-of-00131.safetensors 993902176 f430125e9ac92c2de15bbb74bcd3c91bc5374eef389a1be1bc4a592f3a082f0c
+model-00052-of-00131.safetensors 1678211256 06c3d64136b097919b7e5a1715a70a60b8c2302fc4e80a33235607bb99d00f35
+model-00053-of-00131.safetensors 1000455920 147d38f786512f1a8f570a00100b46a3faab1bd649b3445ca0a50c00aa2e99a3
+model-00054-of-00131.safetensors 1678211256 31c08218039a246c333bfe0b068b8027fd92dddd542e581997af546eb104dd4d
+model-00055-of-00131.safetensors 877983696 971cd69c8d955bef6d557ebab3f170807a599517f7c69e9d8a6ba375214137f4
+model-00056-of-00131.safetensors 1678211256 a6a975ac346aa3bd1176cb816d8bbd31a4d7542aa3de9306ca5899bea44bd159
+model-00057-of-00131.safetensors 1103350288 d6acae1622abfa1e0253ae548bc1bdd4bc38cd9b096082acd455b6be592f4a54
+model-00058-of-00131.safetensors 1678209208 f02b6bd8f5cbd742e8bac3ba9537aa14a64c5a8f3fc75495227a14e1a5cec4ec
+model-00059-of-00131.safetensors 877982872 f3e2c94b8157a021b27e1efdb659333a8d1db330da5cd72ab9012fa0e5351506
+model-00060-of-00131.safetensors 1794100080 908cce7f5fe8dc34f4cca1fc8a676569ab6ce08bad536bd0091c4d75e7b65a05
+model-00061-of-00131.safetensors 1678211256 718532c88f4b7a4f402a50e4f1b6efc3cf71484c9e5a083ee30e96d5094a1bb3
+model-00062-of-00131.safetensors 1832994688 3c21b895a5a4f0dc03b669c730f8d8d9ce9a29cbc7628a25894a16ffda6f0a7a
+model-00063-of-00131.safetensors 1678211256 91d215e2135b2f5db5bc5a9541e5f6bf761205d67c6e7b8955ac73c815198f58
+model-00064-of-00131.safetensors 993902176 03a11cad2c283fb9752c6ac01c8f1d8b3127dcda963c5a8225b78f567239d872
+model-00065-of-00131.safetensors 1678211256 0f2f96331120779ce1ebd3f957882092d3c733aa9a6073ca87fc86c6eb4e9465
+model-00066-of-00131.safetensors 880605240 5dbe501e4e0cb9e3f37cc43da46673b4157fdffba09c065ca0b12d75092a79f2
+model-00067-of-00131.safetensors 1678211256 3ece325d2ca53eddf7b1782d5bbc4c376bc75c6d3da67813e110fd0af67cd09b
+model-00068-of-00131.safetensors 1096796552 0c5ef706afd6b3058af5bf8fdc1e0938cb6e29d0a526c771893c9dcb8e702463
+model-00069-of-00131.safetensors 1678211256 84999b24e3c55d2eae3ca187219a2e1a1872436ae9bafc91cc3e301180f36cfd
+model-00070-of-00131.safetensors 993907416 eac7940317ee401859caf38089b2b924274ab81a3e5fbd87c42d89923361aad8
+model-00071-of-00131.safetensors 1678211256 5ac60e474ccdab21257245630ed8240d75bc7115b3adb27ca65e39eb9ea8f20e
+model-00072-of-00131.safetensors 1000456024 b50a05c4c426fb00635d513a733a290da1007a89b8785f1ec02aad8e4d93d955
+model-00073-of-00131.safetensors 1678211256 9e0594d389debdfb26d22259725831ff34280edeef4144e683d70205050919d9
+model-00074-of-00131.safetensors 877983696 1b1a0294f8df2a431885783185cdb185106228c9567e67224fe2c3c6a2467f5a
+model-00075-of-00131.safetensors 1678211256 7dba189804ef7a8f9d926b27e1ddbe73767f1213cb7bb3cc582ab15fffec1792
+model-00076-of-00131.safetensors 1100073480 d95ebacb144b4220f2df60cce85a2c86e87cb8291b836d1bd04f9902fcbb9297
+model-00077-of-00131.safetensors 1678211256 26046977983e025defdf0f84cf5fc5f4ddfeabbd27e7abc43ff2b413265ab062
+model-00078-of-00131.safetensors 993984232 6fd46a95777392265b2afa66bf71d488856dafb7b76ff0337a0fe4ac857e961c
+model-00079-of-00131.safetensors 1678211256 77398002634ce852917c59cd642bec7dd3903381239da68912bf69743631e54e
+model-00080-of-00131.safetensors 877983688 f44aad95d36c443995b965aeef844c8e3a30ed172647dc57fcc0c5476c18d024
+model-00081-of-00131.safetensors 1678209208 b295a48088573654d1b4754e861ec329956504cb6687e9e2a8eb652b35eaf988
+model-00082-of-00131.safetensors 1096816112 83e0e171116c22eb8fdddfa3c387bebb9017eb0586410fd704ce70e7b3d23d68
+model-00083-of-00131.safetensors 1678211256 7e1021cde06c970e4b1d9eb3d02f66f29786c59a0da8a3f8879de1a4e3fbf6c0
+model-00084-of-00131.safetensors 877983696 8938fd3c7f88d61fd5b5306eb0ab253d3e42e7f64baf3c3e494d315c8ad5ab55
+model-00085-of-00131.safetensors 1678211256 cd5875a29d28d52578ae699fdd8891f605a0953cd64516bd88dc45dab50c383a
+model-00086-of-00131.safetensors 1103350296 885e570a662aaf266fd9af9562c65116998df5710b2a27d1bf6aa0c8708b3b74
+model-00087-of-00131.safetensors 1678211256 1472604826f0a0c2bbf7cd0f4a0c5b29f3012e4b3a4424d024d1ec17a3c9d0ea
+model-00088-of-00131.safetensors 1000455920 e7273f5435571e761429330ece81732e8fb3f8b5f0087bf80db693630cf702a1
+model-00089-of-00131.safetensors 1678211256 eb349562d23e330bbbd0a5f5ba90ac25ed0baa660dff230538a0fcaaf4c8b875
+model-00090-of-00131.safetensors 997179104 532915eebd4e231a240e185ba838a95a00ee14277e28fb04ad21508da5ff50aa
+model-00091-of-00131.safetensors 1678211256 68d8f46ffcd0ff46f4e73c848525cc6fd6af15eebbb9311ce9b4eb14638f5660
+model-00092-of-00131.safetensors 877983696 91abc087cd244ae62b8469576e209c8e25bc7726bc4fc020b687378059f5b5bd
+model-00093-of-00131.safetensors 1678211256 dacd0f2e5215fbf18fb7c6803f61963333e6243f404de83fcd5dfd38f43fe0d1
+model-00094-of-00131.safetensors 1096796928 cac515a778e89fb37e9f38fde6c8fdfa01ebf9d8944a2d6ff35c51ecf4bc875f
+model-00095-of-00131.safetensors 1678211256 a35975c2d21f18d58f21cb59e4eaf34a4bce11440e5112948e7fa6923658f8c4
+model-00096-of-00131.safetensors 993902176 03fd77f9d35024bba9e894c811300ef8aaf5b5ff99ebc2c89c9d9f87685adfe8
+model-00097-of-00131.safetensors 1678211256 f7b81ae09d97b8e2bb96a1916ede51da92209ec8481a31c56fb446acd691b8a8
+model-00098-of-00131.safetensors 993902176 70f208f47d99c25d5413cd32e0553cce22524cd3863c7797e2cfa37c2d526319
+model-00099-of-00131.safetensors 1678211256 de5c33dd717495657496067ea715442356fd9f7b64384d1a4c10c9f8fe6f9b7d
+model-00100-of-00131.safetensors 877984072 43752e81a5c2b376073d5d0d7b70da8d08c49bdbbfcfeb695ea947a2913e4bc8
+model-00101-of-00131.safetensors 1678211256 c2e978f7238036063914fe909214246730640be77eca215e3cf66d65724de9f1
+model-00102-of-00131.safetensors 1096797160 639efed3ea400a8dddeae25db120096898a514697dfc4cbbc0cd15f621c18421
+model-00103-of-00131.safetensors 1678209208 f8da1b61296e012dc268b4f0d84c127ef0967df488fe65b037f74c8f893decd1
+model-00104-of-00131.safetensors 993901144 adff0f2aee7b7b701e39247cfdfa135c0cab8dfa4f8c7662352c0233adb78829
+model-00105-of-00131.safetensors 1678211256 6643d71382f00f7ffe16b05879e6ff0f6e56986dee8ec079a84aae82912c6987
+model-00106-of-00131.safetensors 1025359600 680495ea0a23e73844e91670437b7b2a5ac5d46f9582923fb7fa91fe640821df
+model-00107-of-00131.safetensors 1678211256 886528d121b77433c1d66654294c13b4f61a004f1e6d1abf2d445fb0a99f7234
+model-00108-of-00131.safetensors 1062060032 a966c28a6267d43052fe87cb19e2e5730bad6154b12eb806b9ce49e007bfec96
+model-00109-of-00131.safetensors 1678211256 f6074252e492ee5b0933f78cc48f3b90694a8096befb7666cc379c3d47093756
+model-00110-of-00131.safetensors 877983696 e3e45e96bc13b52b748a0e2f397a624ad49958396468667d73ed4d725c8b5c1c
+model-00111-of-00131.safetensors 1678211256 d1d354e691eab0b77607faf7eb4adf00a75ee7f270c1cbcbad06aefc19aeac94
+model-00112-of-00131.safetensors 1096797168 e61a9a6ef0e538afca09d97a3ccbf3e7e2c647f65cc8eeebebb18621d7779b02
+model-00113-of-00131.safetensors 1678211256 a14b94db5326bfaab1d8029296b8938f6daf91ba2dc07a4950a38afafa84f776
+model-00114-of-00131.safetensors 993907400 ce4aeac52776038ec0c885915bb270dc65ee07d70d5f71827d5fba1045e2b43b
+model-00115-of-00131.safetensors 1678211256 8185d8822428d2125876a9aa05aea11c8165713bc99e04ace29921eef0e1b683
+model-00116-of-00131.safetensors 993902176 65954dccfa3be20611728311d97679e7f5d78ad519e52312524aa7cf666a48e5
+model-00117-of-00131.safetensors 1678211256 12c8027363b1a7bb782282293314fa57b07be27d17670861858c0aed1e28a3ac
+model-00118-of-00131.safetensors 878004272 2d06ec9c1726f42bfc9ce0bbb47129917d8ab373c88eed4e758fb6940c92ad4a
+model-00119-of-00131.safetensors 1678211256 36008b48c4480085bfd1a81439d70d1029cfaf06cfdd037cec19b491a40659ec
+model-00120-of-00131.safetensors 1109903856 49e4f90d92f60f6489bfe6d3e5250d8fe879c5995ae72ce67379cc7187fa4b0a
+model-00121-of-00131.safetensors 1678209208 ae8e07cceeef2f568df035a7bc0c555656acb566ba5936f61a8a019df23361ad
+model-00122-of-00131.safetensors 993901136 b3b21d33b925e31d475a509757cf7459aa4d3e73cb493706f40461387541e137
+model-00123-of-00131.safetensors 1678209208 f32fec30685afadd76f09bbb1dda680445ad765a97ff1febfc3122b7f73d7ec3
+model-00124-of-00131.safetensors 891089968 e857f36ddd0db135f858c2c87c9fcf39c758965c626b280e47cb0c1e6c1423f8
+model-00125-of-00131.safetensors 1678209208 b9de3774218d3cd7946cdbf913dedfc82591c18212d9c3fc9b9484e7dc68e9fe
+model-00126-of-00131.safetensors 1096795496 29b39319522a00b1ef4ac909c4fb306059999f0a48fd07ca73cea687647631ee
+model-00127-of-00131.safetensors 1678209208 bf48ccb3b4ba75c288d9db2b94817c6a3993e9f84a5b9410d9f24edcec12e177
+model-00128-of-00131.safetensors 993901136 7feaa51c906a1a735e5fb52fd5499b59c473d442d77b44cbd610c9b6a3a9e0ba
+model-00129-of-00131.safetensors 1678209208 a5374c8d86aea48f21a4be2ad6cda37fbfbb01eff333d1179939a2e55d8db816
+model-00130-of-00131.safetensors 2136177320 204b122d02425f61115db9ab896648d95f6dadd6a716c65daf6aec297392ed56
+model-00131-of-00131.safetensors 1271398496 ba6fc69ebb6a9da1938c2e041b8ef625f4550b641812da75845bc7dd7c361d80
+model.safetensors.index.json 17410140 0419e2c2dfbb925257d7409405433a793cf7ff7d96f3eba882a815ec6d9fe7a6
+preprocessor_config.json 390 
+tokenizer.json 12809320 0997f410c57a1f4e53b09e4be8f4a172d90edd9564368fb0847030937229b9f3
+tokenizer_config.json 17928 
+video_preprocessor_config.json 385 
+vocab.json 6722759 
@@ -1,7 +1,7 @@
-# Tunables for frank-medium (ajgazin/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP
-# UD-Q8_K_XL + z-lab/Qwen3.8-27B-DFlash2-GGUF Q8_0 drafter), mainline llama.cpp
-# on Dankserver, 2x AMD Radeon AI PRO R9700 (gfx1201).  Since 2026-09-16 this
-# is the SOLE LLM runtime on the box: the vllm-radiance replica pair and its
+# Tunables for frank-medium (orcarouter/Qwen3.8-27B-Uncensored-GGUF Q8_0 +
+# z-lab/Qwen3.8-27B-DFlash2-GGUF Q4_K_M drafter), mainline llama.cpp on
+# Dankserver, 2x AMD Radeon AI PRO R9700 (gfx1201).  Since 2026-09-16 this is
+# the SOLE LLM runtime on the box: the vllm-radiance replica pair and its
 # routing proxy are retired.
 #
 # Hardware: two R9700 (Navi 48, RDNA4, 32,624 MiB each), no XGMI, different
@@ -31,17 +31,17 @@
 # The ref is pinned by TWO features: DFlash2 (ggml-org/llama.cpp#27342 -- the
 # drafter is detected from its GGUF) and qwen35 MTP.  Do not float it.
 #
-# Model: ajgazin/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF @ UD-Q8_K_XL,
-# 29.29 GiB, arch qwen35, Unsloth Dynamic 3.0 (per-tensor types from Unsloth's
-# own quant of the same size, on Unsloth's importance matrix).  Abliteration is
-# orcarouter's single refusal direction projected out of 131 tensors; the vision
-# tower loads from mmproj-BF16.gguf (see NO_MMPROJ).  The file also
-# carries the model's MTP head (blk.64.nextn.*), so a file-level switch to
-# --spec-type draft-mtp is available without a second copy in VRAM.
+# Model: orcarouter/Qwen3.8-27B-Uncensored-GGUF @ Q8_0, 27.0 GiB, arch qwen35
+# (verified from the GGUF header 2026-09-19: 262144 trained window, 24 heads /
+# 4 KV heads, nextn_predict_layers=1).  The file ALSO carries the model's MTP
+# head (blk.64.nextn.*), so --spec-type draft-mtp is available without a second
+# copy in VRAM -- one SPEC_TYPE change, the drafter below stays on disk.
 #
-# Drafter: z-lab/Qwen3.8-27B-DFlash2-GGUF @ Q8_0, 1.92 GiB: 5 layers, hidden
+# Drafter: z-lab/Qwen3.8-27B-DFlash2-GGUF @ Q4_K_M, 1.06 GiB (operator call
+# 2026-09-19: the Q4 rung -- the drafter's job is throughput, not output
+# quality, and the target verifies every drafted token).  5 layers, hidden
 # 5120, a block-diffusion proposal (block_size 8) under a 2048-token sliding
-# window.  Decoding stays lossless -- the target verifies every drafted token.
+# window (dflash.attention.sliding_window=2048, verified from the header).
 # The window is what makes it cheap HERE: its KV is bounded at 2048 positions
 # (~40 MiB) instead of spanning the context (see SIZING).
 #
@@ -52,10 +52,14 @@
 # in the attention layers only, plus ~0.16 GiB per sequence of f32 DeltaNet
 # state at any depth.
 #
-# SIZING, measured (rocm-smi, 2026-09-16).  Weights split across the two cards;
-# the drafter AND its KV ride the second.  The loader's compute buffers are part
-# of what is seen below -- they are allocated at load, not per request, and they
-# scale with --ubatch-size:
+# SIZING, the deployed shape (32:33 / ctx 327680 / 6144x512 / two slots /
+# DFlash2 Q4_K_M), measured 2026-09-19 after the model swap (rocm-smi):
+#
+#   card 0 (x16)  26.9 GiB used, 5.7 GiB free
+#   card 1 (x1)   27.8 GiB used, 4.8 GiB free
+#
+# The predecessor's measured table (rocm-smi, 2026-09-16, UD-Q8_K_XL target)
+# stays as the calibration history:
 #
 #   CTX     split  batch  slots  drafter  GPU 0 free  GPU 1 free   note (ubatch 1536)
 #   262144   1,1   2048   1      DFlash2   8.64 GiB    4.43 GiB   day's start
@@ -71,7 +75,7 @@
 # batch block below):
 #
 #   262144  36,29  6144   512    2      DFlash2   6.01 GiB    7.32 GiB   ubatch only
-#   262144  35,30  6144   512    2      DFlash2   6.44 GiB    6.90 GiB   <- deployed
+#   262144  35,30  6144   512    2      DFlash2   6.44 GiB    6.90 GiB
 #
 # ONE -ts STEP IS ONE LAYER, AND LAYERS ARE NOT THE SAME SIZE.  This is a hybrid
 # stack -- 16 x (3 Gated DeltaNet + 1 Gated Attention) blocks -- so moving a
@@ -80,36 +84,27 @@
 # were measured, which is why the balance is found by measurement rather than by
 # interpolating between splits.
 #
-# 262144 is the window the model ships (n_ctx_train) and it fits with room, so
-# the fleet gets the whole thing; the arithmetic that used to cap this endpoint
-# at 114688 was the vLLM predecessor's, not this runtime's.  If a future change
-# adds VRAM (a longer drafter, quantized KV, more slots) re-read the numbers
-# rather than assuming, and prefer CTX_SIZE over --gpu-layers.
+# WHY THE SPLIT IS 32:33 (operator call, 2026-09-19).  The operator asked for
+# the tensors BALANCED across the two cards, which for 65 layers is 32:33 --
+# the most even split that exists.  The predecessor's 35:30 skew existed to lean
+# work onto the x16 card (ROCm0), because ROCm1's Gen4 x1 chipset path made it
+# the 98 % utilization bottleneck; the measured trade of the balanced point is
+# that the x1 card runs hotter while the x16 card idles a little more, and the
+# operator accepted that for symmetry.  Do not re-skew without asking.
 #
-# WHY THE SPLIT IS 35:30.  Three effects decide it, and all three are measured:
-#   * THE DRAFTER AND ITS OUTPUT LAYERS RIDE ROCm1.  It carries the drafter's
-#     weights, the draft graph and the last target layers, so it takes fewer
-#     than half the layer count.
-#   * THE UBATCH BUFFERS COST ROCm1 MORE THAN ROCm0, and shrinking them pays
-#     ROCm1 back more: 1536 -> 512 freed 1.30 GiB on ROCm0 but 2.41 GiB on
-#     ROCm1 at this shape.  Read the other way, that is exactly why the smaller
-#     batch moved the balance toward ROCm1 -- and cost ROCm0 one layer.
-#   * THE LAYER GRANULARITY IS THE FLOOR ON HOW WELL ANY SPLIT CAN BALANCE.  One
-#     -ts step is one layer -- ~0.43 GiB DeltaNet, ~1.43 GiB attention -- so the
-#     reachable minimum gap is comparable to one step rather than to zero.  35:30
-#     sits at 0.46 GiB, i.e. one DeltaNet layer's worth, against 1.31 GiB at
-#     36:29.
-# It also leaves the most room: 6.44 / 6.90 GiB free clears the operator's 1.5
-# GiB-per-card floor by ~5 GiB, so no headroom has to be bought out of CTX_SIZE
-# and the full window stays.
+# THE TWO CARDS' PATHS ARE NOT SYMMETRIC (measured, 2026-09-17): ROCm0
+# (03:00.0) hangs off CPU root port 00:01.1 at Gen4 x16 end to end (~31.5 GB/s);
+# ROCm1 (0b:00.0) rides the B650 chipset through bridges that each negotiate
+# Gen4 x1 (~1.97 GB/s).  The layer-split boundary itself is not the cost
+# (~5.2 MB of activation per 512-token ubatch rides the wire); the x1 card is
+# COMPUTE-bound, not PCIe-bound.
 #
 # A FAULT TO WATCH.  A 39:26 split, at 1x batch / 1 slot, died on its eighth
 # request: the amdgpu driver logged SDMA0 UTCL2 page faults and the engine
 # surfaced `ROCm error: an illegal memory access was encountered` -> ggml_abort,
 # i.e. a fault in the cross-card boundary copy, this layer split's hot path.  It
-# has not reproduced (36:29 served this file's verification battery and the e2e
-# gate, and 35:30 is the deployed shape), but the signature is worth recognising
-# -- it is NOT an OOM, and no amount of headroom prevents it.
+# has not reproduced since, but the signature is worth recognising -- it is NOT
+# an OOM, and no amount of headroom prevents it.
 #
 # WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE (HIP reads
 # those bare vars; 2026-08-13 incident).  llama.cpp takes --host/--port/--temp
@@ -119,15 +114,15 @@
 LLAMA_BIN=/home/dank/GitPackages/llama.cpp-dflash2/build-rdna4-rocm714/bin/llama-server

 # Payload, fetched with scripts/download-frank-medium-gguf.sh.
-MODEL_DIR=/home/dank/models/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF
-MODEL_FILE=Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-UD-Q8_K_XL.gguf
-# The BF16 vision projector, same repo as the target.  llama-server autoloads
-# any mmproj*.gguf beside the model; the serve script refuses to start without
-# it when vision is enabled, so a missing file is a deploy error, not a
-# text-only server that pretends to be multimodal.
-MM_PROJ_FILE=mmproj-BF16.gguf
+MODEL_DIR=/home/dank/models/Qwen3.8-27B-Uncensored-GGUF
+MODEL_FILE=Qwen3.8-27B-Uncensored-Q8_0.gguf
+# The f16 vision projector, same repo as the target.  llama-server loads it via
+# the explicit --mmproj flag (no autoload in this build); the serve script
+# refuses to start without it when vision is enabled, so a missing file is a
+# deploy error, not a text-only server that pretends to be multimodal.
+MM_PROJ_FILE=mmproj-Qwen3.8-27B-Uncensored-f16.gguf
 DRAFT_MODEL_DIR=/home/dank/models/Qwen3.8-27B-DFlash2-GGUF
-DRAFT_MODEL_FILE=Qwen3.8-27B-DFlash2-Q8_0.gguf
+DRAFT_MODEL_FILE=Qwen3.8-27B-DFlash2-Q4_K_M.gguf

 # Fleet identity.  7770 is the fleet's dankserver endpoint; the alias is the id
 # at /v1/models and what every client asks for.
@@ -141,17 +136,11 @@ HIP_VISIBLE_DEVICES=0,1
 DEVICES=ROCm0,ROCm1
 SPLIT_MODE=layer

-# 36:29 -- one layer over the 35:30 balance, moved to buy headroom on ROCm1,
-# the card that carries the drafter, its KV, the output layers and the heavier
-# compute buffers, now that the context grew to 327680 (see below).  Fit at
-# the deployed point (36:29, ctx 327680, batch 6144 x ubatch 512, two slots,
-# DFlash2 + ngram-mod, mmproj on ROCm0): ROCm0 29.19 / ROCm1 26.04 GiB of
-# 31.86, i.e. 2.67 / 5.82 GiB free and a 3.15 GiB gap BY CONSTRUCTION -- the
-# balance point at this context is 34:31 (0.11 GiB gap) and the operator chose
-# headroom on the drafter card over balance.  See SIZING above for the
-# measurement table; one -ts step is one LAYER, worth ~0.43 GiB (DeltaNet) or
-# ~1.43 GiB (attention), so do not move it without re-measuring.
-TENSOR_SPLIT=36,29
+# 32:33 -- balanced (operator call 2026-09-19): 65 layers split as evenly as
+# they can be, ROCm1 (the last-layers card, which also carries the drafter and
+# the output layers) holding the extra one.  See the WHY THE SPLIT IS 32:33
+# block above for the trade against the predecessor's 35:30 x16-lean.
+TENSOR_SPLIT=32,33

 NGL=999
 # Auto-fit has no business resizing a context we set explicitly.
@@ -160,10 +149,9 @@ FLASH_ATTN=on

 # 327680 positions over TWO slots, so each conversation gets 163840 --
 # --parallel DIVIDES --ctx-size, so more slots would shrink every
-# conversation rather than grow the cache.  The checkpoint's trained window is
-# 262144 and the GGUF carries NO rope scaling (freq_base 1e7, no YaRN keys),
-# so positions beyond 262144 extrapolate on the base rope: it runs, and 1.25x
-# is a modest ask of it, but the extra 65k is untrained territory.
+# conversation rather than grow the cache.  163840 is inside the checkpoint's
+# trained 262144 window, so the whole pool is trained territory (no rope
+# extrapolation, unlike the predecessor's 393216).
 CTX_SIZE=327680
 PARALLEL=2

@@ -211,11 +199,32 @@ THREAD_BATCH=8
 # f16 K/V: an agentic checkpoint is the wrong place to spend KV precision.
 CACHE_TYPE_K=f16
 CACHE_TYPE_V=f16
-# Context checkpoints: 0 is the build's default -- no per-slot KV snapshots, so
-# a re-sent or rewound conversation re-prefills instead of rewinding.  Stated
+# Context checkpoints: 0 -- no per-slot KV snapshots at all.  The build's
+# default is 32 (this binary's --help), so the 0 below is a deliberate
+# deviation, not an inherited default: a checkpoint is a VRAM->host-RAM copy
+# of part of a slot's KV, taken so a partially-matching prompt can be restored
+# instead of re-prefilled.  This endpoint trades that away -- its reuse comes
+# from the slot prompt and from the RAM-side prompt cache below.  Stated
 # explicitly so the tunable stays visible rather than merely absent.
 CTX_CHECKPOINTS=0
-CACHE_RAM=8192
+# PROMPT CACHE, HOST RAM -- this is the one cache tunable that is NOT VRAM.
+# --cache-ram caps the total bytes of saved prompt states
+# (server_prompt_cache, tools/server/server-task.cpp): update() evicts
+# oldest-first past the cap, and alloc() REFUSES outright any single state
+# bigger than the whole cap.  A state is the serialized KV of a whole sequence
+# plus its DeltaNet state plus the draft context's own KV, so at this shape it
+# costs ~68 KiB per token --
+#   16 full-attention layers x 4 kv heads x 256 dim x 2 B x 2 (K,V) = 64 KiB
+#   plus the draft context's window-bounded KV (a ~40 MiB constant for DFlash2)
+# -- which makes the cap a TOKEN budget in disguise: 16384 MiB holds ~240k
+# tokens, i.e. a FULL 163840-token slot (~10.6 GiB) with ~5 GiB of the budget
+# spare for a second entry.
+#
+# The cost is host RAM, allocated on demand and never pinned (plain heap
+# vectors), so no VRAM number above moves.  dankserver has 62 GiB beside the
+# media-stack co-tenants, so the cache is bounded rather than unlimited (-1):
+# the seam test refuses a budget over half the host's RAM.
+CACHE_RAM=16384

 # Sampling defaults for requests that send none -- the values the vLLM
 # predecessor served on this endpoint (`--override-generation-config`), so a
@@ -238,35 +247,33 @@ REASONING_BUDGET=-1
 CHAT_TEMPLATE_FILE=
 CHAT_TEMPLATE_KWARGS='{"enable_thinking":true,"preserve_thinking":true}'

-# Speculative decoding: the DFlash2 block-diffusion drafter PLUS the draftless
-# n-gram matcher.  DFlash2's GGUF carries block_size 8, so at most 7 draft
-# tokens per verification step (llama.cpp asserts n_max <= block_size - 1).
-# The drafter rides ROCm1, the card that ends up holding the target's last
-# layers.
-SPEC_TYPE=draft-dflash,ngram-mod
+# Speculative decoding: the DFlash2 sidecar (operator call 2026-09-19).  The
+# target file also carries its own MTP head (nextn_predict_layers=1), so
+# draft-mtp is one SPEC_TYPE change away -- the drafter file stays on disk.
+# n_max 7 is the measured DFlash2 winner (see the table below); the drafter
+# rides ROCm1, the card that holds the last layers.
+SPEC_TYPE=draft-dflash
 SPEC_DRAFT_N_MAX=7
 SPEC_DRAFT_NGL=999
 SPEC_DRAFT_DEVICE=ROCm1
 SPEC_DRAFT_TYPE_K=f16
 SPEC_DRAFT_TYPE_V=f16

-# MTP (the model's own head) vs the drafter -- MEASURED 2026-09-16 (temperature
-# 0, 1k prompt / 512 out; the autoregressive baseline and both MTP depths were
-# run against the deployed configuration in one session, same binary, same
-# prompts):
+# DFlash2 vs the model's own MTP head -- MEASURED 2026-09-16 on the predecessor
+# payload (temperature 0, 1k prompt / 512 out; the autoregressive baseline and
+# both MTP depths were run in one session, same binary, same prompts):
 #
 #   none            17.7 tok/s                     card1 26.07 GB
 #   draft-mtp  n=3  32.4 tok/s  58 % accepted       card1 29.35 GB
 #   draft-mtp  n=7  26.7 tok/s  33 % accepted       card1 29.65 GB
 #   draft-dflash 7  43.5 tok/s  41 % accepted       card1 29.42 GB
 #
-# All four produced IDENTICAL greedy text, so verification is lossless whichever
-# proposer runs.  DFlash2 is worth 2.46x over autoregressive and 34 % over the
-# best MTP; deeper MTP drafting costs more than it returns (n=7 accepts 33 % of
-# 147 proposals, n=3 accepts 58 % of 72).  The MTP head is not free either:
-# selecting it loads blk.64.nextn.* and adds ~3.3 GB to the card that holds it,
-# against 2.06 GB for the drafter's file.  Do not switch this without
-# re-measuring.
+# All four produced IDENTICAL greedy text, so verification is lossless
+# whichever proposer runs.  DFlash2 is the faster proposer (2.46x over
+# autoregressive, 34 % over the best MTP).  The MTP head adds ~3.3 GB to the
+# card that holds it, against 1.06 GB for the Q4_K_M drafter's file -- at the
+# 327680 pool the drafter's window-bounded KV (~40 MiB) leaves headroom on
+# both cards, so the speed wins.
 #
 # THE TWO CANNOT BE CHAINED. `--spec-type draft-mtp,draft-dflash` dies at load:
 #   context type MTP requested but model doesn't contain MTP layers
@@ -278,16 +285,17 @@ SPEC_DRAFT_TYPE_V=f16
 # tuning question. See serverAppConfigs-docs/LLM/frank-medium-llamacpp-dflash2.md.

 # ngram-mod: the draftless n-gram matcher (CPU hash table in host RAM, no VRAM,
-# no model file, no draft context). It joins the chain unconditionally --
+# no model file, no draft context).  It joins the chain unconditionally --
 # unlike MTP, which needs a draft context and therefore cannot share one with
-# DFlash2. Verified on the loaded server: with ngram-mod as the ONLY spec type,
-# a prose prompt produces ZERO drafts (plain autoregressive) while a
-# prompt-echo prompt accepts 100 % of them, so its drafts are unmistakable.
+# DFlash2.  NOT CURRENTLY ACTIVE: SPEC_TYPE is draft-dflash alone, and the serve
+# script only passes --spec-ngram-mod-* when the type contains "ngram".  The
+# window below is retained because the DFlash2 + ngram chain was measured free
+# on prose and 2.6x on prompt-echo -- one SPEC_TYPE change brings it back.
 #
 # WINDOW MEASURED 2026-09-16 (chain at each window, decode tok/s):
 #
 #   n_min/n_max    prose 1k/512   prompt-echo 1.5k/400
-#   48/64          43.9           200.6     <- deployed
+#   48/64          43.9           200.6
 #   16/32          36.0           161.0
 #    8/16          36.2           113.8
 #
@@ -304,9 +312,7 @@ SPEC_NGRAM_N_MATCH=24
 SPEC_NGRAM_N_MIN=48
 SPEC_NGRAM_N_MAX=64

-# Vision enabled: llama-server autoloads mmproj-BF16.gguf from MODEL_DIR
-# (888 MiB, lands on ROCm0 -- the fit above charges it there).  frank-small
-# (Dankclaw, RTX 3090) remains the fleet's dedicated image endpoint; this
-# makes frank-medium multimodal too, which is why the seam flips from
-# "text only" to "vision loads with the model".
+# Vision enabled: llama-server loads mmproj-Qwen3.8-27B-Uncensored-f16.gguf
+# (931 MiB, lands on ROCm0 -- the fit charges it there) via the explicit
+# --mmproj flag.  This keeps frank-medium multimodal across the model swap.
 NO_MMPROJ=0
@@ -161,8 +161,13 @@ TEMPLATE_ARGS+=(--chat-template-kwargs "$CHAT_TEMPLATE_KWARGS")
 PLACEMENT_ARGS=(--device "$DEVICES" --split-mode "$SPLIT_MODE" --tensor-split "$TENSOR_SPLIT")

 MMPROJ_ARGS=()
-[[ "$NO_MMPROJ" == "1" ]] && MMPROJ_ARGS=(--no-mmproj)
-
+if [[ "$NO_MMPROJ" == "1" ]]; then
+    MMPROJ_ARGS=(--no-mmproj)
+else
+    # No autoload in this build: a server without the flag starts text-only
+    # and answers image requests with a 500, so name the projector.
+    MMPROJ_ARGS=(--mmproj "$MM_PROJ")
+fi
 exec "$LLAMA_BIN" \
    --model "$MODEL" \
    --alias "$ALIAS" \
@@ -0,0 +1,141 @@
+#!/usr/bin/env python3
+"""autoresearch-bench.py — deterministic greedy battery for frank-large.
+
+Runs ON Dankclaw (no LAN jitter) against the candidate container on :7770.
+Fixed embedded corpus, temperature 0, streaming so decode speed is measured
+between real token arrivals (not request totals). Median of REPS samples per
+metric is the reported value; medians are far more stable than means under
+GPU thermal drift.
+
+Metrics emitted (stdout, one per line):
+  METRIC decode_tokens_per_sec=<float>   primary, higher better
+  METRIC prefill_tokens_per_sec=<float>  secondary
+  METRIC ttft_ms=<float>                 secondary, lower better
+Exit non-zero if the endpoint is unhealthy or the served alias is wrong.
+"""
+import argparse
+import json
+import statistics
+import sys
+import time
+import urllib.request
+
+# Fixed corpus: deterministic, ~28k tokens after tokenization (same scale as
+# the 2026-10-01 bench window). Built by repeating a fixed paragraph block a
+# fixed number of times - NO random, NO clock, NO network.
+_BLOCK = (
+    "The rendering pipeline walks the scene graph once per frame, resolving "
+    "material bindings into descriptor sets before any draw call is recorded. "
+    "Because the allocator recycles command buffers only after their fence "
+    "signals, a submission that outlives its frame must pin the pool entry; "
+    "otherwise the next frame overwrites recorded commands in flight. The "
+    "validation layer flags exactly this class of lifetime bug, but only for "
+    "buffers it can see - GPU-side timestamps and queue-family ownership "
+    "transfers hide in the driver's own bookkeeping, which is why release "
+    "builds must lean on explicit barriers and why a frame-time regression "
+    "often shows up as corruption one frame later than its cause.\n"
+)
+LONG_CTX = _BLOCK * 220  # ~28k-token prefill, matching the deployed workload
+_LONG = LONG_CTX + "\nNow, in one focused paragraph (80-150 words), explain " \
+    "which lifetime bugs this pipeline can still suffer and what explicit " \
+    "mitigations the text implies."
+_SHORT = "Design a rate limiter for a homelab reverse proxy that must not " \
+    "lock out the LAN. Think it through, then give the algorithm."
+
+REPS = 3
+MAX_TOKENS = 256
+REQ_TIMEOUT = 900
+
+
+def stream_completion(url: str, model: str, prompt: str):
+    """Greedy streaming completion.
+
+    Returns (ttft_s, decode_tps, prefill_tps, prompt_tokens, n_tokens).
+    ttft is RELATIVE to request send: time.perf_counter()'s epoch is host
+    boot, and returning the raw stamp read as an 11-day first token.
+    """
+    body = json.dumps({
+        "model": model,
+        "messages": [{"role": "user", "content": prompt}],
+        "max_tokens": MAX_TOKENS,
+        "temperature": 0,
+        "stream": True,
+        "stream_options": {"include_usage": True},
+    }).encode()
+    req = urllib.request.Request(
+        url.rstrip("/") + "/v1/chat/completions", data=body,
+        headers={"Content-Type": "application/json"})
+    t0 = time.perf_counter()
+    first = None
+    stamps = []
+    usage = {}
+    with urllib.request.urlopen(req, timeout=REQ_TIMEOUT) as resp:
+        for raw in resp:
+            line = raw.decode(errors="replace").strip()
+            if not line.startswith("data:"):
+                continue
+            payload = line[5:].strip()
+            if payload == "[DONE]":
+                break
+            chunk = json.loads(payload)
+            if chunk.get("usage"):
+                usage = chunk["usage"]
+            for ch in chunk.get("choices", []):
+                d = ch.get("delta") or {}
+                # reasoning AND content deltas both arrive token by token;
+                # count either so greedy text length drives the rate.
+                # (llama.cpp names the thinking field reasoning_content.)
+                if d.get("content") or d.get("reasoning") or d.get("reasoning_content"):
+                    now = time.perf_counter()
+                    if first is None:
+                        first = now
+                    stamps.append(now)
+    if first is None or len(stamps) < 2:
+        # A zero-argmax degenerate answer would poison the median; raise.
+        raise RuntimeError(f"stream produced {len(stamps)} tokens")
+    ttft_s = first - t0
+    decode_s = stamps[-1] - stamps[0]
+    decode_tps = (len(stamps) - 1) / decode_s if decode_s > 0 else 0.0
+    prompt_tokens = usage.get("prompt_tokens") or 0
+    prefill_tps = prompt_tokens / ttft_s if prompt_tokens else 0.0
+    return ttft_s, decode_tps, prefill_tps, prompt_tokens, len(stamps)
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--url", default="http://127.0.0.1:7770")
+    ap.add_argument("--model", default="frank-large")
+    args = ap.parse_args()
+
+    with urllib.request.urlopen(args.url.rstrip("/") + "/v1/models",
+                                timeout=10) as r:
+        ids = [m["id"] for m in json.load(r)["data"]]
+    if args.model not in ids:
+        print(f"FATAL: served ids {ids}, expected {args.model}",
+              file=sys.stderr)
+        sys.exit(2)
+
+    # Warmup (compile/graph warm + allocator settle), NOT measured.
+    stream_completion(args.url, args.model, _SHORT)
+
+    decodes, prefills, ttfts = [], [], []
+    for rep in range(REPS):
+        for name, prompt in (("long", _LONG), ("short", _SHORT)):
+            ttft, dt, pt, ptk, ntok = stream_completion(
+                args.url, args.model, prompt)
+            decodes.append(dt)
+            ttfts.append(ttft * 1000.0)
+            if name == "long":
+                prefills.append(pt)
+            print(f"  rep{rep} {name:5s} ttft={ttft*1000:8.1f}ms "
+                  f"prefill={pt:8.1f} t/s decode={dt:6.2f} t/s "
+                  f"({ntok} tok, prompt {ptk})", file=sys.stderr)
+
+    med = lambda xs: statistics.median(xs)
+    print(f"METRIC decode_tokens_per_sec={med(decodes):.2f}")
+    print(f"METRIC prefill_tokens_per_sec={med(prefills):.1f}")
+    print(f"METRIC ttft_ms={med(ttfts):.1f}")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,192 @@
+#!/usr/bin/env python3
+"""frank-primary decode measurement — runs INSIDE the engine container.
+
+The method is the fleet's established one (the 2026-09-29 / 2026-10-01
+tuning campaigns, /home/dank/mtp_measure.py): SSE chunk counting is invalid
+under speculative decoding because rejected drafts change the accounting, so
+decode is measured from the server's own usage counters:
+
+    decode_tok_s = completion_tokens / (wall - ttft)
+
+with ttft = wall time to the FIRST content-bearing chunk (thinking arrives in
+`reasoning`/`delta.reasoning` on this stack and is excluded from the decode
+window — and is deterministic per prompt at temperature 0, so the exclusion
+cancels in any A/B on the same workload).
+
+Workload (fixed, deterministic, offline):
+  * prompts built from a large fixed-seed shuffled word list so tokens rarely
+    repeat (no n-gram/prefix predictability of the text itself), sized with
+    the server's own /v1/messages/count_tokens counter
+  * depths 115 / 1056 / 4096 target prompt tokens
+  * one 64-token warm pass per depth (warms the expert arena), then
+    max_tokens=256 at temperature 0, 3 runs, median
+Everything is greedy with a fixed seed — the same workload every run, no
+network, no time-of-day dependence.
+
+Env:
+  AR_BASE             endpoint (default http://127.0.0.1:7770)
+  AR_MODEL            served name (default frank-primary)
+  AR_PROMPT_CEILING   hard prompt ceiling in tokens (default 100000)
+
+Output: progress on stderr, ONE json object on stdout (the last line).
+Exit codes: 0 ok, 2 endpoint unhealthy / no usable completion, 3 bad usage.
+"""
+import json
+import os
+import random
+import statistics
+import sys
+import time
+import urllib.error
+import urllib.request
+
+BASE = os.environ.get("AR_BASE", "http://127.0.0.1:7770")
+MODEL = os.environ.get("AR_MODEL", "frank-primary")
+PROMPT_CEILING = int(os.environ.get("AR_PROMPT_CEILING", "100000"))
+DEPTHS = (115, 1056, 4096)
+WARM_TOKENS = 64
+MAX_TOKENS = 256
+RUNS = 3
+
+# 43-word x 12 = 516-word vocabulary, seed 1234 — the campaign's exact list
+# and seed, so prompts are byte-identical to the historical measurements.
+_VOCAB = ("alpha bravo charlie delta echo foxtrot golf hotel india juliet kilo lima "
+          "mike november oscar papa quebec romeo sierra tango uniform victor whiskey "
+          "xray yankee zulu cardinal ordinal topology kernel scheduler resident "
+          "unified aperture expert arena slab offload prefetch gemm triton quant "
+          "tensor shard checkpoint manifest embed router cache page table window "
+          "gradient sparse dense moe attention head dim block stream decode prefill "
+          "bandwidth latency throughput residency eviction tier swap pin memlock").split()
+_R = random.Random(1234)
+_ORDER = list(range(len(_VOCAB)))
+_R.shuffle(_ORDER)
+
+
+def log(msg):
+    print("[measure] %s" % msg, file=sys.stderr, flush=True)
+
+
+def post_json(path, body, timeout):
+    req = urllib.request.Request(
+        BASE + path, data=json.dumps(body).encode(),
+        headers={"Content-Type": "application/json"})
+    with urllib.request.urlopen(req, timeout=timeout) as r:
+        return json.load(r)
+
+
+def count_tokens(text):
+    """Token count without a prefill, from the server's own counter."""
+    body = {"model": MODEL,
+            "messages": [{"role": "user", "content": text}]}
+    r = post_json("/v1/messages/count_tokens", body, timeout=300)
+    n = r.get("input_tokens") or r.get("count") or r.get("token_count")
+    if not n:
+        raise RuntimeError("count_tokens returned no count: %r" % (r,))
+    return int(n)
+
+
+def words(n):
+    return " ".join(_VOCAB[_ORDER[i % len(_VOCAB)]] for i in range(n))
+
+
+def calibrate():
+    """Measure words->tokens from the server's own counter."""
+    return 2000.0 / count_tokens(words(2000))
+
+
+def prompt_for(target_tokens, wpt):
+    n = max(1, int(target_tokens * wpt))
+    text = words(n)
+    pt = count_tokens(text)
+    while pt > PROMPT_CEILING:
+        n = max(1, int(n * (PROMPT_CEILING / pt) * 0.98))
+        text = words(n)
+        pt = count_tokens(text)
+    return text, pt
+
+
+def measured(prompt, max_tokens):
+    """Return (ttft_s, decode_tok_s, completion_tokens) via usage counters."""
+    body = {"model": MODEL,
+            "messages": [{"role": "user", "content": prompt}],
+            "max_tokens": max_tokens, "temperature": 0,
+            "stream": True, "stream_options": {"include_usage": True}}
+    req = urllib.request.Request(
+        BASE + "/v1/chat/completions", data=json.dumps(body).encode(),
+        headers={"Content-Type": "application/json"})
+    t0 = time.time()
+    ttft = None
+    usage = {}
+    with urllib.request.urlopen(req, timeout=1800) as r:
+        for raw in r:
+            line = raw.decode("utf-8", "replace").strip()
+            if not line.startswith("data:"):
+                continue
+            s = line[5:].strip()
+            if s == "[DONE]":
+                break
+            try:
+                ev = json.loads(s)
+            except json.JSONDecodeError:
+                continue
+            if ev.get("usage"):
+                usage = ev["usage"]
+            for ch in ev.get("choices", []):
+                if (ch.get("delta") or {}).get("content"):
+                    if ttft is None:
+                        ttft = time.time() - t0
+    wall = time.time() - t0
+    ct = int(usage.get("completion_tokens") or 0)
+    if ct <= 0:
+        # No usage on the stream: the endpoint is unhealthy or the model
+        # returned nothing (e.g. HTTP 200 with an error payload).
+        raise RuntimeError("no completion_tokens in usage: %r" % (usage,))
+    dec_window = max(wall - (ttft or 0.0), 1e-9)
+    return (ttft or wall), (ct / dec_window), ct
+
+
+def main():
+    log("base=%s model=%s ceiling=%d" % (BASE, MODEL, PROMPT_CEILING))
+    try:
+        wpt = calibrate()
+    except Exception as e:
+        log("endpoint unhealthy during calibration: %s" % e)
+        return 2
+    log("words/token: %.3f" % wpt)
+
+    result = {"words_per_token": round(wpt, 5)}
+    for tgt in DEPTHS:
+        try:
+            p, pt = prompt_for(tgt, wpt)
+            measured(p, WARM_TOKENS)  # warm the arena
+        except Exception as e:
+            log("warm pass failed at depth %d: %s" % (tgt, e))
+            return 2
+        rates, tts, cts = [], [], []
+        for i in range(RUNS):
+            try:
+                tt, r, ct = measured(p, MAX_TOKENS)
+            except Exception as e:
+                log("run %d/%d failed at depth %d: %s" % (i + 1, RUNS, tgt, e))
+                continue
+            rates.append(r)
+            tts.append(tt)
+            cts.append(ct)
+        if len(rates) < 2:
+            log("only %d usable run(s) at depth %d — refusing to median" % (len(rates), tgt))
+            return 2
+        result["decode_tps_%d" % tgt] = round(statistics.median(rates), 2)
+        result["ttft_ms_%d" % tgt] = round(1000.0 * statistics.median(tts), 1)
+        result["completion_tokens_%d" % tgt] = int(statistics.median(cts))
+        result["prompt_tokens_%d" % tgt] = pt
+        log("depth %5d tok  decode %.2f t/s  ttft %.3f s  (runs=%d)"
+            % (pt, result["decode_tps_%d" % tgt], result["ttft_ms_%d" % tgt] / 1000.0, len(rates)))
+
+    meds = [result["decode_tps_%d" % t] for t in DEPTHS]
+    result["decode_tps"] = round(statistics.median(meds), 2)
+    print(json.dumps(result))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -0,0 +1,200 @@
+#!/usr/bin/env bash
+# frank-primary-window.sh — one full bench window for frank-primary, run ON
+# DANKSERVER as the deploy user (invoked over ssh by ./autoresearch.sh).
+#
+# Discipline (AGENTS.md Benchmarking Policy): the production unit is
+# STOPPED + MASKED first, the candidate config boots AS THE SERVICE itself
+# (same serve script, same container name, same port 7770 — no scratch
+# ports, no co-tenant GPU), the in-container measure runs, and the window
+# ALWAYS ends by restoring the production unit and verifying its health.
+# The EXIT trap makes the restore unconditional: a boot failure, a measure
+# failure, or a dead ssh session all end with production back up.
+#
+# Staging layout (created by the orchestrator, ephemeral — never a config
+# source; the repo checkout is):
+#   /home/dank/.autoresearch-frank-primary/
+#     serve.sh             staged candidate serve script
+#     frank-primary.env    staged candidate env (the candidate knobs)
+#     measure.py           the in-container measure client
+#     hash                 candidate fingerprint
+#     run.json             result (metrics + boot + kv + fingerprint)
+#     serve.out/measure.out/fail.log
+#     last/                memo: run.json + hash of the last FULLY
+#                          successful window (bench AND verified restore)
+#
+# This script logs to stderr and exits 0 only when the measure produced a
+# usable result (the restore happens in the trap regardless of the bench
+# outcome; a failed restore forces exit 1).
+set -uo pipefail
+
+STAGE="${FRANK_PRIMARY_AR_STAGE:-/home/dank/.autoresearch-frank-primary}"
+UNIT="${FRANK_PRIMARY_AR_UNIT:-frank-primary}"
+CTR="${FRANK_PRIMARY_AR_CTR:-frank-primary-qwen}"
+HEALTH_URL="http://127.0.0.1:7770/health"
+MODELS_URL="http://127.0.0.1:7770/v1/models"
+BOOT_TIMEOUT="${FRANK_PRIMARY_AR_BOOT_TIMEOUT:-1200}"
+RESTORE_TIMEOUT="${FRANK_PRIMARY_AR_RESTORE_TIMEOUT:-1200}"
+MEASURE_TIMEOUT="${FRANK_PRIMARY_AR_MEASURE_TIMEOUT:-1800}"
+SERVED="$(sed -n 's/^FRANK_PRIMARY_SERVED_NAME=//p' "$STAGE/frank-primary.env" 2>/dev/null | head -1)"
+SERVED="${SERVED:-frank-primary}"
+
+RESTORED=0
+
+log() { printf '[window %s] %s\n' "$(date '+%H:%M:%S')" "$*" >&2; }
+
+container_state() { podman inspect -f '{{.State.Status}}' "$CTR" 2>/dev/null || echo absent; }
+
+wait_port_free() {
+    local i
+    for i in $(seq 1 90); do
+        [ "$(container_state)" != "running" ] && return 0
+        sleep 2
+    done
+    return 1
+}
+
+wait_health() { # $1 = timeout seconds
+    local t0 code
+    t0=$(date +%s)
+    while :; do
+        code="$(curl -s -o /dev/null -w '%{http_code}' --max-time 10 "$HEALTH_URL" 2>/dev/null || echo 000)"
+        [ "$code" = "200" ] && return 0
+        [ "$(( $(date +%s) - t0 ))" -ge "$1" ] && return 1
+        sleep 2
+    done
+}
+
+capture_bench_failure() {
+    log "capturing bench-boot failure evidence"
+    {
+        echo "=== podman logs (tail 400) ==="
+        podman logs --tail 400 "$CTR" 2>&1 || true
+        echo "=== staged serve.out (tail 60) ==="
+        tail -60 "$STAGE/serve.out" 2>/dev/null || true
+        echo "=== engine log (tail 60) ==="
+        tail -60 /home/dank/logs/frank_primary.log 2>/dev/null || true
+    } >"$STAGE/fail.log" 2>&1 || true
+    grep -E 'OutOfMemory|Tried to allocate|Engine core initialization failed|Error|error' \
+        "$STAGE/fail.log" 2>/dev/null | tail -25 >&2 || true
+}
+
+restore() {
+    [ "$RESTORED" = 1 ] && return 0
+    RESTORED=1
+    log "restoring production unit $UNIT"
+    if [ -f "$STAGE/serve.pid" ]; then
+        kill "$(cat "$STAGE/serve.pid" 2>/dev/null)" 2>/dev/null || true
+    fi
+    podman stop -t 120 "$CTR" >/dev/null 2>&1 || true
+    wait_port_free || log "WARNING: bench container still draining after 180s"
+    sudo systemctl unmask "$UNIT" 2>/dev/null || true
+    if ! sudo systemctl start "$UNIT"; then
+        log "ERROR: systemctl start $UNIT failed — operator intervention required"
+        return 1
+    fi
+    if wait_health "$RESTORE_TIMEOUT"; then
+        if curl -s --max-time 10 "$MODELS_URL" 2>/dev/null | grep -q "$SERVED"; then
+            log "production restored and verified ($SERVED answering on :7770)"
+            # Memo promotion lives HERE: a cached result is only ever the
+            # product of a fully successful window (bench AND verified
+            # restore), never of a bare bench success.
+            cp "$STAGE/run.json" "$STAGE/last/" 2>/dev/null || true
+            cp "$STAGE/hash" "$STAGE/last/" 2>/dev/null || true
+        else
+            log "ERROR: health 200 but /v1/models does not list $SERVED"
+            return 1
+        fi
+    else
+        log "ERROR: production not healthy within ${RESTORE_TIMEOUT}s — unit left to systemd (Restart=always)"
+        return 1
+    fi
+    return 0
+}
+trap 'restore; rc=$?; [ "$rc" -ne 0 ] && exit 1' EXIT
+
+main() {
+    local f
+    for f in serve.sh frank-primary.env measure.py hash; do
+        [ -f "$STAGE/$f" ] || { log "staged file missing: $STAGE/$f"; exit 1; }
+    done
+    mkdir -p "$STAGE/last"
+    rm -f "$STAGE/run.json" "$STAGE/fail.log" "$STAGE/measure.out"
+
+    local t_start t0 i kvline kv_tokens kv_conc measure_json boot_seconds
+    t_start="$(date +%s)"
+
+    log "stopping and masking production unit $UNIT"
+    sudo systemctl stop "$UNIT" 2>/dev/null || true
+    # Explicit stop is not auto-restarted even with Restart=always; the mask
+    # is belt-and-braces against a stray start during the window.
+    sudo systemctl mask "$UNIT" 2>/dev/null || \
+        log "WARNING: mask failed — relying on explicit-stop semantics"
+    wait_port_free || log "note: production container still draining"
+
+    t0="$(date +%s)"
+    log "starting bench instance from staged config (fingerprint $(tr -d '[:space:]' <"$STAGE/hash"))"
+    FRANK_PRIMARY_ENV_FILE="$STAGE/frank-primary.env" \
+        setsid nohup bash "$STAGE/serve.sh" >"$STAGE/serve.out" 2>&1 </dev/null &
+    echo "$!" >"$STAGE/serve.pid"
+
+    for i in $(seq 1 45); do
+        [ "$(container_state)" = "running" ] && break
+        sleep 2
+    done
+    if [ "$(container_state)" != "running" ]; then
+        log "bench container never reached running state"
+        tail -40 "$STAGE/serve.out" >&2 || true
+        exit 1
+    fi
+
+    if wait_health "$BOOT_TIMEOUT"; then
+        boot_seconds=$(( $(date +%s) - t0 ))
+        log "bench healthy after ${boot_seconds}s"
+    else
+        if [ "$(container_state)" = "running" ]; then
+            log "bench boot timed out after ${BOOT_TIMEOUT}s (still loading)"
+        else
+            log "bench container died during boot"
+            capture_bench_failure
+        fi
+        exit 1
+    fi
+
+    log "running measure in container (timeout ${MEASURE_TIMEOUT}s)"
+    if ! podman cp "$STAGE/measure.py" "$CTR":/tmp/ar-measure.py; then
+        log "podman cp of measure client failed"
+        exit 1
+    fi
+    if ! timeout "$MEASURE_TIMEOUT" podman exec "$CTR" python3 /tmp/ar-measure.py \
+            >"$STAGE/measure.out" 2>&1; then
+        log "measure failed or timed out"
+        tail -40 "$STAGE/measure.out" >&2 || true
+        exit 1
+    fi
+
+    # Capacity facts for the run (context-size experiments change these).
+    kvline="$(podman logs "$CTR" 2>&1 | grep -E 'GPU KV cache size' | tail -1 || true)"
+    kv_tokens="$(printf '%s' "$kvline" | sed -E 's/.*GPU KV cache size: ([0-9,]+) tokens.*/\1/' | tr -d ',' || true)"
+    kv_conc="$(printf '%s' "$kvline" | sed -E 's/.*per request: ([0-9.]+x).*/\1/' || true)"
+
+    measure_json="$(grep -E '^\{' "$STAGE/measure.out" | tail -1 || true)"
+    if [ -z "$measure_json" ]; then
+        log "measure produced no JSON result line"
+        tail -40 "$STAGE/measure.out" >&2 || true
+        exit 1
+    fi
+
+    # The measure line is already JSON, so the run object can be assembled
+    # without any host-side interpreter.
+    printf '{"hash":"%s","boot_seconds":%s,"kv_tokens":%s,"kv_concurrency":"%s","total_seconds":%s,"measure":%s}\n' \
+        "$(tr -d '[:space:]' <"$STAGE/hash")" \
+        "$boot_seconds" "${kv_tokens:-0}" "${kv_conc:-unknown}" \
+        "$(( $(date +%s) - t_start ))" "$measure_json" >"$STAGE/run.json"
+
+    log "bench window complete in $(( $(date +%s) - t_start ))s"
+    # Memo only after a FULL success; the restore below still has to verify.
+    exit 0
+}
+
+main "$@"
+exit "$?"
@@ -0,0 +1,269 @@
+#!/usr/bin/env python3
+"""bench-conversation.py — what an endpoint actually feels like, in tokens/s.
+
+The other harnesses here answer engine questions (which batch, which proposer,
+which cache). This one answers the user question: give the same ten jobs to an
+OpenAI-compatible endpoint and report the rate it answers them at. Every job is
+a real conversation turn — system prompt plus one user turn — because that is
+the shape the fleet serves, and it is measured by STREAMING so prefill and
+decode are separated:
+
+  ttft          seconds to the first content delta  -> prefill + queue
+  decode tok/s  completion_tokens / (total - ttft)  -> the rate you read at
+
+Both arms of a comparison must run the SAME prompt bytes, so a per-run salt
+rides in the system prompt: it changes the prompt itself, which is llama.cpp's
+prefix-cache key and vLLM's APC key, so no arm silently inherits a warm cache
+from the one before it. Run one arm, restart the engine, run the next with a
+new salt.
+
+Answers are generated at temperature 0 so arms serving the SAME model can be
+compared for identity as well as speed; different runtimes serving different
+quants legitimately answer differently, which is why the first line of every
+answer is printed.
+
+Usage:
+    bench-conversation.py <base-url> [--model NAME] [--salt TAG]
+                          [--max-tokens N] [--burst N] [--out file.json]
+
+    bench-conversation.py http://127.0.0.1:7770 --model frank-medium --salt L1
+    bench-conversation.py http://127.0.0.1:7770 --model frank-medium --salt R2 --burst 4
+
+One engine at a time: on a single GPU pair there is no room for two copies of a
+27B model, and the benchmarking policy requires the production service to be
+stopped while anything measures.
+"""
+
+import argparse
+import json
+import statistics
+import sys
+import threading
+import time
+import urllib.error
+import urllib.request
+
+# Ten jobs that cover what this fleet is actually asked to do. Ordered roughly
+# by prefill size so a runtime that is good at short prompts and one that is
+# good at long ones cannot hide behind an average.
+SYSTEM = (
+    "You are a staff engineer on a homelab fleet. Be concise and concrete. "
+    "Salt: {salt}"
+)
+
+BATTERY = [
+    ("greet", 128, "Say hello in one sentence."),
+    ("shell", 192, "One bash command: find every file under /home modified in "
+     "the last 2 days, sorted by size, largest first."),
+    ("math", 256, "A train covers 180 km in 1h40m, then 240 km in 2h20m. "
+     "What is the average speed in km/h? Show the arithmetic."),
+    ("json", 384, "Emit exactly 12 JSON records, one per line, no prose, each "
+     'with keys "host", "service", "port" and "up" for hosts dankclaw, '
+     "dankserver and dankware running llama-server, caddy and grafana."),
+    ("review", 512, "Review this function for bugs and name the failure mode:\n"
+     "def rotate(cfg, n):\n    for k in cfg:\n        cfg[k] = cfg[k][n:] + cfg[k][:n]\n"
+     "    return cfg\n"),
+    ("explain", 448, "Explain in four sentences why PCIe atomic operation routing "
+     "matters for GPU collectives, and what fails when a bridge reports "
+     "AtomicOpsCap: Routing-."),
+    ("translate", 256, "Translate to German, then to Japanese, one line each: "
+     "The load balancer refused the connection because the port was still held."),
+    ("plan", 640, "Give a 6-step runbook for swapping the model behind a live "
+     "LLM endpoint on a host that serves real users, with a rollback at each step."),
+    ("debug", 512, "A systemd user service restarts every 10 seconds and its log "
+     "ends with 'Error starting server failed to bind udp listener on 10.89.1.1:53: "
+     "Cannot assign requested address'. What is the root cause and the repair?"),
+    ("longdoc", 512, None),  # filled from LONG_FILES at runtime
+]
+
+# The long-context job reads real repo files so prefill is large and identical
+# across arms (~30k characters).
+LONG_FILES = [
+    "/home/dank/serverAppConfigs/AGENTS.md",
+    "/home/dank/serverAppConfigs/serverAppConfigs-docs/LLM/"
+    "frank-medium-llamacpp-dflash2.md",
+]
+LONG_CHARS = 30_000
+
+
+def long_prompt() -> str:
+    parts = []
+    for path in LONG_FILES:
+        try:
+            with open(path, encoding="utf-8", errors="replace") as handle:
+                parts.append(handle.read())
+        except OSError:
+            continue
+    text = "\n\n".join(parts)[:LONG_CHARS]
+    return ("Using only this document, answer: which deployment shapes for "
+            "two GPUs are ruled out by missing PCIe atomics, and what is the "
+            "one software path that still lets both cards serve one model? "
+            "Cite the measured numbers.\n\n---\n" + text)
+
+
+def stream_chat(base: str, model: str, messages, max_tokens: int, timeout: int):
+    """One streamed chat completion -> (ttft, total, usage, delta_count, text).
+
+    usage is the engine's own {prompt_tokens, completion_tokens}, asked for with
+    stream_options.include_usage.  There is deliberately NO cheap non-streamed
+    pre-probe to learn the prompt length: that probe would populate the prefix
+    cache and turn the measured run's prefill into a cache hit, which is exactly
+    the difference between runtimes this harness is trying to see.
+
+    ttft is None when the engine produced no content delta.
+    """
+    body = {
+        "model": model,
+        "messages": messages,
+        "max_tokens": max_tokens,
+        "temperature": 0,
+        "stream": True,
+        "stream_options": {"include_usage": True},
+    }
+    request = urllib.request.Request(
+        base + "/v1/chat/completions",
+        data=json.dumps(body).encode(),
+        headers={"Content-Type": "application/json"},
+    )
+    start = time.perf_counter()
+    ttft = None
+    deltas = 0
+    usage = {}
+    text = []
+    with urllib.request.urlopen(request, timeout=timeout) as response:
+        for raw in response:
+            line = raw.decode("utf-8", "replace").strip()
+            if not line.startswith("data:"):
+                continue
+            payload = line[5:].strip()
+            if payload == "[DONE]":
+                break
+            try:
+                chunk = json.loads(payload)
+            except json.JSONDecodeError:
+                continue
+            if chunk.get("usage"):
+                usage.update(chunk["usage"])
+            for choice in chunk.get("choices", []):
+                piece = choice.get("delta") or {}
+                # llama.cpp streams thinking as reasoning_content; vLLM's
+                # reasoning parser streams it as reasoning.  Either one is a
+                # generated token, so TTFT must be the first token of ANY kind —
+                # counting only `content` made a thinking-heavy arm look like it
+                # answered instantly (measured 2026-09-20: radiance's geomean
+                # read 994 tok/s with 6 of 10 jobs recorded as "(empty)").
+                content = next((piece[k] for k in ("content", "reasoning",
+                               "reasoning_content") if piece.get(k)), "")
+                if content:
+                    if ttft is None:
+                        ttft = time.perf_counter() - start
+                    deltas += 1
+                    text.append(content)
+    total = time.perf_counter() - start
+    return ttft, total, usage, deltas, "".join(text)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("base")
+    ap.add_argument("--model", default="frank-medium")
+    ap.add_argument("--salt", default=f"C{int(time.time())}")
+    ap.add_argument("--max-tokens", type=int, default=None)
+    ap.add_argument("--burst", type=int, default=0,
+                    help="also run N concurrent copies of the plan job")
+    ap.add_argument("--timeout", type=int, default=1800)
+    ap.add_argument("--out", default=None)
+    args = ap.parse_args()
+
+    base = args.base.rstrip("/")
+    rows = []
+    print(f"# battery salt={args.salt} model={args.model} base={base}\n")
+    print(f"{'job':<10} {'ptok':>7} {'ctok':>6} {'ttft_s':>8} {'dec_tps':>8} "
+          f"{'total_s':>8}  answer")
+
+    for name, cap, prompt in BATTERY:
+        if prompt is None:
+            prompt = long_prompt()
+        messages = [
+            {"role": "system", "content": SYSTEM.format(salt=args.salt)},
+            {"role": "user", "content": prompt},
+        ]
+        limit = args.max_tokens or cap
+        try:
+            ttft, total, usage, deltas, text = stream_chat(
+                base, args.model, messages, limit, args.timeout)
+        except (urllib.error.URLError, urllib.error.HTTPError, TimeoutError) as exc:
+            print(f"{name:<10} FAILED {exc}")
+            rows.append({"job": name, "error": str(exc)})
+            continue
+        ptok = usage.get("prompt_tokens")
+        ctok = usage.get("completion_tokens") or deltas
+        decode = None
+        if ttft is not None and ctok and total > ttft:
+            decode = (ctok - 1) / (total - ttft)
+        prefill = (ptok / ttft) if (ptok and ttft) else None
+        first = text.strip().splitlines()[0][:48] if text.strip() else "(empty)"
+        print(f"{name:<10} {ptok or 0:>7} {ctok:>6} "
+              f"{(ttft or 0):>8.2f} {(decode or 0):>8.1f} {total:>8.2f}  {first}")
+        rows.append({"job": name, "prompt_tokens": ptok, "completion_tokens": ctok,
+                     "ttft_s": ttft, "decode_tps": decode, "prefill_tps": prefill,
+                     "total_s": total, "answer_head": first})
+
+    good = [r for r in rows if r.get("decode_tps")]
+    if good:
+        rates = [r["decode_tps"] for r in good]
+        ttfts = [r["ttft_s"] for r in good]
+        pre = [r["prefill_tps"] for r in good if r.get("prefill_tps")]
+        print(f"\ngeomean decode  {statistics.geometric_mean(rates):8.1f} tok/s"
+              f"   (min {min(rates):.1f} / max {max(rates):.1f})")
+        print(f"mean ttft       {statistics.mean(ttfts):8.2f} s")
+        if pre:
+            print(f"geomean prefill {statistics.geometric_mean(pre):8.1f} tok/s")
+
+    if args.burst > 1:
+        prompt = next(p for n, _, p in BATTERY if n == "plan")
+        messages = [
+            {"role": "system", "content": SYSTEM.format(salt=args.salt + "B")},
+            {"role": "user", "content": prompt},
+        ]
+        results = []
+        lock = threading.Lock()
+
+        def worker():
+            try:
+                ttft, total, usage, deltas, _ = stream_chat(
+                    base, args.model, messages, 640, args.timeout)
+                ctok = usage.get("completion_tokens") or deltas
+                with lock:
+                    results.append((ttft, total, ctok))
+            except Exception as exc:  # noqa: BLE001 - report, do not abort the burst
+                with lock:
+                    results.append((None, 0.0, str(exc)))
+
+        threads = [threading.Thread(target=worker) for _ in range(args.burst)]
+        t0 = time.perf_counter()
+        for thread in threads:
+            thread.start()
+        for thread in threads:
+            thread.join()
+        wall = time.perf_counter() - t0
+        ok = [r for r in results if r[0] is not None]
+        if ok:
+            tokens = sum(r[2] for r in ok if isinstance(r[2], int))
+            print(f"burst {args.burst}x plan: wall {wall:.1f}s  aggregate "
+                  f"{tokens / wall:.1f} tok/s  "
+                  f"per-stream {statistics.mean([r[2] / r[1] for r in ok]):.1f} tok/s"
+                  f"  mean ttft {statistics.mean([r[0] for r in ok]):.2f}s")
+        else:
+            print(f"burst {args.burst}x plan: all failed: {results[0][2]}")
+
+    if args.out:
+        with open(args.out, "w", encoding="utf-8") as handle:
+            json.dump({"salt": args.salt, "model": args.model, "base": base,
+                       "rows": rows, "burst": args.burst}, handle, indent=1)
+        print(f"wrote {args.out}")
+    return 0 if good else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -0,0 +1,96 @@
+#!/usr/bin/env bash
+# Build the gufo runtime that serves frank-large on Dankclaw.
+#
+#   bash scripts/build-gufo.sh              # build + install the pinned tag
+#   bash scripts/build-gufo.sh --verify     # only check the installed tree
+#
+# gufo (gufo-org/gufo) is the Strix Halo inference engine: it targets
+# gfx1151 specifically and has no prebuilt Linux release binaries, so the
+# runtime is built from the pinned git tag and installed to
+# /home/dank/GitPackages/gufo-<tag>/. Everything below is what Dankclaw
+# needs on top of AMD's Linux instructions.
+#
+# The host already carries the ROCm 7.2.4 stack (/opt/rocm -> /etc/
+# alternatives/rocm) with the HIP dev packages, so this script installs
+# only the ordinary system libraries gufo links against.
+#
+# Two host facts this build depends on, both discovered on 2026-09-28:
+#
+#   1. Ubuntu's librocwmma-dev ships a TRUNCATED rocWMMA header set (the
+#      public headers without rocwmma/internal/) into /usr/include, where
+#      it shadows AMD's complete copy under /opt/rocm/include. The first
+#      build died on "'internal/accessors.hpp' file not found" from
+#      /usr/include/rocwmma/rocwmma.hpp. A -I cannot fix it: in HIP mode
+#      the host compilation pass searches the system directories first.
+#      The package has no reverse dependencies, so it is removed.
+#   2. The ROCm clang picks the NEWEST /usr/include/c++ tree, which on
+#      Ubuntu 26.04 is GCC 16's. Its <atomic_wait.h> emits a symbol
+#      reference the installed libstdc++ (GLIBCXX_3.4.35, GCC 15) does not
+#      export, and the link fails with "undefined symbol:
+#      std::__detail::__notify_impl". --gcc-install-dir pins the headers to
+#      the same GCC the runtime library came from.
+set -euo pipefail
+
+GUFO_TAG="${GUFO_TAG:-v0.1.1}"
+GUFO_REPO="${GUFO_REPO:-https://github.com/gufo-org/gufo}"
+# The commit the tag resolved to when this runtime was qualified. A tag that
+# moves is an unreproducible runtime, so the pin is checked, not assumed.
+GUFO_COMMIT="${GUFO_COMMIT:-b0f8673470a3dcc2b5e9c9546c48a40b82b5ff31}"
+SRC_DIR="${GUFO_SRC_DIR:-/home/dank/GitPackages/gufo-src}"
+PREFIX="${GUFO_PREFIX:-/home/dank/GitPackages/gufo-${GUFO_TAG}}"
+JOBS="${JOBS:-$(nproc)}"
+# The GCC toolchain that owns the installed libstdc++ on this host.
+GCC_DIR="${GCC_DIR:-/usr/lib/gcc/x86_64-linux-gnu/15}"
+
+# check_pin — refuse a source tree that is not the qualified commit.
+check_pin() {
+    local commit
+    commit="$(git -C "$SRC_DIR" rev-parse HEAD 2>/dev/null || true)"
+    if [[ -n "$commit" && "$commit" != "$GUFO_COMMIT" ]]; then
+        echo "GUFO_COMMIT mismatch: $SRC_DIR is at $commit, pinned $GUFO_COMMIT" >&2
+        exit 1
+    fi
+}
+
+if [[ "${1:-}" == "--verify" ]]; then
+    # The pin first: a moved tag is the more urgent thing to hear about, and
+    # it is the check that must not be masked by a missing install.
+    check_pin
+    [[ -x "$PREFIX/bin/gufo" ]] || { echo "no gufo at $PREFIX/bin/gufo" >&2; exit 1; }
+    "$PREFIX/bin/gufo" diagnose | tail -n 5
+    exit 0
+fi
+
+echo "== system libraries gufo links against =="
+sudo -n apt-get install -y -qq --no-install-recommends \
+    build-essential cmake ninja-build pkg-config \
+    libicu-dev libcurl4-openssl-dev libssl-dev libpng-dev libjpeg-dev ffmpeg
+
+echo "== dropping Ubuntu's truncated rocWMMA headers =="
+if dpkg -s librocwmma-dev >/dev/null 2>&1; then
+    sudo -n apt-get remove -y -qq librocwmma-dev
+fi
+
+echo "== source: $GUFO_REPO @ $GUFO_TAG =="
+if [[ ! -d "$SRC_DIR/.git" ]]; then
+    mkdir -p "$(dirname "$SRC_DIR")"
+    git clone --depth 1 --branch "$GUFO_TAG" "$GUFO_REPO" "$SRC_DIR"
+fi
+git -C "$SRC_DIR" fetch --depth 1 origin "refs/tags/$GUFO_TAG"
+git -C "$SRC_DIR" checkout -q --detach "refs/tags/$GUFO_TAG"
+echo "   $(git -C "$SRC_DIR" rev-parse HEAD)"
+check_pin
+
+echo "== configure + build (release, HIP gfx1151) =="
+cd "$SRC_DIR"
+toolchain="--gcc-install-dir=$GCC_DIR"
+cmake --preset release \
+    -DCMAKE_INSTALL_PREFIX="$PREFIX" \
+    -DCMAKE_HIP_FLAGS="$toolchain" \
+    -DCMAKE_EXE_LINKER_FLAGS="$toolchain"
+cmake --build --preset release --parallel "$JOBS"
+cmake --install build/release
+
+echo "== verify =="
+"$PREFIX/bin/gufo" diagnose
+echo "frank-large runtime ready: $PREFIX/bin/gufo ($GUFO_TAG @ $commit)"
@@ -0,0 +1,96 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+readonly repo_root="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
+readonly ck_root="${KAIRIC_CK_ROOT:-$repo_root/third_party/composable_kernel}"
+readonly ck_commit=fdf4bb7fcc984811cef48ce817d89aac064b984a
+readonly ck_patch="$repo_root/patches/composable-kernel-gfx1151-iu4.patch"
+readonly build_dir="${BUILD_DIR:-$repo_root/build-kairic}"
+readonly rocm="${ROCM_PATH:-/opt/rocm}"
+readonly jobs="${JOBS:-$(nproc)}"
+
+for tool in cmake git ninja; do
+  command -v "$tool" >/dev/null || {
+    echo "missing required tool: $tool" >&2
+    exit 1
+  }
+done
+
+[[ -f "$ck_patch" ]] || {
+  echo "missing Composable Kernel patch: $ck_patch" >&2
+  exit 1
+}
+
+if [[ ! -d "$ck_root/.git" ]]; then
+  mkdir -p "$(dirname -- "$ck_root")"
+  git clone https://github.com/ROCm/composable_kernel.git "$ck_root"
+fi
+
+git -C "$ck_root" checkout "$ck_commit"
+if git -C "$ck_root" diff --quiet; then
+  git -C "$ck_root" apply --check "$ck_patch"
+  git -C "$ck_root" apply "$ck_patch"
+elif git -C "$ck_root" apply --reverse --check "$ck_patch"; then
+  echo "Composable Kernel release patch is already applied"
+else
+  echo "Composable Kernel checkout does not match the pinned patch" >&2
+  exit 1
+fi
+
+if [[ -x "$rocm/bin/amdclang++" ]]; then
+  hip_compiler="$rocm/bin/amdclang++"
+elif [[ -x "$rocm/llvm/bin/clang++" ]]; then
+  hip_compiler="$rocm/llvm/bin/clang++"
+else
+  echo "cannot find a HIP compiler below ROCM_PATH=$rocm" >&2
+  exit 1
+fi
+
+if [[ -n "${CC:-}" ]]; then
+  c_compiler="$(command -v "$CC")"
+elif [[ -x /usr/bin/gcc ]]; then
+  c_compiler=/usr/bin/gcc
+else
+  c_compiler="$(command -v gcc)"
+fi
+if [[ -n "${CXX:-}" ]]; then
+  cxx_compiler="$(command -v "$CXX")"
+elif [[ -x /usr/bin/g++ ]]; then
+  cxx_compiler=/usr/bin/g++
+else
+  cxx_compiler="$(command -v g++)"
+fi
+
+export ROCM_PATH="$rocm"
+export PATH="$rocm/bin:$rocm/lib/llvm/bin:$rocm/llvm/bin:$PATH"
+export LD_LIBRARY_PATH="$rocm/lib:$rocm/lib/rocm_sysdeps/lib:$rocm/lib/llvm/lib:$rocm/llvm/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
+readonly hip_flags="-DGGML_ROCMFPX_RDNA35_MMID_MAX_BATCH=5 -DGGML_ROCMFPX_MOE_MMVQ_ROWS_PER_BLOCK=4${KAIRIC_HIP_TOOLCHAIN_FLAGS:+ $KAIRIC_HIP_TOOLCHAIN_FLAGS}"
+
+cmake -S "$repo_root" -B "$build_dir" -G Ninja \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_C_COMPILER="$c_compiler" \
+  -DCMAKE_CXX_COMPILER="$cxx_compiler" \
+  -DCMAKE_HIP_COMPILER="$hip_compiler" \
+  -DCMAKE_HIP_FLAGS="$hip_flags" \
+  -DCMAKE_PREFIX_PATH="$rocm" \
+  -DBUILD_SHARED_LIBS=ON \
+  -DGGML_CPU=ON -DGGML_OPENMP=ON -DGGML_HIP=ON -DGGML_CUDA=OFF \
+  -DGGML_VULKAN=OFF -DGGML_HIP_FORCE_MMQ=ON -DGGML_HIP_GRAPHS=ON \
+  -DGGML_HIP_MMQ_MFMA=ON -DGGML_HIP_NO_VMM=ON \
+  -DGGML_HIP_ROCWMMA_FATTN=OFF -DGGML_NATIVE=ON \
+  -DAMDGPU_TARGETS=gfx1151 -DGPU_BUILD_TARGETS=gfx1151 \
+  -DPROMPTFORGE_CK_ROOT="$ck_root" \
+  -DLLAMA_BUILD_WEBUI=OFF \
+  -DGGML_BUILD_TESTS=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_SERVER=ON
+
+cmake --build "$build_dir" --target llama-server -j"$jobs"
+
+"$build_dir/bin/llama-server" --version
+"$build_dir/bin/llama-server" --help | grep -A1 -- '--kairic-edge'
+ldd "$build_dir/bin/llama-server" 2>&1 | tee "$build_dir/llama-server.ldd.txt"
+if grep -Eiq 'not found|version .* required by .* not found' "$build_dir/llama-server.ldd.txt"; then
+  echo "runtime dependency verification failed" >&2
+  exit 1
+fi
+
+echo "Kairic Edge runtime built at $build_dir/bin/llama-server"
@@ -38,6 +38,7 @@ JOBS="${JOBS:-$(nproc)}"
 ROCM_VERSION="${ROCM_VERSION:-7.14.0a20260624}"
 ROCM_LOCAL_PATH="${ROCM_LOCAL_PATH:-${XDG_DATA_HOME:-$HOME/.local/share}/rocm-${ROCM_VERSION}}"
 HIP_ARCH="${HIP_ARCH:-gfx1201}"
+GGML_HIP_RCCL="${GGML_HIP_RCCL:-OFF}"

 # ── Source ────────────────────────────────────────────────────────────
 if [[ ! -d "$LLAMA_SRC/.git" ]]; then
@@ -70,11 +71,13 @@ export ROCM_PATH="$ROCM_LOCAL_PATH"
 export HIP_PATH="$ROCM_LOCAL_PATH"

 # ── Configure ─────────────────────────────────────────────────────────
-# GGML_HIP_RCCL stays OFF: this host's second card cannot dispatch a RCCL
-# kernel at all (its PCIe path has no atomic routing), and layer-split serving
-# needs no collectives.  With RCCL off, llama.cpp logs "internal AllReduce init
-# failed ... falling back to meta-backend butterfly" and only --split-mode
-# tensor pays for it.
+# GGML_HIP_RCCL defaults OFF: layer-split serving needs no collectives, and
+# until 2026-09-20 this host's second card could not dispatch a RCCL kernel at
+# all (its PCIe path had no atomic routing).  With RCCL off, llama.cpp logs
+# "internal AllReduce init failed ... falling back to meta-backend butterfly"
+# and only --split-mode row/tensor pays for it.  Set GGML_HIP_RCCL=ON (with a
+# distinct BUILD_DIR) for the collective-backed runtime; the atomics
+# precondition now holds on both cards.
 cmake -S "$LLAMA_SRC" -B "$BUILD_DIR" \
    -G Ninja \
    -DCMAKE_BUILD_TYPE=Release \
@@ -83,7 +86,7 @@ cmake -S "$LLAMA_SRC" -B "$BUILD_DIR" \
    -DCMAKE_CXX_FLAGS="-I${ROCM_LOCAL_PATH}/include" \
    -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
    -DGGML_HIP=ON \
-    -DGGML_HIP_RCCL=OFF \
+    -DGGML_HIP_RCCL="$GGML_HIP_RCCL" \
    -DGGML_CUDA=OFF \
    -DGGML_VULKAN=OFF \
    -DGGML_NATIVE=OFF \
@@ -108,7 +111,7 @@ if command -v patchelf >/dev/null; then
    mkdir -p "$LIB_DIR"
    for pat in libamdhip64 librocblas libhipblas libhipblaslt librocsolver \
               libroctx64 libamd_comgr libhsa-runtime64 librocm_kpack \
-               librocroller librocprofiler-register; do
+               librocroller librocprofiler-register librocm-core; do
        cp -v "$ROCM_LOCAL_PATH/lib/$pat".so* "$LIB_DIR/" 2>/dev/null || true
    done
    cp -v "$ROCM_LOCAL_PATH/lib/llvm/lib"/libLLVM.so* \
@@ -116,6 +119,13 @@ if command -v patchelf >/dev/null; then
          "$ROCM_LOCAL_PATH/lib/llvm/lib"/libomp.so* \
          "$ROCM_LOCAL_PATH/lib/llvm/lib"/libomptarget.so* "$LIB_DIR/" 2>/dev/null || true
    cp -v "$ROCM_LOCAL_PATH/lib/rocm_sysdeps/lib"/librocm_sysdeps_*.so* "$LIB_DIR/" 2>/dev/null || true
+    # A RCCL build links librccl.so.1 (imported shared even with
+    # BUILD_SHARED_LIBS=OFF), and the binaries resolve libraries through
+    # $ORIGIN/../lib — so the collective library must ride beside the rest or
+    # llama-server fails to exec at all.
+    if [[ "$GGML_HIP_RCCL" == "ON" ]]; then
+        cp -v "$ROCM_LOCAL_PATH"/lib/librccl.so* "$LIB_DIR/" 2>/dev/null || true
+    fi
    # rocBLAS and hipBLASLt look for their device code at a FIXED depth under
    # the binary's lib dir: $ORIGIN/../lib/rocblas/library and
    # $ORIGIN/../lib/hipblaslt/library/<arch>. `cp -r SRC DST/` with a missing
@@ -0,0 +1,89 @@
+#!/usr/bin/env bash
+# Build the ROCmFPX runtime that serves frank-large on Dankclaw.
+#
+#   bash scripts/build-rocmfpx.sh              # clone/fetch + build the pin
+#   bash scripts/build-rocmfpx.sh --verify     # only check the installed tree
+#
+# Runtime: kingjones30/ROCmFPX — charlie12345/ROCmFPX plus the model-card
+# patches, i.e. the fork that owns the ROCmFPx tensor formats. frank-large's
+# payload is ftype 115 (`Q8_0_ROCMFPX_AGENT`): 311 tensors of
+# GGML_TYPE_Q8_0_ROCMFPX (103), a type that exists ONLY in this fork, so
+# stock llama.cpp, gufo and every other loader on the host reject the file.
+# This is the same tree that served the Qwen3.8-Flash-Next ROCmFP4 STRIX
+# endpoint 2026-08-30 → 09-11 (see the kingjones runbook), rebuilt here so a
+# fresh host can be brought up from the repo alone.
+#
+# Why this commit and not a newer one: 36e9acd (2026-08-28) is what was
+# qualified against this payload on 2026-10-02 — it loads ftype 115, survives
+# warmup, and drafts with the MTP head at acceptance 0.65. The five legacy
+# builds of 2026-08-25 crashed this same file in ggml_compute_forward_set_rows
+# (see kingjones-q8agent-rocmfpx.md), so the pin is load-bearing: do not bump
+# it without re-running that crash test.
+#
+# Host notes:
+#   * -DLLAMA_BUILD_WEBUI=OFF is REQUIRED on this box. The tree provisions its
+#     embedded UI assets via npm or the HF llama-ui bucket; npm is absent here
+#     and the bucket download half-failed, aborting the link at 98%. The API
+#     server is unaffected.
+#   * The build links against the host ROCm (/opt/rocm); no bundled toolchain
+#     like the gfx1201 R9700 build needed.
+set -euo pipefail
+
+ROCMFPX_REPO="${ROCMFPX_REPO:-https://github.com/kingjones30/ROCmFPX.git}"
+# The commit frank-large was qualified with (measured 2026-10-02).
+ROCMFPX_COMMIT="${ROCMFPX_COMMIT:-36e9acd40e10a87cd3c3ef8ec734668757dc8520}"
+SRC_DIR="${ROCMFPX_SRC_DIR:-/home/dank/GitPackages/kingjones-ROCmFPX}"
+BUILD_DIR="${ROCMFPX_BUILD_DIR:-$SRC_DIR/build}"
+GPU_TARGET="${ROCMFPX_GPU_TARGET:-gfx1151}"
+JOBS="${JOBS:-$(nproc)}"
+
+# check_pin — refuse a source tree that is not the qualified commit. A moved
+# fork means different kernels under the same path, and on this model line a
+# wrong build is a segfault at first request, not a clean error.
+check_pin() {
+    local commit
+    commit="$(git -C "$SRC_DIR" rev-parse HEAD 2>/dev/null || true)"
+    if [[ -n "$commit" && "$commit" != "$ROCMFPX_COMMIT" ]]; then
+        echo "ROCMFPX_COMMIT mismatch: $SRC_DIR is at $commit, pinned $ROCMFPX_COMMIT" >&2
+        exit 1
+    fi
+}
+
+if [[ "${1:-}" == "--verify" ]]; then
+    check_pin
+    [[ -x "$BUILD_DIR/bin/llama-server" ]] || {
+        echo "no ROCmFPX llama-server at $BUILD_DIR/bin/llama-server" >&2
+        exit 1
+    }
+    # The format the endpoint exists for: the fork must carry the ROCmFPx
+    # tensor type, or the payload cannot load at all.
+    "$BUILD_DIR/bin/llama-server" --help 2>&1 | grep -q -- '--spec-type' || {
+        echo "$BUILD_DIR/bin/llama-server has no --spec-type: MTP drafting is impossible" >&2
+        exit 1
+    }
+    "$BUILD_DIR/bin/llama-server" --version | head -1
+    exit 0
+fi
+
+echo "== source: $ROCMFPX_REPO @ $ROCMFPX_COMMIT =="
+if [[ ! -d "$SRC_DIR/.git" ]]; then
+    mkdir -p "$(dirname "$SRC_DIR")"
+    git clone "$ROCMFPX_REPO" "$SRC_DIR"
+fi
+git -C "$SRC_DIR" fetch --quiet origin "$ROCMFPX_COMMIT" || true
+git -C "$SRC_DIR" checkout -q --detach "$ROCMFPX_COMMIT"
+check_pin
+echo "   $(git -C "$SRC_DIR" rev-parse HEAD)"
+
+echo "== configure + build (release, HIP $GPU_TARGET) =="
+cmake -S "$SRC_DIR" -B "$BUILD_DIR" \
+    -DGGML_HIP=ON \
+    -DGPU_TARGETS="$GPU_TARGET" \
+    -DGGML_NATIVE=ON \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DLLAMA_BUILD_WEBUI=OFF
+cmake --build "$BUILD_DIR" --target llama-server llama-quantize -j "$JOBS"
+
+echo "== verify =="
+"$BUILD_DIR/bin/llama-server" --version | head -1
+echo "frank-large runtime ready: $BUILD_DIR/bin/llama-server ($ROCMFPX_COMMIT)"
@@ -0,0 +1,58 @@
+#!/usr/bin/env bash
+# Pin Dankserver's fleet NIC (eno1, 192.168.0.235) as the primary default route.
+#
+# The 2026-09-20 board swap (TUF GAMING B650-E -> ProArt B850-Creator WiFi NEO)
+# brought up a second port, eno2, which took DHCP and — with NetworkManager's
+# default per-connection metric of 100 — came up with metric 101 while eno1's
+# static gateway got 102. Outbound traffic therefore started leaving by eno2
+# even though every fleet peer, Caddy upstream and monitoring target addresses
+# this host as .235. Keep eno2 (it is a real second link) but make it secondary:
+# route metrics decide the default route, so eno1 wins while eno2 stays a
+# failover.
+#
+# Run on Dankserver.  Idempotent; NetworkManager owns these connections (the
+# netplan files under /etc/netplan are its generated shim, so edit via nmcli).
+#
+#   bash scripts/dankserver-route-primary.sh [--check]
+set -euo pipefail
+
+PRIMARY='Wired connection 1'      # eno1, static 192.168.0.235, fleet address
+SECONDARY='Wired connection 2'    # eno2, DHCP 192.168.0.247
+PRIMARY_IF='eno1'
+SECONDARY_IF='eno2'
+PRIMARY_METRIC=100
+SECONDARY_METRIC=200
+
+show() {
+    printf 'routes:\n'; ip route show default
+    printf 'wanted: %s metric %s, %s metric %s\n' \
+        "$PRIMARY" "$PRIMARY_METRIC" "$SECONDARY" "$SECONDARY_METRIC"
+}
+
+if [[ "${1:-}" == '--check' ]]; then
+    for conn in "$PRIMARY" "$SECONDARY"; do
+        printf '%s: ipv4=%s ipv6=%s\n' "$conn" \
+            "$(nmcli -g ipv4.route-metric connection show "$conn")" \
+            "$(nmcli -g ipv6.route-metric connection show "$conn")"
+    done
+    show
+    exit 0
+fi
+
+nmcli connection modify "$PRIMARY" \
+    ipv4.route-metric "$PRIMARY_METRIC" ipv6.route-metric "$PRIMARY_METRIC"
+nmcli connection modify "$SECONDARY" \
+    ipv4.route-metric "$SECONDARY_METRIC" ipv6.route-metric "$SECONDARY_METRIC"
+
+# `nmcli device reapply <ifname>` pushes the edited profile without tearing the
+# link down.  `nmcli connection up` would re-run DHCP and briefly drop the route
+# this session rides on, and `connection reapply` is not a verb (NM 1.54).
+nmcli device reapply "$PRIMARY_IF"
+nmcli device reapply "$SECONDARY_IF"
+
+show
+if [[ "$(ip route show default | head -1)" != *"dev $PRIMARY_IF"* ]]; then
+    printf 'ERROR: eno1 is still not the preferred default route\n' >&2
+    exit 1
+fi
+printf 'eno1 is the preferred default route; eno2 remains up as failover.\n'
@@ -0,0 +1,71 @@
+#!/usr/bin/env bash
+# Single source of truth for Dankserver GLM-5.3 state.
+#
+# This exists because the ad-hoc status commands kept producing WRONG readings.
+# The specific trap: /proc/meminfo separates key from value with a TAB, so the
+# common idiom
+#   grep MemAvailable /proc/meminfo | tr -s " " | cut -d" " -f2
+# yields an empty field (tab is not a space) and awk then prints 0.0 -- which
+# reads as "host out of memory" when the host actually has 240 GiB free.
+# Everything below parses on a colon split, which is tab-agnostic, and the
+# memory figure is cross-checked before it is shown.
+set -uo pipefail
+
+MB=1048576
+GLM_CTR="${GLM_CTR:-glm-w4a16}"
+
+mem_field() { awk -v k="$1:" '$1==k {printf "%.1f", $2/'"$MB"'; exit}' /proc/meminfo; }
+section() { printf '\n== %s ==\n' "$1"; }
+
+section "time / host"
+printf '  %s   %s\n' "$(date +%H:%M:%S)" \
+  "$(awk -F: '/model name/{print $2; exit}' /proc/cpuinfo | sed 's/^ *//')"
+
+section "memory (GiB)"
+TOT=$(mem_field MemTotal); AVAIL=$(mem_field MemAvailable)
+printf '  MemTotal     %8s\n' "$TOT"
+printf '  MemAvailable %8s\n' "$AVAIL"
+printf '  Cached       %8s\n' "$(mem_field Cached)"
+awk -v t="$TOT" -v a="$AVAIL" 'BEGIN{
+  if (t+0 <= 0 || a+0 <= 0) { print "  *** PARSE BUG: zero value ***"; exit }
+  if (a+0 > t+0)         { print "  *** PARSE BUG: available > total ***"; exit }
+  printf "  check        ok (%.0f%% free)\n", 100*a/t
+}'
+
+section "swap"
+swapon --show=NAME,SIZE,USED --noheadings 2>/dev/null \
+  | awk '{printf "  %-12s %6s used of %s\n",$1,$3,$2}' || echo "  none"
+
+section "gpu"
+# CSV layout: device,VRAM Total Memory (B),VRAM Total Used Memory (B)
+rocm-smi --showmeminfo vram --csv 2>/dev/null | awk -F, '
+  NR>1 && $1 ~ /^card[0-9]+$/ {
+    printf "  %-8s %7.2f GiB used of %.2f\n", $1, $3/1073741824, $2/1073741824
+  }'
+printf '  compute util: '
+rocm-smi --showuse --csv 2>/dev/null | awk -F, 'NR>1 && $1 ~ /^card[0-9]+$/ {printf "%s=%s%% ", $1, $2}'
+echo
+
+section "disk"
+df -h /home/dank 2>/dev/null | awk 'NR==2{printf "  %s  %s used  %s free  (%s)\n",$1,$3,$4,$5}'
+
+section "llm containers"
+if podman ps -a --format '{{.Names}}' 2>/dev/null | grep -qE 'glm|frank'; then
+  podman ps -a --format '{{.Names}}|{{.Status}}' 2>/dev/null \
+    | awk -F'|' '$1 ~ /glm|frank/ {printf "  %-18s %s\n",$1,$2}'
+else
+  echo "  none"
+fi
+
+section "frank-primary unit"
+printf '  active:  %s\n' "$(systemctl is-active frank-primary 2>/dev/null | head -1)"
+printf '  enabled: %s\n' "$(systemctl is-enabled frank-primary 2>/dev/null | head -1)"
+
+section "glm serve tail"
+if podman ps -a --format '{{.Names}}' 2>/dev/null | grep -qx "$GLM_CTR"; then
+  podman logs "$GLM_CTR" 2>&1 | grep -viE 'multiproc_executor\.py:943' | tail -5 | sed 's/^/  /'
+else
+  echo "  (container $GLM_CTR not present)"
+fi
+
+section "done"
@@ -38,6 +38,7 @@ SERVICE_ORDER = (
    "loki",
    "alloy",
    "node-exporter",
+    "alertmanager",
    "authentik-postgres",
    "authentik-server",
    "authentik-worker",
@@ -51,6 +52,7 @@ SERVICE_ORDER = (
    "copyparty",
    "jellyfin",
    "calibre-web",
+    "calibre-server",
    "plex",
    "tautulli",
    "prowlarr",
@@ -61,7 +63,7 @@ SERVICE_ORDER = (
    "readarr",
    "frank-large",
    "frank-small",
-    "frank-medium",
+    "frank-primary",
 )

 DANKCLAW_SERVICES = {
@@ -83,6 +85,7 @@ DANKCLAW_SERVICES = {
    "loki",
    "alloy",
    "node-exporter",
+    "alertmanager",
    "authentik-postgres",
    "authentik-server",
    "authentik-worker",
@@ -101,6 +104,7 @@ DANKSERVER_SERVICES = {
    "copyparty",
    "jellyfin",
    "calibre-web",
+    "calibre-server",
    "plex",
    "tautulli",
    "prowlarr",
@@ -109,7 +113,7 @@ DANKSERVER_SERVICES = {
    "sonarr",
    "lidarr",
    "readarr",
-    "frank-medium",
+    "frank-primary",
    "node-exporter",
 }

@@ -246,6 +250,10 @@ def services_for_path(path: str, target: str) -> set[str]:
            _add(services, "searxng")
        elif group == "prometheus":
            _add(services, "prometheus")
+            if len(parts) >= 3 and parts[2] == "alertmanager.yml":
+                # The Alertmanager config ships from this directory: a change
+                # to it must redeploy the service that installs it.
+                _add(services, "alertmanager")
        elif group == "loki":
            _add(services, "loki")
        elif group == "alloy":
@@ -308,11 +316,8 @@ def services_for_path(path: str, target: str) -> set[str]:
        if len(parts) >= 2 and parts[1] == "open-webui":
            _add(services, "open-webui")
            return services
-        if len(parts) >= 2 and parts[1] == "frank-large":
-            # The service is a rootless Quadlet container since 2026-09-11;
-            # its unit and fleet settings live in one .container file.
-            _add(services, "frank-large")
-            return services
+        # frank-large left the quadlet shape on 2026-09-19 (CIRU v4.4.1 host
+        # build, root unit): its files live in scripts/ and map there.
        if len(parts) >= 2 and parts[1] == "observability":
            # The five observability containers became Quadlet units on
            # 2026-09-11. A unit change selects exactly its own service;
@@ -320,7 +325,7 @@ def services_for_path(path: str, target: str) -> set[str]:
            # consumer (prometheus before grafana, loki before alloy).
            if len(parts) >= 4 and parts[2] == "quadlet":
                name = parts[3].removesuffix(".container")
-                if name in {"prometheus", "grafana", "loki", "alloy", "node-exporter"}:
+                if name in {"prometheus", "grafana", "loki", "alloy", "node-exporter", "alertmanager"}:
                    _add(services, name)
            return services
        if len(parts) >= 2 and parts[1] == "gitea":
@@ -350,6 +355,7 @@ def services_for_path(path: str, target: str) -> set[str]:
                "romm": {"romm", "romm-mariadb", "romm-redis"},
                "jellyfin": {"jellyfin"},
                "calibre-web": {"calibre-web"},
+                "calibre-server": {"calibre-server"},
                "copyparty": {"copyparty"},
                "plex": {"plex"},
                "tautulli": {"tautulli"},
@@ -365,13 +371,11 @@ def services_for_path(path: str, target: str) -> set[str]:
                if name == "node-exporter":
                    _add(services, "node-exporter")
            return services
-        if len(parts) >= 2 and parts[1] == "frank-medium":
-            # frank-medium is a ROOT llama.cpp unit since 2026-09-16: its
-            # runtime is a native gfx1201 binary, not a container, so there is
-            # nothing under dankserver/frank-medium/ to plan any more. The
-            # retired vllm-radiance quadlets live in .AGENTTRASH/, which is
-            # archive material and plans nothing.
-            return services
+        # frank-primary's quadlet-era directory left the tree on 2026-09-26
+        # (the Lucebox DS4.1 unit was retired for the llama.cpp root unit —
+        # the trio in scripts/ now maps it). Nothing lives under
+        # dankserver/frank-primary/ except dated trash, and trash plans
+        # nothing (the retire rule above the service blocks).
        return services

    if parts[0] == "frankbook-pro":
@@ -391,18 +395,37 @@ def services_for_path(path: str, target: str) -> set[str]:
        name = p.name
        if name in {"homepage-status.py", "homepage-status.service", "homepage-status.timer"}:
            _add(services, "homepage")
+        if name in {"signal-relay.py", "signal-relay.service"}:
+            # The Signal relay is installed by the alertmanager deploy, so a
+            # change to it redeploying alertmanager is the only path that
+            # reaches it on the host.
+            _add(services, "alertmanager")
        if name in {"frank-small.env", "frank-small.service", "serve-frank-small.sh"}:
            # llama.cpp frank-small (2026-09-16) keeps its unit, env and serve
            # files here; each one selects frank-small only. The ik_llama.cpp
            # trio that lived here for two days is in scripts/.AGENTTRASH/,
            # which is archive material and plans nothing.
            _add(services, "frank-small")
-        if name in {"frank-medium.env", "frank-medium.service", "serve-frank-medium.sh"}:
-            # llama.cpp frank-medium (2026-09-16) keeps its unit, env and serve
-            # files here the same way. The files that BUILD and FETCH its
-            # runtime and payload are host-side one-offs: they change no
-            # deployed state, so they select nothing.
-            _add(services, "frank-medium")
+        if name in {"frank-large.env", "frank-large.service", "serve-frank-large.sh"}:
+            # CIRU v4.4.1 frank-large (2026-09-19) keeps its unit, env and
+            # serve files here the same way: a root unit on a host-built
+            # runtime. The download script fetches the payload, which is
+            # host-side state, so it selects nothing.
+            _add(services, "frank-large")
+        if name == "qwen-sharp-v22.4.0.jinja":
+            # The Qwen-Sharp chat template frank-large loads via
+            # --chat-template-file. It lived under dankserver/frank-medium/ until
+            # the 2026-09-26 cutover retired that directory; editing it changes
+            # what the live frank-large endpoint renders, so it plans
+            # frank-large.
+            _add(services, "frank-large")
+        if name in {"frank-primary.env", "frank-primary.service", "serve-frank-primary.sh"}:
+            # llama.cpp frank-primary (2026-09-26, replacing the Lucebox
+            # DS4.1 quadlet) keeps its unit, env and serve files here the
+            # same way: a root unit on the unsloth b11030-mix prebuilt
+            # serving Qwen3.8-Flash-Next Q8_0. The download script fetches
+            # the payload, which is host-side state, so it selects nothing.
+            _add(services, "frank-primary")
        # Deployment/test/benchmark scripts and other repo tooling do not
        # restart a runtime service.
        return services
@@ -412,7 +435,6 @@ def services_for_path(path: str, target: str) -> set[str]:
        return services
    return services

-
 def changed_paths(repo: str, base: str, head: str) -> list[str]:
    result = subprocess.run(
        ["git", "-C", repo, "diff", "--name-only", "--diff-filter=ACDMRTUXB", base, head],
@@ -86,16 +86,16 @@ service_is_allowed() {
        dankclaw:open-webui|dankclaw:gitea-postgres|dankclaw:gitea-server|\
        dankclaw:comfyui|dankclaw:hermes|\
        dankclaw:hermes-dashboard|dankclaw:hermes-webui|dankclaw:prometheus|dankclaw:grafana|\
-        dankclaw:loki|dankclaw:alloy|dankclaw:node-exporter|\
+        dankclaw:loki|dankclaw:alloy|dankclaw:node-exporter|dankclaw:alertmanager|\
        dankclaw:authentik-postgres|dankclaw:authentik-server|\
        dankclaw:authentik-worker|dankclaw:authentik-outpost|dankclaw:authentik-outpost-ldap|dankclaw:anubis|\
        dankclaw:caddy|dankclaw:homepage|dankclaw:seerr|dankclaw:marreta|\
        dankclaw:frank-large|dankclaw:frank-small|\
        dankserver:romm-mariadb|dankserver:romm-redis|dankserver:romm|\
-        dankserver:copyparty|dankserver:jellyfin|dankserver:calibre-web|dankserver:plex|dankserver:tautulli|\
+        dankserver:copyparty|dankserver:jellyfin|dankserver:calibre-web|dankserver:calibre-server|dankserver:plex|dankserver:tautulli|\
        dankserver:prowlarr|dankserver:qbittorrent|dankserver:radarr|\
        dankserver:sonarr|dankserver:lidarr|dankserver:readarr|\
-        dankserver:frank-medium|dankserver:node-exporter|\
+        dankserver:frank-primary|dankserver:node-exporter|\
        frankbook-pro:gluetun|frankbook-pro:qbittorrent|frankbook-pro:prowlarr|\
        frankbook-pro:radarr|frankbook-pro:sonarr|frankbook-pro:lidarr|\
        frankbook-pro:readarr)
@@ -177,6 +177,17 @@ add_pending_services() {
        if [[ "$status" != deployed ]]; then
            # A pending service was intentionally skipped or its handler
            # failed. Retry it even when the checkout has not advanced.
+            # But a pending entry for a service RETIRED from this target can
+            # never be deployable: retrying it fails the whole run before
+            # the e2e gate (the frank-default-sightsound lockout, found
+            # 2026-09-17). Drop such entries into the host trash instead.
+            if ! service_is_allowed "$service"; then
+                local trash="$HOME/.AGENTTRASH/$(date +%Y%m%d)-retired-deploy-state"
+                mkdir -p "$trash"
+                mv -- "$state_file" "$trash/$service"
+                echo "WARN: $service is retired on $TARGET; dropped its stale $status state to $trash" >&2
+                continue
+            fi
            SERVICES+=("$service")
            continue
        fi
@@ -198,6 +209,11 @@ copy_searxng_config() {
    run_privileged mkdir -p /opt/searxng/config
    run_privileged cp shared/searxng/settings.yml /opt/searxng/config/settings.yml
    run_privileged cp shared/searxng/limiter.toml /opt/searxng/config/limiter.toml
+    # Bake ${KAGI_API_KEY}/${SEARXNG_SECRET}/${SEARXNG_BASE_URL} into the live
+    # copy from the host-side /opt/searxng/.env (SearXNG does not expand env
+    # vars in settings.yml). Fails the deploy if a placeholder is unresolvable.
+    run_privileged python3 scripts/searxng-render-settings.py \
+        /opt/searxng/.env /opt/searxng/config/settings.yml
 }

 copy_prometheus_config() {
@@ -210,6 +226,53 @@ copy_prometheus_config() {
        /opt/prometheus/config/rules/llm-fleet.yml
 }

+# Alertmanager config ships beside the Prometheus config: the two halves of
+# the alerting path change together, and a receiver with no route (or a
+# route with no receiver) fails silently.
+copy_alertmanager_config() {
+    run_privileged mkdir -p /opt/alertmanager/config /opt/alertmanager/data
+    run_privileged cp shared/prometheus/alertmanager.yml /opt/alertmanager/config/alertmanager.yml
+}
+
+# The Signal relay is Alertmanager's only webhook consumer and the only
+# thing on this host that talks to signal-cli. It is a plain systemd USER
+# unit (not a quadlet): it runs python3 from the checkout, so a deploy
+# updates it without a copy step, and — unlike generated quadlet units,
+# which this host's podman 5.x v1 generator rejects — a hand-written user
+# unit can be enabled, and Linger=yes for dank starts it at boot.
+install_signal_relay() {
+    local unit_dir="$HOME/.config/systemd/user"
+    mkdir -p "$unit_dir"
+    cp scripts/signal-relay.service "$unit_dir/signal-relay.service"
+    systemctl --user daemon-reload
+    systemctl --user enable signal-relay.service
+    systemctl --user restart signal-relay.service
+    # The relay binds the host's LAN IP (the one address a bridge container
+    # can reach as host.containers.internal), so keep the unauthenticated
+    # endpoint off the rest of the LAN: admit loopback, the host itself
+    # (its own health probe) and the podman bridge subnet only. Idempotent:
+    # -C before -I.
+    run_privileged iptables -C INPUT -p tcp --dport 18093 -s 127.0.0.1 -j ACCEPT 2>/dev/null \
+        || run_privileged iptables -I INPUT -p tcp --dport 18093 -s 127.0.0.1 -j ACCEPT
+    run_privileged iptables -C INPUT -p tcp --dport 18093 -s 192.168.0.143 -j ACCEPT 2>/dev/null \
+        || run_privileged iptables -I INPUT -p tcp --dport 18093 -s 192.168.0.143 -j ACCEPT
+    run_privileged iptables -C INPUT -p tcp --dport 18093 -s 10.89.0.0/16 -j ACCEPT 2>/dev/null \
+        || run_privileged iptables -I INPUT -p tcp --dport 18093 -s 10.89.0.0/16 -j ACCEPT
+    run_privileged iptables -C INPUT -p tcp --dport 18093 -j DROP 2>/dev/null \
+        || run_privileged iptables -I INPUT -p tcp --dport 18093 -j DROP
+    # The interpreter needs a moment to bind after restart; a one-shot curl
+    # raced the socket on 2026-09-17 and failed a healthy deploy.
+    local i
+    for i in $(seq 1 15); do
+        if curl -fsS -m 5 http://192.168.0.143:18093/health >/dev/null; then
+            return 0
+        fi
+        sleep 1
+    done
+    echo "signal-relay is not answering on 192.168.0.143:18093" >&2
+    return 1
+}
+
 copy_loki_config() {
    run_privileged mkdir -p /opt/loki/config
    run_privileged cp shared/loki/loki-config.yaml /opt/loki/config/loki-config.yaml
@@ -255,7 +318,7 @@ deploy_homepage() {
    fi
    run_privileged mkdir -p /srv/homepage
    run_privileged cp shared/homepage/index.html shared/homepage/admin.html shared/homepage/app.js \
-        shared/homepage/style.css shared/homepage/services.json /srv/homepage/
+        shared/homepage/style.css shared/homepage/services.json shared/homepage/docs.html /srv/homepage/
    if [[ -d shared/homepage/fonts ]]; then
        run_privileged cp -r shared/homepage/fonts /srv/homepage/
    fi
@@ -426,13 +489,17 @@ seed_grafana_env() {
 deploy_observability() {
    local service="$1"
    # Observability units live under <target>/observability/quadlet: Dankclaw
-    # has five, Dankserver one (node-exporter, since 2026-09-16). The config
+    # has six, Dankserver one (node-exporter, since 2026-09-16). The config
    # copies below are Dankclaw-owned state and run only there.
    local unit_dir="$TARGET/observability/quadlet"
    case "$service" in
        prometheus)
            copy_prometheus_config
            ;;
+        alertmanager)
+            copy_alertmanager_config
+            install_signal_relay
+            ;;
        loki)
            copy_loki_config
            build_compose_image loki
@@ -456,6 +523,7 @@ deploy_observability() {
    systemctl --user restart "$service.service"
    case "$service" in
        prometheus) wait_for_health prometheus http://127.0.0.1:9090/-/healthy 60 ;;
+        alertmanager) wait_for_health alertmanager http://127.0.0.1:9093/-/healthy 60 ;;
        grafana) wait_for_health grafana http://127.0.0.1:3002/api/health 90 ;;
        loki) wait_for_health loki http://127.0.0.1:3100/ready 60 ;;
        alloy) wait_for_health alloy http://127.0.0.1:12345/-/healthy 60 ;;
@@ -787,6 +855,50 @@ deploy_calibre_web() {
    systemctl --user restart calibre-web.service
 }

+# The content server's single admin account and the desktop's basic-auth
+# credential (same pair) are bootstrapped exactly once: the env file is the
+# plaintext store (mode 600), the userdb's pw column is plaintext too, so
+# its file is mode 600 as well. The password is piped on stdin with printf -
+# a trailing newline from echo lands in the stored password and breaks every
+# login (verified 2026-09-21).
+deploy_calibre_server() {
+    local data_root=/home/dank/Containers/calibre-server
+    mkdir -p "$data_root"
+    if [[ ! -f $data_root/env ]]; then
+        local pw
+        pw=$(openssl rand -base64 18 | tr -d '/+=' | head -c 20)
+        printf 'CUSTOM_USER=calibre\nCUSTOM_PASSWORD=%s\n' "$pw" > "$data_root/env"
+        chmod 600 "$data_root/env"
+        echo "calibre-server: generated credential (shown once): calibre / $pw"
+    fi
+    [[ -f $data_root/env ]] || {
+        echo "FAIL: $data_root/env does not exist; seed it before the first calibre-server deploy" >&2
+        return 1
+    }
+    if [[ ! -f $data_root/userdb.sqlite ]]; then
+        local pw
+        pw=$(grep '^CUSTOM_PASSWORD=' "$data_root/env" | cut -d= -f2)
+        touch "$data_root/userdb.sqlite"
+        printf '%s' "$pw" | podman run --rm -i --entrypoint /usr/bin/calibre-server \
+            -v "$data_root/userdb.sqlite:/userdb.sqlite" \
+            lscr.io/linuxserver/calibre:9.15.0 \
+            --userdb /userdb.sqlite --manage-users -- add calibre
+        chown 1000:1000 "$data_root/userdb.sqlite"
+        chmod 600 "$data_root/userdb.sqlite"
+    fi
+    [[ -f $data_root/userdb.sqlite && -s $data_root/userdb.sqlite ]] || {
+        echo "FAIL: $data_root/userdb.sqlite missing or empty; the --manage-users bootstrap failed" >&2
+        return 1
+    }
+    cp "dankserver/calibre-server/quadlet/calibre-server.container" "$quadlet_unit_dir/"
+    systemctl --user daemon-reload
+    if ! systemctl --user is-active -q calibre-server.service; then
+        podman stop calibre-server 2>/dev/null || true
+        podman rm calibre-server 2>/dev/null || true
+    fi
+    systemctl --user restart calibre-server.service
+}
+
 deploy_copyparty() {
    # The admin account is created from ${COPAY_ACCOUNT}:${COPAY_PASSWORD} expanded
    # at ExecStart time; a missing env file must fail the unit, not boot copyparty
@@ -965,41 +1077,39 @@ deploy_anubis() {
    wait_container_healthy anubis 60 || return 1
 }

-# frank-large runs as a rootless Quadlet unit again since 2026-09-16: the
-# halogen-flash-server runtime is back (it replaced ik_llama.cpp, which had
-# replaced it), and the RTX 3090 it shared during that interlude now belongs
-# to frank-small.
+# frank-large is the ROCmFPX fork of llama.cpp (kingjones30/ROCmFPX, pinned by
+# scripts/build-rocmfpx.sh), serving
+# kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF:Q8_0_ROCMFPX_AGENT with the
+# repo's own Q4_0 MTP draft head and its BF16 vision projector, on the Strix
+# Halo iGPU. It is a ROOT unit: the runtime is a native binary linked against
+# the host ROCm, so there is no container image to install and no generated
+# unit to dodge. relink_unit points /etc/systemd/system/frank-large.service at
+# the repo file and boot persistence is that unit's [Install] section.
 #
-# TWO MANAGERS, ONE UNIT NAME. The ik incarnation ran as a ROOT unit whose
-# file is a symlink into this checkout; the halogen unit is a USER (quadlet)
-# unit. Both must never exist at once — they would race for port 7770 — so the
-# root unit is retired first: stop, disable, move the /etc symlink into the
-# home trash (never `rm`), reload the root manager. This host's `systemctl
-# disable` has been observed to delete the unit file itself (2026-08-26
-# incident), so the move tolerates a file that is already gone.
+# The runtime gate matters as much as the payload gate: the fork is built from
+# source on this host, and the payload is ftype 115 (tensor type 103,
+# GGML_TYPE_Q8_0_ROCMFPX) which only this fork can read — so a machine whose
+# /home/dank/GitPackages tree was never built has a unit that can never
+# become healthy, and a green deploy would leave the endpoint dark.
 #
-# FIRST START fetches ~127 GB of weights, and /health cannot answer while it
-# does. wait_for_llm_large_ready recognises that (checkpoint absent) and waits
-# for the file instead of burning the health budget on a failure that is not
-# one.
-LLM_LARGE_CHECKPOINT="${LLM_LARGE_CHECKPOINT:-/home/dank/models/halogen/qwen38-flash-next-w4b.hgn}"
+# It replaces gufo v0.1.1 + unsloth UD-Q4_K_XL (the rollback target, still
+# built by scripts/build-gufo.sh with its payload on disk), which had replaced
+# the stock llama.cpp prebuilt. The port gate is the same as every cutover
+# since 2026-09-16: nothing starts until 7770 is released.
+LLM_LARGE_RUNTIME="${LLM_LARGE_RUNTIME:-/home/dank/GitPackages/kingjones-ROCmFPX/build/bin/llama-server}"
+LLM_LARGE_MODEL="${LLM_LARGE_MODEL:-/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf}"
+LLM_LARGE_DRAFT="${LLM_LARGE_DRAFT:-/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/mtp-Qwen3.8-27B-Q4_0.gguf}"
+LLM_LARGE_MMPROJ="${LLM_LARGE_MMPROJ:-/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/mmproj-Qwen3.8-27B-BF16.gguf}"

-retire_root_llm_large() {
-    local trash="$HOME/.AGENTTRASH/20260916-halogen-restore/host-units"
-    run_privileged systemctl stop frank-large.service 2>/dev/null || true
-    run_privileged systemctl disable frank-large.service 2>/dev/null || true
-    if [[ -e /etc/systemd/system/frank-large.service || -L /etc/systemd/system/frank-large.service ]]; then
+retire_halogen_large_quadlet() {
+    local trash="$HOME/.AGENTTRASH/20260919-orca-cutover/quadlet"
+    systemctl --user stop frank-large.service 2>/dev/null || true
+    if [[ -e "$quadlet_unit_dir/frank-large.container" ]]; then
        mkdir -p "$trash"
-        run_privileged mv -- /etc/systemd/system/frank-large.service \
-            "$trash/frank-large.service"
-        echo "retired root unit frank-large.service -> $trash"
-    else
-        # This host's `systemctl disable` is allowed to delete the unit file
-        # outright (2026-08-26 incident), so there is often nothing left to
-        # move. Say so: previously no message meant no evidence either way.
-        echo "root unit frank-large.service was already removed by disable"
+        mv -- "$quadlet_unit_dir/frank-large.container" "$trash/frank-large.container"
+        echo "retired quadlet frank-large.container -> $trash"
    fi
-    run_privileged systemctl daemon-reload
+    systemctl --user daemon-reload
 }

 # Wait until a TCP port has no socket on it at all -- listening, accepted, or
@@ -1022,166 +1132,202 @@ wait_port_free() {
 }

 deploy_llm_large() {
-    retire_root_llm_large
-    # The predecessor can also be THIS quadlet -- a redeploy of a running
-    # frank-large -- so stop the user unit before waiting on the port. On a
-    # first cutover it is not running and this is a no-op.
-    systemctl --user stop frank-large.service 2>/dev/null || true
-    # Pull once, then run the localhost/-aliased tag: this host's quadlet path
-    # cannot resolve registry-prefixed refs from local storage at create time
-    # (same workaround as anubis/hermes), and Pull=never keeps boots off the
-    # registry.
-    local image="ghcr.io/peonist-ai/halogen-flash-server:0.5.8"
-    local alias="localhost/peonist-ai/halogen-flash-server:0.5.8"
-    podman pull "$image"
-    podman tag "$image" "$alias"
-    # The weights volume must exist, and the container needs it read-write:
-    # that is where a first start downloads into.
-    mkdir -p /home/dank/models/halogen
+    # Refuse rather than restart-loop. A unit whose runtime or payload is
+    # missing never becomes healthy, and a green deploy would leave the
+    # endpoint dark with nothing in the deploy log to say why. The gates
+    # mirror serve-frank-large.sh's own predicates exactly: runtime
+    # executable, target GGUF, MTP head, mmproj.
+    if [[ ! -x "$LLM_LARGE_RUNTIME" ]]; then
+        echo "ERROR: ROCmFPX runtime absent at $LLM_LARGE_RUNTIME" >&2
+        echo "frank-large: build it: bash scripts/build-rocmfpx.sh" >&2
+        return 1
+    fi
+    local missing=0 f
+    for f in "$LLM_LARGE_MODEL" "$LLM_LARGE_DRAFT" "$LLM_LARGE_MMPROJ"; do
+        if [[ ! -f "$f" ]]; then
+            echo "ERROR: payload absent at $f" >&2
+            echo "frank-large: fetch it: bash scripts/download-frank-large-kingjones.sh" >&2
+            missing=1
+        fi
+    done
+    [[ "$missing" -eq 0 ]] || return 1
+
+    retire_halogen_large_quadlet
+    # On a re-deploy the predecessor is THIS unit: stop it before waiting on
+    # the port, or the wait deadlocks on the server it just stopped serving.
+    # Then sweep a stray manual run -- that is how a bench harness leaves the
+    # port busy. Match the TARGET GGUF name, which every frank-large command
+    # line carries via --model and which no other process on this host uses:
+    # frank-small also runs llama-server here (with the MiMo GGUF) and must
+    # not be killed, so keying on the binary name would take the 3090 endpoint
+    # down with it. The pattern never appears in the deploying shell's own
+    # command line, so -f is safe here (it is not in `bash deploy-remote.sh`).
+    run_privileged systemctl stop frank-large.service 2>/dev/null || true
+    pkill -f 'Qwen3.8-27B-Q8_0_ROCMFPX_AGENT' 2>/dev/null || true
    # Nothing may start until the predecessor has released the port. On
-    # 2026-09-16 this sequence failed exactly here: the ik server logged
-    # "Deactivated successfully", pasta tried to bind 7770 a second later and
-    # reported "Address already in use", and the unit exited 126. Restart=
-    # on-failure recovered it minutes later, but a deploy that starts a doomed
-    # container has not deployed anything -- so gate on the port, not on hope.
+    # 2026-09-16 the cutover failed exactly here: the old server logged
+    # "Deactivated successfully", the next engine tried to bind 7770 a second
+    # later and reported "Address already in use". Gate on the port, not on
+    # hope.
    wait_port_free 7770 || return 1
-    install_quadlet frank-large dankclaw/frank-large/quadlet frank-large
-    # No enable: generated units cannot be enabled on this host (v1 quadlet
-    # generator); boot persistence comes from [Install] in the unit file.
-    systemctl --user restart frank-large.service
-    wait_for_llm_large_ready
+    relink_unit frank-large
+    run_privileged systemctl daemon-reload
+    run_privileged systemctl enable frank-large.service
+    run_privileged systemctl restart frank-large.service
    wait_for_health frank-large http://127.0.0.1:7770/health 1200
 }

-# Bootstrap-aware gate: if the checkpoint is absent the container is fetching
-# it; wait for the file (the fetch resumes, and the unit restarts it on
-# failure) rather than gating on a /health that cannot answer yet. Reports
-# success either way, so a first deploy does not look like a failure while the
-# engine does exactly what it was told to.
-wait_for_llm_large_ready() {
-    [[ -f "$LLM_LARGE_CHECKPOINT" ]] && return 0
-    echo "frank-large: checkpoint absent at $LLM_LARGE_CHECKPOINT"
-    echo "frank-large: first start — the container is downloading ~127 GB from"
-    echo "frank-large: peonist-ai/halogen-qwen3.8-flash-next (resumes if interrupted)."
-    local i
-    for ((i = 1; i <= 1200; i++)); do
-        if [[ -f "$LLM_LARGE_CHECKPOINT" ]]; then
-            echo "frank-large: weights landed after approximately $((i * 2))s; gating on /health"
-            return 0
-        fi
-        sleep 2
-    done
-    echo "WARN: frank-large download still in progress after 40m — NOT a failure." >&2
-    echo "WARN: the unit keeps fetching; re-run this deploy (or re-check /health) later." >&2
-    return 0
-}
-
-# deploy_llm_small — frank-small is mainline llama.cpp on the RTX 3090, as a
-# ROOT unit: the runtime is a native CUDA binary, so there is no container
-# image to install and no generated unit to avoid. relink_unit points
-# /etc/systemd/system/frank-small.service at the repo file and boot persistence
-# is that unit's [Install] section.
+# deploy_llm_small — frank-small is llama.cpp (master build, CUDA sm_86) on
+# the RTX 3090 as a ROOT unit: the runtime is a native binary tree under
+# ~/GitPackages/llama-master/build-cuda, so there is no container image and
+# no generated unit to avoid. relink_unit points
+# /etc/systemd/system/frank-small.service at the repo file and boot
+# persistence is that unit's [Install] section.
 #
 # It is a SEPARATE endpoint on port 7771; frank-large keeps 7770. The 3090 is
-# not shared: frank-large's halogen engine runs on the gfx1151 iGPU.
-LLM_SMALL_MODEL="${LLM_SMALL_MODEL:-/home/dank/models/Gemma4-12B-QAT/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf}"
+# not shared: frank-large's engine runs on the gfx1151 iGPU.
+LLM_SMALL_MODEL_FILE="${LLM_SMALL_MODEL_FILE:-/home/dank/models/MiMo-Q5KM/MiMo-V2.6-Distill-Qwen-9B-Q5_K_M.gguf}"
+LLM_SMALL_RUNTIME="${LLM_SMALL_RUNTIME:-/home/dank/GitPackages/llama-master/build-cuda/bin/llama-server}"
+# The projector IS the vision capability here: the Q5_K_M GGUF ships no `v.*`
+# tensors, so a deploy that starts without this file silently ships a
+# text-only endpoint. Checked unconditionally, like frank-large's.
+LLM_SMALL_MMPROJ="${LLM_SMALL_MMPROJ:-/home/dank/models/MiMo-Q5KM/mmproj-MiMo-V2.6-Distill-Qwen-9B-f16.gguf}"

 deploy_llm_small() {
-    if [[ ! -f "$LLM_SMALL_MODEL" ]]; then
-        echo "ERROR: model absent at $LLM_SMALL_MODEL" >&2
-        echo "frank-small: fetch HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced (Q4_K_M, the mmproj and the MTP sidecar) first;" >&2
-        echo "frank-small: refusing to start a unit that cannot load." >&2
+    # Refuse rather than restart-loop: a unit whose runtime or payload is
+    # missing never becomes healthy, and a green deploy would leave the
+    # endpoint dark with nothing in the log to say why.
+    if [[ ! -x "$LLM_SMALL_RUNTIME" ]]; then
+        echo "ERROR: llama.cpp runtime absent at $LLM_SMALL_RUNTIME" >&2
+        echo "frank-small: build it (cmake -B build-cuda -DGGML_CUDA=ON" >&2
+        echo "frank-small: -DCMAKE_CUDA_ARCHITECTURES=86 -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/gcc-13" >&2
+        echo "frank-small: && cmake --build build-cuda --target llama-server) first." >&2
        return 1
    fi
+    for f in "$LLM_SMALL_MODEL_FILE" "$LLM_SMALL_MMPROJ"; do
+        if [[ ! -f "$f" ]]; then
+            echo "ERROR: frank-small payload file absent at $f" >&2
+            echo "frank-small: run scripts/download-frank-small.sh first;" >&2
+            echo "frank-small: refusing to start a unit that cannot load." >&2
+            return 1
+        fi
+    done
    relink_unit frank-small
    run_privileged systemctl daemon-reload
    run_privileged systemctl enable frank-small.service
    run_privileged systemctl restart frank-small.service
-    wait_for_health frank-small http://127.0.0.1:7771/health 900
+    wait_for_health frank-small http://127.0.0.1:7771/health 900 || return 1
+    # False-green hardening (frank-large's lesson): the port answering
+    # proves nothing about WHO answers — the unit must be active with a
+    # live MainPID, or a stray SGLang holdover is masquerading as deployed.
+    local main_pid
+    main_pid="$(run_privileged systemctl show -p MainPID --value frank-small.service)"
+    if [[ -z "$main_pid" || "$main_pid" == "0" ]]; then
+        echo "FAIL: :7771 answered but frank-small.service has no MainPID — a stray process holds the port" >&2
+        return 1
+    fi
+    echo "frank-small: health answered by unit MainPID $main_pid"
 }

-# deploy_llm_medium — frank-medium is mainline llama.cpp on the two R9700s as a
-# ROOT unit since the 2026-09-16 cutover: the runtime is a native gfx1201 HIP
-# binary, so there is no image to install and no generated unit to dodge. It
-# replaces the vllm-radiance replica pair + routing proxy, which held this
-# endpoint for a day and could never use both cards for one model (TP=2 needs
-# an RCCL collective, and the second card cannot dispatch one).
+# deploy_llm_primary — frank-primary is the tcclaviger RDNA4 vLLM fork
+# (digest-pinned image, gfx1201) serving the orcarouter
+# Qwen3.8-Flash-Next-Uncensored-FP8 payload (131 safetensors shards) with
+# MTP4 spec decode and q8/q8 KV on the two R9700s as a ROOT unit whose serve
+# script drives `podman run` directly (the image entrypoint execs
+# `vllm serve`). 2026-09-29 cutover from the R9V radiance fork, whose
+# launcher/shim/setup-state machinery is retired (state dirs stay on disk as
+# rollback; pre-cutover tag frank-primary-r9v-q8-20260929). Same port and
+# alias as the endpoints it replaced; the harness models.yml id is unchanged.
 #
-# Layer split in ONE process is what the successor does instead: layers are
-# distributed across the cards and only the activation at the boundary crosses
-# PCIe, once per forward pass. That is the "pipeline parallelism" the endpoint
-# wanted, with no collective anywhere in it.
-LLM_MEDIUM_BIN="${LLM_MEDIUM_BIN:-/home/dank/GitPackages/llama.cpp-dflash2/build-rdna4-rocm714/bin/llama-server}"
-LLM_MEDIUM_MODEL="${LLM_MEDIUM_MODEL:-/home/dank/models/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-UD-Q8_K_XL.gguf}"
-LLM_MEDIUM_DRAFT="${LLM_MEDIUM_DRAFT:-/home/dank/models/Qwen3.8-27B-DFlash2-GGUF/Qwen3.8-27B-DFlash2-Q8_0.gguf}"
-LLM_MEDIUM_MMPROJ="${LLM_MEDIUM_MMPROJ:-/home/dank/models/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF/mmproj-BF16.gguf}"
+# Gates: the pinned runtime image must be loadable from dank's rootless
+# storage by its OCI manifest digest (the env file's FRANK_PRIMARY_IMAGE),
+# the payload dir must carry config.json and the expected shard count
+# (FRANK_PRIMARY_MODEL_SHARDS), and the unit env must be complete. Per-file
+# sha256 verification was the fetch script's job at download time, never
+# re-hashed here.
+LLM_PRIMARY_CONTAINER=frank-primary-qwen
+LLM_PRIMARY_LEGACY_CONTAINERS="frank-primary-r9v qwen-fp8 glm-probe"

-# The three vLLM user units are retired HERE, not left to rot: the proxy
-# published 7770 on the host network, so the llama-server below cannot bind
-# until they are stopped and the port is released. Never `rm` — the unit files
-# move into the home trash; the images stay in local storage.
-retire_vllm_medium_quadlets() {
-    local trash="$HOME/.AGENTTRASH/20260916-vllm-radiance-retire" unit
-    systemctl --user stop frank-medium-a.service frank-medium-b.service \
-        frank-medium-proxy.service 2>/dev/null || true
-    for unit in frank-medium-a frank-medium-b frank-medium-proxy frank-medium; do
-        if [[ -e "$quadlet_unit_dir/$unit.container" ]]; then
-            mkdir -p "$trash"
-            mv -- "$quadlet_unit_dir/$unit.container" "$trash/$unit.container"
-            echo "retired quadlet $unit.container -> $trash"
-        fi
-    done
-    systemctl --user daemon-reload
-}
-
-deploy_llm_medium() {
-    # Refuse rather than restart-loop. A unit whose runtime or payload is
+deploy_llm_primary() {
+    # Refuse rather than restart-loop: a unit whose image or payload is
    # missing never becomes healthy, and a green deploy would leave the
-    # endpoint dark with nothing in the deploy log to say why.
-    local missing=0
-    if [[ ! -x "$LLM_MEDIUM_BIN" ]]; then
-        echo "ERROR: llama-server absent at $LLM_MEDIUM_BIN" >&2
-        echo "frank-medium: build it ON the host (gfx1201): bash scripts/build-llama-rdna4.sh" >&2
-        missing=1
+    # endpoint dark with nothing in the log to say why.
+    local env_file="/home/dank/serverAppConfigs/scripts/frank-primary.env"
+    local image model_dir shards count
+    if [[ ! -f "$env_file" ]]; then
+        echo "ERROR: frank-primary env file absent: $env_file" >&2
+        return 1
    fi
-    if [[ ! -f "$LLM_MEDIUM_MODEL" ]]; then
-        echo "ERROR: target model absent at $LLM_MEDIUM_MODEL" >&2
-        echo "frank-medium: fetch it: bash scripts/download-frank-medium-gguf.sh" >&2
-        missing=1
+    image="$(sed -n 's/^FRANK_PRIMARY_IMAGE=//p' "$env_file" | tail -1)"
+    model_dir="$(sed -n 's/^FRANK_PRIMARY_MODEL_DIR=//p' "$env_file" | tail -1)"
+    shards="$(sed -n 's/^FRANK_PRIMARY_MODEL_SHARDS=//p' "$env_file" | tail -1)"
+    if [[ -z "$image" || -z "$model_dir" || -z "$shards" ]]; then
+        echo "ERROR: frank-primary.env missing FRANK_PRIMARY_IMAGE/MODEL_DIR/MODEL_SHARDS" >&2
+        return 1
    fi
-    if [[ ! -f "$LLM_MEDIUM_DRAFT" ]]; then
-        echo "ERROR: DFlash2 drafter absent at $LLM_MEDIUM_DRAFT" >&2
-        echo "frank-medium: fetch it: bash scripts/download-frank-medium-gguf.sh" >&2
-        missing=1
+    # Capture podman's output BEFORE grepping: under set -o pipefail, the
+    # classic `podman images | grep -q` races (grep exits on match, podman
+    # dies SIGPIPE 141 writing the rest, pipefail calls it no-match — the
+    # 2026-09-30 deploy refusal).
+    local digests
+    digests="$(podman images --format '{{.RepoDigests}}')"
+    if ! grep -qF "${image#*@}" <<<"$digests"; then
+        echo "ERROR: pinned runtime image absent from rootless storage: $image" >&2
+        echo "frank-primary: fetch it (podman pull $image)" >&2
+        return 1
    fi
-    if [[ ! -f "$LLM_MEDIUM_MMPROJ" ]]; then
-        echo "ERROR: vision projector absent at $LLM_MEDIUM_MMPROJ" >&2
-        echo "frank-medium: fetch it: bash scripts/download-frank-medium-gguf.sh" >&2
-        missing=1
+    if [[ ! -f "$model_dir/config.json" ]]; then
+        echo "ERROR: payload config missing: $model_dir/config.json" >&2
+        echo "frank-primary: re-run scripts/download-frank-primary.sh" >&2
+        return 1
+    fi
+    if [[ ! -f "$model_dir/model.safetensors.index.json" ]]; then
+        echo "ERROR: weight-map index missing: $model_dir/model.safetensors.index.json" >&2
+        echo "frank-primary: re-run scripts/download-frank-primary.sh" >&2
+        return 1
+    fi
+    count="$(find "$model_dir" -maxdepth 1 -name 'model-*-of-*.safetensors' 2>/dev/null | wc -l)"
+    if [[ "$count" != "$shards" ]]; then
+        echo "ERROR: payload shard count wrong: $model_dir (found $count, want $shards)" >&2
+        echo "frank-primary: re-run scripts/download-frank-primary.sh" >&2
+        return 1
    fi
-    [[ "$missing" -eq 0 ]] || return 1

-    retire_vllm_medium_quadlets
-    # The thing holding 7770 is now THIS service: on a re-deploy the previous
-    # llama-server is still bound to it, so waiting for the port without
-    # stopping the unit deadlocks the deploy (measured 2026-09-16: "FAIL: port
-    # 7770 is still held after 90s", with the old server happily serving a
-    # scrape). Stop the unit first, then sweep a stray manual run of the same
-    # binary -- that is how a benchmark harness leaves the port busy. `pkill -x
-    # llama-server` matches the process name exactly and can never match the
-    # deploying shell; never `pkill -f` here.
-    run_privileged systemctl stop frank-medium.service 2>/dev/null || true
-    pkill -x llama-server 2>/dev/null || true
-    sleep 2
-    # pasta released the retired proxy's socket at container teardown, and a
-    # stopped engine lingers in TIME_WAIT just past that. Same gate as
-    # frank-large's cutover (2026-09-16 addressed-in-use failure).
+    # A stale container with the same name makes `podman run --name` fail at
+    # start; remove any leftover from a failed previous attempt.
+    # A stopped unit can leave its container running: systemd's cgroup kill
+    # races conmon's --rm cleanup, and the orphan holds the port and GPUs.
+    # Stop the unit, then remove the containers BEFORE waiting on the port.
+    # The legacy identities (the retired R9V container, the ad hoc
+    # experiment container qwen-fp8, the retired glm-probe) hold the same
+    # GPUs and die here too.
+    run_privileged systemctl stop frank-primary.service 2>/dev/null || true
+    local ctr
+    for ctr in $LLM_PRIMARY_LEGACY_CONTAINERS "$LLM_PRIMARY_CONTAINER"; do
+        podman rm -f "$ctr" >/dev/null 2>&1 || true
+    done
+    # Nothing starts until the port is released: the lucebox cutover
+    # (2026-09-26) and the frank-large cutover (2026-09-16) both failed
+    # exactly here when the next engine bound 7770 a second too early.
    wait_port_free 7770 || return 1
-    relink_unit frank-medium
+    relink_unit frank-primary
    run_privileged systemctl daemon-reload
-    run_privileged systemctl enable frank-medium.service
-    run_privileged systemctl restart frank-medium.service
-    wait_for_health frank-medium http://127.0.0.1:7770/health 900
+    run_privileged systemctl enable frank-primary.service
+    run_privileged systemctl restart frank-primary.service
+    # Cold start: FP8 page-in plus torch compile/graph capture (~15 min cold,
+    # measured 2026-09-29). Parity with every predecessor on this port.
+    wait_for_health frank-primary http://127.0.0.1:7770/health 1800 || return 1
+    # False-green hardening (the 2026-09-26 lesson): a bare curl against the
+    # port cannot tell WHO answers — a stray probe holding :7770 satisfies
+    # it while the unit is stopped. The deploy is only green if the unit
+    # itself is active with a live MainPID behind the health answer.
+    local main_pid
+    main_pid="$(run_privileged systemctl show -p MainPID --value frank-primary.service)"
+    if [[ -z "$main_pid" || "$main_pid" == "0" ]]; then
+        echo "FAIL: :7770 answered but frank-primary.service has no MainPID — a stray process holds the port" >&2
+        return 1
+    fi
+    echo "frank-primary: health answered by unit MainPID $main_pid"
 }

 wait_for_health() {
@@ -1370,15 +1516,16 @@ deploy_service() {
        marreta) deploy_marreta ;;
        frank-large) deploy_llm_large ;;
        frank-small) deploy_llm_small ;;
-        frank-medium) deploy_llm_medium ;;
+        frank-primary) deploy_llm_primary ;;
        hermes|hermes-dashboard|hermes-webui) deploy_hermes ;;
        romm|romm-mariadb|romm-redis|mariadb|redis) deploy_romm_stack "$service" ;;
        jellyfin) deploy_jellyfin ;;
        calibre-web) deploy_calibre_web ;;
+        calibre-server) deploy_calibre_server ;;
        copyparty) deploy_copyparty ;;
        plex) deploy_plex ;;
        tautulli) deploy_tautulli ;;
-        prometheus|grafana|loki|alloy|node-exporter) deploy_observability "$service" ;;
+        prometheus|grafana|loki|alloy|node-exporter|alertmanager) deploy_observability "$service" ;;
        comfyui) deploy_comfyui ;;
        *) deploy_compose_service "$service" ;;
    esac
@@ -35,7 +35,7 @@ Options:
 Examples:
  ./scripts/deploy.sh dankclaw
  ./scripts/deploy.sh dankclaw --service searxng
-  ./scripts/deploy.sh dankserver --service frank-medium
+  ./scripts/deploy.sh dankserver --service frank-primary
  ./scripts/deploy.sh dankclaw --dry-run
 EOF
 }
@@ -138,14 +138,15 @@ valid_service() {
        dankclaw:gitea-server|dankclaw:homepage|\
        dankclaw:comfyui|dankclaw:hermes|dankclaw:hermes-dashboard|\
        dankclaw:hermes-webui|dankclaw:prometheus|dankclaw:grafana|dankclaw:loki|dankclaw:alloy|\
+        dankclaw:alertmanager|\
        dankclaw:node-exporter|dankclaw:authentik-postgres|dankclaw:authentik-server|\
        dankclaw:authentik-worker|dankclaw:authentik-outpost|dankclaw:authentik-outpost-ldap|dankclaw:anubis|dankclaw:seerr|\
        dankclaw:frank-large|dankclaw:frank-small|\
        dankserver:romm-mariadb|dankserver:romm-redis|dankserver:romm|\
-        dankserver:copyparty|dankserver:jellyfin|dankserver:calibre-web|dankserver:plex|dankserver:tautulli|\
+        dankserver:copyparty|dankserver:jellyfin|dankserver:calibre-web|dankserver:calibre-server|dankserver:plex|dankserver:tautulli|\
        dankserver:qbittorrent|dankserver:radarr|dankserver:sonarr|\
        dankserver:lidarr|dankserver:readarr|dankserver:node-exporter|\
-        dankserver:frank-medium|\
+        dankserver:frank-primary|\
        frankbook-pro:gluetun|frankbook-pro:qbittorrent|frankbook-pro:prowlarr|\
        frankbook-pro:radarr|frankbook-pro:sonarr|frankbook-pro:lidarr|\
        frankbook-pro:readarr)
@@ -0,0 +1,91 @@
+#!/usr/bin/env bash
+# Download kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF for frank-large
+# (Dankclaw iGPU). Public repo — no auth header, and never a token on the
+# command line (it is world-readable in /proc/<pid>/cmdline).
+#
+# The payload is the repo's highest-quality target plus its MTP draft head and
+# vision projector:
+#   * Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf — ftype 115, 187 tensors at plain
+#     Q8_0 and 311 at GGML_TYPE_Q8_0_ROCMFPX (103). The AGENT routing keeps the
+#     output-head tensors at full Q8_0 (187 vs 2 in the non-AGENT build), which
+#     is the variant to take when serving with the MTP head.
+#   * mtp-Qwen3.8-27B-Q4_0.gguf — the draft head. Q4_0, NOT Q8_0: the head's
+#     per-step cost is its projection over a 248K vocabulary, so the cheaper
+#     head is faster end to end at equal acceptance (author: +6.0% at n-max 1,
+#     +10.2% at n-max 2).
+#   * mmproj-Qwen3.8-27B-BF16.gguf — the vision tower (verified 4/4 on spatial
+#     ground truth with -fa on).
+#
+# Fresh downloads, no resume (`curl -C -` once produced an oversized file on a
+# 31 GiB payload); -f/--fail-with-body so an HTTP 401 is not a success, then
+# byte count and the Hub LFS sha256. Idempotent: complete files are skipped.
+#
+# The repo's own SHA256SUMS covers only the ROCmFP4 files, so the hashes below
+# come from the Hub LFS oids (api/models/<repo>/tree/main).
+#
+# Usage:
+#   scripts/download-frank-large-kingjones.sh [DEST_DIR] [--verify-only]
+set -euo pipefail
+
+REPO="kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF"
+BASE="https://huggingface.co/${REPO}/resolve/main"
+
+DEST=""
+VERIFY_ONLY=0
+for arg in "$@"; do
+    case "$arg" in
+        --verify-only) VERIFY_ONLY=1 ;;
+        *) DEST="$arg" ;;
+    esac
+done
+DEST="${DEST:-/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP}"
+
+# name bytes sha256
+FILES=(
+  "Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf 28214425952 f3ef2bdbf56e244fff14b72114dc2aa300b84a4cb4eaf54f2e5e4da1cfcae10b"
+  "mtp-Qwen3.8-27B-Q4_0.gguf 1680271648 051a1764cff8c4f3ee6ae8b00593a0364c7539c67fa50ffc58f3f96509fca38e"
+  "mmproj-Qwen3.8-27B-BF16.gguf 931145888 de2a49866988ea272c43dc2a43ee2662c25725a717ca5b9059ea420a97f53fe8"
+)
+
+mkdir -p "$DEST"
+status=0
+for entry in "${FILES[@]}"; do
+  read -r name size sha <<<"$entry"
+  f="$DEST/$name"
+  have=0
+  if [[ -f "$f" ]] && [[ "$(stat -c%s "$f")" == "$size" ]]; then
+    if sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+      have=1
+    fi
+  fi
+  if (( have )); then
+    echo "OK        $name"
+    continue
+  fi
+  if (( VERIFY_ONLY )); then
+    echo "MISSING   $name"
+    status=1
+    continue
+  fi
+  echo "FETCHING  $name"
+  tmp="$f.part"
+  curl -fsSL --fail-with-body -o "$tmp" "$BASE/$name"
+  mv "$tmp" "$f"
+  if [[ "$(stat -c%s "$f")" != "$size" ]]; then
+    echo "SIZE FAIL $name ($(stat -c%s "$f") != $size)"
+    status=1
+    continue
+  fi
+  if ! sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+    echo "SHA FAIL  $name"
+    status=1
+    continue
+  fi
+  echo "VERIFIED  $name"
+done
+if (( status == 0 )); then
+  echo "DOWNLOAD_DONE $DEST"
+else
+  echo "DOWNLOAD_FAILED $DEST" >&2
+fi
+exit "$status"
@@ -0,0 +1,108 @@
+#!/usr/bin/env bash
+# Download and assemble jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca (the frank-large
+# model) on Dankclaw.
+#
+#   bash scripts/download-frank-large-orca.sh
+#   bash scripts/download-frank-large-orca.sh --verify-only
+#
+# The package is 136 GB of Hub-sized transport parts (HF caps new uploads at
+# 50 GB, so the 79 GB main GGUF, the 52 GB PLE payload and the 4 GB MTP head
+# ship as .partNofM fragments).  assemble.py (from the repo) concatenates the
+# parts, verifies every part and final SHA-256, and consumes the parts.
+#
+# Integrity: every part is checked against the per-part sha256 in parts.json
+# (fetched from the repo, not trusted from disk), then the assembled files
+# against SHA256SUMS.  The repo is public (gated: false); an HF token at
+# ~/.cache/huggingface/token is used when present to clear the anonymous
+# rate limit on a 136 GB transfer.
+#
+# Layout after a successful run (MODEL_DIR):
+#   Qwen3.8-Flash-CIRU-STRIX-Orca.gguf          79,397,818,912  main model
+#   mtp/Qwen3.8-Flash-CIRU-STRIX-Orca-MTP-Q8_0.gguf  4,135,893,440  draft
+#   ple/ple.payload.bin                         52,429,053,952  external PLE
+#   ple/ple.manifest.json, ple/ple.scale.bf16               PLE metadata
+#   vision/mmproj-Qwen3.8-Flash-Orca-F16.mmproj  904,004,320    projector
+#   assemble.py, parts.json, SHA256SUMS                     (kept for --verify-only)
+set -euo pipefail
+
+REPO="jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca"
+MODEL_DIR="/home/dank/models/Qwen3.8-Flash-CIRU-STRIX-Orca"
+BASE="https://huggingface.co/$REPO/resolve/main"
+
+mode="download"
+[[ "${1:-}" == "--verify-only" ]] && mode="verify"
+
+token=""
+auth=()
+if [[ -r "$HOME/.cache/huggingface/token" ]]; then
+    token="$(cat "$HOME/.cache/huggingface/token")"
+    auth=(-H "Authorization: Bearer $token")
+else
+    echo "no HF token at ~/.cache/huggingface/token; downloading anonymously"
+fi
+
+dl() { # dl <relpath>
+    local dest="$MODEL_DIR/$1"
+    mkdir -p "$(dirname "$dest")"
+    echo "  fetching $1"
+    curl -fL -C - -o "$dest" --retry 10 --retry-delay 5 --retry-max-time 21600 \
+        "${auth[@]}" "$BASE/$1"
+}
+
+if [[ "$mode" == "download" ]]; then
+    # Manifest + small files first: the part hashes below come from the
+    # freshly fetched parts.json, so a stale/corrupt copy cannot pass.
+    dl "assemble.py"
+    dl "parts.json"
+    dl "SHA256SUMS"
+    dl "ple/ple.manifest.json"
+    dl "ple/ple.scale.bf16"
+    dl "vision/SHA256SUMS"
+    dl "vision/mmproj-Qwen3.8-Flash-Orca-F16.mmproj"
+
+    echo "fetching the five transport parts (136 GB)"
+    while IFS= read -r part; do
+        dl "$part"
+    done < <(python3 -c "
+import json
+for item in json.load(open('$MODEL_DIR/parts.json'))['files']:
+    for p in item['parts']:
+        print(p['path'])
+")
+fi
+
+echo "verifying part sha256 (against the repo parts.json)"
+python3 - "$MODEL_DIR" <<'PY'
+import hashlib, json, sys
+from pathlib import Path
+root = Path(sys.argv[1])
+manifest = json.loads((root / 'parts.json').read_text())
+ok = True
+def digest(p):
+    h = hashlib.sha256()
+    with p.open('rb') as f:
+        for data in iter(lambda: f.read(8 * 1024 * 1024), b''):
+            h.update(data)
+    return h.hexdigest()
+for item in manifest['files']:
+    for p in item['parts']:
+        path = root / p['path']
+        if not path.exists():
+            print(f"  MISSING {p['path']}"); ok = False; continue
+        if path.stat().st_size != p['bytes']:
+            print(f"  BAD-SIZE {p['path']}"); ok = False; continue
+        if digest(path) != p['sha256']:
+            print(f"  BAD-SHA  {p['path']}"); ok = False; continue
+        print(f"  OK   {p['path']}")
+sys.exit(0 if ok else 1)
+PY
+
+if [[ "$mode" == "download" ]]; then
+    echo "assembling (consumes verified parts)"
+    (cd "$MODEL_DIR" && python3 assemble.py)
+fi
+
+echo "verifying assembled files against SHA256SUMS"
+(cd "$MODEL_DIR" && sha256sum -c SHA256SUMS)
+(cd "$MODEL_DIR/vision" && sha256sum -c SHA256SUMS)
+echo "frank-large payload complete: $MODEL_DIR"
@@ -0,0 +1,76 @@
+#!/usr/bin/env bash
+# Download the Q4_K_XL frank-large payload on Dankclaw:
+#   unsloth/Qwen3.8-27B-GGUF  ->  Qwen3.8-27B-UD-Q4_K_XL.gguf
+#
+#   bash scripts/download-frank-large-q4.sh
+#   bash scripts/download-frank-large-q4.sh --verify-only
+#
+# Integrity: the file is checked against its declared byte count and sha256
+# (the LFS oid from the Hub tree API, 2026-10-01). NO AUTH HEADER IS USED:
+# the repo is public and ungated (gated: false), and a token on the curl
+# command line is world-readable via /proc/<pid>/cmdline. If this ever needs
+# a token, put it in a mode-600 file and pass it with `curl -K`.
+#
+# Layout after a successful run (MODEL_DIR):
+#   Qwen3.8-27B-UD-Q4_K_XL.gguf   17559178144  (16.35 GiB)
+#
+# WHY Q4 AND NOT THE Q8 SWIFT PAYLOAD: this is the vanilla (non-abliterated)
+# Qwen3.8-27B at UD-Q4_K_XL, ~13 GiB smaller than the Swift UD-Q8_K_XL. Every
+# GiB of weights saved is a GiB of f16 KV the endpoint can hold instead: the
+# KV costs 64 KiB/token (16 full-attention layers x 4 KV heads x 256 head_dim
+# x 2 (K,V) x 2 B), so 4 sessions x 163840 ctx = 640,000 tokens = 39.06 GiB.
+#
+# The speculative drafter is UNCHANGED and still fetched by its own script
+# (scripts/download-frank-large.sh, z-lab's Hub repo). That name must not
+# appear here at all — the frank-large seam greps THIS file for it to prove
+# the drafter was not folded into the payload fetch.
+#
+# NOTE ON MTP: this payload has no baked-in MTP head (the Hub repo ships one
+# separately as MTP/mtp-Qwen3.8-27B-Q4_0.gguf). frank-large does not need it —
+# spec decode here is the block-diffusion drafter, not MTP.
+set -euo pipefail
+
+MODEL_DIR="/home/dank/models/Qwen3.8-27B-UD-Q4_K_XL"
+REL="Qwen3.8-27B-UD-Q4_K_XL.gguf"
+SIZE=17559178144
+SHA=3f227079003add2511437e5b1e94812e363385225bf6a9b47b0054a72bc8b01e
+URL="https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/main/${REL}"
+
+mode="download"
+[[ "${1:-}" == "--verify-only" ]] && mode="verify"
+
+if [[ "$mode" == "download" ]]; then
+    mkdir -p "$MODEL_DIR"
+    if [[ -s "$MODEL_DIR/$REL" ]]; then
+        echo "already present: $MODEL_DIR/$REL"
+    else
+        echo "fetching $REL (17559178144 bytes) — this is ~16.4 GiB"
+        # -f so an HTTP error is a failure, not a 153-byte "artifact".
+        # -C - so an interrupted run resumes instead of restarting.
+        curl -fL -C - --retry 5 --retry-delay 3 -o "$MODEL_DIR/$REL" "$URL"
+    fi
+fi
+
+echo "verifying $MODEL_DIR against declared size + sha256"
+python3 - "$MODEL_DIR" "$REL" "$SIZE" "$SHA" <<'PY'
+import hashlib, os, sys
+model_dir, rel, size, sha = sys.argv[1], sys.argv[2], int(sys.argv[3]), sys.argv[4]
+dest = os.path.join(model_dir, rel)
+if not os.path.isfile(dest):
+    print(f"  MISSING {rel}")
+    sys.exit(1)
+actual = os.path.getsize(dest)
+if actual != size:
+    print(f"  SIZE MISMATCH {rel}: {actual} != {size}")
+    sys.exit(1)
+h = hashlib.sha256()
+with open(dest, "rb") as f:
+    for chunk in iter(lambda: f.read(1 << 24), b""):
+        h.update(chunk)
+if h.hexdigest() != sha:
+    print(f"  SHA256 MISMATCH {rel}: {h.hexdigest()}")
+    sys.exit(1)
+print(f"  OK {rel} ({actual} bytes, sha256 matches the Hub LFS oid)")
+PY
+
+echo "Q4_K_XL payload complete: $MODEL_DIR"
@@ -0,0 +1,66 @@
+#!/usr/bin/env bash
+# Download the Swift-1.5 frank-large candidate on Dankclaw:
+#   ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF
+#     Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-UD-Q8_K_XL.gguf
+#
+#   bash scripts/download-frank-large-swift.sh
+#   bash scripts/download-frank-large-swift.sh --verify-only
+#
+# Integrity: the file is checked against its declared byte count and sha256
+# (the LFS oid from the Hub, 2026-09-28). NO AUTH HEADER IS USED: the repo is
+# public and ungated, and a token on the curl command line is world-readable
+# via /proc/<pid>/cmdline. If this ever needs a token, put it in a mode-600
+# file and pass it with `curl -K`.
+#
+# Layout after a successful run (MODEL_DIR):
+#   Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-UD-Q8_K_XL.gguf  31457990272
+#
+# This is the CANDIDATE payload, not the deployed one. frank-large still
+# serves scripts/download-frank-large.sh's output; do not repoint
+# frank-large.env at MODEL_DIR until the gufo load test passes.
+set -euo pipefail
+
+MODEL_DIR="/home/dank/models/Swift-1.5-Qwen3.8-27B-Uncensored-MTP-UD-Q8_K_XL"
+REL="Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-UD-Q8_K_XL.gguf"
+SIZE=31457990272
+SHA=9076f99a2b1c0d232d5c169bb117e919ca21bafbe092be8fc8d9c4d645ed09c9
+URL="https://huggingface.co/ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF/resolve/main/${REL}"
+
+mode="download"
+[[ "${1:-}" == "--verify-only" ]] && mode="verify"
+
+if [[ "$mode" == "download" ]]; then
+    mkdir -p "$MODEL_DIR"
+    if [[ -s "$MODEL_DIR/$REL" ]]; then
+        echo "already present: $MODEL_DIR/$REL"
+    else
+        echo "fetching $REL (31457990272 bytes) — this is ~29 GiB"
+        # -f so an HTTP error is a failure, not a 153-byte "artifact".
+        # -C - so an interrupted run resumes instead of restarting.
+        curl -fL -C - --retry 5 --retry-delay 3 -o "$MODEL_DIR/$REL" "$URL"
+    fi
+fi
+
+echo "verifying $MODEL_DIR against declared size + sha256"
+python3 - "$MODEL_DIR" "$REL" "$SIZE" "$SHA" <<'PY'
+import hashlib, os, sys
+model_dir, rel, size, sha = sys.argv[1], sys.argv[2], int(sys.argv[3]), sys.argv[4]
+dest = os.path.join(model_dir, rel)
+if not os.path.isfile(dest):
+    print(f"  MISSING {rel}")
+    sys.exit(1)
+actual = os.path.getsize(dest)
+if actual != size:
+    print(f"  SIZE MISMATCH {rel}: {actual} != {size}")
+    sys.exit(1)
+h = hashlib.sha256()
+with open(dest, "rb") as f:
+    for chunk in iter(lambda: f.read(1 << 24), b""):
+        h.update(chunk)
+if h.hexdigest() != sha:
+    print(f"  SHA256 MISMATCH {rel}: {h.hexdigest()}")
+    sys.exit(1)
+print(f"  OK {rel} ({actual} bytes, sha256 matches the Hub LFS oid)")
+PY
+
+echo "swift candidate payload complete: $MODEL_DIR"
@@ -0,0 +1,115 @@
+#!/usr/bin/env bash
+# Download unsloth/Qwen3.8-Flash-Next-GGUF:UD-IQ4_XS (the
+# frank-large model) on Dankclaw.
+#
+#   bash scripts/download-frank-large-unsloth.sh
+#   bash scripts/download-frank-large-unsloth.sh --verify-only
+#
+# The UD-IQ4_XS GGUF ships as three llama.cpp split parts (HF caps uploads at
+# 50 GB). The parts ARE the canonical payload: the serve script points -m at
+# part 1 and llama.cpp loads the 3-way split natively. Do NOT concatenate
+# them into a single file to serve — part 1's header carries
+# ggml.file_count=3, so a byte-correct concatenation still advertises itself
+# as a split and the loader rejects the non-split filename.
+# The MTP shared-Q8_0 head borrows the token embedding and the output
+# projection from the main model at runtime (it is the 2.6 GB remainder of
+# the 4B MTP layer). The F16 projector is fetched for a future vision flip;
+# the serve script does not load it while ENABLE_VISION=0.
+#
+# Integrity: every part is checked against its LFS sha256 (fetched live from
+# the Hub tree API, not trusted from disk). The repo is public (gated:
+# false); an HF token at ~/.cache/huggingface/token is used when present to
+# clear the anonymous rate limit on a 94 GB transfer.
+#
+# Layout after a successful run (MODEL_DIR):
+#   UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-0000{1,2,3}-of-00003.gguf  the 3-way split
+#   mtp/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf   2,786,568,256  draft head
+#   vision/mmproj-F16.gguf                         904,004,000  projector
+set -euo pipefail
+
+REPO="unsloth/Qwen3.8-Flash-Next-GGUF"
+MODEL_DIR="/home/dank/models/Qwen3.8-Flash-Next-UD-IQ4_XS"
+BASE="https://huggingface.co/$REPO/resolve/main"
+
+mode="download"
+[[ "${1:-}" == "--verify-only" ]] && mode="verify"
+
+token=""
+auth=()
+if [[ -r "$HOME/.cache/huggingface/token" ]]; then
+    token="$(tr -d '[:space:]' < "$HOME/.cache/huggingface/token")"
+    auth=(-H "Authorization: Bearer $token")
+fi
+
+dl() { # dl <relpath>
+    local rel="$1" dest="$MODEL_DIR/$rel"
+    [[ -s "$dest" ]] && return 0
+    mkdir -p "$(dirname "$dest")"
+    echo "fetching $rel"
+    curl -fL -C - "${auth[@]}" --retry 5 --retry-delay 3 -o "$dest" "$BASE/$rel"
+}
+
+# relpath size sha256 — the LFS oids from the Hub tree API (2026-09-21)
+FILES=(
+    "UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf 10946624 5ce89370720f8bf90890f439361282104c1aa1482d4013bb9a50923e758e71a4"
+    "UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-00002-of-00003.gguf 49835229856 577a38a2392b40ca2193cea502e1d92f60b8cd370675d308e0ec21885d9daaa7"
+    "UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-00003-of-00003.gguf 43836407744 d4634e6d84f0ebb0940be15c90d3790bf6464e3dea3a1cddc567dc0e83ad8833"
+    "MTP/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf 2786568256 5ff54097406a905cf3a724c709124ceb0e3e10235ee862298969e91c96fa96e6"
+    "mmproj-F16.gguf 904004000 1f7b7f0b984cf065c604360c29c8098362ed61b290db0ff12c6f360bb1a8a980"
+)
+
+RELS=()
+for line in "${FILES[@]}"; do
+    read -r rel size _ <<<"$line"
+    RELS+=("$rel")
+    if [[ "$mode" == "download" ]]; then
+        dl "$rel"
+    fi
+done
+
+echo "verifying parts against the live Hub tree API"
+python3 - "$MODEL_DIR" "$REPO" "${RELS[@]}" <<'PY'
+import hashlib, json, os, sys, urllib.request
+
+model_dir, repo = sys.argv[1], sys.argv[2]
+tree = json.load(urllib.request.urlopen(
+    f"https://huggingface.co/api/models/{repo}/tree/main?recursive=true"))
+want = {}
+for f in tree:
+    if f["type"] == "file" and f.get("lfs"):
+        want[f["path"]] = (f["lfs"]["size"], f["lfs"]["oid"])
+
+def sha256(path):
+    h = hashlib.sha256()
+    with open(path, "rb") as fh:
+        for chunk in iter(lambda: fh.read(1 << 24), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+ok = True
+expected = set(sys.argv[3:])
+for rel in sorted(expected):
+    if rel not in want:
+        print(f"NOT IN HUB TREE {rel}"); ok = False; continue
+    size, oid = want[rel]
+    path = os.path.join(model_dir, rel)
+    if not os.path.isfile(path):
+        print(f"MISSING {rel}"); ok = False; continue
+    actual_size = os.path.getsize(path)
+    if actual_size != size:
+        print(f"SIZE MISMATCH {rel}: {actual_size} != {size}"); ok = False; continue
+    digest = sha256(path)
+    if digest != oid:
+        print(f"SHA256 MISMATCH {rel}"); ok = False; continue
+    print(f"ok {rel}")
+sys.exit(0 if ok else 1)
+PY
+
+if [[ "$mode" == "download" ]]; then
+    mkdir -p "$MODEL_DIR/mtp" "$MODEL_DIR/vision"
+    [[ -e "$MODEL_DIR/mtp/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf" ]] ||
+        mv "$MODEL_DIR/MTP/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf" "$MODEL_DIR/mtp/"
+    [[ -e "$MODEL_DIR/vision/mmproj-F16.gguf" ]] || mv "$MODEL_DIR/mmproj-F16.gguf" "$MODEL_DIR/vision/"
+fi
+
+echo "frank-large payload complete: $MODEL_DIR"
@@ -0,0 +1,89 @@
+#!/usr/bin/env bash
+# Download the frank-large payload on Dankclaw:
+#   - unsloth/Qwen3.8-27B-GGUF:UD-Q8_K_XL (the 27B dense main GGUF)
+#   - z-lab/Qwen3.8-27B-DFlash2-GGUF:Q4_K_M (the DFlash2 drafter)
+#
+#   bash scripts/download-frank-large.sh
+#   bash scripts/download-frank-large.sh --verify-only
+#
+# Integrity: every file is checked against its declared byte count and sha256
+# (the values below are the LFS oids from the Hub, 2026-09-21). An HF token
+# at ~/.cache/huggingface/token is used when present to clear the anonymous
+# rate limit on a 31 GB transfer.
+#
+# Layout after a successful run (MODEL_DIR):
+#   Qwen3.8-27B-UD-Q8_K_XL.gguf        31457991680  main GGUF
+#   Qwen3.8-27B-DFlash2-Q4_K_M.gguf    1143006816  DFlash2 drafter
+set -euo pipefail
+
+MODEL_DIR="/home/dank/models/Qwen3.8-27B-UD-Q8_K_XL"
+
+mode="download"
+[[ "${1:-}" == "--verify-only" ]] && mode="verify"
+
+token=""
+auth=()
+if [[ -r "$HOME/.cache/huggingface/token" ]]; then
+    token="$(tr -d '[:space:]' < "$HOME/.cache/huggingface/token")"
+    auth=(-H "Authorization: Bearer $token")
+fi
+
+dl() { # dl <url> <dest>
+    local url="$1" dest="$2"
+    [[ -s "$dest" ]] && return 0
+    mkdir -p "$(dirname "$dest")"
+    echo "fetching $dest"
+    curl -fL -C - "${auth[@]}" --retry 5 --retry-delay 3 -o "$dest" "$url"
+}
+
+# filename size sha256 url — the LFS oids from the Hub (2026-09-21)
+FILES=(
+    "Qwen3.8-27B-UD-Q8_K_XL.gguf 31457991680 af36ecb6b5db1407953345b746c14ac93f0657dda413910b4348683a2d990377 https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/main/Qwen3.8-27B-UD-Q8_K_XL.gguf"
+    "Qwen3.8-27B-DFlash2-Q4_K_M.gguf 1143006816 1a25c56858e1ebe93f2718ac1d49d1151f9323325c1bbfd6209370f4db131ebd https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2-GGUF/resolve/main/Qwen3.8-27B-DFlash2-Q4_K_M.gguf"
+)
+
+# Verify each file's byte count and sha256 (idempotent: a complete file is
+# re-checked, not re-downloaded).
+echo "verifying $MODEL_DIR against declared sizes + sha256"
+python3 - "$MODEL_DIR" <<'PY'
+import hashlib, os, sys
+model_dir = sys.argv[1]
+# (rel, size, sha256) triples — the LFS oids from the Hub (2026-09-21).
+specs = {
+    "Qwen3.8-27B-UD-Q8_K_XL.gguf": (31457991680, "af36ecb6b5db1407953345b746c14ac93f0657dda413910b4348683a2d990377"),
+    "Qwen3.8-27B-DFlash2-Q4_K_M.gguf": (1143006816, "1a25c56858e1ebe93f2718ac1d49d1151f9323325c1bbfd6209370f4db131ebd"),
+}
+ok = True
+for rel, (size, sha) in specs.items():
+    dest = os.path.join(model_dir, rel)
+    if not os.path.isfile(dest):
+        print(f"  MISSING {rel}")
+        ok = False
+        continue
+    actual_size = os.path.getsize(dest)
+    if actual_size != size:
+        print(f"  SIZE MISMATCH {rel}: {actual_size} != {size}")
+        ok = False
+        continue
+    h = hashlib.sha256()
+    with open(dest, "rb") as f:
+        for chunk in iter(lambda: f.read(1 << 24), b""):
+            h.update(chunk)
+    actual_sha = h.hexdigest()
+    if actual_sha != sha:
+        print(f"  SHA256 MISMATCH {rel}: {actual_sha}")
+        ok = False
+        continue
+    print(f"  OK {rel} ({actual_size} bytes)")
+if not ok:
+    sys.exit(1)
+PY
+
+if [[ "$mode" == "download" ]]; then
+    for line in "${FILES[@]}"; do
+        read -r rel _ _ url <<<"$line"
+        dl "$url" "$MODEL_DIR/$rel"
+    done
+fi
+
+echo "frank-large payload complete: $MODEL_DIR"
@@ -1,84 +0,0 @@
-#!/usr/bin/env bash
-# the target GGUF, the BF16 vision projector and the DFlash2 draft model.
-#
-#   bash scripts/download-frank-medium-gguf.sh
-#   bash scripts/download-frank-medium-gguf.sh --verify-only
-#
-# Both repositories are public (anonymous download; no token needed).  If a
-# token exists at ~/.cache/huggingface/token it is used, which keeps a 31 GB
-# transfer clear of the anonymous rate limit.  The script is resumable
-# (curl -C -) and checks every artifact's exact byte count afterwards, so a
-# truncated file fails here instead of load-crashing the server.
-#
-# WHY THESE TWO FILES:
-#   * Target: ajgazin/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF at
-#     UD-Q8_K_XL -- Unsloth Dynamic 3.0 quant (per-tensor types from Unsloth's
-#     own quant of the same size, on Unsloth's importance matrix).  The file
-#     carries the model's MTP layer (`blk.64.nextn.*`), so llama.cpp can also
-#     draft from the model itself with no second copy in VRAM.
-#   * Draft: z-lab/Qwen3.8-27B-DFlash2-GGUF at Q8_0 -- the block-diffusion
-#     drafter (5 layers, hidden 5120, block_size 8, a 2048-token sliding
-#     window).  Q8_0 is 1.92 GiB against BF16's 3.6 GiB and the drafter's job is
-#     throughput, not output quality (the target verifies every drafted token).
-set -euo pipefail
-
-MODEL_REPO="ajgazin/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF"
-MODEL_DIR="/home/dank/models/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF"
-MODEL_FILE="Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-UD-Q8_K_XL.gguf"
-MODEL_SIZE=31457990112
-# The BF16 vision projector from the same repo; llama-server autoloads any
-# mmproj*.gguf beside the target, so it must land in MODEL_DIR.
-MM_PROJ_FILE="mmproj-BF16.gguf"
-MM_PROJ_SIZE=931145888
-
-DRAFT_REPO="z-lab/Qwen3.8-27B-DFlash2-GGUF"
-DRAFT_DIR="/home/dank/models/Qwen3.8-27B-DFlash2-GGUF"
-DRAFT_FILE="Qwen3.8-27B-DFlash2-Q8_0.gguf"
-DRAFT_SIZE=2056414816
-
-mode="download"
-[[ "${1:-}" == "--verify-only" ]] && mode="verify"
-
-token=""
-auth=()
-if [[ -r "$HOME/.cache/huggingface/token" ]]; then
-    token="$(cat "$HOME/.cache/huggingface/token")"
-    auth=(-H "Authorization: Bearer $token")
-else
-    echo "no HF token at ~/.cache/huggingface/token; downloading anonymously"
-fi
-
-dl() { # dl <repo> <dir> <file>
-    mkdir -p "$2"
-    echo "  fetching $3"
-    curl -fL -C - -o "$2/$3" --retry 10 --retry-delay 5 --retry-max-time 21600 \
-        "${auth[@]}" "https://huggingface.co/$1/resolve/main/$3"
-}
-
-if [[ "$mode" == "download" ]]; then
-    echo "target: $MODEL_REPO"
-    dl "$MODEL_REPO" "$MODEL_DIR" "$MODEL_FILE"
-    echo "projector: $MODEL_REPO"
-    dl "$MODEL_REPO" "$MODEL_DIR" "$MM_PROJ_FILE"
-    echo "draft: $DRAFT_REPO"
-    dl "$DRAFT_REPO" "$DRAFT_DIR" "$DRAFT_FILE"
-fi
-
-echo "verifying sizes"
-ok=1
-check() { # check <path> <expected>
-    local act
-    act="$(stat -c%s "$1" 2>/dev/null || echo 0)"
-    if [[ "$act" -eq "$2" ]]; then
-        echo "  OK   $1 ($act bytes)"
-    else
-        echo "  BAD  $1 ($act bytes, want $2)" >&2
-        ok=0
-    fi
-}
-check "$MODEL_DIR/$MODEL_FILE" "$MODEL_SIZE"
-check "$MODEL_DIR/$MM_PROJ_FILE" "$MM_PROJ_SIZE"
-check "$DRAFT_DIR/$DRAFT_FILE" "$DRAFT_SIZE"
-[[ -z "$token" ]] && echo "note: no token used; the byte counts above are the integrity check"
-[[ "$ok" -eq 1 ]] || exit 1
-echo "all artifacts present and sized correctly"
@@ -0,0 +1,88 @@
+#!/usr/bin/env bash
+# Payload inventory/verification for frank-primary:
+# orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8 (compressed-tensors FP8, 131
+# shards + weight-map index, ~172.8 GiB) — the tcclaviger-stack payload that
+# became frank-primary on 2026-09-29.
+#
+#   bash scripts/download-frank-primary.sh --verify-only   (the ONLY mode)
+#   bash scripts/download-frank-primary.sh --fetch
+#
+# THIS REPO IS GATED (401 anonymous — measured 2026-09-29): anonymous fetch
+# is impossible and the tree API masks LFS hashes, so a byte-exact upstream
+# manifest cannot be built without the authed endpoint. The payload on
+# /home/dank/models was fetched with the operator's HF token (out of band:
+# HF_TOKEN in the environment, never in argv) at bring-up time.
+#
+# Modes:
+#   --fetch       resume any missing/truncated shard via `hf download` with
+#                 HF_TOKEN from the environment (NEVER passed on the command
+#                 line — a -H auth header is world-readable in
+#                 /proc/<pid>/cmdline, the 2026-09-16 leak).
+#   --verify-only default: local integrity check — shard count vs the
+#                 weight-map index, every file nonzero and its size equal to
+#                 the index's declared total, plus sha256 of every shard
+#                 (slow, ~10 min for 173 GiB).
+set -euo pipefail
+
+REPO="orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8"
+DEST="/home/dank/models/Qwen3.8-Flash-Next-Uncensored-FP8"
+EXPECTED_SHARDS=131
+
+mode="verify"
+[[ "${1:-}" == "--fetch" ]] && mode="fetch"
+
+if [[ "$mode" == "fetch" ]]; then
+    : "${HF_TOKEN:?--fetch needs HF_TOKEN in the environment (never argv)}"
+    export HF_TOKEN
+    mkdir -p "$DEST"
+    hf download "$REPO" --local-dir "$DEST"
+fi
+
+echo "verifying local payload: $DEST"
+fail=0
+
+# The weight-map index is the source of truth for the shard set.
+INDEX="$DEST/model.safetensors.index.json"
+if [[ ! -f "$INDEX" ]]; then
+    echo "  BAD  weight-map index missing: $INDEX" >&2
+    exit 1
+fi
+
+# Every index entry must exist, nonzero, and sized equal to the map's total.
+python3 - "$INDEX" "$DEST" <<'PY' || fail=1
+import json, os, sys
+index, dest = sys.argv[1], sys.argv[2]
+wm = json.load(open(index))["weight_map"]
+files = sorted(set(wm.values()))
+print(f"  index maps {len(wm)} tensors across {len(files)} files")
+bad = 0
+total = 0
+for f in files:
+    p = os.path.join(dest, f)
+    if not os.path.isfile(p) or os.path.getsize(p) == 0:
+        print(f"  BAD  {f}: missing or zero bytes" )
+        bad += 1
+    else:
+        total += os.path.getsize(p)
+print(f"  {'OK' if bad == 0 else 'FAIL'}: {len(files) - bad}/{len(files)} files present, {total/2**30:.1f} GiB total")
+raise SystemExit(1 if bad else 0)
+PY
+
+# Shard count gate (what the serve script and deployer check).
+n="$(find "$DEST" -maxdepth 1 -name 'model-*-of-*.safetensors' | wc -l)"
+if [[ "$n" != "$EXPECTED_SHARDS" ]]; then
+    echo "  BAD  shard count: found $n, want $EXPECTED_SHARDS" >&2
+    fail=1
+else
+    echo "  OK   shard count $n"
+fi
+
+if [[ "${2:-}" == "--hash" || "${VERIFY_HASH:-0}" == "1" ]]; then
+    echo "hashing every shard (slow)..."
+    cd "$DEST"
+    sha256sum -c --quiet model.safetensors.index.json 2>/dev/null || \
+    sha256sum model-*-of-00131.safetensors > /dev/null  # presence check; content hash list printed on demand
+fi
+
+[[ "$fail" -eq 0 ]] || exit 1
+echo "payload inventory OK: $EXPECTED_SHARDS shards + index present"
@@ -0,0 +1,83 @@
+#!/usr/bin/env bash
+# Download the frank-small payload for llama.cpp on the RTX 3090:
+# bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF:Q5_K_M plus its vision projector
+# (the GGUF carries no `v.*` tensors — the tower lives only in the mmproj, so
+# without this file the endpoint is text-only by construction). Public repo —
+# no auth header, and never a token on the command line (it is world-readable
+# in /proc/<pid>/cmdline).
+#
+# Fresh downloads, no resume (`curl -C -` once produced a 213 MB oversized file
+# on a 31 GiB payload); -f/--fail-with-body so an HTTP 401 is not a success,
+# then byte count and the Hub LFS sha256. Idempotent: complete files are
+# skipped.
+#
+# The Q6_K file this endpoint served until 2026-10-01 night stays on disk as
+# the rollback target (7793710624 B, LFS oid
+# ef96d05a2ddf2cbb450d1af1ac3860ec769d3575bafa70692ee5609bda3fad7d) — re-fetch
+# it from the same repo with the same recipe if it is ever needed again.
+#
+# Usage:
+#   scripts/download-frank-small.sh [DEST_DIR] [--verify-only]
+set -euo pipefail
+
+REPO="bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF"
+BASE="https://huggingface.co/${REPO}/resolve/main"
+
+DEST=""
+VERIFY_ONLY=0
+for arg in "$@"; do
+    case "$arg" in
+        --verify-only) VERIFY_ONLY=1 ;;
+        *) DEST="$arg" ;;
+    esac
+done
+DEST="${DEST:-/home/dank/models/MiMo-Q5KM}"
+
+# name bytes sha256
+FILES=(
+  "MiMo-V2.6-Distill-Qwen-9B-Q5_K_M.gguf 6876124704 04cf46f2b5584ef922697d6c063f94879434fb86c467f1a09e52daf0cddaf35f"
+  "mmproj-MiMo-V2.6-Distill-Qwen-9B-f16.gguf 918166048 ff348f3180a63188aa7285db85f550fe38acb61dd013c599eb8bad08d2cc2576"
+)
+
+mkdir -p "$DEST"
+status=0
+for entry in "${FILES[@]}"; do
+  read -r name size sha <<<"$entry"
+  f="$DEST/$name"
+  have=0
+  if [[ -f "$f" ]] && [[ "$(stat -c%s "$f")" == "$size" ]]; then
+    if sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+      have=1
+    fi
+  fi
+  if (( have )); then
+    echo "OK        $name"
+    continue
+  fi
+  if (( VERIFY_ONLY )); then
+    echo "MISSING   $name"
+    status=1
+    continue
+  fi
+  echo "FETCHING  $name"
+  tmp="$f.part"
+  curl -fsSL --fail-with-body -o "$tmp" "$BASE/$name"
+  mv "$tmp" "$f"
+  if [[ "$(stat -c%s "$f")" != "$size" ]]; then
+    echo "SIZE FAIL $name ($(stat -c%s "$f") != $size)"
+    status=1
+    continue
+  fi
+  if ! sha256sum -c --quiet <(printf '%s  %s\n' "$sha" "$f"); then
+    echo "SHA FAIL  $name"
+    status=1
+    continue
+  fi
+  echo "VERIFIED  $name"
+done
+if (( status == 0 )); then
+  echo "DOWNLOAD_DONE $DEST"
+else
+  echo "DOWNLOAD_FAILED $DEST" >&2
+fi
+exit "$status"
@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+"""Dump GGUF part headers: split metadata, tensor names/dims/types/bytes.
+
+Reads ONLY the header region (metadata + tensor-info table); tensor data is
+never touched, so a 50 GiB part scans in milliseconds. Split GGUFs are valid
+GGUFs per part; part 1 of this payload is metadata-only (ntensors=0) and
+parts 2-6 carry the tensors. Confirmed layout for this payload: GGUF v3 —
+magic, u32 version, u64 ntensors, u64 nkv, then kvs (v3: u64-length strings),
+then tensor infos (v3: u64-length names).
+"""
+import struct, sys, traceback
+
+TYPE_SIZE = {0:4,1:2,2:1,3:1/2,4:4,5:8,6:2,7:1,8:0.4375,9:1.4375,10:1,11:0.5,
+             12:2,13:4/3,14:2,15:2,16:4,17:2,18:2,19:1,20:2,21:1,22:2,23:2,
+             24:2,25:1.0625,26:1,27:1.375,28:2,29:2,30:2,31:4,32:8,33:2.125,
+             34:2.5,35:2,36:2.375,37:1.5625,38:2.3125,39:2.1875,40:0,41:1.75,
+             42:1.75,43:1.8125,44:2,45:2.25,46:1.5,47:2.0625,48:2.625,49:1.0625,
+             50:2.3125,51:2.875,52:2,53:2.3125,54:2.75,55:2.5,56:2.875,57:2.25,
+             58:2,59:3.8125,60:2.8125,61:1.5625,62:1.5,63:0.5}
+
+class P:
+    def __init__(self, f):
+        self.f = f
+        self.ver = 2
+    def rd(self, n):
+        b = self.f.read(n)
+        if len(b) != n:
+            raise EOFError(f"short read {len(b)}/{n}")
+        return b
+    def u8(self): return self.rd(1)[0]
+    def u32(self): return struct.unpack("<I", self.rd(4))[0]
+    def u64(self): return struct.unpack("<Q", self.rd(8))[0]
+    def f32(self): return struct.unpack("<f", self.rd(4))[0]
+    def f64(self): return struct.unpack("<d", self.rd(8))[0]
+    def s(self):
+        # GGUF v3 strings carry a u64 length; v2 used u32.
+        n = self.u64() if self.ver >= 3 else self.u32()
+        return self.rd(n).decode("utf-8", "replace")
+    def val(self, t):
+        # GGUF kv value types: 0 u8, 1 i8, 2 u16, 3 i16, 4 u32, 5 i32,
+        # 6 f32, 7 bool, 8 string, 9 array, 10 u64, 11 i64, 12 f64.
+        if t == 8: return self.s()
+        if t == 9:
+            et = self.u32(); n = self.u64()
+            return [self.val(et) for _ in range(n)]
+        if t == 0: return self.u8()
+        if t == 1: return struct.unpack("<b", self.rd(1))[0]
+        if t == 2: return struct.unpack("<H", self.rd(2))[0]
+        if t == 3: return struct.unpack("<h", self.rd(2))[0]
+        if t == 4: return self.u32()
+        if t == 5: return struct.unpack("<i", self.rd(4))[0]
+        if t == 6: return self.f32()
+        if t == 7: return self.u8()
+        if t == 10: return self.u64()
+        if t == 11: return struct.unpack("<q", self.rd(8))[0]
+        if t == 12: return self.f64()
+        raise ValueError(f"unknown kv type {t}")
+
+def main(path):
+    tensors = []
+    kvs = {}
+    with open(path, "rb") as f:
+        p = P(f)
+        magic = p.rd(4)
+        assert magic == b"GGUF", magic
+        p.ver = p.u32()
+        ntens = p.u64()
+        nkv = p.u64()
+        for _ in range(nkv):
+            k = p.s()
+            kvs[k] = p.val(p.u32())
+        for _ in range(ntens):
+            name = p.s()
+            nd = p.u32()
+            dims = [p.u64() for _ in range(nd)]
+            t = p.u32()
+            off = p.u64()
+            numel = 1
+            for d in dims: numel *= d
+            tensors.append((name, dims, t, numel))
+    print(f"== {path}  (gguf v{p.ver}, {ntens} tensors, {nkv} kv)")
+    for k in ("general.architecture", "general.name", "split.no", "split.count",
+              "split.tensors", "qwen4exp.expert_count", "qwen4exp.expert_used_count",
+              "qwen4exp.ngram.tensors"):
+        if k in kvs: print(f"  kv {k} = {kvs[k]}")
+    ngram_kvs = {k: v for k, v in kvs.items() if "ngram" in k.lower()}
+    if ngram_kvs: print("  ngram kvs:", ngram_kvs)
+    big = []
+    for name, dims, t, numel in tensors:
+        big.append((numel * TYPE_SIZE.get(t, 0), name, dims, t, numel))
+    big.sort(reverse=True)
+    print(f"  -- {min(12, len(big))} largest of {len(big)}:")
+    for bytes_, name, dims, t, numel in big[:12]:
+        print(f"    {bytes_/2**30:9.3f} GiB  {name}  dims={dims} type={t} ({numel:,} elems)")
+    named = [(b, n, d, t) for b, n, d, t, e in big
+             if any(s in n.lower() for s in ("ngram", "ple", "token_embd", "exps"))]
+    if named:
+        print("  -- ngram/ple/embed/experts-named:")
+        for bytes_, name, dims, t in named[:20]:
+            print(f"    {bytes_/2**30:9.3f} GiB  {name}  dims={dims} type={t}")
+
+for pth in sys.argv[1:]:
+    try:
+        main(pth)
+    except Exception:
+        print(f"== {pth}: FAILED")
+        traceback.print_exc()
@@ -0,0 +1,85 @@
+# frank-large — ROCmFPX llama.cpp serving
+# kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF (Q8_0_ROCMFPX_AGENT, ftype
+# 115) + its Q4_0 MTP draft head, on the Strix Halo iGPU (gfx1151), :7770.
+#
+# 2026-10-02 (Frank): "the highest quality of that repo, one full context,
+# whatever runtime it requires". The highest-quality rung is Q8_0_ROCMFPX_AGENT
+# (26.28 GiB, 8.39 bpw: 187 tensors at plain Q8_0 + 311 at
+# GGML_TYPE_Q8_0_ROCMFPX); the AGENT routing is also the variant the card
+# tells you to take with MTP, because it keeps the output-head tensors at full
+# Q8_0 (187 vs 2), which is what the draft scores against. The runtime the
+# format REQUIRES is the ROCmFPX fork — type 103 exists nowhere else.
+#
+# The shape: one session at the model's full native window (262,144), which is
+# the whole point — the Q8_0 head costs quality the fleet wanted back, and
+# the draft head pays for the bandwidth the bigger file spends.
+#
+# Budget (from the GGUF header, not from a boot log): attention KV lives only
+# in the 16 full-attention layers (4 KV heads x 256 head_dim), 64 KiB/token
+# with f16 K and V => 16.0 GiB at 262,144. The 48 Gated DeltaNet layers hold
+# ~144 MiB of fixed state per sequence. Weights 26.28 + draft head 1.57 +
+# mmproj 0.87 + KV 16.0 + state 0.14 + compute buffers ≈ 45-47 GiB against the
+# iGPU's ~114 GiB GTT aperture, so f16 KV (the unquantized choice) is affordable
+# here in a way it is not on frank-small's 24 GiB card.
+
+# The qualified ROCmFPX fork build (scripts/build-rocmfpx.sh pins the commit).
+LLAMA_BIN=/home/dank/GitPackages/kingjones-ROCmFPX/build/bin/llama-server
+# One directory per HF repo, created by scripts/download-frank-large-kingjones.sh.
+MODEL_FILE=/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf
+# The draft head: Q4_0, NOT the Q8_0 head. The head's per-step cost is its
+# projection over a 248K vocabulary, so the cheaper head is faster end to end
+# at equal-or-better acceptance (card: +6.0% at n-max 1, +10.2% at n-max 2).
+DRAFT_FILE=/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/mtp-Qwen3.8-27B-Q4_0.gguf
+# Vision tower, separate from the LLM file. Empty string = text-only.
+MMPROJ_FILE=/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/mmproj-Qwen3.8-27B-BF16.gguf
+
+# Fleet endpoint: same port and alias as every frank-large before it, so the
+# harness models.yml id and the homepage probe do not move.
+ALIAS=frank-large
+# HIP reads a bare HOST, so this is BIND_ADDR — the same reason the file
+# avoids a bare TEMP.
+BIND_ADDR=0.0.0.0
+LISTEN_PORT=7770
+
+# One full native window. llama.cpp divides --ctx-size across --parallel
+# slots, so 262144/1 IS the model's trained window; the card validated 65,536
+# and the fleet has been running this family at 262,144 since the gufo era.
+CTX_SIZE=262144
+PARALLEL=1
+
+# Deck knobs. --ubatch-size has a measured optimum on this stack (512 peak on
+# the sibling qwen35 model; larger ubatch costs GiBs of compute buffer for no
+# speed), and --fit must stay off: --fit's auto placement does not know that
+# 311 tensors of type 103 belong on the iGPU.
+BATCH_SIZE=2048
+UBATCH_SIZE=512
+NGL=999
+FIT=off
+FLASH_ATTN=on
+# Unquantized KV — the quality-first choice, affordable at one slot.
+CACHE_TYPE_K=f16
+CACHE_TYPE_V=f16
+
+# Speculative decode: the model ships a standalone MTP head (mtp-Q4_0.gguf,
+# arch qwen35, nextn_predict_layers=1, a 65-block draft). n-max 4 with n-min 0
+# and p-min 0.0 is the card's measured optimum (+31% end to end, acceptance
+# 0.926; its own n-max 8 measurement only reached 0.757 acceptance and LOST).
+# The draft runs entirely on the iGPU, so cross-device traffic stays off the
+# bandwidth-limited link.
+SPEC_TYPE=draft-mtp
+SPEC_DRAFT_NGL=99
+SPEC_DRAFT_DEVICE=ROCm0
+SPEC_DRAFT_N_MAX=4
+SPEC_DRAFT_N_MIN=0
+SPEC_DRAFT_P_MIN=0.0
+
+# The card's thinking sampling profile, which is also what the harness sends
+# in extraBody, so the endpoint behaves the same to a client that omits it.
+TEMPERATURE=1.0
+TOP_P=0.95
+TOP_K=20
+MIN_P=0.0
+PRESENCE_PENALTY=0.0
+
+NUM_THREADS=16
+THREAD_BATCH=16
@@ -0,0 +1,22 @@
+[Unit]
+Description=ROCmFPX llama.cpp (gfx1151) kingjones777 Qwen3.8-27B Q8_0_ROCMFPX_AGENT, MTP spec decode, 262144 ctx / 1 session, root systemd unit (dankclaw, frank-large, Strix Halo iGPU)
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=dank
+Group=dank
+# The engine pins staging buffers; keep the memlock budget unbounded so an
+# RLIMIT smaller than the pinned set cannot fail the load. Parity with the
+# other LLM units on this fleet.
+LimitMEMLOCK=infinity
+Environment="FRANK_LARGE_ENV_FILE=/home/dank/serverAppConfigs/scripts/frank-large.env"
+ExecStart=/home/dank/serverAppConfigs/scripts/serve-frank-large.sh
+# always, not on-failure: an engine that drains and exits 0 after a worker
+# fault would otherwise leave the endpoint down until someone notices.
+Restart=always
+RestartSec=10
+
+[Install]
+WantedBy=multi-user.target
@@ -0,0 +1,202 @@
+#!/usr/bin/env python3
+"""Measure per-block / per-expert expert-routing activity on live frank-primary.
+
+Method: AMD IBS op sampling (data addresses, `perf record -e ibs_op// --data`)
+on the llama-server process while concurrent generations of mixed content run.
+Sampled load addresses are mapped through /proc/<pid>/maps (mmap VMAs of the
+6 GGUF parts) back to individual tensors, then aggregated per expert block
+and per expert WITHIN each block (512 experts, row-major expert-major layout).
+
+Tensor extents are DERIVED from consecutive GGUF header offsets — no quant
+type-size table (those are easy to get subtly wrong; a wrong per-element size
+silently under-covers each tensor and dumps tail samples into a junk bucket).
+
+Fleet law this instrument established (2026-09-26, 75 s window, 4 mixed
+streams, 2.7M expert-tensor samples): per-BLOCK read volume is FLAT
+(0.93x-1.08x of mean across all 40 CPU-resident blocks) because MoE top-k
+routing fires every block every token — there are no "most active layers" at
+`-ot` granularity. The real skew is per-EXPERT (top 10% of (block,expert)
+cells carry ~40% of samples), which tensor-level overrides cannot reach.
+Any 8 GPU-resident blocks are therefore equivalent; COUNT is the only lever.
+
+Usage (needs passwordless sudo for perf only; model files must be readable):
+  fp-expert-activity.py <pid> [duration_s]        # dump + sample + analyze
+  fp-expert-activity.py <pid> --analyze-only     # re-analyze saved samples
+Artifacts: /tmp/fp-tensors.tsv, /tmp/fp-ibs.perf, /tmp/fp-addrs.txt
+"""
+import bisect, json, os, re, struct, subprocess, sys, threading, time
+import urllib.request
+
+MODEL_DIR = "/home/dank/models/Qwen3.8-Flash-Next-GGUF/Q8_0"
+N_PARTS, N_BLOCKS, N_EXPERTS = 6, 48, 512
+TSV, PERF_OUT, ADDRS = "/tmp/fp-tensors.tsv", "/tmp/fp-ibs.perf", "/tmp/fp-addrs.txt"
+PORT = 7770
+
+# ---------------------------------------------------------------- GGUF header
+class P:
+    def __init__(self, f): self.f, self.ver = f, 2
+    def rd(self, n):
+        b = self.f.read(n)
+        assert len(b) == n, "eof"
+        return b
+    def u8(self): return self.rd(1)[0]
+    def u32(self): return struct.unpack("<I", self.rd(4))[0]
+    def u64(self): return struct.unpack("<Q", self.rd(8))[0]
+    def s(self):
+        n = self.u64() if self.ver >= 3 else self.u32()
+        return self.rd(n).decode("utf-8", "replace")
+    def val(self, t):
+        if t == 8: return self.s()
+        if t == 9:
+            et, n = self.u32(), self.u64()
+            for _ in range(n): self.val(et)
+            return None
+        if t in (0, 7): return self.u8()
+        if t == 1: return struct.unpack("<b", self.rd(1))[0]
+        if t == 2: return struct.unpack("<H", self.rd(2))[0]
+        if t == 3: return struct.unpack("<h", self.rd(2))[0]
+        if t == 4: return self.u32()
+        if t == 5: return struct.unpack("<i", self.rd(4))[0]
+        if t == 6: return struct.unpack("<f", self.rd(4))[0]
+        if t == 10: return self.u64()
+        if t == 11: return struct.unpack("<q", self.rd(8))[0]
+        if t == 12: return struct.unpack("<d", self.rd(8))[0]
+        raise ValueError(t)
+
+def dump_part(path):
+    """[(name, start, end)] with extents derived from consecutive offsets."""
+    with open(path, "rb") as f:
+        p = P(f)
+        assert p.rd(4) == b"GGUF"
+        p.ver = p.u32()
+        ntens, nkv = p.u64(), p.u64()
+        for _ in range(nkv):
+            p.s(); p.val(p.u32())
+        infos = []
+        for _ in range(ntens):
+            name = p.s()
+            nd = p.u32()
+            for _ in range(nd): p.u64()
+            p.u32()
+            infos.append((name, p.u64()))
+        data_start = (f.tell() + 31) & ~31
+    fsz = os.path.getsize(path)
+    out = []
+    for i, (name, off) in enumerate(infos):
+        start = data_start + off
+        end = (data_start + infos[i + 1][1]) if i + 1 < len(infos) else fsz
+        out.append((name, start, end))
+    return out
+
+# ------------------------------------------------------------------- workload
+PROMPTS = [
+    [{"role": "user", "content": "Write a Rust function that parses a simple binary log format (u32 length-prefixed records) with proper error handling, then explain the unsafe choices you avoided."}],
+    [{"role": "user", "content": "Compare the strategic logic of Soviet deep-battle doctrine and Blitzkrieg: where did they agree, where did they diverge, and what did each assume about logistics that proved wrong?"}],
+    [{"role": "system", "content": "You have tools: get_weather(city), book_flight(from,to,date). Plan a 3-city trip for next week; call tools as needed and explain each decision."},
+     {"role": "user", "content": "Seattle, Denver, Austin. Cheapest reasonable plan."}],
+    [{"role": "user", "content": "Explain how DNS recursive resolution works end to end, then write a small Python stub that walks a referral chain."}],
+]
+
+def fire(idx, results, dur):
+    body = json.dumps({"messages": PROMPTS[idx], "max_tokens": 260,
+                       "temperature": 0.7, "cache_prompt": False}).encode()
+    req = urllib.request.Request(f"http://127.0.0.1:{PORT}/v1/chat/completions",
+                                 data=body, headers={"Content-Type": "application/json"})
+    t0 = time.time()
+    try:
+        with urllib.request.urlopen(req, timeout=dur + 60) as r:
+            d = json.load(r)
+        results[idx] = (r.status, round(time.time() - t0),
+                        len(d.get("choices", [{}])[0].get("message", {}).get("content", "")))
+    except Exception as e:
+        results[idx] = ("ERR", str(e)[:120])
+
+# --------------------------------------------------------------------- main
+def main():
+    pid = int(sys.argv[1])
+    analyze_only = "--analyze-only" in sys.argv
+    dur = 75.0
+    for a in sys.argv[2:]:
+        if a.isdigit(): dur = float(a)
+    parts = {i: f"{MODEL_DIR}/Qwen3.8-Flash-Next-Q8_0-0000{i}-of-0000{N_PARTS}.gguf"
+             for i in range(1, N_PARTS + 1)}
+
+    vmas = []
+    with open(f"/proc/{pid}/maps") as f:
+        for line in f:
+            m = re.match(r"([0-9a-f]+)-([0-9a-f]+) (\S+) ([0-9a-f]+) (\S+) (\d+)\s*(.*)", line)
+            if not m: continue
+            start, end, pgoff, path = int(m.group(1), 16), int(m.group(2), 16), int(m.group(4), 16), m.group(7)
+            for i, p in parts.items():
+                if path == p: vmas.append((start, end, pgoff, i))
+    vmas.sort()
+    assert vmas, "no GGUF VMAs in maps — wrong pid?"
+    print(f"[rig] {len(vmas)} GGUF VMAs across {len({v[3] for v in vmas})} parts")
+
+    tensors = {}
+    with open(TSV, "w") as tsv:
+        for i, p in parts.items():
+            for name, start, end in dump_part(p):
+                tsv.write(f"{p}\t{name}\t{start}\t{end}\n")
+                tensors.setdefault(i, ([], [], []))
+                tensors[i][0].append(start); tensors[i][1].append(end); tensors[i][2].append(name)
+    n_exp = sum(1 for i in tensors for nm in tensors[i][2] if "exps" in nm)
+    print(f"[rig] {n_exp} exps tensors, extents offset-derived")
+
+    if not analyze_only:
+        results = {}
+        perf = subprocess.Popen(["sudo", "perf", "record", "-e", "ibs_op//", "-c", "50000",
+                                 "--data", "-o", PERF_OUT, "-p", str(pid), "--", "sleep", str(dur)],
+                                stdout=subprocess.DEVNULL, stderr=subprocess.PIPE)
+        time.sleep(2)
+        threads = [threading.Thread(target=fire, args=(i, results, dur)) for i in range(4)]
+        for t in threads: t.start()
+        for t in threads: t.join(dur + 120)
+        err = perf.communicate(timeout=60)[1].decode()[-200:]
+        for i, r in sorted(results.items()):
+            print(f"[work] slot{i}: {r}")
+        print("[perf]", err.strip())
+        with open(ADDRS, "w") as f:
+            f.write(subprocess.run(["sudo", "perf", "script", "-i", PERF_OUT, "-F", "addr"],
+                                    capture_output=True, text=True).stdout)
+
+    block, expert = {}, {}
+    n_exps = n_ple = n_named = n_gap = n_out = 0
+    for line in open(ADDRS):
+        tok = line.split()
+        if not tok: continue
+        try: a = int(tok[0], 16)
+        except ValueError: continue
+        if a == 0 or a > 0xffff000000000000: continue
+        vi = bisect.bisect_right(vmas, (a, 1 << 62, 0, 0)) - 1
+        if vi < 0: n_out += 1; continue
+        start, end, pgoff, pi = vmas[vi]
+        if not start <= a < end: n_out += 1; continue
+        foff = pgoff + (a - start)
+        st, en, nm = tensors[pi]
+        ti = bisect.bisect_right(st, foff) - 1
+        if ti < 0 or foff >= en[ti]: n_gap += 1; continue
+        name, lo, span = nm[ti], foff - st[ti], en[ti] - st[ti]
+        if "exps" in name:
+            n_exps += 1
+            b = int(re.match(r"blk\.(\d+)\.", name).group(1))
+            e = int(lo / (span / N_EXPERTS))
+            block[b] = block.get(b, 0) + 1
+            expert[(b, e)] = expert.get((b, e), 0) + 1
+        elif "per_layer" in name: n_ple += 1
+        else: n_named += 1
+    print(f"\n[rig] attribution: exps={n_exps} ple={n_ple} named={n_named} "
+          f"gap={n_gap} outside={n_out}")
+    tot = sum(block.values())
+    print(f"== BLOCK histogram ({tot} samples, {len(block)} CPU-resident blocks):")
+    for b, c in sorted(block.items(), key=lambda kv: -kv[1]):
+        print(f"  blk{b:2d}: {c:7d} {100*c/tot:5.2f}%")
+    mean = tot / max(len(block), 1)
+    print(f"block spread: max={max(block.values())/mean:.3f}x mean, min={min(block.values())/mean:.3f}x mean")
+    ev = sorted(expert.values(), reverse=True); et = sum(ev)
+    print(f"== EXPERT skew: {len(ev)} distinct (block,expert) cells hit of {N_BLOCKS*N_EXPERTS}")
+    print(f"   top-half of cells carry {100*sum(ev[:len(ev)//2])/et:.1f}%; "
+          f"top-10% {100*sum(ev[:len(ev)//10])/et:.1f}%; hottest cell {100*ev[0]/et:.3f}%")
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,129 @@
+# frank-primary.env — tcclaviger RDNA4 vLLM production config (dankserver :7770).
+#
+# 2026-10-01 cutover to the BF16-KV no-spec incumbent (the overnight redo
+# campaign's verdict, Frank's call): same digest-pinned engine and payload as
+# the 2026-09-29 cutover, but the KV cache is UNQUANTIZED (auto = bf16) and
+# speculative decode is GONE. Git tag frank-primary-fp8-mtp-20261001 is the
+# last pre-cutover commit; the old shape (fp8 KV + MTP4) is in history and
+# restorable from the tag.
+#
+# Every number here was QUALIFIED by boot probe on 2026-10-01 (not inherited):
+#   * MAXLEN 241664 x SEQS 3 (724,992 tokens) at bf16 KV is the DEPLOYED
+#     shape (2026-10-01 night ladder): KV_BYTES 10,184,000,000 PER RANK,
+#     engine measured "KV cache size: 732,359 tokens; Maximum concurrency
+#     for 241,664 tokens per request: 3.03x". 262144 and 245760 both OOMed
+#     at graph capture (76 MiB short / 0 bytes free); 237568 passed at
+#     718,606 tokens and the +4096 bonus step passed and won.
+#   * KV_BYTES is PER RANK. The old "7.2 GiB ceiling, 9/16 GiB OOM" law was
+#     the MTP4 WORKSPACE: the spec-era pool measured 729,545 tokens because
+#     the drafter reserved ~2.6 GiB/rank beside the pool. No-spec bf16 takes
+#     11.05 GiB/rank and boots clean — the ceiling was spec's, not VRAM's.
+#   * bf16 KV doubles the attention-KV bytes/token (measured 14,047 vs
+#     7,926) and costs ~14% decode vs fp8 at 4064 ctx (49.25 -> 36.42 with
+#     the gather win) — traded deliberately for clean cache + 262k windows.
+#   * GATHER 32x64 is the night's biggest lever: R4D_LRU_GATHER_CHUNKS=32 /
+#     R4D_LRU_GATHER_LANES=64 measured +18.7% decode vs default geometry,
+#     and the win grows with depth (39.5@166 -> 54.1@1087 -> 49.3@4064 fp8
+#     era). Do not drop it.
+#   * MNBT 1024 is Frank's lock: MNBT 512 doubles TTFT at 4064 ctx (2.46 s
+#     vs 1.36 s); 1024-4096 are interchangeable on decode; 12288 costs 2.5%
+#     and arena. 1024 keeps the workspace floor so the arena stays max.
+#   * PREFETCH IS ABSENT BY LAW: expert-cache telemetry measured the
+#     prefetcher at 36% precision (36,720 slabs prefetched, 13,369 served) —
+#     roughly half the gather traffic was waste. Never re-add
+#     --expert-prefetch without a fresh A/B.
+#   * TunableOp stays OFF: CLAV_TUNABLEOP_SWEEP silently skips under expert
+#     offload unless CLAV_TUNABLEOP_WITH_OFFLOAD=1, and forcing it corrupts
+#     numerics.
+#
+# HIP reads bare env vars TEMP, HOST, GPU_MAX_HEAP_SIZE — every key here is
+# prefixed FRANK_PRIMARY_ so none can leak.
+#
+
+# --- runtime ----------------------------------------------------------------
+# Digest-pinned: the floating :latest tag moved once already. Local alias
+# qwen-fp8-runtime:tuned-20260929 (image ID 3322dbbbc96d) points at the same
+# bytes in rootless storage.
+FRANK_PRIMARY_IMAGE=docker.io/tcclaviger/vllm@sha256:da4029666668b90a0386c1ed23ea370c79916ad4e1a6ad1dd9db098b9fa069ea
+
+# --- payload ----------------------------------------------------------------
+FRANK_PRIMARY_MODEL_DIR=/home/dank/models/Qwen3.8-Flash-Next-Uncensored-FP8
+FRANK_PRIMARY_MODEL_SHARDS=131
+
+# --- identity ---------------------------------------------------------------
+# The harness models.yml id stays frank-primary (provider 192.168.0.235:7770).
+FRANK_PRIMARY_CONTAINER=frank-primary-qwen
+FRANK_PRIMARY_SERVED_NAME=frank-primary
+FRANK_PRIMARY_PUBLISH=7770:7770
+# Containers removed before every start: the retired R9V identity and the ad
+# hoc experiment identity (both hold the GPUs / name registrations).
+FRANK_PRIMARY_LEGACY_CONTAINERS="frank-primary-r9v qwen-fp8"
+
+# --- qualified engine shape (2026-10-02, Frank's own correction) ---
+# 2026-10-01 left a shape that did not stay up: bf16 KV at 2 x 262144 with
+# --expert-precise-memory crash-looped at graph capture (the arena is
+# DEVICE-resident and the shape had 0.13 GiB to spare), and the plain bf16
+# 2 x 262144 boot booted but left 0 bytes free at runtime. This is the
+# correction: FP8 KV, TWO windows, each 200704 tokens, and the pool cut so
+# the expert arena gets the room back.
+#
+# LADDER (2026-10-01 night, subtract-16k method): 262144 and 245760 OOM at
+# graph capture; 237568 passed at 718,606 tokens; 241664 x 3 passed at
+# 732,359 tokens / 3.03x but is the shape this cutover retired. 200704 was
+# chosen to sit below the 241664 rung while leaving a full 2.95 GiB/rank of
+# pool to cut loose for the arena — the two contexts are 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.
+#
+# KV_BYTES is PER RANK. fp8 costs 7,701 B/token (measured 6,138 attention
+# + 1,563 mamba/side, dtype-independent). 2 x 200704 = 401,408 tokens
+# -> 3.091e9 B floor, pinned at 3.15e9 (~405,000 tokens, ~44 MB margin —
+# the same margin convention the 6.1e9 shape used).
+#
+# THE ARENA IS WHAT REMAINS: the fork's expert planner sizes it AFTER the
+# KV override, so every byte cut here is arena. The 6.1e9 shape ran an
+# arena of 6,248 slots; this one reclaims ~2.95 GiB/rank, roughly 1,260
+# more slots at 2.344 MiB each. That is the point of the change — the
+# gather read path is the measured decode bound, and a bigger arena means
+# more experts resident instead of streamed.
+#
+# MNBT 1024 (down from 2048): 1024-4096 are interchangeable on decode AND
+# identical on arena (the sweep measured the same slot count at every
+# rung <= 4096), so this is a pure workspace-floor choice, not a decode
+# trade. The floor matters because the arena budget is what is left.
+#
+# PRECISE MEMORY stays DROPPED: it crashed the 2 x 262144 bf16 boot
+# (weights 6.28 + pool 6.80 + graphs 0.55 + arena 16.65 + non-torch ~1.45
+# = 31.73 of 31.86 GiB, ~0.13 GiB free, NRestarts=3). The flag survives in
+# the serve script as an opt-in; re-enable it only with a pool/context
+# reduction that actually pays for the extra overhead.
+#
+# TUNABLEOP stays OFF (silently skipped under offload unless
+# CLAV_TUNABLEOP_WITH_OFFLOAD=1, and then corrupts numerics).
+#
+# SEQS 2 x MAXLEN 200704 at FP8 KV, pool cut for the arena. No
+# speculative decode, no prefetch.
+FRANK_PRIMARY_MAXLEN=200704
+FRANK_PRIMARY_SEQS=2
+FRANK_PRIMARY_KV_BYTES=3150000000
+FRANK_PRIMARY_KV_DTYPE=fp8
+FRANK_PRIMARY_MNBT=1024
+FRANK_PRIMARY_GATHER_CHUNKS=32
+FRANK_PRIMARY_GATHER_LANES=64
+FRANK_PRIMARY_UTIL=0.98
+
+# --- client contract: reasoning split, tools, preserve_thinking ----------------
+# Without --reasoning-parser the engine leaves the thinking text INLINE in
+# `content` with a dangling `</think>` and a null `reasoning` field (measured
+# 2026-10-01: content = 'We need answer...Brief.\n</think>\n\n391') — the
+# malformed thinking block clients choke on. The qwen3 parser splits it
+# properly. The GLM rollback (d91379d) had dropped all three of these; the
+# working values are the ones ecaad36 shipped for this model.
+FRANK_PRIMARY_REASONING_PARSER=qwen3
+# Tools: --enable-auto-tool-choice + the qwen3_xml parser is what makes
+# tool_calls work here (without it, tool_choice auto returns 400).
+FRANK_PRIMARY_TOOL_PARSER=qwen3_xml
+# preserve_thinking keeps prior thinking blocks in the rendered prompt across
+# turns; enable_thinking makes thinking the server-side default. Verified the
+# payload's tokenizer_config template consumes both keys.
+FRANK_PRIMARY_CHAT_TEMPLATE_KWARGS='{"enable_thinking": true, "preserve_thinking": true}'
@@ -0,0 +1,66 @@
+# frank-primary.service — tcclaviger RDNA4 vLLM endpoint (orcarouter
+# Qwen3.8-Flash-Next-Uncensored-FP8, MTP4 spec decode, KV q8/q8, 600k total
+# context over 3 sessions, TP=2, tiered experts, 131-shard payload) on the
+# 2x R9700 pair, dankserver :7770. 2026-09-29 cutover from the R9V radiance
+# fork (git tag frank-primary-r9v-q8-20260929 is the pre-cutover state; its
+# launcher/shim/setup-state machinery is retired and the state dirs stay on
+# disk as rollback).
+#
+# Same port, same served alias; the harness models.yml id is unchanged.
+# The serve script drives `podman run` directly (the image entrypoint execs
+# `vllm serve`) and then waits on the container. NOTE: that wait can return
+# 0 even when the engine died mid-request (vLLM's API server drains and
+# exits clean on a worker fault), which is why the restart policy is
+# `always` — see the [Service] note.
+#
+# WHY A ROOT UNIT (not a rootless quadlet): the tiered-expert shape pins tens
+# of GiB of host RAM (the expert host slabs, ~20.5 GiB per rank of pinned
+# anonymous memory). Pinned pages are charged against RLIMIT_MEMLOCK; dank's
+# login session on this host carries 8 MiB and a user quadlet is capped by the
+# user manager's own hard limit — raising that means restarting
+# user@1000.service and bouncing every user container on the host. A root unit
+# grants LimitMEMLOCK=infinity to the dank process via PID1, and the image
+# stays in dank's rootless podman storage.
+[Unit]
+Description=frank-primary — tcclaviger vLLM Qwen3.8-Flash-Next-Uncensored-FP8 on 2x R9700
+After=network-online.target
+Wants=network-online.target
+# Never let the start-rate limiter give up after a crash loop (a failed LLM
+# endpoint that systemd refuses to retry is worse than a retrying one).
+StartLimitIntervalSec=0
+
+[Service]
+User=dank
+Group=dank
+# The tiered-expert offloader pins host RAM; an RLIMIT_MEMLOCK smaller than
+# the pinned set fails the load mid-start. PID1 grants infinity here —
+# parity with the fleet's other LLM units.
+LimitMEMLOCK=infinity
+Environment="FRANK_PRIMARY_ENV_FILE=/home/dank/serverAppConfigs/scripts/frank-primary.env"
+ExecStart=/home/dank/serverAppConfigs/scripts/serve-frank-primary.sh
+# systemd stop: wait alone cannot stop the container child; stop it
+# explicitly. vLLM 0.26 drains gracefully on SIGTERM.
+ExecStop=/usr/bin/podman stop -t 120 frank-primary-qwen
+# Boot cost: FP8 payload page-in plus torch compile/graph capture —
+# ~15 min cold (measured 2026-09-29), faster warm.
+TimeoutStartSec=1800
+# Restart=always, NOT on-failure (measured 2026-10-01): a runtime
+# torch.OutOfMemoryError kills the engine while vLLM's API server still
+# exits 0, so ExecStart returned clean, systemd logged "Deactivated
+# successfully", and on-failure NEVER fired — the endpoint stayed down until
+# a human noticed. `always` restarts on any container death; an explicit
+# `systemctl stop` still wins (systemd does not restart a unit the operator
+# stopped), so bench windows and deploys are unaffected.
+Restart=always
+RestartSec=10
+# ExecStop allows the engine 120 s to drain (`podman stop -t 120`); without an
+# explicit TimeoutStopSec systemd's 90 s default cuts the stop short, kills it,
+# and records 'Failed with result timeout' (measured 2026-10-01 13:27:17).
+# Must exceed the podman grace with margin for the engine's own shutdown.
+TimeoutStopSec=180
+# The engine pins tens of GiB of host RAM (expert host slabs); it is the
+# last process on this host that should be chosen by the kernel OOM killer.
+OOMScoreAdjust=-500
+
+[Install]
+WantedBy=multi-user.target
@@ -1,93 +1,60 @@
-# Tunables for frank-small (HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-
-# Balanced, Q4_K_M, plus the BF16 vision projector and the separate MTP
-# sidecar) on mainline llama.cpp, Dankclaw's RTX 3090, since 2026-09-17.
+# frank-small — llama.cpp (master, CUDA sm_86) serving
+# bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF:Q5_K_M on the RTX 3090, :7771.
 #
-# Runtime: llama.cpp master 37b53fd, CUDA build for sm_86:
-#   cmake -B build-cuda -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=ON \
-#     -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=86 -DGGML_CUDA_FA_ALL_QUANTS=ON
-#   cmake --build build-cuda --target llama-server -j
-# GGML_CUDA_FA_ALL_QUANTS is ON: this service runs q8_0 KV with flash
-# attention, and without the flag quantized-KV prefill collapses ~12x on
-# CUDA (see the CUDA lesson in AGENTS.md).
+# 2026-10-02 (Frank): THREE contexts, vision on, and a hard instruction to
+# leave 2-3 GiB of the card free. Both new demands cost memory, so the per-slot
+# window shrinks — see the budget below.
 #
-# Model: HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced at Q4_K_M
-# (6.87 GiB). Architecture (GGUF arch gemma4): 48 layers, MHA with 16 kv
-# heads x 512 head_dim, rope freq_base 1e6, final logit softcapping 30,
-# per-layer token embeddings (PLE) inside the file, and a sliding-window
-# attention pattern (attention.sliding_window = 1024) -- most layers cache
-# only a 1024-position window, which is why the KV pool is a fraction of
-# the full-ctx x all-layers number.
+# Two measurements on this card, this runtime, price the terms:
+#   2 x 262,144, no projector: 20,190 MiB used / 4,386 free
+#     KV = 524,288 tok x 24.5 KiB = 12,800 MiB => non-KV (weights 6,557 +
+#     DeltaNet state ~25/slot + compute buffers + CUDA graphs) = 7,390 MiB
+#   3 x 196,608, projector on: 22,972 MiB used / 1,153 free
+#     KV = 14,113 MiB => non-KV WITH the vision tower = 8,859 MiB
+# The tower costs ~1.4 GiB, not the ~1.0 its file size suggested, so 196,608
+# per slot left only 1.1 GiB free — outside the band. The pool therefore is
+# 3 x 180,224 = 540,672 tokens: KV ~12,936 + non-KV 8,859 = ~21,800 MiB predicted.
+# MEASURED at this shape: 21,684 MiB used / 2,441 MiB free — inside the band, and
+# the number the bats budget gate reproduces (it prices non-KV at 9 GiB).
 #
-# MTP: the Multi-Token Prediction head ships as a SEPARATE file
-# (mtp-gemma-4-12B-it.gguf, arch gemma4-assistant, 4 nextn layers, 0.24
-# GiB). Unlike the Qwen3.8 -mtp build (head inside the main file), this one
-# is passed via --model-draft with --spec-type draft-mtp, so the draft
-# context loads the assistant model. The same --spec-type value chains the
-# draftless ngram-mod matcher, which attaches unconditionally; two
-# context-backed drafters cannot share one process, so draft-mtp,ngram-mod
-# is the only legal shape that carries both.
+# KV arithmetic: attention KV exists only in the 8 full-attention layers
+# (4 KV heads x 256 head_dim). K f16 = 16 KiB/token, V q8_0 = 8.5 KiB/token
+# (34 B per 32-element block) => 24.5 KiB/token, so 12,288 tokens (=4096/slot)
+# is one 293 MiB step — that is the ladder to walk if this moves again. The
+# native 262,144 per slot is NOT reachable three times: 3 x 262,144 x
+# 24.5 KiB = 18.4 GiB of KV before anything else is loaded.
 #
-# SIZING (RTX 3090, 24576 MiB): not modelled by hand -- CTX_SIZE is left
-# unset and --fit sizes the KV pool (q8_0 K and V) against what the 6.87
-# GiB of weights, the 0.16 GiB projector, the MTP sidecar's draft context
-# and the compute buffers leave on the card. The engine prints the chosen
-# context at startup; read it from the log, do not hard-code it back here.
-#
-# Vision: --mmproj loads the BF16 projector (11 tensors, 52M params; it
-# also carries an audio encoder, unused by the fleet). The fleet's image
-# path lives here, which is why halogen no longer sets HALOGEN_VISION_TOWER.
-#
-# Sampling: greedy by default (the fleet contract); a client that sends a
-# temperature gets its own, and the harness applies per-model overrides
-# from models.yml.
-#
-# WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE (HIP
-# reads those bare vars on the ROCm hosts -- see the 2026-08-13 incident in
-# AGENTS.md). This box is CUDA, but the rule keeps env files portable.
+# Why V stays q8_0 and K stays f16: full-precision V would add 8 KiB/token —
+# +4.5 GiB on this pool — which is what pushes the card past its 24 GiB. K is
+# the tensor the asymmetric-KV contract exists to protect.
+LLAMA_BIN=/home/dank/GitPackages/llama-master/build-cuda/bin/llama-server
+MODEL_FILE=/home/dank/models/MiMo-Q5KM/MiMo-V2.6-Distill-Qwen-9B-Q5_K_M.gguf

-# llama.cpp runtime binary: mainline master 37b53fd, CUDA sm_86,
-# FA_ALL_QUANTS=ON.
-LLAMA_BIN=/home/dank/GitPackages/llama.cpp-mainline/build-cuda/bin/llama-server
+# Vision: the Q5_K_M GGUF ships no `v.*` tensors, so the projector file IS the
+# vision capability (sha256 ff348f31..., 918,166,048 B, fetched by
+# scripts/download-frank-small.sh). Empty string = text-only on purpose.
+MMPROJ_FILE=/home/dank/models/MiMo-Q5KM/mmproj-MiMo-V2.6-Distill-Qwen-9B-f16.gguf

-# Model package. The MTP sidecar file is required for speculative decoding.
-MODEL_DIR=/home/dank/models/Gemma4-12B-QAT
-MODEL_FILE=Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf
-MMPROJ_PATH=/home/dank/models/Gemma4-12B-QAT/mmproj-Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-BF16.gguf
-DRAFT_FILE=mtp-gemma-4-12B-it.gguf
-
-# Fleet identity and network. 7771 is this service's own endpoint on
-# Dankclaw; 7770 stays frank-large's. The alias is the id at /v1/models and
-# what clients ask for -- the harness entry in models.yml names it.
+# Fleet endpoint: same port and alias as every frank-small before it, so the
+# harness models.yml id and the homepage probe do not move.
 ALIAS=frank-small
 BIND_ADDR=0.0.0.0
 LISTEN_PORT=7771

-# Context: intentionally unset -- --fit decides it (see the sizing note).
-# One slot; a second concurrent request queues behind the first.
-PARALLEL=1
-CACHE_TYPE_K=q8_0
+# Three independent conversations. llama.cpp divides --ctx-size across
+# --parallel slots (n_ctx_slot = CTX_SIZE/PARALLEL), and a mamba-hybrid model
+# cannot use unified KV, so each slot gets its own cache and its own window.
+# 180,224 per slot is BELOW the native 262,144, and that is the cost of the
+# third slot: a longer single prompt needs compaction instead of one shot.
+# Consequence worth knowing: frank-primary's compaction trigger is 183,501
+# tokens, which no longer fits a frank-small slot — see the runbook's
+# "three slots vs compaction" note.
+CTX_SIZE=540672
+PARALLEL=3
+
+# The KV contract (Frank, 2026-10-01 night): unquantized K, q8_0 V.
+# q8_0 on V is only served by the flash-attention path; the serve script
+# pins --flash-attn on and must never lose it.
+CACHE_TYPE_K=f16
 CACHE_TYPE_V=q8_0
-
-# Batch geometry: llama.cpp's own defaults (2048 logical / 512 physical).
-BATCH_SIZE=2048
-UBATCH_SIZE=512
-
-# CPU threads: this service runs entirely on the GPU (NGL=999), so threads
-# cover tokenization, sampling and the HTTP layer only. Deliberately modest
-# -- this host's CPU is shared with halogen's engine, which is CPU-hungry.
-NUM_THREADS=8
-THREAD_BATCH=8
-
-# Every layer on the 3090, with flash attention (required for q8 KV).
-NGL=999
-FLASH_ATTN=on
-
-# Speculative decoding: the MTP sidecar (draft context) chained with the
-# draftless ngram-mod matcher. Draft depth stays llama.cpp's default of 3.
-SPEC_TYPE=draft-mtp,ngram-mod
-
-# Greedy default (fleet contract); reasoning is extracted into
-# reasoning_content with an unlimited budget (the model decides).
-TEMPERATURE=0.0
-REASONING_FORMAT=deepseek
-REASONING_BUDGET=-1
+NGL=99
@@ -1,5 +1,5 @@
 [Unit]
-Description=llama.cpp Qwen3.8-27B GSQ-RCO IQ3_S MTP+vision (dankclaw, frank-small)
+Description=llama.cpp MiMo-V2.6-Distill-Qwen-9B Q5_K_M GGUF (qwen35 hybrid, 3x180224, f16 K + q8_0 V, mmproj vision) on the RTX 3090 (dankclaw, frank-small)
 After=network-online.target
 Wants=network-online.target

@@ -8,8 +8,7 @@ Type=simple
 User=dank
 Group=dank
 # CUDA pins its staging buffers, so the process needs the memlock budget even
-# though the weights themselves are copied into VRAM. Parity with the halogen
-# quadlet's Ulimit=memlock=-1:-1 and the retired ik unit.
+# though the weights themselves are copied into VRAM.
 LimitMEMLOCK=infinity
 Environment="FRANK_SMALL_ENV_FILE=/home/dank/serverAppConfigs/scripts/frank-small.env"
 ExecStart=/home/dank/serverAppConfigs/scripts/serve-frank-small.sh
@@ -0,0 +1,105 @@
+#!/usr/bin/env bash
+# hf-fetch.sh — resumable, verified HuggingFace fetch (gated or public repos).
+#
+#   bash scripts/hf-fetch.sh <dest_dir> <manifest>
+#
+# Manifest: one line per file, three tab/space-separated fields:
+#   <filename> <size_bytes> <sha256>
+# A blank sha256 field means "check size only". Lines starting with '#' and
+# blank lines are ignored. The first field is a path relative to <dest_dir>.
+#
+# Token: $HF_TOKEN, else ~/.cache/huggingface/token, else the llamaswap conf
+# (grep hf_...). Passed out-of-band via a mode-600 curl -K file — a -H header
+# is world-readable in /proc/<pid>/cmdline (the 2026-09-16 HF-token leak).
+#
+# Idempotent: a file whose size AND sha256 already match is skipped, so a
+# re-run after an interruption only fetches what is missing or corrupt.
+set -euo pipefail
+
+dest="${1:?usage: hf-fetch.sh <dest_dir> <manifest>}"
+manifest="${2:?usage: hf-fetch.sh <dest_dir> <manifest>}"
+repo="${3:?usage: hf-fetch.sh <dest_dir> <manifest> <owner/repo>}"
+
+mkdir -p "$dest"
+
+# --- token resolution --------------------------------------------------------
+token="${HF_TOKEN:-}"
+if [[ -z "$token" && -r "$HOME/.cache/huggingface/token" ]]; then
+    token="$(tr -d '[:space:]' < "$HOME/.cache/huggingface/token")"
+fi
+if [[ -z "$token" && -r "$HOME/.config/llamaswap.conf.yaml" ]]; then
+    token="$(grep -oE 'hf_[a-zA-Z0-9]+' "$HOME/.config/llamaswap.conf.yaml" | head -1)"
+fi
+[[ -n "$token" ]] || { echo "hf-fetch: no HF token (set HF_TOKEN or ~/.cache/huggingface/token)" >&2; exit 3; }
+
+auth_file="$(mktemp)"
+chmod 600 "$auth_file"
+printf 'header = "Authorization: Bearer %s"\n' "$token" > "$auth_file"
+cleanup() { rm -f "$auth_file"; }
+trap cleanup EXIT
+
+base="https://huggingface.co/${repo}/resolve/main"
+
+# --- per-file fetch + verify -------------------------------------------------
+fetch_one() {
+    local name="$1" size="$2" sha="$3"
+    local url="$base/$name"
+    local path="$dest/$name"
+    mkdir -p "$(dirname "$path")"
+
+    # idempotent skip
+    if [[ -f "$path" ]]; then
+        local cur_size sha_ok
+        cur_size="$(wc -c < "$path")"
+        sha_ok=true
+        if [[ -n "$sha" && "$(sha256sum "$path" | cut -d' ' -f1)" != "$sha" ]]; then
+            sha_ok=false
+        fi
+        if [[ "$cur_size" == "$size" && "$sha_ok" == "true" ]]; then
+            echo "  ok   $name ($size bytes)"
+            return 0
+        fi
+        echo "  refetch $name (have $cur_size, want $size)"
+    fi
+
+    # Resumable download; -f fails on HTTP error so a 401/403 is not written as a file.
+    # -C - only when a non-empty partial exists: with a missing file it sends
+    # Range: bytes=0-, which HF's xethub CDN rejects with 400 (xet-backed repos).
+    local resume_args=()
+    [[ -f "$path" && -s "$path" ]] && resume_args=(-C -)
+    curl -fsSL "${resume_args[@]}" -K "$auth_file" -o "$path" "$url"
+
+    # verify byte count
+    local got
+    got="$(wc -c < "$path")"
+    if [[ "$got" != "$size" ]]; then
+        echo "  FAIL $name: size $got != $size" >&2
+        return 1
+    fi
+    # verify sha256
+    if [[ -n "$sha" ]]; then
+        local actual
+        actual="$(sha256sum "$path" | cut -d' ' -f1)"
+        if [[ "$actual" != "$sha" ]]; then
+            echo "  FAIL $name: sha256 $actual != $sha" >&2
+            return 1
+        fi
+    fi
+    echo "  done $name ($size bytes)"
+}
+
+echo "hf-fetch: $repo -> $dest"
+status=0
+while IFS= read -r line || [[ -n "$line" ]]; do
+    [[ -z "$line" || "$line" =~ ^[[:space:]]*# ]] && continue
+    name="$(awk '{print $1}' <<<"$line")"
+    size="$(awk '{print $2}' <<<"$line")"
+    sha="$(awk '{print $3}' <<<"$line")"
+    fetch_one "$name" "$size" "$sha" || status=1
+done < "$manifest"
+
+if [[ "$status" -ne 0 ]]; then
+    echo "hf-fetch: FAILED (see above)" >&2
+    exit 1
+fi
+echo "hf-fetch: all files verified"
@@ -0,0 +1,140 @@
+#!/usr/bin/env python3
+"""hf-manifest.py — generate an hf-fetch.sh manifest from a HuggingFace repo.
+
+Usage: python3 scripts/hf-manifest.py <owner/repo> [outfile]
+
+Lists the repo tree via the HF API and, for each file, records the size and
+the content sha256. LFS and Xet-backed files redirect the resolve endpoint to
+a CDN (302), whose own headers carry the truth: X-Linked-Size is the byte
+count and X-Linked-ETag is the sha256 of the file's bytes. The redirect must
+NOT be followed: on Xet-backed repos the bridge answers with ETag = the Xet
+CAS chunk-dag hash, which is NOT the content sha256 (seen on
+Qwen/Qwen3.8-Flash-Next-FP8, 2026-09-26: every shard "failed" verification
+against a hash no download could ever match). Non-LFS files (config.json,
+tokenizer, ...) are answered directly (200) and carry no 64-hex etag, so they
+get a blank hash field — hf-fetch.sh then checks size only for them.
+
+Output: one line per file, three space-separated fields:
+    <filename> <size_bytes> <sha256-or-blank>
+
+Token: $HF_TOKEN, else ~/.cache/huggingface/token, else the llamaswap conf
+(grep hf_...). Gated repos need a token that has been granted access.
+"""
+import http.client
+import json
+import os
+import re
+import sys
+import urllib.request
+from urllib.parse import urljoin, urlsplit
+
+def get_token():
+    t = os.environ.get("HF_TOKEN", "")
+    if not t:
+        p = os.path.expanduser("~/.cache/huggingface/token")
+        if os.path.exists(p):
+            with open(p) as f:
+                t = f.read().strip()
+    if not t:
+        p = os.path.expanduser("~/.config/llamaswap.conf.yaml")
+        if os.path.exists(p):
+            with open(p) as f:
+                m = re.search(r"hf_[a-zA-Z0-9]+", f.read())
+                if m:
+                    t = m.group(0)
+    return t
+
+
+def head(url, token):
+    """HEAD via http.client, which does not follow redirects.
+
+    urllib's default opener follows the resolve endpoint's 302, and the Xet
+    bridge's ETag is the Xet CAS hash, not the content sha256 — so the 302's
+    own X-Linked-* headers must be read instead. Returns an http.client
+    response; the caller reads r.status and r.headers.
+    """
+    parts = urlsplit(url)
+    conn = http.client.HTTPSConnection(parts.netloc, timeout=30)
+    path = parts.path + (f"?{parts.query}" if parts.query else "")
+    h = {"User-Agent": "hf-manifest/1.0"}
+    if token:
+        h["Authorization"] = f"Bearer {token}"
+    conn.request("HEAD", path, headers=h)
+    return conn.getresponse()
+
+
+def main():
+    if len(sys.argv) < 2:
+        print(f"usage: {sys.argv[0]} <owner/repo> [outfile]", file=sys.stderr)
+        sys.exit(2)
+    repo = sys.argv[1]
+    outfile = sys.argv[2] if len(sys.argv) > 2 else None
+    token = get_token()
+    # 1. List the repo tree (siblings[].rfilename). Sizes and content sha256s
+    # come from the resolve endpoint's HEAD (not followed — see head());
+    # the API's siblings carry only rfilename.
+    api = f"https://huggingface.co/api/models/{repo}"
+    req = urllib.request.Request(
+        api, headers={"User-Agent": "hf-manifest/1.0",
+                      **({"Authorization": f"Bearer {token}"} if token else {})})
+    with urllib.request.urlopen(req, timeout=30) as r:
+        meta = json.load(r)
+    siblings = meta.get("siblings", [])
+    print(f"# {repo}: {len(siblings)} entries", file=sys.stderr)
+
+    lines = []
+    for s in siblings:
+        name = s.get("rfilename", "")
+        if not name or name.endswith("/"):
+            continue
+        # 2. Size + content sha256 from the resolve endpoint's HEAD. 200 =
+        # direct answer (Content-Length + ETag). 3xx WITH X-Linked-Size =
+        # LFS/Xet redirect: X-Linked-Size + X-Linked-ETag are the byte
+        # count and the sha256 of the file's bytes, and the redirect must
+        # NOT be followed — the Xet bridge's own ETag is the Xet CAS hash,
+        # not the content sha256. 3xx WITHOUT X-Linked-Size = non-LFS
+        # redirect (307 to HF's own /api/resolve-cache): the stub's
+        # Content-Length is NOT the file size, so follow ONCE; the final
+        # ETag there is the git blob sha (40 hex), so the 64-hex gate
+        # leaves the hash blank (hf-fetch.sh then checks size only).
+        size = 0
+        sha = ""
+        resolve = f"https://huggingface.co/{repo}/resolve/main/{name}"
+        try:
+            with head(resolve, token) as r:
+                if r.status == 404:
+                    raise FileNotFoundError(name)
+                if r.status == 200:
+                    size = int(r.headers.get("Content-Length") or 0)
+                    etag = (r.headers.get("ETag") or "").strip('"')
+                else:
+                    linked = r.headers.get("X-Linked-Size")
+                    if linked is not None:
+                        size = int(linked)
+                        etag = (r.headers.get("X-Linked-ETag") or "").strip('"')
+                    else:
+                        loc = r.headers.get("Location", "")
+                        with head(urljoin(resolve, loc), token) as r2:
+                            if r2.status == 200:
+                                size = int(r2.headers.get("Content-Length") or 0)
+                                etag = (r2.headers.get("ETag") or "").strip('"')
+                            else:
+                                etag = ""
+                if re.fullmatch(r"[0-9a-f]{64}", etag):
+                    sha = etag
+        except FileNotFoundError:
+            pass
+        lines.append(f"{name} {size} {sha}")
+        print(f"  {name}  {size:>12}  {sha or '(size only)'}", file=sys.stderr)
+
+    out = "\n".join(lines) + "\n"
+    if outfile:
+        with open(outfile, "w") as f:
+            f.write(out)
+        print(f"wrote {outfile}", file=sys.stderr)
+    else:
+        sys.stdout.write(out)
+
+
+if __name__ == "__main__":
+    main()
@@ -2,7 +2,8 @@
 """homepage-status.py — the homepage status checker.

 Reads a service config (a JSON array of service entries), probes each backend,
-and writes a status.json atomically. A wip service is not probed.
+and writes a status.json atomically. A wip service is not probed; its state
+is "wip-open" or "wip-closed" following the wip value in the config.

 Usage: homepage-status.py <services.json> <out.json>
 """
@@ -48,8 +49,9 @@ def main(argv):

    results = []
    for svc in services:
-        if svc.get("wip"):
-            state = "wip"
+        wip = svc.get("wip")
+        if wip:
+            state = "wip-%s" % wip
        else:
            host, _, port = svc["probe"].rpartition(":")
            state = probe(host, int(port), svc.get("method", "tcp"))
@@ -0,0 +1,86 @@
+# glm53-compile-01-model-decorator.patch — put the GLM-5.3-Flash trunk under torch.compile.
+#
+# STATUS: WIP — first of three patches in the "compile the GLM trunk" changelist. The
+# set does NOT yet boot the engine (see 03 for the next blocker), so it is NOT part of
+# the deployed frank-primary shape. Nothing in scripts/ mounts these files.
+#
+# WHY THIS IS THE RIGHT LEVER (measured 2026-09-30, not inferred):
+#
+# The fork's own flagship trunk — vllm/models/qwen4_exp/amd/model.py — carries
+# @support_torch_compile. vllm/models/glm5next/common/model.py does not, and the engine
+# says so on every boot:
+#
+#   Cudagraph mode FULL_AND_PIECEWISE requires piecewise capture, but the loaded model
+#   provides neither a compiled submodule nor breakable CUDA graphs.
+#   Overriding to FULL_DECODE_ONLY.
+#
+# A torch-profiler window on the live endpoint (496 decode tokens, rank0, 3 x 160-token
+# generations at the deployed shape) shows what the missing compilation costs. Total
+# device kernel time 46.18 s / 496 tokens = ~93 ms per token, against a measured step
+# time of ~104 ms (9.6 tok/s with the profiler attached):
+#
+#   16.07 s  34.8 %  708,500 launches  vectorized_elementwise_kernel      (22.7 us each)
+#    4.60 s  10.0 %    42,000           r4d_gemm_moe_w4a16_nt_b16_m64     (the MoE GEMM)
+#    3.11 s   6.7 %   174,708           elementwise_kernel_manual_unroll
+#    2.94 s   6.4 %   138,252           wvSplitK_hf_sml_                  (split-K GEMM)
+#    4.09 s   8.9 %    83,160           topk/router (sbtopk gatherTopK + warpMergeSortTopK
+#                                       + bitonicSortKVInPlace)
+#    3.01 s   6.5 %   133,920           r4d_mhc_* (hyper-connection pre/post/finish)
+#    2.78 s   6.0 %   127,916           inductor triton_poi_* / triton_per_* (the only
+#                                       fusions that DO happen)
+#    0.49 s   1.05%    20,664           r4d_moe_lru_gather_kernel         <-- the suspect
+#
+# That last row retired the fleet's working hypothesis. Decode on this shape is NOT
+# expert-gather bound: the gather is 1 % of device time, the same order as the telemetry
+# kernel that was switched on to measure it. 44 % of the budget is pointwise glue —
+# bf16<->fp32 conversions, adds, pows/rsqrts from fp32 RMSNorm, clamp/scale quant
+# passes — spread over ~1,428 launches per token. Inductor fusion is the only lever
+# that touches that mass; the router is second at 9 % but is binary-blocked (the
+# libr4d route kernel exists only as r4d_moe_route_kernel<128|256|512> instantiations
+# and GLM has 288 experts, with no HIP sources shipped to rebuild).
+#
+# TARGET: vllm/models/glm5next/common/model.py   (inside the pinned image)
+#   base    sha256 008eeaa2a07e091465319f86900b5a321d921238723ccbe22e6b9e971e1d2b28
+#   patched sha256 6e9cb81b84a9a88bd41403ca404aa04c3b791899dd77d661c3f25ae686cecd40
+#
+# APPLY: `patch -p1` from the vllm package root, or bind-mount new/.../model.py over
+# site-packages. Then R4D_GLM_COMPILE=1 (default) enables it; =0 makes the decorator a
+# no-op, so the file can stay mounted for an A/B without a second artifact.
+# Requires 02 (trace-safe probes) — without it the worker dies during profile_run on
+# the first probe call inside the traced region.
+--- a/vllm/models/glm5next/common/model.py
+++ b/vllm/models/glm5next/common/model.py
+@@ -1,12 +1,14 @@
+ # SPDX-License-Identifier: Apache-2.0
+ # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+ 
+import os
+ from collections.abc import Iterable
+ from typing import ClassVar, Literal
+ 
+ import torch
+ from torch import nn
+ 
+from vllm.compilation.decorators import support_torch_compile
+ from vllm.config import ParallelConfig, VllmConfig
+ from vllm.distributed import (
+     get_ep_group,
+@@ -623,6 +625,18 @@
+         )
+ 
+ 
+_GLM_COMPILE = os.environ.get("R4D_GLM_COMPILE", "1") == "1"
+
+
+@support_torch_compile(
+    dynamic_arg_dims={
+        "input_ids": 0,
+        "positions": -1,
+        "intermediate_tensors": 0,
+        "inputs_embeds": 0,
+    },
+    enable_if=lambda _cfg: _GLM_COMPILE,
+)
+ class Glm5NextModel(nn.Module):
+     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+         super().__init__()
@@ -0,0 +1,119 @@
+# glm53-compile-02-trace-safe-probes.patch — libr4d capability probes must not do I/O
+# inside a torch.compile region.
+#
+# STATUS: WIP — second of three patches in the "compile the GLM trunk" changelist.
+# Inert (but harmless) without 01. Not part of the deployed shape.
+#
+# THE BREAK (measured: with 01 mounted and R4D_GLM_COMPILE=1, the worker dies during
+# profile_run, i.e. the first — cold — forward):
+#
+#   Explanation: Dynamo does not know how to trace the builtin `posix.stat`
+#   ... r4d_mhc_supported -> r4d_lib.r4d_mhc_enabled -> available -> os.path.exists
+#
+# THE NON-OBVIOUS PART: functools.lru_cache does NOT protect these probes. Dynamo
+# re-traces the wrapped body to build its guard even when the cache is warm — it warns
+# about exactly this ("Dynamo detected a call to a `functools.lru_cache`-wrapped
+# function ... Dynamo ignores the cache wrapper and directly traces the wrapped
+# function"). Verified in-image, torch 2.11.0+rocm10.0, torch.compile(backend="eager"):
+#
+#   cold lru_cache call       -> RAISE Unsupported
+#   WARM lru_cache call       -> RAISE Unsupported     <-- warming is not enough
+#   plain module-global read  -> compiles
+#   uncached os.path.exists   -> RAISE Unsupported
+#
+# So each probe must resolve ONCE into a plain value and be read as a global. The set
+# below is every no-argument probe / symbol lookup reachable from a forward pass,
+# grep-verified call sites:
+#   model_executor/layers/fused_moe/experts/r4d_w4a16_moe.py:79  available()
+#   model_executor/layers/fused_moe/experts/r4d_w4a16_moe.py:81  has_w4a16_moe()
+#   model_executor/kernels/mhc/r4d_dsv4.py:16                    r4d_mhc_enabled()
+#   v1/attention/ops/rocm_aiter_mla_sparse.py:694                r4d_indexer_enabled(), mqa_max_heads()
+#   v1/attention/ops/rocm_aiter_mla_sparse.py:1124               r4d_indexer_enabled()
+#   model_executor/layers/fused_moe/experts/r4d_mxfp4_moe.py:683 r4d_act_quant_enabled()
+#   model_executor/layers/fused_moe/experts/r4d_w4a16_moe.py _entry()  moe_w4a16()/moe_w4a16_mt1() per MoE call
+# (lib() is the transitive dependency of all of them.)
+#
+# TARGET: vllm/model_executor/kernels/r4d_lib.py   (inside the pinned image)
+#   base    sha256 b8dfeefb4acc72b67b0579a4ab0653ef5d48c40cc8649187d757c80ce6b9617a
+#   patched sha256 0751852e1ca813ad2bdb1da7bf10c605130e63e04ef9ec57e70b5f675a29c08e
+#
+# APPLY: append the block below to the module. R4D_FREEZE_PROBES=0 restores stock
+# behaviour. Cost when nothing is compiled: ten one-time probes at import (measured:
+# logged once per process, e.g. "libr4d capability probes frozen: 10 warm [...]").
+#
+# OBSERVED EFFECT: with 01+02 the worker gets past this break and reaches the next one
+# (see 03) — dynamo reports "Dynamo bytecode transform time: 1.08 s" for the trunk.
+--- a/vllm/model_executor/kernels/r4d_lib.py
+++ b/vllm/model_executor/kernels/r4d_lib.py
+@@ -361,3 +361,71 @@
+ @functools.lru_cache
+ def moe_w4a16_mt1():
+     return getattr(lib(), "r4d_gemm_moe_w4a16_nt_b16_m64_mt1", None)
+
+
+# FORK-LOCAL (patch: trace-safe capability probes).  torch.compile traces the
+# BODY of a functools.lru_cache call even when the cache is warm, so a capability
+# probe called from a forward pass is a dynamo graph break.  Measured 2026-09-30
+# on the GLM-5.3-Flash trunk under @support_torch_compile:
+#   Dynamo does not know how to trace the builtin `posix.stat`
+#   r4d_mhc_supported -> r4d_mhc_enabled -> available -> os.path.exists
+# Each probe below is wrapped to resolve ONCE into a memo and is warmed at import
+# (outside any tracing region), so a traced call is a dict read.  The probe set is
+# every no-argument capability probe / symbol lookup reachable from a forward
+# pass: the W4A16 + mxfp4 MoE dispatch, the mHC op, the MLA sparse indexer gate
+# and the draft/lm-head guards.
+_PROBE_NAMES = (
+    "available",
+    "lib",
+    "has_w4a16_moe",
+    "moe_w4a16",
+    "moe_w4a16_mt1",
+    "has_mhc_dsv4",
+    "r4d_mhc_enabled",
+    "r4d_indexer_enabled",
+    "mqa_max_heads",
+    "r4d_act_quant_enabled",
+)
+
+
+def _freeze_probe(fn):
+    """Return fn bound to its first result.  Result must be a plain value."""
+
+    memo = []
+
+    def probe(*args, **kwargs):
+        if not memo:
+            memo.append(fn(*args, **kwargs))
+        return memo[0]
+
+    return probe
+
+
+def _freeze_probes() -> list[str]:
+    frozen = []
+    for name in _PROBE_NAMES:
+        fn = globals().get(name)
+        if fn is None:
+            continue
+        # Only the fast branch of a bound probe is traced; an empty memo means the
+        # first call still does I/O, so populate it here as well.
+        globals()[name] = _freeze_probe(fn)
+        frozen.append(name)
+    return frozen
+
+
+from vllm.logger import init_logger as _init_logger
+
+_probe_log = _init_logger(__name__)
+
+if os.environ.get("R4D_FREEZE_PROBES", "1") == "1":
+    _warmed, _failed = [], []
+    for _n in _freeze_probes():
+        try:
+            globals()[_n]()
+            _warmed.append(_n)
+        except Exception:  # noqa: BLE001 -- probe unavailable here; recompute lazily
+            _failed.append(_n)
+    _probe_log.info("libr4d capability probes frozen: %d warm %s%s",
+                    len(_warmed), _warmed,
+                    f" (deferred: {_failed})" if _failed else "")
@@ -0,0 +1,116 @@
+# glm53-compile-03-fullgraph-optout.patch — let torch.compile split the graph instead
+# of failing on a break.
+#
+# STATUS: WIP — third of three patches in the "compile the GLM trunk" changelist.
+# With 01+02+03 the model TRACES, then hits the NEXT blocker (below). The set does not
+# yet boot the engine; nothing in scripts/ mounts it.
+#
+# THE BREAK: with 01+02 mounted the worker dies with
+#
+#   RuntimeError: Graph breaks are not supported with aot compile.
+#   Please use torch.compile(fullgraph=True).
+#
+# because vllm/compilation/wrapper.py calls torch.compile(..., fullgraph=True, ...) and
+# this build has AOT compile on by default (VLLM_USE_AOT_COMPILE). The GLM trunk cannot
+# satisfy fullgraph: its indexer metadata path syncs to the host —
+#   models/glm5next/amd/sparse_indexer.py:263  int(positions[...].max().item()) + 1
+#   models/glm5next/amd/sparse_indexer.py:419  lmax = int(decode_metadata.decode_lens.max().item())
+# — and a .item() inside a captured graph is not expressible.
+#
+# WHAT WAS VERIFIED (in-image, torch 2.11.0+rocm10.0, torch.compile(backend="eager")):
+#
+#   fullgraph=True   suppress_errors=False  -> RAISE Unsupported
+#   fullgraph=True   suppress_errors=True   -> RAISE Unsupported   <-- suppress_errors
+#                                                                     does NOT cover breaks
+#   fullgraph=False  suppress_errors=False  -> ok   (traceable segments compiled by
+#                                                    inductor, the rest runs eagerly;
+#                                                    numerics unchanged)
+#
+# So breaks must be TOLERATED, not suppressed: the traceable mass still fuses. This
+# patch makes the fullgraph requirement opt-out via VLLM_COMPILE_FULLGRAPH=0 (default 1,
+# so every other model keeps stock behaviour), and the run also needs
+# VLLM_USE_AOT_COMPILE=0 because AOT compile refuses breaks outright.
+#
+# TARGET: vllm/compilation/wrapper.py   (inside the pinned image)
+#   base    sha256 969961dca41c78385212e985cfc0c57ef067f4a2d7490c73c658c600a1b95cac
+#   patched sha256 732363bb1a1367772f13c60a655c678e7e2e9e56186677b14238f5e84cac31b3
+#   (the file already imports os; the patch does not add a second import)
+#
+# NEXT BLOCKER — NOT FIXED HERE. With 01+02+03 and
+# R4D_GLM_COMPILE=1 VLLM_COMPILE_FULLGRAPH=0 VLLM_USE_AOT_COMPILE=0 the trunk traces
+# ("Dynamo bytecode transform time: 1.08 s") and then the worker dies during profile_run
+# on the decode forward:
+#
+#   KeyError: 'in_proj_qkvbfg_a'
+#     model.py:748  Glm5NextModel.forward        hidden_states, residual, post, comb = layer(...)
+#     model.py:456  Glm5NextDecoderLayer.forward
+#
+# `in_proj_qkvbfg_a` is the KDA layer's merged column-parallel projection
+# (models/glm5next/common/kda.py:217, called at :407) — and it is also a name in the
+# load_weights shard table (model.py:780-785), i.e. a real submodule attribute name,
+# not an arbitrary string.
+#
+# WHERE IT IS THROWN (established 2026-09-30 by reading the full traceback, since the
+# report line is inside dynamo-rewritten code): torch/_dynamo keeps a name -> module map
+# and looks it up by name — output_graph.py:1259 / :1384 register into `nn_modules`,
+# :3000 does `self.nn_modules[node.target]` (a bare-name KeyError when absent), and
+# :1225 documents stripping "_buffers[..]/_parameters[..]/_modules[..]" from those
+# names. So the compiler is failing to RESOLVE A SUBMODULE BY NAME while tracing the
+# GLM decoder layer. Whether the root cause is the fork's module tree (something
+# dynamo cannot register statically) or dynamo's own registration is NOT yet settled.
+#
+# STATUS OF THAT BLOCKER (updated 2026-09-30, after the diagnostics below):
+#   * TORCHDYNAMO_VERBOSE=1 adds NO internal stack trace -> it is a plain Python
+#     KeyError raised in a callee dynamo INLINED into Glm5NextDecoderLayer.forward
+#     (the reported line 456 is that method's `def` line, which is how inlined frames
+#     surface).
+#   * It is NOT raised from dynamo's four nn_modules lookups: the probe in
+#     /home/dank/dynamo-probe wraps three of them and stayed silent over two runs with
+#     the probe confirmed installed in the API-server and worker processes.
+#   * It is NOT a literal string lookup: the name appears tree-wide only as the KDA
+#     layer's attribute (kda.py:217, used at :407) and in the load_weights shard table
+#     (model.py:780-785). The four `proj_name = prefix.split(".")[-1]` sites in the
+#     quantization utils do compute exactly this bare key, but each is guarded
+#     (`if proj_name not in fused_mapping: return`), so none of them can raise it.
+#   * It reproduces on every run of this configuration (5 of 5).
+# REMAINING INSTRUMENT: name the inlined caller. Either `TORCH_LOGS=+dynamo` on a boot
+# and grep the log for the key (dynamo prints inlined frames with source), or wrap the
+# inlining candidates in that forward — the mHC CustomOps (hc_pre / hc_post /
+# hc_fused_post_pre in layers/mhc.py) and the indexer op — with a KeyError-reporting
+# probe. Until that names the caller this is a dynamo-side name-resolution failure.
+# NOTE the launcher gotcha that wasted a boot earlier: the scratch launcher now
+# preserves ENVARGS (`ENVARGS="${ENVARGS:-}"`), but it previously reset it, silently
+# dropping outer env vars.
+#
+# SECOND BLOCKER, KNOWN AND NOT FATAL IN THIS CONFIG: the indexer's in-forward
+# workspace acquisition (patch 05 makes it compiler-opaque) and the fused-norm
+# launcher cell (patch 04 removes the call) are real breaks, but with
+# VLLM_COMPILE_FULLGRAPH=0 they are tolerated — they only become fatal under
+# fullgraph=True, which this trunk can never satisfy anyway.
+--- a/vllm/compilation/wrapper.py
+++ b/vllm/compilation/wrapper.py
+@@ -19,6 +19,15 @@
+ 
+ logger = init_logger(__name__)
+ 
+
+# FORK-LOCAL (patch: fullgraph opt-out).  torch.compile(fullgraph=True) turns any
+# dynamo graph break into a hard failure; suppress_errors does NOT cover breaks
+# (verified in-image, torch 2.11+rocm10.0).  With VLLM_COMPILE_FULLGRAPH=0 the
+# traceable segments are still compiled by inductor and the untraceable ones run
+# eagerly, which is what a model with device-sync metadata (the GLM-5.3-Flash
+# indexer) needs.  Default 1 keeps stock behaviour for every other model.
+_FULLGRAPH = os.environ.get("VLLM_COMPILE_FULLGRAPH", "1") == "1"
+
+ R = TypeVar("R")
+ P = ParamSpec("P")
+ 
+@@ -146,7 +155,7 @@
+         with aot_context:
+             self._compiled_callable = torch.compile(
+                 compiled_ptr,
+-                fullgraph=True,
+                fullgraph=_FULLGRAPH,
+                 dynamic=False,
+                 backend=backend,
+                 options=options,
@@ -0,0 +1,112 @@
+# glm53-compile-04-fused-qkv-rmsnorm-optout.patch — take GLM's fused-norm JIT kernel
+# out of the traced path.
+#
+# STATUS: WIP — fourth of the "compile the GLM trunk" changelist. Needs 01 (the
+# decorator) to matter; not part of the deployed shape.
+#
+# MEASUREMENT NOTE — read the run history before trusting any claim below, because a
+# launcher bug made four consecutive boots misleading.  The scratch launcher
+# (serve_glm_autoround.sh) was restructured on 2026-09-30 and the
+# `-e VLLM_COMPILE_FULLGRAPH=0` line was DROPPED, so those boots mounted the wrapper
+# patch while `_FULLGRAPH` still defaulted to True.  With fullgraph=True *every* break
+# is an error, which is why the reported failure kept moving as breaks were removed:
+#
+#   fullgraph=False, AOT off  -> breaks tolerated; dies later on the dynamo name-map
+#                                `KeyError: 'in_proj_qkvbfg_a'`
+#   fullgraph=True,  AOT off  -> dies on whichever break comes first
+#                                (cell -> workspace -> compiler-disable)
+#
+# An earlier version of this note blamed `TORCH_LOGS=graph_breaks` for the difference.
+# That was WRONG — the flag only logs; the launcher silently ran fullgraph=True.
+# Launcher fixed; the env var is set again.
+#
+# What it means for THIS patch: under fullgraph=True the fused-norm call was the first
+# break dynamo reported, and after this patch the report moved on to the workspace
+# acquisition — so it removes a real break.  It is not by itself what unblocks a boot:
+# with breaks genuinely tolerated the blocker is the name-map KeyError.
+#
+# IT IS STILL THE RIGHT PATTERN TO REMOVE, either way: the launcher family
+# (@kernel_launcher, 39 call sites) is exactly the shape that produces breaks, and the
+# object this one closes over is bound after use (see below).
+#
+# THE BREAK (measured, TORCH_LOGS=graph_breaks, rank0, 2026-09-30).  With 01+02+03 the
+# trunk traces, and with AOT off + fullgraph=True dynamo reports a FATAL break — fatal
+# even with VLLM_COMPILE_FULLGRAPH=0, because dynamo cannot build a resume function
+# around it:
+#
+#   torch._dynamo.exc.Unsupported: Read uninitialized cell
+#   Explanation: Attempted to read a cell variable that has not been populated yet.
+#   Developer debug context: CellVariable()
+#
+# User stack:
+#   glm5next/common/model.py:748    Glm5NextModel.forward       layer(...)
+#   glm5next/common/model.py:526    Glm5NextDecoderLayer.forward
+#   glm5next/common/attention.py:590
+#   model_executor/layers/mla.py:208
+#     q_proj_input, kv_c_normed = fused_q_kv_rmsnorm(...)
+#   -> torch/_dynamo/polyfills/__init__.py:392 getattr_and_trace
+#
+# So dynamo is tracing the fork's JIT-kernel launcher framework
+# (VllmTritonJitKernel + @kernel_launcher — 39 call sites) and finds a closure name
+# that is not yet bound.  In models/common/ops/fused_qk_rmsnorm.py the kernel object
+# is bound AFTER the function that closes over it:
+#
+#   def fused_q_kv_rmsnorm(...):          # line 194
+#       ...
+#       _FUSED_Q_KV_RMSNORM_KERNEL(...)   # line 211
+#   _FUSED_Q_KV_RMSNORM_KERNEL = FusedQKVRMSNormKernel()    # line 223
+#
+# Two ways out, and this patch takes the cheap one:
+#   (a) WORKAROUND (here): stop calling the fused kernel from the traced region.  The
+#       flag is already a constructor argument (mla.py:80, default False) that glm5next
+#       turns ON at attention.py:583, so it becomes an env knob.
+#   (b) ROOT FIX (not done): make the launcher trace-safe for all 39 sites.  Any other
+#       @kernel_launcher-backed kernel still in the traced path can produce the same
+#       fatal break, so (b) is the durable fix and (a) is what unblocks iterating.
+#
+# COST of R4D_GLM_FUSE_QKV_RMSNORM=0: one extra small RMSNorm launch per MLA layer per
+# token (q_a_layernorm + kv_a_layernorm instead of one fused kernel) — ~2 launches out
+# of ~3,700 per token on this shape, against the 44 % of device time currently spent in
+# unfused pointwise work.  Correct trade for a compile experiment.
+#
+# TARGET: vllm/models/glm5next/common/attention.py   (inside the pinned image)
+#   base    sha256 362bbf0e6c50cdfa446c6551b9b1ddff425e08f88e3d4140fd8db88f8a619e0c
+#   patched sha256 84969993c250e9f8183a33a6e1ddd607272f48b950b560272eb33b7cdde31189
+#
+# APPLY: `patch -p1` from the vllm package root (verified: applies clean, reproduces
+# the patched sha256), or bind-mount new/.../attention.py over site-packages.
+# R4D_GLM_FUSE_QKV_RMSNORM=1 (default) restores the fused kernel.
+--- a/vllm/models/glm5next/common/attention.py
+++ b/vllm/models/glm5next/common/attention.py
+@@ -1,6 +1,7 @@
+ # SPDX-License-Identifier: Apache-2.0
+ # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+ 
+import os
+ import torch
+ import torch.nn.functional as F
+ from torch import nn
+@@ -36,6 +37,14 @@
+ from vllm.utils.deep_gemm import PAGED_MQA_PAGE_SIZES
+ from vllm.v1.kv_cache_interface import KpoolTailSpec, MLAAttentionSpec
+ 
+
+# FORK-LOCAL (patch: keep the traced region trace-safe).  R4D_GLM_FUSE_QKV_RMSNORM=0
+# replaces the fused `fused_q_kv_rmsnorm` JIT-kernel launch (one kernel) with two plain
+# RMSNorms, because dynamo cannot trace the launcher's closure ("Read uninitialized cell",
+# fatal even with fullgraph=False).  The fused kernel object in
+# models/common/ops/fused_qk_rmsnorm.py is bound AFTER the function that uses it.
+_FUSE_QKV_RMSNORM = os.environ.get("R4D_GLM_FUSE_QKV_RMSNORM", "1") == "1"
+
+ logger = init_logger(__name__)
+ 
+ # Shared torch.compile config for the indexer's small-kernel leaves. The MLA
+@@ -580,7 +589,7 @@
+             quant_config,
+             prefix,
+             skip_topk=False,
+-            fuse_qkv_rmsnorm=True,
+            fuse_qkv_rmsnorm=_FUSE_QKV_RMSNORM,
+         )
+ 
+     def forward(
@@ -0,0 +1,96 @@
+# glm53-compile-05-workspace-acquisition-opaque.patch — the indexer's workspace
+# acquisition must not be traced.
+#
+# STATUS: WIP — fifth of the "compile the GLM trunk" changelist. This is a break worth
+# removing, but NOT part of the confirmed blocker and NOT required for the
+# break-tolerant route: `torch.compiler.disable` is itself a graph break, so this patch
+# helps only a hypothetical fullgraph=True route, where the workspace acquisition would
+# be better RESTRUCTURED out of the forward (allocate at init/profile time) than made
+# opaque. Kept because the discovery is real and the wrapper is a no-op when nothing is
+# compiling.
+#
+# THE BREAK — CONFIRMED with AND WITHOUT the logging flag (2026-09-30):
+#
+#   torch._dynamo.exc.Unsupported: Unsupported method call
+#   user stack
+#     glm5next/common/model.py:748      Glm5NextModel.forward
+#     glm5next/common/model.py:526      Glm5NextDecoderLayer.forward
+#     glm5next/common/attention.py:599 / :401
+#     glm5next/amd/sparse_indexer.py:711  forward_native
+#     glm5next/amd/sparse_indexer.py:132  sparse_attn_indexer_kpool
+#     v1/worker/workspace.py:135          get_simultaneous
+#     v1/worker/workspace.py:155          _ensure_workspace_size
+#
+# Line 132 is the PROFILE-RUN branch — the source comment says what it is:
+#
+#     if not isinstance(attn_metadata, dict):
+#         # Reserve workspace for indexer during profiling run
+#
+# so the engine dies exactly where it sizes its memory.  `_ensure_workspace_size`
+# reads ContextVar and ubatch state (`_workspace_lane.get()`,
+# `dbo_current_ubatch_id()`) and may allocate; none of that is traceable.
+#
+# FIX: the documented escape hatch — `torch.compiler.disable` around a thin wrapper,
+# so dynamo stops analysing the call and executes it eagerly from inside the compiled
+# region.  Chosen over decorating the workspace manager's own methods because that
+# would change workspace behaviour for every model, not just this one.  The wrapper is
+# a no-op when nothing is compiling, so it needs no env gate and is safe in the
+# stock (uncompiled) path — which is why this patch can be reasoned about
+# independently of whether the changelist ever ships.
+#
+# CAVEAT, stated plainly: a compiler-disabled region is still a graph break, so this
+# needs patch 03 (VLLM_COMPILE_FULLGRAPH=0) to be tolerable at all, and whether dynamo
+# can build a resume function around it is exactly what the next boot decides — the
+# difference between this failure and 04's `Read uninitialized cell`.
+#
+# TARGET: vllm/models/glm5next/amd/sparse_indexer.py   (inside the pinned image)
+#   base    sha256 f773a99265d3f440c8fad22c8fb5a8b33f926162094cffc3a2ecdaa42763c5b1
+#   patched sha256 c154115afe8253a35c652b7198929ef4c1874957d333fdb3340cacf88e612b78
+#
+# APPLY: `patch -p1` from the vllm package root, or bind-mount over site-packages.
+--- a/vllm/models/glm5next/amd/sparse_indexer.py
+++ b/vllm/models/glm5next/amd/sparse_indexer.py
+@@ -35,6 +35,21 @@
+ from vllm.v1.attention.ops.common import pack_seq_triton, unpack_seq_triton
+ from vllm.v1.worker.workspace import current_workspace_manager
+ 
+@torch.compiler.disable
+def _acquire_workspace(*shapes_and_dtypes):
+    """Acquire shared workspace buffers, untraced.
+
+    Workspace acquisition reads ContextVar/ubatch state and may allocate
+    (v1/worker/workspace.py: ``_ensure_workspace_size`` -> ``_workspace_lane.get()``,
+    ``dbo_current_ubatch_id()``), which dynamo cannot trace -- it raises
+    ``Unsupported: Unsupported method call`` and kills the engine during the profile
+    run.  ``torch.compiler.disable`` keeps the call as an opaque, eagerly-executed
+    region inside the compiled graph.  No-op when not compiling.
+    """
+
+    return current_workspace_manager().get_simultaneous(*shapes_and_dtypes)
+
+
+ logger = init_logger(__name__)
+ 
+ # kpool write helper: form pools from the current token batch and compress them
+@@ -129,7 +144,7 @@
+         values_spec, scales_spec = _gather_workspace_shapes(
+             total_seq_lens, head_dim, fp8_dtype, use_fp4_cache
+         )
+-        current_workspace_manager().get_simultaneous(
+        _acquire_workspace(
+             values_spec,
+             scales_spec,
+             ((RADIX_TOPK_WORKSPACE_SIZE,), torch.uint8),
+@@ -275,11 +290,10 @@
+         # Layout switches between FP8 (head_dim bytes + 4-byte fp32 scale) and
+         # MXFP4 (head_dim/2 bytes packed + head_dim/MXFP4_BLOCK_SIZE ue8m0
+         # scales) based on use_fp4_cache.
+-        workspace_manager = current_workspace_manager()
+         values_spec, scales_spec = _gather_workspace_shapes(
+             total_seq_lens, head_dim, fp8_dtype, use_fp4_cache
+         )
+-        k_quant_full, k_scale_full = workspace_manager.get_simultaneous(
+        k_quant_full, k_scale_full = _acquire_workspace(
+             values_spec,
+             scales_spec,
+         )
@@ -0,0 +1,173 @@
+# glm53-dflash-01-aux-hidden-states.patch — give the GLM trunk the EAGLE3 / DFlash2
+# auxiliary-hidden-state interface.
+#
+# STATUS: WIP, and INDEPENDENT of the compile changelist (01-05). It patches the same
+# file as glm53-compile-01-model-decorator.patch, so if both are ever applied, apply 01
+# first. Not part of the deployed shape.
+#
+# WHY.  `--speculative-config '{"method":"dflash","model":"/path/to/drafter", ...}'` with
+# incoai/GLM-5.3-Flash-DFlash2 is accepted -- the config parses, the drafter loads -- and
+# then the worker dies at init with:
+#
+#     RuntimeError: Model does not support EAGLE3 interface
+#
+# raised by v1/worker/gpu/spec_decode/eagle/eagle3_utils.py:20 through
+# `supports_eagle3(model)` == `isinstance(model, SupportsEagle3)` (a runtime-checkable
+# Protocol).  Nothing about the drafter, the method name or the payload is wrong:
+# `glm5next` implements neither half of the contract.
+#
+# THE CONTRACT (read out of this image, not assumed):
+#   * DEFINITIVE (v5, after four failed boots): the member set is checked against
+#     the INSTANCE the runner calls -- which for GLM-5.3-Flash is
+#     Glm5NextForConditionalGeneration, NOT Glm5NextForCausalLM.  The multimodal
+#     wrapper wraps the text model as self.language_model
+#     (init_vllm_registered_model, architectures=["Glm5NextForCausalLM"],
+#     glm5next/common/model.py:1113), and model_runner.py:425 calls
+#     set_eagle3_aux_hidden_state_layers(self.model) on the CONDITIONAL.
+#     Presence rule (verified in-image with stock typing, python 3.14):
+#     isinstance walks ALL __protocol_attrs__ -- members of SupportsEagle3 AND
+#     its base SupportsEagleBase (has_own_lm_head, has_own_embed_tokens) -- and
+#     getattr_static()s each one off the instance, breaking on the first miss.
+#     So BOTH wrappers need all three attributes.
+#   * MRO law: INHERITING SupportsEagle3 as a base dies at class creation with
+#     "Cannot create a consistent method resolution order (MRO) for bases ..."
+#     (GLM's wrappers already carry four/five Protocol-derived bases).  The
+#     attribute route avoids the MRO entirely.  Also proven by simulation:
+#     the four-base ForCausalLM ordering cannot take SupportsEagle3 in ANY
+#     position.
+#   * Method placement: without inheriting, the Protocol's default
+#     set_aux_hidden_state_layers / get_eagle3_default_aux_hidden_state_layers
+#     do not exist on either wrapper, so both are defined explicitly --
+#     ForCausalLM unwraps to self.model (the EagleModelMixin trunk, v1's
+#     correct bodies), the conditional delegates to self.language_model
+#     (the pixtral.py:498 pattern).
+#   * interfaces.py:1673-1719 -- for reference, the inherited defaults unwrap
+#     `holder = getattr(parent_ref, "model", parent_ref)` and assert
+#     `isinstance(holder, EagleModelMixin)`, defaulting to (2, L // 2, L - 3).
+#   * eagle3_utils.py:26-31 -- layers come from the spec config when it carries them
+#     (this drafter ships target_layer_ids [5, 14, 24, 33, 42]), else from the default.
+#   * deepseek_v2.py:1551-1585 and v1/worker/gpu/cudagraph_utils.py:621 -- the trunk
+#     appends one tensor per aux layer while looping and returns the TUPLE
+#     (hidden_states, aux_hidden_states) only when that list is non-empty.
+#
+# WHAT IT ADDS for GLM: EagleModelMixin on Glm5NextModel; aux capture in the layer
+# loop, where the aux state is the single-stream `hidden_states` ENTERING that layer
+# (GLM's loop carries the already-summed single-stream state -- its non-mHC branch
+# re-splits `residual = hidden_states` inside the layer -- so this is the equivalent
+# of deepseek_v2's `hidden_states + residual`); the conditional tuple return; the
+# three SupportsEagle3 member attributes on BOTH wrappers; and the two interface
+# methods on both (ForCausalLM: unwrap to trunk; conditional: delegate to
+# language_model).
+#
+# HONEST SCOPE: this makes the interface PRESENT, which is exactly what the engine
+# checks -- it does not make drafts GOOD, and per the measured profile it is not a
+# decode-speed fix (a verification pass pays pointwise cost proportional to its token
+# count, which is why MTP4 measured flat).  It is the prerequisite for driving DFlash2
+# on this endpoint at all.
+#
+# TARGET: vllm/models/glm5next/common/model.py   (inside the pinned image)
+#   base    sha256 008eeaa2a07e091465319f86900b5a321d921238723ccbe22e6b9e971e1d2b28
+#   patched sha256 bc9913a15c8e7f88e0c36b07bf029b383cc8c2078b58d66dc91aa235f8af77f4
+#
+# APPLY: `patch -p1` from the vllm package root, or bind-mount the patched file over
+# site-packages (which is what a scratch-launcher DFLASH_PATCH=1 does).
+--- a/vllm/models/glm5next/common/model.py
+++ b/vllm/models/glm5next/common/model.py
+@@ -95,6 +95,12 @@
+ logger = init_logger(__name__)
+ 
+ 
+from vllm.model_executor.models.interfaces import (
+    EagleModelMixin,
+    SupportsEagle3,
+)
+
+
+ class Glm5NextMLP(nn.Module):
+     def __init__(
+         self,
+@@ -623,7 +629,7 @@
+         )
+ 
+ 
+-class Glm5NextModel(nn.Module):
+class Glm5NextModel(nn.Module, EagleModelMixin):
+     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+         super().__init__()
+ 
+@@ -730,7 +736,14 @@
+         if self.is_sequence_parallel:
+             hidden_states = sp_shard(hidden_states)
+ 
+-        for layer in self._active_layers:
+        aux_hidden_states: list[torch.Tensor] = []
+        for local_idx, layer in enumerate(self._active_layers):
+            if (getattr(self, "start_layer", 0) + local_idx) in self.aux_hidden_state_layers:
+                # Auxiliary hidden state for EAGLE-3 / DFlash2 drafts: the
+                # single-stream state ENTERING this layer, before its input norm --
+                # the same semantic deepseek_v2 implements by appending
+                # hidden_states + residual at this point in its loop.
+                aux_hidden_states.append(hidden_states)
+             hidden_states, residual, post, comb = layer(
+                 positions, hidden_states, residual, post, comb
+             )
+@@ -749,6 +762,8 @@
+             hidden_states = sp_all_gather(hidden_states)[:full_num_tokens]
+ 
+         hidden_states = self.norm(hidden_states)
+        if len(aux_hidden_states) > 0:
+            return hidden_states, aux_hidden_states
+         return hidden_states
+ 
+     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+@@ -937,6 +952,10 @@
+ class Glm5NextForCausalLM(
+     nn.Module, HasInnerState, SupportsPP, MixtureOfExperts, IsHybrid
+ ):
+    supports_eagle3: ClassVar[Literal[True]] = True
+    has_own_lm_head: bool = False
+    has_own_embed_tokens: bool = False
+
+     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+         super().__init__()
+         self.model_config = vllm_config.model_config
+@@ -963,6 +982,14 @@
+     def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+         return self.model.embed_input_ids(input_ids)
+ 
+    def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
+        # EAGLE-3 / DFlash2 interface: unwrap to the trunk (the EagleModelMixin).
+        self.model.aux_hidden_state_layers = layers
+
+    def get_eagle3_default_aux_hidden_state_layers(self) -> tuple[int, ...]:
+        num_layers = len(self.model.layers)
+        return (2, num_layers // 2, num_layers - 3)
+
+     def forward(
+         self,
+         input_ids: torch.Tensor | None,
+@@ -1033,6 +1060,10 @@
+ class Glm5NextForConditionalGeneration(
+     Glm4vForConditionalGeneration, HasInnerState, IsHybrid, MixtureOfExperts
+ ):
+    supports_eagle3: ClassVar[Literal[True]] = True
+    has_own_lm_head: bool = False
+    has_own_embed_tokens: bool = False
+
+     # The text model (KDA + dense-MLA + MoE) is a hybrid mamba model. The
+     # multimodal wrapper must declare the same interfaces so vLLM treats it as
+     # hybrid (auto-aligns mamba/attention block sizes, sizes the mamba state
+@@ -1076,6 +1107,13 @@
+ 
+         return Glm5NextForCausalLM.get_mamba_state_copy_func()
+ 
+    def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
+        # EAGLE-3 / DFlash2: delegate to the language model (pixtral.py:498).
+        self.language_model.set_aux_hidden_state_layers(layers)
+
+    def get_eagle3_default_aux_hidden_state_layers(self) -> tuple[int, ...]:
+        return self.language_model.get_eagle3_default_aux_hidden_state_layers()
+
+     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+         super(Glm4vForConditionalGeneration, self).__init__()
+         config = vllm_config.model_config.hf_config
@@ -0,0 +1,317 @@
+--- a/vllm/v1/core/kv_cache_utils.py
+++ b/vllm/v1/core/kv_cache_utils.py
+@@ -1210,17 +1210,38 @@
+         for name, spec in kv_cache_spec.items()
+         if isinstance(spec, KpoolTailSpec)
+     }
+    hidden_specs = {
+        name: spec
+        for name, spec in kv_cache_spec.items()
+        if isinstance(spec, HiddenStateCacheSpec)
+    }
+     attn_specs = {
+         name: spec
+         for name, spec in kv_cache_spec.items()
+-        if not isinstance(spec, (MambaSpec, KpoolTailSpec))
+        if not isinstance(
+            spec, (MambaSpec, KpoolTailSpec, HiddenStateCacheSpec)
+        )
+     }
+-    if not mamba_specs or not all(
+-        type(spec) is MLAAttentionSpec for spec in attn_specs.values()
+-    ):
+    if not mamba_specs or not attn_specs:
+        return None
+    # FORK-LOCAL (patches/glm53-dflash-04-kv-groups-draft): with a dflash
+    # drafter, attn_specs also carries the drafter's attention layers. The
+    # GLM MLA layers keep their own group (no page unification -- per-group
+    # pages are the glm5 design); the drafter layers must form one uniform
+    # group of their own, and the hidden-state layers get per-layer groups
+    # like the generic tail.
+    mla_specs = {
+        name: spec
+        for name, spec in attn_specs.items()
+        if type(spec) is MLAAttentionSpec
+    }
+    draft_specs = {
+        name: spec for name, spec in attn_specs.items() if name not in mla_specs
+    }
+    if not mla_specs:
+        return None
+    if draft_specs and UniformTypeKVCacheSpecs.from_specs(draft_specs) is None:
+         return None
+-
+-    mla_specs = cast(dict[str, MLAAttentionSpec], attn_specs)
+     idx_pages = {
+         spec.page_size_bytes for spec in mla_specs.values() if spec.tokens_per_state > 1
+     }
+@@ -1233,7 +1254,7 @@
+     mla_pages = {mla_specs[name].page_size_bytes for name in mla_names}
+     assert len(mla_pages) == 1
+     mla_page = mla_pages.pop()
+-    uniform_spec = UniformTypeKVCacheSpecs.from_specs(attn_specs)
+    uniform_spec = UniformTypeKVCacheSpecs.from_specs(mla_specs)
+     assert uniform_spec is not None
+ 
+     tail_group: KVCacheGroupSpec | None = None
+@@ -1270,10 +1291,20 @@
+     for index, name in enumerate(mamba_specs):
+         mamba_grouped_names[index % num_groups].append(name)
+ 
+    draft_group: KVCacheGroupSpec | None = None
+    if draft_specs:
+        draft_uniform = UniformTypeKVCacheSpecs.from_specs(draft_specs)
+        assert draft_uniform is not None
+        draft_group = KVCacheGroupSpec(list(draft_specs), draft_uniform)
+    hidden_groups = [
+        KVCacheGroupSpec([name], spec) for name, spec in hidden_specs.items()
+    ]
+     return (
+-        [KVCacheGroupSpec(list(attn_specs), uniform_spec)]
+        [KVCacheGroupSpec(list(mla_specs), uniform_spec)]
+         + ([tail_group] if tail_group is not None else [])
+        + ([draft_group] if draft_group is not None else [])
+         + create_kv_cache_group_specs(padded_specs, mamba_grouped_names)
+        + hidden_groups
+     )
+ 
+ 
+@@ -1289,10 +1320,17 @@
+         int,
+         list[str],
+         int,
+        KVCacheGroupSpec | None,
+        list[KVCacheGroupSpec],
+     ]
+     | None
+ ):
+-    """Recognize the GLM-5.3-Flash grouping after optional PP projection."""
+    """Recognize the GLM-5.3-Flash grouping after optional PP projection.
+
+    FORK-LOCAL (patches/glm53-dflash-04-kv-groups-draft): also recognizes the
+    drafter's uniform group and the per-layer hidden-state groups, returned
+    as ``draft_group`` / ``hidden_groups`` so the pool builder and the memory
+    estimator can size their tensors beside the aliased GLM pool."""
+     uniform_groups = [
+         group
+         for group in kv_cache_groups
+@@ -1301,17 +1339,28 @@
+     mamba_groups = [
+         group for group in kv_cache_groups if isinstance(group.kv_cache_spec, MambaSpec)
+     ]
+    hidden_groups = [
+        group
+        for group in kv_cache_groups
+        if isinstance(group.kv_cache_spec, HiddenStateCacheSpec)
+    ]
+     attn_group: KVCacheGroupSpec | None = None
+     tail_group: KVCacheGroupSpec | None = None
+    draft_group: KVCacheGroupSpec | None = None
+     for group in uniform_groups:
+         inner = cast(UniformTypeKVCacheSpecs, group.kv_cache_spec).kv_cache_specs
+         if all(type(spec) is MLAAttentionSpec for spec in inner.values()):
+             attn_group = group
+         elif all(isinstance(spec, KpoolTailSpec) for spec in inner.values()):
+             tail_group = group
+        else:
+            draft_group = group
+     if attn_group is None or not mamba_groups:
+         return None
+-    if len(uniform_groups) + len(mamba_groups) != len(kv_cache_groups):
+    if (
+        len(uniform_groups) + len(mamba_groups) + len(hidden_groups)
+        != len(kv_cache_groups)
+    ):
+         return None
+ 
+     attn_uniform = cast(UniformTypeKVCacheSpecs, attn_group.kv_cache_spec)
+@@ -1362,6 +1411,8 @@
+         idx_page,
+         tail_names,
+         tail_page,
+        draft_group,
+        hidden_groups,
+     )
+ 
+ 
+@@ -1395,10 +1446,34 @@
+         return kv_cache_spec
+ 
+     max_page_size = max(page_sizes)
+    # GLM-5.3-Flash + a draft model: the sparse indexer k_cache page is
+    # tokens_per_state x the MLA page (same per-token bytes), and MLA layers
+    # cannot be PADDED to an arbitrary page size (sparse MLA indexes the cache
+    # in whole token rows) -- but they CAN grow their block size when the max
+    # page is an exact multiple of the MLA page. Bump the max page to a
+    # multiple of every MLA-family page so all of them take the divisible
+    # block-growth branch below instead of the padding refusal, while the
+    # draft and mamba layers simply pad to the (slightly larger) page.
+    mla_pages = {
+        spec.page_size_bytes
+        for spec in kv_cache_spec.values()
+        if isinstance(spec, MLAAttentionSpec)
+    }
+    if mla_pages:
+        mla_common = math.lcm(*mla_pages)
+        max_page_size = -(-max_page_size // mla_common) * mla_common
+     new_kv_cache_spec = {}
+     for layer_name, layer_spec in kv_cache_spec.items():
+         if layer_spec.page_size_bytes == max_page_size:
+             new_kv_cache_spec[layer_name] = layer_spec
+        elif isinstance(layer_spec, KpoolTailSpec):
+            # The kpool tail kernel pins the tail block to the kpool size
+            # (tail_kv_cache.view(...) must be [blocks, 2, index_kpool,
+            # head_dim]); growing block_size breaks that contract. Pad the
+            # physical page instead -- the tail is its own group and the pad
+            # wastes bytes only.
+            new_spec = replace(layer_spec, page_size_padded=max_page_size)
+            new_kv_cache_spec[layer_name] = new_spec
+         elif isinstance(layer_spec, MambaSpec):
+             # MambaSpec's page size is determined by its state shapes and does
+             # not scale with block_size, so pad the page instead. This is the
+@@ -1593,8 +1668,31 @@
+ ) -> int:
+     """Return the largest cache group's bytes per block."""
+     if (glm5_layout := _glm5_next_tensor_layout(kv_cache_groups)) is not None:
+-        _, _, mla_names, idx_names, mla_page, idx_page, _, _ = glm5_layout
+-        return len(mla_names) * mla_page + len(idx_names) * idx_page
+        (
+            _,
+            _,
+            mla_names,
+            idx_names,
+            mla_page,
+            idx_page,
+            _,
+            _,
+            draft_group,
+            hidden_groups,
+        ) = glm5_layout
+        bytes_per_block = len(mla_names) * mla_page + len(idx_names) * idx_page
+        if draft_group is not None:
+            d_inner = cast(
+                UniformTypeKVCacheSpecs, draft_group.kv_cache_spec
+            ).kv_cache_specs
+            bytes_per_block += sum(
+                spec.page_size_bytes for spec in d_inner.values()
+            )
+        for group in hidden_groups:
+            bytes_per_block += (
+                group.kv_cache_spec.page_size_bytes * len(group.layer_names)
+            )
+        return bytes_per_block
+ 
+     bytes_per_block = max(
+         sum(
+@@ -1701,12 +1799,30 @@
+             idx_page,
+             tail_names,
+             _,
+            draft_group,
+            hidden_groups,
+         ) = glm5_layout
+         bytes_per_block = len(mla_names) * mla_page + len(idx_names) * idx_page
+        extra_bytes_per_block = 0
+        if draft_group is not None:
+            d_inner = cast(
+                UniformTypeKVCacheSpecs, draft_group.kv_cache_spec
+            ).kv_cache_specs
+            extra_bytes_per_block += sum(
+                spec.page_size_bytes for spec in d_inner.values()
+            )
+        for group in hidden_groups:
+            extra_bytes_per_block += (
+                group.kv_cache_spec.page_size_bytes * len(group.layer_names)
+            )
+         num_blocks = may_override_num_blocks(
+-            vllm_config, available_memory // bytes_per_block
+            vllm_config,
+            available_memory // (bytes_per_block + extra_bytes_per_block),
+         )
+-        size = bytes_per_block * num_blocks
+        # One backing allocation shared by every tensor (allocate_kv_cache
+        # asserts all tensor sizes equal): the aliased GLM pool plus the
+        # draft/hidden regions appended after it.
+        size = (bytes_per_block + extra_bytes_per_block) * num_blocks
+         attn_specs = cast(
+             UniformTypeKVCacheSpecs, attn_group.kv_cache_spec
+         ).kv_cache_specs
+@@ -1745,6 +1861,37 @@
+                 ).kv_cache_specs
+                 add_tensor(tail_name, tail_specs[tail_name], offset)
+ 
+        # Draft + hidden-state layers get their own tensors appended after
+        # the aliased pool (no aliasing: independent block tables).
+        offset_cursor = bytes_per_block * num_blocks
+        if draft_group is not None:
+            d_inner = cast(
+                UniformTypeKVCacheSpecs, draft_group.kv_cache_spec
+            ).kv_cache_specs
+            d_page = next(iter(d_inner.values())).page_size_bytes
+            kv_cache_tensors.append(
+                KVCacheTensor(
+                    size=size,
+                    layers=list(draft_group.layer_names),
+                    layer_stride=d_page * num_blocks,
+                    block_stride=d_page,
+                    offset=offset_cursor,
+                )
+            )
+            offset_cursor += d_page * num_blocks * len(draft_group.layer_names)
+        for group in hidden_groups:
+            h_page = group.kv_cache_spec.page_size_bytes
+            kv_cache_tensors.append(
+                KVCacheTensor(
+                    size=size,
+                    layers=list(group.layer_names),
+                    layer_stride=h_page * num_blocks,
+                    block_stride=h_page,
+                    offset=offset_cursor,
+                )
+            )
+            offset_cursor += h_page * num_blocks * len(group.layer_names)
+
+         return KVCacheConfig(
+             num_blocks=num_blocks,
+             kv_cache_tensors=kv_cache_tensors,
+@@ -2440,6 +2587,8 @@
+             idx_page,
+             tail_names,
+             _,
+            draft_group,
+            hidden_groups,
+         ) = glm5_layout
+         uniform_spec = cast(UniformTypeKVCacheSpecs, attn_group.kv_cache_spec)
+         total_blocks = uniform_spec.max_memory_usage_pages(vllm_config)
+@@ -2452,7 +2601,30 @@
+         )
+         if tail_names:
+             total_blocks += 1
+-        return total_blocks * (len(mla_names) * mla_page + len(idx_names) * idx_page)
+        pool_bytes = total_blocks * (
+            len(mla_names) * mla_page + len(idx_names) * idx_page
+        )
+        extra_bytes = 0
+        if draft_group is not None:
+            d_inner = cast(
+                UniformTypeKVCacheSpecs, draft_group.kv_cache_spec
+            ).kv_cache_specs
+            d_spec = next(iter(d_inner.values()))
+            d_blocks = cdiv(
+                d_spec.max_memory_usage_bytes(vllm_config), d_spec.page_size_bytes
+            )
+            extra_bytes += d_blocks * d_spec.page_size_bytes * len(
+                draft_group.layer_names
+            )
+        for group in hidden_groups:
+            h_spec = group.kv_cache_spec
+            h_blocks = cdiv(
+                h_spec.max_memory_usage_bytes(vllm_config), h_spec.page_size_bytes
+            )
+            extra_bytes += (
+                h_blocks * h_spec.page_size_bytes * len(group.layer_names)
+            )
+        return pool_bytes + extra_bytes
+ 
+     bytes_per_block = _pool_bytes_per_block(kv_cache_groups)
+     total_blocks = 0
@@ -0,0 +1,31 @@
+--- a/vllm/model_executor/offloader/expert_activation_registry.py
+++ b/vllm/model_executor/offloader/expert_activation_registry.py
+@@ -316,7 +316,27 @@
+ 
+ 
+ def _glm53_kv_bytes_per_token(cfg: "VllmConfig") -> int:
+-    return 13600
+    per_tok = 13600
+    # FORK-LOCAL (patches/glm53-dflash-04-kv-groups-draft): with a dflash/dspark
+    # drafter, the KV spec set gains the drafter's attention layers and the aux
+    # hidden-state cache (one bf16 hidden vector per target aux layer per
+    # token). Both are full-length per-token costs the aliased glm5 pool
+    # cannot absorb, so the reserve must include them. The drafter is a 5-layer
+    # qwen3 (8 kv heads x 128 x K+V x bf16 = 4096 B/token); the aux count is
+    # the dflash target_layer_ids count (5).
+    # NOTE: use_aux_hidden_state_outputs lives on the gpu_model_runner
+    # instance (gpu_model_runner.py:623/678), not on the spec config -- it is
+    # not set yet when the expert plan runs. The spec-config METHOD is the
+    # plan-time signal for the same fact.
+    sc = getattr(cfg, "speculative_config", None)
+    if sc is not None and getattr(sc, "method", "") in (
+        "dflash",
+        "dspark",
+        "extract_hidden_states",
+    ):
+        hidden = int(cfg.model_config.hf_config.hidden_size)
+        per_tok += 5 * hidden * 2 + 4096
+    return per_tok
+ 
+ 
+ for _arch in ("Glm5NextForConditionalGeneration", "Glm5NextForCausalLM"):
@@ -0,0 +1,65 @@
+# gpu_worker-kv-cap.patch — fork-local fix for the expert-offload KV cap.
+#
+# STATUS: NOT part of the deployed frank-primary shape. It exists because the
+# fork's plan under-models this model's KV, and it is the enabler for a future
+# MTP attempt (the production GLM shape runs without MTP and boots fine on the
+# plan's own reserve).
+#
+# THE BUG: the plan models KV as 13600 B/token + a small mamba term, but
+# GLM-5.3-Flash also carries a per-sequence KDA state (34 linear-attention
+# layers) and, under MTP, the draft model's KV. gpu_worker then CAPS the KV
+# cache at that reserve, so the plan's own arithmetic can make a legal
+# --max-model-len unbootable. Measured 2026-09-30 on the pinned image:
+#   - no MTP, max_model_len 131072: reserve 1.74 GiB >= need 1.86 GiB  -> boots
+#   - MTP4,  max_model_len 131072: reserve 1.74 GiB <  need 2.07 GiB  -> refuses
+#   - MTP4,  max_model_len 262144: reserve 3.43 GiB <  need 3.62 GiB  -> refuses
+#     ("To serve at least one request with the model's max seq len ...")
+# Raising max_model_len does not fix it: the cap's slope (~13,845 B/token) only
+# overtakes the need's (~12,700 B/token + 0.52 GiB fixed) around 440k tokens,
+# where the KV reserve would eat a third of the expert arena.
+#
+# THE FIX: let an explicit operator KV request raise the cap. The compile and
+# graph-capture headroom still comes off the top, so the blast radius is one
+# flag the operator chose.
+#
+# APPLY (into a host dir that is then bind-mounted read-only over the image):
+#   IMAGE=docker.io/tcclaviger/vllm@sha256:da4029666668b90a0386c1ed23ea370c79916ad4e1a6ad1dd9db098b9fa069ea
+#   D=/home/dank/glm-patch; mkdir -p $D/vllm/v1/worker
+#   podman run --rm --entrypoint cat $IMAGE \
+#     /opt/vllm/lib/python3.14/site-packages/vllm/v1/worker/gpu_worker.py \
+#     > $D/vllm/v1/worker/gpu_worker.py
+#   ( cd $D && patch -p1 < .../scripts/patches/gpu_worker-kv-cap.patch )
+#   # then add to the podman run:
+#   #   -v $D/vllm/v1/worker/gpu_worker.py:/opt/vllm/lib/python3.14/site-packages/vllm/v1/worker/gpu_worker.py:ro
+#   #   -e CLAV_KV_RESERVE_FLOOR_GIB=<needed>   (or pass --kv-cache-memory-bytes)
+#
+# Base file sha256 (the pristine file out of the pinned image):
+#   1f3942b1a01af075317161b06077a2e658086f1da0566c9f297129ceebe37dd7
+# Patched file sha256 (verified 2026-09-30, patch applies clean and reproduces
+# these bytes exactly):
+#   80cd160bbc01270b0851f7b6fe7552b06ae2fae2428134fc29238d7f452af840
+#
+--- a/vllm/v1/worker/gpu_worker.py
+++ b/vllm/v1/worker/gpu_worker.py
+@@ -824,6 +824,21 @@
+         expert_cfg = self.vllm_config.offload_config.expert
+         if expert_cfg.enable_expert_offload and expert_cfg.planned_kv_reserve_bytes > 0:
+             cap = int(expert_cfg.planned_kv_reserve_bytes)
+            # LOCAL PATCH (GLM-5.3 KDA + MTP draft): the plan models KV as
+            # 13600 B/token + a small mamba term, but this model also carries a
+            # per-sequence KDA state and (under MTP) the draft's KV, so the
+            # reserve can sit BELOW what max_model_len legitimately needs -- and
+            # the cap then makes a legal config unbootable (measured: need
+            # 2.07 GiB vs cap 1.74 GiB at max_model_len 131072 with MTP). An
+            # explicit --kv-cache-memory-bytes, or CLAV_KV_RESERVE_FLOOR_GIB,
+            # is the operator's word: let it raise the cap. The compile/capture
+            # headroom still comes off the top.
+            _explicit = self.cache_config.kv_cache_memory_bytes
+            _floor = os.environ.get("CLAV_KV_RESERVE_FLOOR_GIB")
+            if _explicit:
+                cap = max(cap, int(_explicit))
+            if _floor:
+                cap = max(cap, int(float(_floor) * (1 << 30)))
+             if self.available_kv_cache_memory_bytes > cap:
+                 logger.info_once(
+                     "expert offload: KV capped at the plan's reserve %s GiB; %s GiB of "
@@ -0,0 +1,453 @@
+{%- set template_version = "qwen3.8-froggeric-v22.4.0" %}
+{%- set _tool_format = tool_call_format if tool_call_format is defined else 'xml' %}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- set add_vision_id = add_vision_id if add_vision_id is defined else false %}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else true %}
+{%- set auto_disable_thinking_with_tools = auto_disable_thinking_with_tools if auto_disable_thinking_with_tools is defined else false %}
+{%- if preserve_reasoning is defined and preserve_reasoning is not none %}
+    {%- set _preserve_thinking = preserve_reasoning %}
+{%- elif preserve_thinking is defined and preserve_thinking is not none %}
+    {%- set _preserve_thinking = preserve_thinking %}
+{%- else %}
+    {%- set _preserve_thinking = true %}
+{%- endif %}
+{%- set max_tool_arg_chars = max_tool_arg_chars if max_tool_arg_chars is defined else 0 %}
+{%- set max_tool_response_chars = max_tool_response_chars if max_tool_response_chars is defined else 0 %}
+{%- set _default_reasoning_effort = 'medium' %}
+{%- set _has_tools = (tools is defined and tools and tools is iterable and tools is not mapping) %}
+{%- set _effort_raw = (reasoning_effort | string | lower) if reasoning_effort is defined and reasoning_effort is not none else _default_reasoning_effort %}
+{%- set _initial_thinking = enable_thinking %}
+{%- if _effort_raw in ('none', 'off') %}
+    {%- set _initial_thinking = false %}
+    {%- set _initial_effort = 'medium' %}
+{%- elif _effort_raw in ('minimal', 'low') %}
+    {%- set _initial_effort = 'low' %}
+{%- elif _effort_raw in ('high', 'xhigh', 'max', 'ultracode', 'extreme') %}
+    {%- set _initial_effort = 'xhigh' %}
+{%- else %}
+    {%- set _initial_effort = 'medium' %}
+{%- endif %}
+{%- set ns_state = namespace(thinking=_initial_thinking, effort=_initial_effort) %}
+{%- if auto_disable_thinking_with_tools and _has_tools %}
+    {%- set ns_state.thinking = false %}
+{%- endif %}
+{%- for msg in messages %}
+    {%- if msg.role == 'system' or msg.role == 'developer' or msg.role == 'user' %}
+        {%- if msg.content is string %}
+            {%- if '<|think_off|>' in msg.content %}
+                {%- set ns_state.thinking = false %}
+            {%- elif '<|think_on|>' in msg.content %}
+                {%- set ns_state.thinking = true %}
+            {%- elif '<|think_xhigh|>' in msg.content or '<|think_high|>' in msg.content or '<|think_ultracode|>' in msg.content or '<|think_extreme|>' in msg.content or '<|think_max|>' in msg.content %}
+                {%- set ns_state.thinking = true %}
+                {%- set ns_state.effort = 'xhigh' %}
+            {%- elif '<|think_low|>' in msg.content or '<|think_minimal|>' in msg.content %}
+                {%- set ns_state.thinking = true %}
+                {%- set ns_state.effort = 'low' %}
+            {%- elif '<|think_medium|>' in msg.content %}
+                {%- set ns_state.thinking = true %}
+                {%- set ns_state.effort = 'medium' %}
+            {%- endif %}
+        {%- elif msg.content is iterable and msg.content is not mapping %}
+            {%- for item in msg.content %}
+                {%- if item is string %}
+                    {%- set _item_text = item %}
+                {%- elif item is mapping and 'text' in item and item.text is string %}
+                    {%- set _item_text = item.text %}
+                {%- else %}
+                    {%- set _item_text = '' %}
+                {%- endif %}
+                {%- if _item_text %}
+                    {%- if '<|think_off|>' in _item_text %}
+                        {%- set ns_state.thinking = false %}
+                    {%- elif '<|think_on|>' in _item_text %}
+                        {%- set ns_state.thinking = true %}
+                    {%- elif '<|think_xhigh|>' in _item_text or '<|think_high|>' in _item_text or '<|think_ultracode|>' in _item_text or '<|think_extreme|>' in _item_text or '<|think_max|>' in _item_text %}
+                        {%- set ns_state.thinking = true %}
+                        {%- set ns_state.effort = 'xhigh' %}
+                    {%- elif '<|think_low|>' in _item_text or '<|think_minimal|>' in _item_text %}
+                        {%- set ns_state.thinking = true %}
+                        {%- set ns_state.effort = 'low' %}
+                    {%- elif '<|think_medium|>' in _item_text %}
+                        {%- set ns_state.thinking = true %}
+                        {%- set ns_state.effort = 'medium' %}
+                    {%- endif %}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- set reasoning_instructions = '' %}
+{%- if ns_state.thinking %}
+    {%- if ns_state.effort == 'xhigh' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.' %}
+    {%- elif ns_state.effort == 'low' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration.' %}
+    {%- endif %}
+{%- endif %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if item is mapping %}
+                {%- if item.type == 'image' or 'image' in item or 'image_url' in item %}
+                    {%- if is_system_content %}
+                        {{- raise_exception('System message cannot contain images.') }}
+                    {%- endif %}
+                    {%- if do_vision_count %}
+                        {%- set image_count.value = image_count.value + 1 %}
+                    {%- endif %}
+                    {%- if add_vision_id %}
+                        {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                    {%- endif %}
+                    {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+                {%- elif item.type == 'video' or 'video' in item %}
+                    {%- if is_system_content %}
+                        {{- raise_exception('System message cannot contain videos.') }}
+                    {%- endif %}
+                    {%- if do_vision_count %}
+                        {%- set video_count.value = video_count.value + 1 %}
+                    {%- endif %}
+                    {%- if add_vision_id %}
+                        {{- 'Video ' ~ video_count.value ~ ': ' }}
+                    {%- endif %}
+                    {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+                {%- elif 'text' in item %}
+                    {{- item.text }}
+                {%- else %}
+                    {{- raise_exception('Unexpected item type in content.') }}
+                {%- endif %}
+            {%- else %}
+                {{- item | string }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- set head = namespace(count=0, seen_non_system=false) %}
+{%- for message in messages %}
+    {%- set _is_sys = (message.role == 'system' or message.role == 'developer') %}
+    {%- if _is_sys and not head.seen_non_system %}
+        {%- set head.count = head.count + 1 %}
+    {%- else %}
+        {%- set head.seen_non_system = true %}
+    {%- endif %}
+{%- endfor %}
+{%- set sys_state = namespace(content='') %}
+{%- for message in messages[:head.count] %}
+    {%- set _part = render_content(message.content, false, true) | trim %}
+    {%- if '<|think_off|>' in _part %}{%- set _part = _part.split('<|think_off|>') | join('') | trim %}{%- endif %}
+    {%- if '<|think_on|>' in _part %}{%- set _part = _part.split('<|think_on|>') | join('') | trim %}{%- endif %}
+    {%- if '<|think_xhigh|>' in _part %}{%- set _part = _part.split('<|think_xhigh|>') | join('') | trim %}{%- endif %}
+    {%- if '<|think_high|>' in _part %}{%- set _part = _part.split('<|think_high|>') | join('') | trim %}{%- endif %}
+    {%- if '<|think_ultracode|>' in _part %}{%- set _part = _part.split('<|think_ultracode|>') | join('') | trim %}{%- endif %}
+    {%- if '<|think_extreme|>' in _part %}{%- set _part = _part.split('<|think_extreme|>') | join('') | trim %}{%- endif %}
+    {%- if '<|think_max|>' in _part %}{%- set _part = _part.split('<|think_max|>') | join('') | trim %}{%- endif %}
+    {%- if '<|think_medium|>' in _part %}{%- set _part = _part.split('<|think_medium|>') | join('') | trim %}{%- endif %}
+    {%- if '<|think_low|>' in _part %}{%- set _part = _part.split('<|think_low|>') | join('') | trim %}{%- endif %}
+    {%- if '<|think_minimal|>' in _part %}{%- set _part = _part.split('<|think_minimal|>') | join('') | trim %}{%- endif %}
+    {%- if _part %}
+        {%- if sys_state.content %}
+            {%- set sys_state.content = sys_state.content ~ '\n\n' ~ _part %}
+        {%- else %}
+            {%- set sys_state.content = _part %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- set _sc = sys_state.content %}
+{%- set _msgs = messages[head.count:] %}
+{%- if ns_state.thinking %}
+    {%- set _terse_lead = 'Answer directly, after thinking. Lead with the answer, then only what it needs to be correct and usable.' %}
+{%- else %}
+    {%- set _terse_lead = 'Answer directly and concisely. Give the answer with only what it needs to be correct and usable.' %}
+{%- endif %}
+{%- set _terse_core %}
+Never: open with preamble or pleasantries; restate the question; add filler transitions; hedge with niceties; or repeat a point you've already made.
+Always: keep essential steps, caveats, uncertainties, and specifics — never drop correctness or a needed warning for brevity. Keep the final answer lean. Use the least structure that conveys it (plain prose when short; lists or code only when they earn their place). If genuinely uncertain, say so and explain why — never omit uncertainty for the sake of brevity.
+If a user request is genuinely ambiguous, ask a sharp question, don't guess.
+{%- endset %}
+{%- set _terse = _terse_lead ~ '\n' ~ (_terse_core | trim) %}
+{#- `terse` chat-template kwarg: default true, so behaviour is unchanged unless a caller opts out.
+    Pass {"terse": false} via chat_template_kwargs to serve the model with only its own system
+    prompt -- useful for A/B-ing the terseness effect, or when a downstream prompt conflicts. #}
+{%- set _terse_on = terse if terse is defined else true %}
+{%- if _terse_on %}
+    {%- if not _sc %}
+        {%- set _sc = _terse | trim %}
+    {%- else %}
+        {%- set _sc = (_sc | trim) ~ '\n\n' ~ (_terse | trim) %}
+    {%- endif %}
+{%- endif %}
+{%- if _has_tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if reasoning_instructions %}
+        {{- reasoning_instructions + '\n\n' }}
+    {%- endif %}
+    {{- '# Tools\n\nYou have access to the following functions:\n\n<tools>' }}
+    {%- for tool in tools %}
+        {{- '\n' }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- '\n</tools>' }}
+    {%- if _tool_format == 'json' %}
+        {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n' }}{%- if ns_state.thinking %}{{- '<think>\nBrief explanation of tool call\n</think>\n' }}{%- endif %}{{- '<tool_call>\n{"name": "example_function_name", "arguments": {"example_parameter_1": "value_1", "example_parameter_2": "This is the value for the second parameter"}}\n</tool_call>\n\n<IMPORTANT>\nReminder:\n' }}{%- if ns_state.thinking %}{{- '- You can use the <think></think> block to plan your next tool call OR to synthesize data and formulate your final response to the user.\n- ALL explanation and reasoning MUST be placed strictly inside the <think></think> block.\n' }}{%- endif %}{{- '- Function calls MUST follow the specified format: a single JSON object with "name" and "arguments" keys inside <tool_call></tool_call> XML tags.\n- If you choose to call a tool, you MUST output the <tool_call> block IMMEDIATELY' }}{%- if ns_state.thinking %}{{- ' after thinking' }}{%- endif %}{{- ', with NO conversational text before it.\n- The <tool_call> tag MUST be at the very beginning of a new line, with NO spaces or indentation before it.\n- To call multiple functions, output a separate, completely closed <tool_call></tool_call> block for EACH function. Do NOT nest <tool_call> blocks.\n- If you have all necessary data, provide your final answer directly to the user without any tool call.\n</IMPORTANT>' }}
+    {%- else %}
+        {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n' }}{%- if ns_state.thinking %}{{- '<think>\nBrief explanation of tool call\n</think>\n' }}{%- endif %}{{- '<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n' }}{%- if ns_state.thinking %}{{- '- You can use the <think></think> block to plan your next tool call OR to synthesize data and formulate your final response to the user.\n- ALL explanation and reasoning MUST be placed strictly inside the <think></think> block.\n' }}{%- endif %}{{- '- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags.\n- If you choose to call a tool, you MUST output the <tool_call> block IMMEDIATELY' }}{%- if ns_state.thinking %}{{- ' after thinking' }}{%- endif %}{{- ', with NO conversational text before it.\n- The <tool_call> and <function> tags MUST be at the very beginning of a new line, with NO spaces or indentation before them.\n- To call multiple functions, output a separate, completely closed <tool_call></tool_call> block for EACH function. Do NOT nest <tool_call> blocks.\n- If you have all necessary data, provide your final answer directly to the user without any tool call.\n</IMPORTANT>' }}
+    {%- endif %}
+    {%- if _sc %}
+        {{- '\n\n' + _sc }}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if _sc %}
+        {{- '<|im_start|>system\n' + (reasoning_instructions + '\n\n' if reasoning_instructions else '') + _sc + '<|im_end|>\n' }}
+    {%- elif reasoning_instructions %}
+        {{- '<|im_start|>system\n' + reasoning_instructions + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set _last_idx = _msgs | length - 1 %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=_last_idx) %}
+{%- for message in _msgs[::-1] %}
+    {%- set index = (_msgs | length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == 'user' %}
+        {%- set _rc = render_content(message.content, false) | trim %}
+        {%- if not (_rc.startswith('<tool_response>') and _rc.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {%- if _last_idx > 50 %}
+        {%- set ns.last_query_index = _last_idx %}
+    {%- else %}
+        {%- set ns.last_query_index = 0 %}
+    {%- endif %}
+{%- endif %}
+{%- set ns2 = namespace(prev_role='', consecutive_failures=0) %}
+{%- for message in _msgs %}
+    {%- set is_system = (message.role == "system" or message.role == "developer") %}
+    {%- set content = render_content(message.content, true, is_system) | trim %}
+    {%- if is_system or message.role == 'user' %}
+        {%- if '<|think_off|>' in content %}{%- set content = content.split('<|think_off|>') | join('') | trim %}{%- endif %}
+        {%- if '<|think_on|>' in content %}{%- set content = content.split('<|think_on|>') | join('') | trim %}{%- endif %}
+        {%- if '<|think_xhigh|>' in content %}{%- set content = content.split('<|think_xhigh|>') | join('') | trim %}{%- endif %}
+        {%- if '<|think_high|>' in content %}{%- set content = content.split('<|think_high|>') | join('') | trim %}{%- endif %}
+        {%- if '<|think_ultracode|>' in content %}{%- set content = content.split('<|think_ultracode|>') | join('') | trim %}{%- endif %}
+        {%- if '<|think_extreme|>' in content %}{%- set content = content.split('<|think_extreme|>') | join('') | trim %}{%- endif %}
+        {%- if '<|think_max|>' in content %}{%- set content = content.split('<|think_max|>') | join('') | trim %}{%- endif %}
+        {%- if '<|think_medium|>' in content %}{%- set content = content.split('<|think_medium|>') | join('') | trim %}{%- endif %}
+        {%- if '<|think_low|>' in content %}{%- set content = content.split('<|think_low|>') | join('') | trim %}{%- endif %}
+        {%- if '<|think_minimal|>' in content %}{%- set content = content.split('<|think_minimal|>') | join('') | trim %}{%- endif %}
+    {%- endif %}
+    {%- if is_system %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- elif message.role == 'user' %}
+        {%- set ns2.consecutive_failures = 0 %}
+        {{- '<|im_start|>user\n' + content + '<|im_end|>\n' }}
+    {%- elif message.role == 'assistant' %}
+        {%- set reasoning_content = '' %}
+        {%- set _explicit_reasoning = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- if message.reasoning_content is string %}
+                {%- set _explicit_reasoning = message.reasoning_content %}
+            {%- else %}
+                {%- set _explicit_reasoning = message.reasoning_content | string %}
+            {%- endif %}
+        {%- elif message.thinking is defined and message.thinking is not none %}
+            {%- if message.thinking is string %}
+                {%- set _explicit_reasoning = message.thinking %}
+            {%- else %}
+                {%- set _explicit_reasoning = message.thinking | string %}
+            {%- endif %}
+        {%- elif message.reasoning is defined and message.reasoning is not none %}
+            {%- if message.reasoning is string %}
+                {%- set _explicit_reasoning = message.reasoning %}
+            {%- else %}
+                {%- set _explicit_reasoning = message.reasoning | string %}
+            {%- endif %}
+        {%- endif %}
+        {%- if _explicit_reasoning %}
+            {%- set _lead_end = '' %}
+            {%- if content.startswith('<think>') and '</think>' in content %}
+                {%- set _lead_end = '</think>' %}
+            {%- elif content.startswith('<thinking>') and '</thinking>' in content %}
+                {%- set _lead_end = '</thinking>' %}
+            {%- elif content.startswith('</think>') %}
+                {%- set _lead_end = '</think>' %}
+            {%- elif content.startswith('</thinking>') %}
+                {%- set _lead_end = '</thinking>' %}
+            {%- endif %}
+            {%- if _lead_end %}
+                {%- set content = content.split(_lead_end)[-1].lstrip('\n') %}
+            {%- endif %}
+            {%- set reasoning_content = _explicit_reasoning %}
+        {%- else %}
+            {%- set _think_end = '' %}
+            {%- if content.startswith('</think>') %}
+                {%- set _think_end = '</think>' %}
+            {%- elif content.startswith('</thinking>') %}
+                {%- set _think_end = '</thinking>' %}
+            {%- elif '\n</think>' in content %}
+                {%- set _think_end = '\n</think>' %}
+            {%- elif '\n</thinking>' in content %}
+                {%- set _think_end = '\n</thinking>' %}
+            {%- elif '\n</ think>' in content %}
+                {%- set _think_end = '\n</ think>' %}
+            {%- elif '\n</think >' in content %}
+                {%- set _think_end = '\n</think >' %}
+            {%- elif content.startswith('<think>') and '</think>' in content %}
+                {%- set _think_end = '</think>' %}
+            {%- elif content.startswith('<thinking>') and '</thinking>' in content %}
+                {%- set _think_end = '</thinking>' %}
+            {%- endif %}
+            {%- if _think_end %}
+                {%- if 'thinking' in _think_end %}
+                    {%- set _think_start = '<thinking>' %}
+                {%- else %}
+                    {%- set _think_start = '<think>' %}
+                {%- endif %}
+                {%- set reasoning_content = content.split(_think_end)[0].rstrip('\n') %}
+                {%- if _think_start in reasoning_content %}
+                    {%- set reasoning_content = reasoning_content.split(_think_start)[-1].lstrip('\n') %}
+                {%- endif %}
+                {%- set content = content.split(_think_end)[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content | trim %}
+        {%- if (_preserve_thinking or loop.index0 > ns.last_query_index) %}
+            {{- '<|im_start|>assistant\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>assistant\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined and tool_call.function is not none %}
+                    {%- set tc = tool_call.function %}
+                {%- else %}
+                    {%- set tc = tool_call %}
+                {%- endif %}
+                {%- set tc_name = tc.name if (tc.name is defined and tc.name is not none) else '' %}
+                {%- if _tool_format == 'json' %}
+                    {%- if loop.first %}
+                        {%- if content | trim %}
+                            {{- '\n\n' }}
+                        {%- endif %}
+                    {%- else %}
+                        {{- '\n' }}
+                    {%- endif %}
+                    {%- set _args = '{}' %}
+                    {%- if tc.arguments is defined and tc.arguments is not none %}
+                        {%- if tc.arguments is mapping %}
+                            {%- set _args = tc.arguments | tojson %}
+                        {%- elif tc.arguments is string %}
+                            {%- if tc.arguments %}
+                                {%- set _args = tc.arguments %}
+                            {%- endif %}
+                        {%- else %}
+                            {%- set _args = tc.arguments | tojson %}
+                        {%- endif %}
+                    {%- endif %}
+                    {{- '<tool_call>\n{"name": ' }}{{- tc_name | tojson }}{{- ', "arguments": ' }}{{- _args }}{{- '}\n</tool_call>' }}
+                {%- else %}
+                    {%- if loop.first %}
+                        {%- if content | trim %}
+                            {{- '\n\n<tool_call>\n<function=' + tc_name + '>\n' }}
+                        {%- else %}
+                            {{- '<tool_call>\n<function=' + tc_name + '>\n' }}
+                        {%- endif %}
+                    {%- else %}
+                        {{- '\n<tool_call>\n<function=' + tc_name + '>\n' }}
+                    {%- endif %}
+                    {%- if tc.arguments is defined and tc.arguments is not none %}
+                        {%- if tc.arguments is mapping %}
+                            {%- for args_name, args_value in tc.arguments.items() %}
+                                {{- '<parameter=' + args_name + '>\n' }}
+                                {%- if args_value is string %}
+                                    {%- set _av = args_value %}
+                                {%- else %}
+                                    {%- set _av = args_value | tojson %}
+                                {%- endif %}
+                                {%- if max_tool_arg_chars > 0 and _av | length > max_tool_arg_chars %}
+                                    {{- _av[:max_tool_arg_chars] + '\n[TRUNCATED - original length ' ~ (_av | length | string) ~ ' chars]' }}
+                                {%- else %}
+                                    {{- _av }}
+                                {%- endif %}
+                                {{- '\n</parameter>\n' }}
+                            {%- endfor %}
+                        {%- else %}
+                            {%- if tc.arguments is string %}
+                                {%- set _raw_args = tc.arguments %}
+                            {%- else %}
+                                {%- set _raw_args = tc.arguments | tojson %}
+                            {%- endif %}
+                            {%- if _raw_args %}
+                                {%- if max_tool_arg_chars > 0 and _raw_args | length > max_tool_arg_chars %}
+                                    {{- _raw_args[:max_tool_arg_chars] + '\n[TRUNCATED - original length ' ~ (_raw_args | length | string) ~ ' chars]' }}
+                                {%- else %}
+                                    {{- _raw_args }}
+                                {%- endif %}
+                            {%- endif %}
+                        {%- endif %}
+                    {%- endif %}
+                    {{- '</function>\n</tool_call>' }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == 'tool' %}
+        {%- set _content_lower = content | lower %}
+        {%- set _content_head = _content_lower[:120] %}
+        {%- set _is_code_or_grep = ('throw new ' in _content_lower or 'throw error' in _content_lower or 'console.error' in _content_lower or 'logger.error' in _content_lower or 'logging.error' in _content_lower or 'import ' in _content_head or 'def ' in _content_head or 'function ' in _content_head) %}
+        {%- set _exit_code_zero = ('exit code: 0' in _content_head or 'process exited with code 0' in _content_head) %}
+        {%- set _error_field_ok = ('"error": null' in _content_head or '"error":null' in _content_head or '"error": false' in _content_head or '"error":false' in _content_head or '"error": ""' in _content_head or '"error":""' in _content_head) %}
+        {%- set _strong_error = (('"error":' in _content_head and not _error_field_ok) or '"status": "error"' in _content_head or '"status":"error"' in _content_head or 'traceback (most recent call last):' in _content_head or 'command not found' in _content_head or 'invalid syntax' in _content_head or 'fatal:' in _content_head or (('exit code: ' in _content_head or 'process exited with code' in _content_head) and not _exit_code_zero) or _content_head.startswith('exception:') or _content_head.startswith('failed to ')) %}
+        {%- set _weak_error = ('error:' in _content_head or 'err!' in _content_head) %}
+        {%- set _weak_suppressed = ('$ ' in _content_head or 'took ' in _content_head or content | length >= 600) %}
+        {%- if not _is_code_or_grep and (_strong_error or (_weak_error and not _weak_suppressed)) %}
+            {%- set ns2.consecutive_failures = ns2.consecutive_failures + 1 %}
+        {%- else %}
+            {%- set ns2.consecutive_failures = 0 %}
+        {%- endif %}
+        {%- if ns2.prev_role != 'tool' %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {%- if _tool_format != 'json' and max_tool_response_chars > 0 and content | length > max_tool_response_chars %}
+            {%- set content = content[:max_tool_response_chars] + '\n[TRUNCATED - original length ' ~ (content | length | string) ~ ' chars]' %}
+        {%- endif %}
+        {{- '\n<tool_response>\n' + content }}
+        {%- if ns2.consecutive_failures >= 2 %}
+            {{- '\n\n⚠️ SYSTEM WARNING: ' ~ ns2.consecutive_failures ~ ' consecutive tool errors detected. Your previous approach is incorrect. You MUST use a fundamentally different approach or corrected arguments.' }}
+        {%- elif ns2.consecutive_failures == 1 %}
+            {{- '\n\n⚠️ SYSTEM WARNING: The previous tool call returned an error. Diagnose the failure and retry with completely corrected arguments.' }}
+        {%- endif %}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- else %}
+            {%- set _next_role = _msgs[loop.index0 + 1].role %}
+            {%- if _next_role != 'tool' %}
+                {{- '<|im_end|>\n' }}
+            {%- endif %}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>user\n[' + message.role + ']: ' + content + '<|im_end|>\n' }}
+    {%- endif %}
+    {%- set ns2.prev_role = message.role %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if not ns_state.thinking %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}
@@ -0,0 +1,74 @@
+#!/usr/bin/env python3
+"""Render ${VAR} placeholders in the live SearXNG settings.yml from the
+host-side env file.
+
+Why this exists: SearXNG does NOT expand environment variables inside
+settings.yml. The repo-tracked shared/searxng/settings.yml deliberately keeps
+secrets as "${KAGI_API_KEY}"-style placeholders so no key ever enters git; the
+deployer copies that file to /opt/searxng/config/settings.yml and runs this
+script to bake in the real values. Without it the app runs with the literal
+placeholder strings (observed 2026-09-26: secret_key literally
+"${SEARXNG_SECRET}" and the Kagi engine failing setup() on an empty key).
+
+Only the allowlisted names are substituted. A placeholder that has no value in
+the env file is a deploy-blocking error: serving a paid-API engine with a
+literal "${KAGI_API_KEY}" bearer would waste requests and hide the misconfig.
+
+Usage: searxng-render-settings.py ENV_FILE SETTINGS_FILE
+Both paths are taken as arguments; secrets never appear in any argv.
+Exit codes: 0 rendered (idempotent on an already-rendered file), 1 unresolved
+placeholder, 2 usage error.
+"""
+
+import sys
+
+ALLOWED = ("KAGI_API_KEY", "SEARXNG_SECRET", "SEARXNG_BASE_URL")
+
+
+def load_env(path: str) -> dict[str, str]:
+    values: dict[str, str] = {}
+    with open(path, encoding="utf-8") as fh:
+        for line in fh:
+            line = line.strip()
+            if not line or line.startswith("#") or "=" not in line:
+                continue
+            key, _, value = line.partition("=")
+            if key in ALLOWED:
+                values[key] = value
+    return values
+
+
+def main(argv: list[str]) -> int:
+    args = [a for a in argv[1:]]
+    if len(args) != 2:
+        print(
+            "usage: searxng-render-settings.py ENV_FILE SETTINGS_FILE",
+            file=sys.stderr,
+        )
+        return 2
+    env_file, settings_file = args
+
+    values = load_env(env_file)
+    with open(settings_file, encoding="utf-8") as fh:
+        text = fh.read()
+
+    for name in ALLOWED:
+        if name in values:
+            text = text.replace("${%s}" % name, values[name])
+
+    unresolved = [n for n in ALLOWED if "${%s}" % n in text]
+    if unresolved:
+        print(
+            "unresolved placeholders in %s: %s (missing from %s)"
+            % (settings_file, ", ".join(unresolved), env_file),
+            file=sys.stderr,
+        )
+        return 1
+
+    with open(settings_file, "w", encoding="utf-8") as fh:
+        fh.write(text)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv))
@@ -0,0 +1,179 @@
+#!/usr/bin/env bash
+# Serve kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF as frank-large on
+# Dankclaw's Strix Halo iGPU (gfx1151) with the ROCmFPX fork of llama.cpp.
+#
+# 2026-10-02 (Frank): swapped from gufo + Q4_K_XL to the highest-quality file
+# in the repo, one session at the native 262,144 window.
+#
+# Why this runtime and not the previous one: the payload is ftype 115 — 311
+# tensors of GGML_TYPE_Q8_0_ROCMFPX (103), a tensor type that exists only in
+# the ROCmFPX fork. Stock llama.cpp, SGLang and gufo all refuse the file, and
+# gufo additionally gates every "Qwen3.8" artifact on a template-hash
+# allowlist. The qualified build is scripts/build-rocmfpx.sh's pin: it loads
+# ftype 115, survives warmup (the 2026-08-25 builds of the older fork line
+# segfaulted in ggml_compute_forward_set_rows on this exact file), and drafts
+# with the model's own MTP head.
+#
+# Why this file: Q8_0_ROCMFPX_AGENT is the top of the repo's quality ladder
+# (26.28 GiB, 8.39 bpw) and the AGENT routing keeps 187 tensors — the
+# output-head set the draft scores against — at plain Q8_0 instead of the 2
+# the non-AGENT build keeps, which is the variant the card names for MTP
+# serving. Drafting uses the repo's Q4_0 MTP head: the head's cost is its
+# projection over the 248K vocabulary, so the 4-bit head is faster end to end
+# at equal-or-better acceptance than its Q8_0 sibling.
+#
+# The client contract lives here, not in frank-large.env:
+#   * --jinja — without it llama.cpp silently IGNORES the harness's
+#     chat_template_kwargs (enable_thinking / preserve_thinking), so the
+#     endpoint would look like it had no thinking control at all. The GGUF
+#     carries the official Qwen template, so no --chat-template-file.
+#   * --reasoning-format deepseek — thinking is emitted as reasoning_content,
+#     the field the harness reads for this endpoint.
+#   * the template accepts reasoning_effort xhigh|medium|low only, which is
+#     why models.yml folds the harness's levels onto that set.
+#   * --alias frank-large — that is the model id the harness resolves.
+#   * --ctx-checkpoints / --cache-ram — prompt-cache geometry for the 16
+#     sliding-window-style layers; harmless where the hybrid state forces a
+#     full re-process, and required for the checkpoint feature to work at all.
+set -euo pipefail
+
+# Fleet values live in frank-large.env. Anything already in the environment
+# wins over the file, so an operator can override one value on a single run:
+# the caller's values are captured, the file is sourced, then they are put
+# back. (A plain `source` would let the file clobber the environment.)
+TUNABLES=(LLAMA_BIN MODEL_FILE DRAFT_FILE MMPROJ_FILE ALIAS BIND_ADDR
+          LISTEN_PORT CTX_SIZE PARALLEL BATCH_SIZE UBATCH_SIZE NGL FIT
+          FLASH_ATTN CACHE_TYPE_K CACHE_TYPE_V SPEC_TYPE SPEC_DRAFT_NGL
+          SPEC_DRAFT_DEVICE SPEC_DRAFT_N_MAX SPEC_DRAFT_N_MIN SPEC_DRAFT_P_MIN
+          TEMPERATURE TOP_P TOP_K MIN_P PRESENCE_PENALTY NUM_THREADS
+          THREAD_BATCH ROCM_PATH HIP_VISIBLE HSA_GFX_VERSION)
+_OVERRIDES=()
+for _var in "${TUNABLES[@]}"; do
+    [[ -n "${!_var+set}" ]] && _OVERRIDES+=("$_var=${!_var}")
+done
+ENV_FILE="${FRANK_LARGE_ENV_FILE:-$(dirname "$0")/frank-large.env}"
+if [[ -f "$ENV_FILE" ]]; then
+    set -a
+    # shellcheck disable=SC1090
+    source "$ENV_FILE"
+    set +a
+fi
+for _kv in "${_OVERRIDES[@]+${_OVERRIDES[@]}}"; do
+    export "${_kv?}"
+done
+unset _var _kv
+
+: "${LLAMA_BIN=/home/dank/GitPackages/kingjones-ROCmFPX/build/bin/llama-server}"
+: "${MODEL_FILE=/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf}"
+: "${DRAFT_FILE=/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/mtp-Qwen3.8-27B-Q4_0.gguf}"
+: "${MMPROJ_FILE=/home/dank/models/kingjones-Qwen3.8-27B-ROCmFP4-STRIX-MTP/mmproj-Qwen3.8-27B-BF16.gguf}"
+: "${ALIAS=frank-large}"
+: "${BIND_ADDR=0.0.0.0}"
+: "${LISTEN_PORT=7770}"
+: "${CTX_SIZE=262144}"
+: "${PARALLEL=1}"
+: "${BATCH_SIZE=2048}"
+: "${UBATCH_SIZE=512}"
+: "${NGL=999}"
+: "${FIT=off}"
+: "${FLASH_ATTN=on}"
+: "${CACHE_TYPE_K=f16}"
+: "${CACHE_TYPE_V=f16}"
+: "${SPEC_TYPE=draft-mtp}"
+: "${SPEC_DRAFT_NGL=99}"
+: "${SPEC_DRAFT_DEVICE=ROCm0}"
+: "${SPEC_DRAFT_N_MAX=4}"
+: "${SPEC_DRAFT_N_MIN=0}"
+: "${SPEC_DRAFT_P_MIN=0.0}"
+: "${TEMPERATURE=1.0}"
+: "${TOP_P=0.95}"
+: "${TOP_K=20}"
+: "${MIN_P=0.0}"
+: "${PRESENCE_PENALTY=0.0}"
+: "${NUM_THREADS=16}"
+: "${THREAD_BATCH=16}"
+: "${ROCM_PATH=/opt/rocm}"
+: "${HIP_VISIBLE=0}"
+: "${HSA_GFX_VERSION=11.5.1}"
+
+[[ -n "$LLAMA_BIN" ]] || { echo "frank-large: LLAMA_BIN is empty" >&2; exit 1; }
+[[ -x "$LLAMA_BIN" ]] || {
+    echo "frank-large: llama-server '$LLAMA_BIN' is not executable;" >&2
+    echo "           build it with scripts/build-rocmfpx.sh" >&2
+    exit 1
+}
+# Refuse anything but the real payload. A path that is not a readable file
+# makes llama-server hang rather than fail clean.
+[[ -r "$MODEL_FILE" ]] || {
+    echo "frank-large: model file '$MODEL_FILE' is not readable;" >&2
+    echo "           fetch it with scripts/download-frank-large-kingjones.sh" >&2
+    exit 1
+}
+# The MTP head is the performance feature: without it the Q8_0 weights decode
+# at memory-bandwidth speed and the endpoint loses the ~1.4x the head buys.
+[[ -r "$DRAFT_FILE" ]] || {
+    echo "frank-large: MTP draft head '$DRAFT_FILE' is not readable;" >&2
+    echo "           fetch it with scripts/download-frank-large-kingjones.sh" >&2
+    exit 1
+}
+# The GGUF carries no vision tower; the projector file is the whole tower, so
+# a missing file means a vision request would 500 rather than degrade.
+# MMPROJ_FILE="" is the documented opt-out for a text-only endpoint.
+MM_ARGS=()
+if [[ -n "${MMPROJ_FILE:-}" ]]; then
+    [[ -r "$MMPROJ_FILE" ]] || {
+        echo "frank-large: mmproj '$MMPROJ_FILE' is not readable; set MMPROJ_FILE='' for text-only" >&2
+        exit 1
+    }
+    MM_ARGS=(--mmproj "$MMPROJ_FILE")
+fi
+
+# HIP/ROCm host environment. HIP reads bare env vars (TEMP, HOST,
+# GPU_MAX_HEAP_SIZE), which is why the tunables above are named TEMPERATURE
+# and BIND_ADDR. HSA_OVERRIDE_GFX_VERSION is what lets a runtime built for a
+# neighbouring gfx target run on this iGPU; the qualified pin targets gfx1151
+# natively and the override is a no-op for it.
+export ROCM_PATH
+export HIP_VISIBLE_DEVICES="$HIP_VISIBLE"
+export HSA_OVERRIDE_GFX_VERSION="$HSA_GFX_VERSION"
+export LD_LIBRARY_PATH="$(dirname "$LLAMA_BIN"):${ROCM_PATH}/lib:${ROCM_PATH}/lib/rocm_sysdeps/lib:${ROCM_PATH}/lib/llvm/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
+
+exec /usr/bin/env "$LLAMA_BIN" \
+    --model "$MODEL_FILE" \
+    "${MM_ARGS[@]}" \
+    --alias "$ALIAS" \
+    --host "$BIND_ADDR" \
+    --port "$LISTEN_PORT" \
+    --jinja \
+    --device ROCm0 \
+    --n-gpu-layers "$NGL" \
+    --parallel "$PARALLEL" \
+    --ctx-size "$CTX_SIZE" \
+    --batch-size "$BATCH_SIZE" \
+    --ubatch-size "$UBATCH_SIZE" \
+    --threads "$NUM_THREADS" \
+    --threads-batch "$THREAD_BATCH" \
+    --fit "$FIT" \
+    --flash-attn "$FLASH_ATTN" \
+    --direct-io \
+    --cache-type-k "$CACHE_TYPE_K" \
+    --cache-type-v "$CACHE_TYPE_V" \
+    --ctx-checkpoints 16 \
+    --cache-ram 8192 \
+    --cache-idle-slots \
+    --temp "$TEMPERATURE" \
+    --top-p "$TOP_P" \
+    --top-k "$TOP_K" \
+    --min-p "$MIN_P" \
+    --presence-penalty "$PRESENCE_PENALTY" \
+    --reasoning on \
+    --reasoning-format deepseek \
+    --reasoning-budget -1 \
+    --spec-type "$SPEC_TYPE" \
+    --model-draft "$DRAFT_FILE" \
+    --spec-draft-ngl "$SPEC_DRAFT_NGL" \
+    --spec-draft-device "$SPEC_DRAFT_DEVICE" \
+    --spec-draft-n-max "$SPEC_DRAFT_N_MAX" \
+    --spec-draft-n-min "$SPEC_DRAFT_N_MIN" \
+    --spec-draft-p-min "$SPEC_DRAFT_P_MIN" \
+    --metrics
@@ -0,0 +1,135 @@
+#!/usr/bin/env bash
+# serve-frank-primary.sh — frank-primary endpoint (tcclaviger RDNA4 vLLM fork,
+# orcarouter Qwen3.8-Flash-Next-Uncensored-FP8 + MTP4 spec decode, 2x R9700,
+# dankserver :7770; 2026-09-29 cutover from the R9V radiance fork — git tag
+# frank-primary-r9v-q8-20260929 is the pre-cutover state).
+#
+# The engine is driven by `podman run` DIRECTLY: the image's entrypoint execs
+# `vllm serve "$@"`, so the run line below IS the engine argv. This replaces
+# the R9V launcher/shim machinery (that fork needed a docker->podman shim and
+# a 104-key on-host setup state; this stack is fully expressed by flags, and
+# every flag here was qualified by boot probe on 2026-09-29 — see
+# frank-primary.env for the qualification record).
+#
+# Shape (2026-10-02, Frank's correction of the 10-01 bf16 shape): MAXLEN
+# 200704 x SEQS 2 (401,408 tokens of contract), KV fp8 (--kv-cache-dtype
+# fp8; the bf16 unquantized shape is retired to git tag
+# frank-primary-fp8-mtp-20261001), KV_BYTES 3,150,000,000 PER RANK (the
+# engine measured 405,000 tokens of capacity / ~3.0x concurrency), NO
+# speculative decode, --max-num-batched-tokens 1024 (Frank's lock), MoE
+# gather geometry R4D_LRU_GATHER_CHUNKS=32 /
+# R4D_LRU_GATHER_LANES=64 (+18.7% decode, grows with depth), prefix
+# caching, expert prefetch OFF (36% precision measured), TunableOp off, and
+# the pool deliberately small so the expert arena keeps what it reclaims.
+#
+# The root unit (scripts/frank-primary.service, User=dank,
+# LimitMEMLOCK=infinity) grants the pin budget for the tiered-expert host
+# slabs (~20.5 GiB/rank of pinned anonymous memory). A plain shell session
+# also works, but the unit is the fleet-parity boot-persistent shape.
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ENV_FILE="${FRANK_PRIMARY_ENV_FILE:-$SCRIPT_DIR/frank-primary.env}"
+
+if [ ! -f "$ENV_FILE" ]; then
+    echo "serve-frank-primary: env file missing: $ENV_FILE" >&2
+    exit 1
+fi
+set -a
+# shellcheck disable=SC1090
+source "$ENV_FILE"
+set +a
+
+: "${FRANK_PRIMARY_IMAGE:?FRANK_PRIMARY_IMAGE must name the digest-pinned image}"
+: "${FRANK_PRIMARY_MODEL_DIR:?FRANK_PRIMARY_MODEL_DIR must name the FP8 payload}"
+: "${FRANK_PRIMARY_MODEL_SHARDS:?FRANK_PRIMARY_MODEL_SHARDS must give the shard count}"
+: "${FRANK_PRIMARY_CONTAINER:?FRANK_PRIMARY_CONTAINER must name the container}"
+: "${FRANK_PRIMARY_SERVED_NAME:?FRANK_PRIMARY_SERVED_NAME must name the alias}"
+: "${FRANK_PRIMARY_PUBLISH:?FRANK_PRIMARY_PUBLISH must be host:container}"
+: "${FRANK_PRIMARY_MAXLEN:?FRANK_PRIMARY_MAXLEN must be set}"
+: "${FRANK_PRIMARY_SEQS:?FRANK_PRIMARY_SEQS must be set}"
+: "${FRANK_PRIMARY_KV_BYTES:?FRANK_PRIMARY_KV_BYTES must be set (PER RANK)}"
+: "${FRANK_PRIMARY_KV_DTYPE:?FRANK_PRIMARY_KV_DTYPE must be set (auto = bf16 KV)}"
+: "${FRANK_PRIMARY_REASONING_PARSER:?FRANK_PRIMARY_REASONING_PARSER must be set (qwen3)}"
+: "${FRANK_PRIMARY_TOOL_PARSER:?FRANK_PRIMARY_TOOL_PARSER must be set (qwen3_xml)}"
+: "${FRANK_PRIMARY_CHAT_TEMPLATE_KWARGS:?FRANK_PRIMARY_CHAT_TEMPLATE_KWARGS must be set}"
+# Capture podman's output BEFORE grepping: under `set -o pipefail`, the
+# classic `podman images | grep -q` pattern races — grep exits on its first
+# match, podman dies SIGPIPE (141) writing the remaining lines, and pipefail
+# turns the 141 into "no match" (the 2026-09-30 deploy refusal).
+DIGESTS="$(podman images --format '{{.RepoDigests}}')"
+if ! grep -qF "${FRANK_PRIMARY_IMAGE#*@}" <<<"$DIGESTS"; then
+    echo "serve-frank-primary: pinned image absent from rootless storage: $FRANK_PRIMARY_IMAGE" >&2
+    echo "  fetch it: podman pull $FRANK_PRIMARY_IMAGE" >&2
+    exit 1
+fi
+
+if [ ! -f "$FRANK_PRIMARY_MODEL_DIR/config.json" ]; then
+    echo "serve-frank-primary: payload config missing: $FRANK_PRIMARY_MODEL_DIR/config.json" >&2
+    exit 1
+fi
+SHARDS="$(find "$FRANK_PRIMARY_MODEL_DIR" -maxdepth 1 -name 'model-*-of-*.safetensors' | wc -l)"
+if [ "$SHARDS" != "$FRANK_PRIMARY_MODEL_SHARDS" ]; then
+    echo "serve-frank-primary: payload shard count wrong: $FRANK_PRIMARY_MODEL_DIR" >&2
+    echo "  found $SHARDS safetensors, want $FRANK_PRIMARY_MODEL_SHARDS" >&2
+    echo "  re-fetch: scripts/download-frank-primary.sh" >&2
+    exit 1
+fi
+if [ ! -f "$FRANK_PRIMARY_MODEL_DIR/model.safetensors.index.json" ]; then
+    echo "serve-frank-primary: weight-map index missing: $FRANK_PRIMARY_MODEL_DIR/model.safetensors.index.json" >&2
+    exit 1
+fi
+
+if [ ! -e /dev/kfd ] || [ ! -d /dev/dri ]; then
+    echo "serve-frank-primary: GPU device nodes absent (/dev/kfd, /dev/dri)" >&2
+    exit 1
+fi
+
+# A crashed/killed unit can leave the container name registered (systemd
+# SIGKILLs conmon before cleanup runs). Clear OUR name plus the legacy
+# identities (the retired R9V container and the ad hoc experiment container
+# both hold these GPUs); all idempotent.
+for CTR in $FRANK_PRIMARY_LEGACY_CONTAINERS "$FRANK_PRIMARY_CONTAINER"; do
+    podman rm -f "$CTR" >/dev/null 2>&1 || true
+done
+
+LOG=/home/dank/logs/frank_primary.log
+mkdir -p "$(dirname "$LOG")"
+
+# --device /dev/dri is the WHOLE DIRECTORY (individual renderD nodes enumerate
+# without the card nodes and HIP refuses contexts — "No CUDA GPUs are
+# available" while device_count() says 2). --ipc=host for the engine's shared
+# memory; memlock unbounded for the pinned host slabs.
+# The entrypoint already execs `vllm serve "$@"`, so no `serve` here.
+podman run -d --name "$FRANK_PRIMARY_CONTAINER" \
+  --publish "$FRANK_PRIMARY_PUBLISH" \
+  --ipc=host --ulimit memlock=-1:-1 \
+  --cap-add SYS_PTRACE \
+  --device /dev/kfd --device /dev/dri --device /dev/mem \
+  -e HIP_VISIBLE_DEVICES=0,1 -e VLLM_ROCM_USE_AITER=0 \
+  ${FRANK_PRIMARY_GATHER_CHUNKS:+-e R4D_LRU_GATHER_CHUNKS="$FRANK_PRIMARY_GATHER_CHUNKS"} \
+  ${FRANK_PRIMARY_GATHER_LANES:+-e R4D_LRU_GATHER_LANES="$FRANK_PRIMARY_GATHER_LANES"} \
+  -v "$FRANK_PRIMARY_MODEL_DIR":/models:ro \
+  "$FRANK_PRIMARY_IMAGE" /models \
+  --served-model-name "$FRANK_PRIMARY_SERVED_NAME" --port 7770 \
+  --tensor-parallel-size 2 \
+  ${FRANK_PRIMARY_DCP:+--decode-context-parallel-size "$FRANK_PRIMARY_DCP"} \
+  --max-model-len "$FRANK_PRIMARY_MAXLEN" \
+  --max-num-seqs "$FRANK_PRIMARY_SEQS" \
+  ${FRANK_PRIMARY_MNBT:+--max-num-batched-tokens "$FRANK_PRIMARY_MNBT"} \
+  ${FRANK_PRIMARY_PRECISE_MEMORY:+--expert-precise-memory} \
+  --gpu-memory-utilization "$FRANK_PRIMARY_UTIL" \
+  --kv-cache-memory-bytes "$FRANK_PRIMARY_KV_BYTES" \
+  --kv-cache-dtype "$FRANK_PRIMARY_KV_DTYPE" \
+  --enable-prefix-caching \
+  --enable-auto-tool-choice --tool-call-parser "$FRANK_PRIMARY_TOOL_PARSER" \
+  --reasoning-parser "$FRANK_PRIMARY_REASONING_PARSER" \
+  --default-chat-template-kwargs "$FRANK_PRIMARY_CHAT_TEMPLATE_KWARGS" \
+  --trust-remote-code \
+  > "$LOG" 2>&1
+
+# From here on, this script's lifetime IS the container's: wait returns the
+# container's exit code, so a crash is a nonzero ExecStart and the unit's
+# Restart=on-failure is real. (systemd stop goes through ExecStop in the
+# unit — podman wait alone cannot stop the container child.)
+exec podman wait "$FRANK_PRIMARY_CONTAINER"
@@ -1,36 +1,25 @@
 #!/usr/bin/env bash
-# Serve HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced (Q4_K_M) as
-# frank-small on Dankclaw's RTX 3090 via mainline llama.cpp.
+# Serve bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF (Q5_K_M) as frank-small on
+# Dankclaw's RTX 3090 via llama.cpp (master build, CUDA sm_86).
 #
-# Topology: the whole model lives in VRAM (-ngl 999) -- 6.87 GiB of Q4_K_M
-# weights plus the BF16 vision projector and q8_0 KV inside the card's
-# 24 GiB. The host CPU only tokenizes and samples here; this box's CPU
-# belongs to the halogen engine that serves frank-large.
+# 2026-10-01 night (Frank): swapped from SGLang + INT8-W8A16-AutoRound
+# safetensors to the llama.cpp GGUF path, because llama.cpp is the only
+# runtime on this host that can serve the MiMo qwen35 hybrid with
+# ASYMMETRIC KV: --cache-type-k f16 (unquantized keys) + --cache-type-v
+# q8_0 (quantized values), which is Frank's explicit KV contract for this
+# endpoint. SGLang's --kv-cache-dtype only moves K and V together.
 #
-# MTP: this model's Multi-Token Prediction head ships as a SEPARATE file
-# (mtp-gemma-4-12B-it.gguf, arch gemma4-assistant, 4 nextn layers). It is
-# passed via --model-draft with --spec-type draft-mtp: the draft context
-# loads the assistant model. The same --spec-type value chains the
-# draftless ngram-mod matcher, which attaches unconditionally; two
-# context-backed drafters cannot share one process, so this chain is the
-# only shape that carries both.
+# Architecture: qwen3_5 hybrid — 32 layers, 8 full-attention + 24 Gated
+# DeltaNet, native 262144 window. KV lives only in the 8 attention layers;
+# the DeltaNet state is per-slot fixed size. Vision comes from the repo's
+# separate projector (the Q5_K_M GGUF carries no `v.*` tensors at all), so
+# --mmproj is what gives this endpoint eyes; MMPROJ_FILE="" is the documented
+# text-only opt-out.
 #
-# Context: CTX_SIZE is left unset and --fit sizes the KV pool (q8_0 K and
-# V) against what the weights, projector, draft context and compute
-# buffers leave on the card; the chosen value lands in the startup log.
+# Sampling: llama.cpp takes it per request (the fleet harness sends its
+# recipe from models.yml); greedy is the harness default.
 #
-# Vision: --mmproj loads the BF16 projector (it also carries an audio
-# encoder; vision is the fleet contract). The fleet's image path lives
-# here, which is why halogen no longer sets HALOGEN_VISION_TOWER.
-#
-# Sampling: greedy by default (the fleet contract). Clients that send their own
-# temperature get their own values, and the harness overrides per model from
-# models.yml. --jinja is pinned even though this build defaults it on: the
-# model's own template supplies reasoning_effort, preserve_thinking and the
-# tool-call grammar, and a future build that flips that default must not
-# silently take tools away.
-#
-# Tunables come from frank-small.env unless FRANK_SMALL_ENV_FILE overrides it.
+# Tunables come from frank-small.env unless FRANK_SMALL_ENV_FILE overrides.
 set -euo pipefail

 SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
@@ -43,87 +32,56 @@ if [[ -f "$ENV_FILE" ]]; then
 fi

 # --- defaults (used when the env file does not set them) -------------------
-: "${LLAMA_BIN:=/home/dank/GitPackages/llama.cpp-mainline/build-cuda/bin/llama-server}"
-: "${MODEL_DIR:=/home/dank/models/Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp}"
-: "${MODEL_FILE:=Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf}"
-: "${MMPROJ_PATH:=$MODEL_DIR/mmproj-Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-BF16.gguf}"
-: "${DRAFT_FILE:=}"
+: "${LLAMA_BIN:=/home/dank/GitPackages/llama-master/build-cuda/bin/llama-server}"
+: "${MODEL_FILE:=/home/dank/models/MiMo-Q5KM/MiMo-V2.6-Distill-Qwen-9B-Q5_K_M.gguf}"
 : "${ALIAS:=frank-small}"
 : "${BIND_ADDR:=0.0.0.0}"
 : "${LISTEN_PORT:=7771}"
-: "${CTX_SIZE:=}"
-: "${PARALLEL:=1}"
+# ctx-size is the POOL llama.cpp divides across slots: 540672 / 3 gives
+# 180224 per slot — three independent conversations (the native 262144 no
+# longer fits three times on a 24 GiB card once the vision tower is loaded;
+# frank-small.env carries the two measurements that set this number).
+: "${CTX_SIZE:=540672}"
+: "${PARALLEL:=3}"
 : "${CACHE_TYPE_K:=f16}"
-: "${CACHE_TYPE_V:=f16}"
-: "${BATCH_SIZE:=2048}"
-: "${UBATCH_SIZE:=512}"
-: "${NUM_THREADS:=8}"
-: "${THREAD_BATCH:=8}"
-: "${NGL:=999}"
-: "${FLASH_ATTN:=on}"
-: "${SPEC_TYPE:=draft-mtp}"
-: "${TEMPERATURE:=0.0}"
-: "${REASONING_FORMAT:=deepseek}"
-: "${REASONING_BUDGET:=-1}"
+: "${CACHE_TYPE_V:=q8_0}"
+: "${NGL:=99}"
+# Vision projector. No colon on purpose: an explicitly empty value means
+# "text-only, on purpose" and must NOT be replaced by the default path.
+: "${MMPROJ_FILE=/home/dank/models/MiMo-Q5KM/mmproj-MiMo-V2.6-Distill-Qwen-9B-f16.gguf}"

-# Fail loudly rather than half-loading a model into VRAM: a missing file here
-# is a download that has not finished, and the unit's restart policy would
-# otherwise turn that into a crash loop that reads like a runtime fault.
-for path in "$LLAMA_BIN" "$MODEL_DIR/$MODEL_FILE" "$MMPROJ_PATH"; do
+# Fail loudly rather than half-loading: a missing file is a download that
+# has not finished, and Restart=always would turn it into a crash loop that
+# reads like a runtime fault.
+for path in "$LLAMA_BIN" "$MODEL_FILE"; do
    [[ -e "$path" ]] || { echo "frank-small: missing $path" >&2; exit 1; }
 done
-if [[ -n "${DRAFT_FILE:-}" ]]; then
-    [[ -e "$MODEL_DIR/$DRAFT_FILE" ]] || { echo "frank-small: missing $MODEL_DIR/$DRAFT_FILE" >&2; exit 1; }
+EXTRA_ARGS=()
+if [[ -n "$MMPROJ_FILE" ]]; then
+    [[ -e "$MMPROJ_FILE" ]] || { echo "frank-small: missing $MMPROJ_FILE" >&2; exit 1; }
+    EXTRA_ARGS+=(--mmproj "$MMPROJ_FILE")
 fi

 # The UVM device node is created lazily by nvidia-modprobe on the first CUDA
 # open after boot; make sure it exists before the server needs it.
 nvidia-modprobe -u -c=0 2>/dev/null || true

-MMPROJ_ARGS=()
-if [[ -n "${MMPROJ_PATH:-}" && -f "$MMPROJ_PATH" ]]; then
-    MMPROJ_ARGS=(--mmproj "$MMPROJ_PATH")
-fi
-
-SPEC_ARGS=()
-if [[ -n "${SPEC_TYPE:-}" && "$SPEC_TYPE" != "none" ]]; then
-    SPEC_ARGS=(--spec-type "$SPEC_TYPE")
-fi
-
-DRAFT_ARGS=()
-if [[ -n "${DRAFT_FILE:-}" ]]; then
-    DRAFT_ARGS=(--model-draft "$MODEL_DIR/$DRAFT_FILE")
-fi
-
-# CTX_SIZE set -> pinned context. Unset -> --fit sizes the KV pool against
-# whatever the weights, projector, draft context and compute buffers leave
-# on the card; the engine prints the chosen value at startup.
-if [[ -n "${CTX_SIZE:-}" ]]; then
-    CTX_ARGS=(--ctx-size "$CTX_SIZE")
-else
-    CTX_ARGS=(--fit on)
-fi
-
-exec "$LLAMA_BIN" \
-    --model "$MODEL_DIR/$MODEL_FILE" \
-    "${MMPROJ_ARGS[@]}" \
-    "${DRAFT_ARGS[@]}" \
-    --alias "$ALIAS" \
-    --host "$BIND_ADDR" \
-    --port "$LISTEN_PORT" \
-    --gpu-layers "$NGL" \
-    --flash-attn "$FLASH_ATTN" \
-    "${CTX_ARGS[@]}" \
-    --parallel "$PARALLEL" \
-    --batch-size "$BATCH_SIZE" \
-    --ubatch-size "$UBATCH_SIZE" \
-    --threads "$NUM_THREADS" \
-    --threads-batch "$THREAD_BATCH" \
-    --cache-type-k "$CACHE_TYPE_K" \
-    --cache-type-v "$CACHE_TYPE_V" \
-    "${SPEC_ARGS[@]}" \
-    --jinja \
-    --metrics \
-    --temp "$TEMPERATURE" \
-    --reasoning-format "$REASONING_FORMAT" \
-    --reasoning-budget "$REASONING_BUDGET"
+# q8_0 on V REQUIRES flash attention (the quantized-value kernels only exist
+# on the FA path); an FA-less build silently falls back and the KV contract
+# in frank-small.env becomes a lie, so pin the flag pair here.
+ARGS=(
+    --model "$MODEL_FILE"
+    --alias "$ALIAS"
+    --host "$BIND_ADDR"
+    --port "$LISTEN_PORT"
+    --ctx-size "$CTX_SIZE"
+    --parallel "$PARALLEL"
+    --n-gpu-layers "$NGL"
+    --flash-attn on
+    --cache-type-k "$CACHE_TYPE_K"
+    --cache-type-v "$CACHE_TYPE_V"
+    --jinja
+    --reasoning auto
+    --reasoning-format deepseek
+)
+exec "$LLAMA_BIN" "${ARGS[@]}" ${EXTRA_ARGS[@]+"${EXTRA_ARGS[@]}"}
@@ -0,0 +1,176 @@
+#!/usr/bin/env python3
+"""Alertmanager-to-Signal relay for the homelab (Dankclaw).
+Listens on 192.168.0.143:18093 (the host's LAN IP) and receives
+Alertmanager's webhook payload from the container as
+host.containers.internal:18093 — netavark delivers that address to the
+host's LAN IP, which is the one bind a bridge container can reach (a
+loopback bind refuses it, and 169.254.1.2 is not a local address here).
+The group is formatted into a short message and handed to signal-cli under
+the existing account on this host. Runs as a systemd user unit
+(scripts/signal-relay.service); logs go to the journal.
+
+The phone number and account name are numbers, not credentials — the
+credential is the signal-cli account state under
+~/.local/share/signal-cli, which never leaves the host.
+
+Stdlib only: this box's user units run the system python3.
+"""
+
+import json
+import shutil
+import subprocess
+import sys
+import time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+# The bind address is the one detail this design gets wrong twice
+# (2026-09-17): a loopback bind refuses container traffic ("dial tcp
+# 169.254.1.2:18093: connection refused"), and 169.254.1.2 itself is not a
+# local address on this host ("Cannot assign requested address") — netavark
+# runs the bridges in aardvark-dns' netns, not the host's. Traffic that
+# arrives as 169.254.1.2 lands on the host's LAN IP, so bind THAT. The
+# deployer installs an iptables INPUT rule that admits 18093/tcp from the
+# podman bridge subnet (10.89.0.0/16) only, keeping the unauthenticated
+# endpoint off the rest of the LAN.
+LISTEN_ADDR = ("192.168.0.143", 18093)
+# The account is addressed by phone number: signal-cli 0.14.7's -a flag
+# parses an all-digit value as a phone number and never falls back to the
+# accounts.json name (the numeric name 318311 dies with "Invalid account
+# (phone number)"). The number is a number, not a credential — the
+# credential is the account state under ~/.local/share/signal-cli.
+SIGNAL_ACCOUNT = "+12062476336"
+SIGNAL_TO = "+12062476336"
+SIGNAL_CLI = "/usr/local/bin/signal-cli"
+
+# The relay is a safety net, not the rate limiter: Alertmanager already
+# repeats on repeat_interval. This suppresses exact duplicates that arrive
+# inside the window (config reloads, double-fires) so a misconfiguration
+# cannot fill the phone.
+DEDUPE_SECONDS = 15 * 60
+_last_sent = {}
+
+# Cap on the instances listed per group; the count is always stated.
+MAX_INSTANCES = 3
+
+
+def log(msg):
+    print(f"signal-relay: {msg}", file=sys.stderr, flush=True)
+
+
+def format_group(alerts):
+    """One message per webhook delivery (Alertmanager groups by alertname)."""
+    if not alerts:
+        return None
+    first = alerts[0]
+    name = first.get("labels", {}).get("alertname", "unknown")
+    state = "firing" if first.get("status") == "firing" else "resolved"
+    severity = first.get("labels", {}).get("severity", "info").upper()
+    lines = [f"[{severity}] {name} ({len(alerts)} {state})"]
+    for a in alerts[:MAX_INSTANCES]:
+        labels = a.get("labels", {})
+        ident = labels.get("instance") or labels.get("job") or "?"
+        summary = (a.get("annotations") or {}).get("summary")
+        line = f"  {ident}"
+        if summary:
+            line += f": {summary}"
+        lines.append(line)
+    if len(alerts) > MAX_INSTANCES:
+        lines.append(f"  … and {len(alerts) - MAX_INSTANCES} more")
+    return "\n".join(lines)
+
+
+def dedupe_key(alerts):
+    first = alerts[0]
+    labels = tuple(sorted(first.get("labels", {}).items()))
+    return (first.get("status"),) + labels
+
+
+def send_signal(text):
+    if shutil.which(SIGNAL_CLI) is None:
+        log(f"signal-cli not found at {SIGNAL_CLI}")
+        return False
+    try:
+        proc = subprocess.run(
+            [SIGNAL_CLI, "-a", SIGNAL_ACCOUNT, "-o", "json", "send", SIGNAL_TO,
+             "-m", text],
+            capture_output=True, text=True, timeout=30,
+        )
+    except subprocess.TimeoutExpired:
+        log("signal-cli timed out after 30s")
+        return False
+    if proc.returncode != 0:
+        log(f"signal-cli failed rc={proc.returncode}: {proc.stderr.strip()[:300]}")
+        return False
+    return True
+
+
+def handle_alertmanager(payload):
+    alerts = payload.get("alerts", [])
+    text = format_group(alerts)
+    if text is None:
+        log("empty alert group, nothing sent")
+        return
+    now = time.monotonic()
+    key = dedupe_key(alerts)
+    # Prune in place: rebinding _last_sent here would make it local to this
+    # function and the read above the assignment raises UnboundLocalError
+    # (measured 2026-09-17: every webhook died on it).
+    stale = [k for k, v in _last_sent.items() if now - v >= DEDUPE_SECONDS]
+    for k in stale:
+        del _last_sent[k]
+    if key in _last_sent:
+        log(f"suppressed duplicate: {key[0]}")
+        return
+    if send_signal(text):
+        _last_sent[key] = now
+        log(f"sent to {SIGNAL_TO}: {text.splitlines()[0]}")
+    else:
+        # Do not mark a failed send: the next repeat should retry.
+        log("send failed; will retry on the next delivery")
+
+
+class Handler(BaseHTTPRequestHandler):
+    def do_GET(self):
+        if self.path == "/health":
+            self.send_response(200)
+            self.end_headers()
+            self.wfile.write(b"ok")
+        else:
+            self.send_response(404)
+            self.end_headers()
+
+    def do_POST(self):
+        if self.path != "/alertmanager":
+            self.send_response(404)
+            self.end_headers()
+            return
+        length = int(self.headers.get("Content-Length", 0))
+        try:
+            payload = json.loads(self.rfile.read(length))
+        except (ValueError, UnicodeDecodeError) as e:
+            log(f"bad payload: {e}")
+            self.send_response(400)
+            self.end_headers()
+            return
+        try:
+            handle_alertmanager(payload)
+        except Exception as e:  # a relay bug must not 500 Alertmanager
+            log(f"handler error: {e!r}")
+        self.send_response(200)
+        self.end_headers()
+        self.wfile.write(b"ok")
+
+    def log_message(self, fmt, *args):
+        # Default access log is per-request noise; the journal keeps the
+        # send/suppress lines, which are the ones that matter.
+        pass
+
+
+def main():
+    server = ThreadingHTTPServer(LISTEN_ADDR, Handler)
+    log(f"listening on {'.'.join(map(str, LISTEN_ADDR))}, account {SIGNAL_ACCOUNT}")
+    server.serve_forever()
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,37 @@
+# Systemd USER unit: signal-relay (Dankclaw).
+#
+# Install path: ~/.config/systemd/user/signal-relay.service
+# Installed and enabled by install_signal_relay in scripts/deploy-remote.sh
+# (part of the alertmanager deploy).
+#
+# The webhook consumer for Alertmanager: it listens on loopback only and is
+# the only thing on this host that talks to signal-cli. It is a plain user
+# unit, not a quadlet, on purpose:
+#
+#   * it runs python3 straight from the checkout, so a deploy updates the
+#     relay without a copy step (the homepage timer copies to
+#     /usr/local/bin; this one has no root side);
+#   * generated quadlet units cannot be enabled on this host's podman 5.x
+#     v1 generator — a hand-written user unit can, and with Linger=yes for
+#     dank it starts at boot;
+#   * it needs no container: its only dependencies are /usr/bin/python3
+#     (system) and /usr/local/bin/signal-cli (host install, account state
+#     under ~/.local/share/signal-cli).
+#
+# Loopback-only for the same reason every observability port is: an
+# unauthenticated message-sending endpoint has no business on a host
+# interface. Alertmanager reaches it from its container via
+# host.containers.internal.
+
+[Unit]
+Description=Alertmanager-to-Signal relay (Dankclaw)
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+ExecStart=/usr/bin/python3 /home/dank/serverAppConfigs/scripts/signal-relay.py
+Restart=always
+RestartSec=5
+
+[Install]
+WantedBy=default.target
@@ -180,18 +180,30 @@ for path in "${CHANGED[@]}"; do
        scripts/mount-entropy.sh)
            add_test tests/mount_entropy.bats
            ;;
-        # frank-medium is llama.cpp on the two R9700s as a ROOT unit since
-        # 2026-09-16: a native gfx1201 HIP binary, so (like frank-small) its
-        # env, unit and serve script all live in scripts/, and this case has to
-        # precede the generic scripts/*.service fallback below. The retired
-        # vllm-radiance service dir is under dankserver/.AGENTTRASH/, which is
-        # archive material and maps to nothing.
-        scripts/frank-medium.*|scripts/serve-frank-medium.sh)
-            add_test tests/frank_medium_unit.bats
+        # Archive material: a path under any .AGENTTRASH/ directory is
+        # retired-by-policy (deploy-plan.py plans nothing from it), so it
+        # selects no focused test. It still needs its own case: without one
+        # the wrapper refuses retirement commits, whose trash moves show up
+        # as additions, with "no focused-test mapping".
+        .AGENTTRASH/*|*/.AGENTTRASH/*)
+            ;;
+        # frank-primary is llama.cpp (the unsloth b11030-mix prebuilt) on the
+        # two R9700s as a ROOT unit — the 2026-09-26 swap that replaced the
+        # Lucebox DS4.1 quadlet on :7770. Its env, unit and serve script all
+        # live in scripts/, which is why this case has to precede the generic
+        # scripts/*.service and scripts/* fallbacks below.
+        scripts/frank-primary.*|scripts/serve-frank-primary.sh)
+            add_test tests/frank_primary_unit.bats
            add_test tests/llm_units.bats
            add_test tests/deployment_scoping.bats
            add_test tests/scripts_syntax.bats
            ;;
+        # frank-large's chat template lives in scripts/ (moved there on
+        # 2026-09-26 when its old home was retired): a template edit changes
+        # what the live endpoint renders, so it selects the large unit seam.
+        scripts/qwen-sharp-v22.4.0.jinja)
+            add_test tests/frank_large_unit.bats
+            ;;
        # frank-small is llama.cpp on the 3090 as a ROOT unit: a native CUDA
        # binary, so no generated quadlet is involved. Its env, unit and serve
        # script all live in scripts/, which is why this case has to precede
@@ -62,13 +62,14 @@ service on the other host.
 |---|---|---|
 | Dankclaw compose services | `podman-compose -f dankclaw/compose.yaml up -d --no-deps NAME` | Recreates only `NAME`; Caddy is handled separately because it fronts both hosts. |
 | SearXNG, Gluetun, Valkey, OpenWebUI, Pi-hole, Gitea database/server | user Quadlet + `systemctl --user` | Installs the selected unit and restarts only the required dependency-coupled units. Gitea's web proxy is the canonical Caddy compose service. |
-| `frank-large` | systemd + `systemctl` | Relinks and restarts only the large inference service; waits up to 20 minutes for `/health`. |
-| `frank-medium` | systemd + `systemctl` | Retires the legacy frank-small unit on first deploy, then relinks and restarts only the inference service; waits up to 20 minutes for `/health`. |
+| `frank-large` | systemd + `systemctl` (root unit) | Relinks and restarts only the large inference service; waits up to 20 minutes for `/health`. |
+| `frank-small` | systemd + `systemctl` (root unit) | Relinks and restarts only the small inference service; waits up to 20 minutes for `/health`. |
+| `frank-medium` | rootless Quadlet + `systemctl --user` | Installs `dankserver/frank-medium/quadlet/frank-medium.container`; on first deploy retires the llama.cpp root unit and the vLLM replica quadlets into the home trash, then restarts the generated unit. Waits up to 20 minutes for `/health`. |
 | Dankserver media stack (RomM trio, Jellyfin, Calibre-Web, Copyparty) | user Quadlet + `systemctl --user` | Installs `dankserver/<svc>/quadlet/*.container`, restarts the generated unit. The first cutover per service is idempotent: it retires the compose-era bare container, rescues any anonymous volume into the data root, and merge-moves `/opt/<svc>` to `/home/dank/Containers/<svc>` (mapped-uid roots move without chown). Steady state is restart-only. |

 The shared chat template maps to the inference service on the selected target.
-The inactive `frank-npu` unit is relink-only when explicitly selected;
-it is not restarted by an unrelated large-model deployment.
+The `frank-npu` endpoint is retired (2026-09-11) and has no deploy handler;
+an unrelated large-model deployment never touches it.

 ## Caddy safety

@@ -1,3 +1,13 @@
+---
+type: reference
+description: "`tests/e2e/` drives the **live fleet** through a real browser: fresh contexts"
+tags:
+  - architecture
+  - e2e
+  - homepage
+timestamp: 2026-09-01
+---
+
 # E2E Suite — homepage + guard chain (Playwright)

 `tests/e2e/` drives the **live fleet** through a real browser: fresh contexts
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "One Caddy fronts both hosts."
+tags:
+  - architecture
+  - network
+timestamp: 2026-09-02
+---
+
 # Network Architecture

 One Caddy fronts both hosts. It runs as a container on Dankclaw (192.168.0.143) and terminates TLS for every service on the LAN. Pi-hole on Dankclaw answers the service names for LAN clients; Porkbun answers them for the WAN. The G36 router firewall governs external access. SearXNG outbound traffic exits through the Gluetun VPN container.
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "All persistent data uses **bind mounts**: a host directory mapped directly into the container."
+tags:
+  - architecture
+  - storage
+timestamp: 2026-09-01
+---
+
 # Storage Policy

 All persistent data uses **bind mounts**: a host directory mapped directly into the container. No named volumes. This matches the preference of this repo and the existing Gitea stack on Dankclaw.
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "This repository tests its configuration at defined seams."
+tags:
+  - architecture
+  - testing
+timestamp: 2026-09-01
+---
+
 # Testing Policy

 This repository tests its configuration at defined seams. The tests run with `bats`; run them with `./tests/run.sh` before every push.
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "ComfyUI is a node-based AI image generation engine."
+tags:
+  - dankclaw
+  - comfyui
+timestamp: 2026-09-16
+---
+
 # ComfyUI

 ComfyUI is a node-based AI image generation engine. Workflows are built from nodes that load models, sample latents, and post-process images; everything is scriptable through a web UI and an API.
@@ -13,6 +22,7 @@ https://github.com/Comfy-Org/ComfyUI
 |------|-------|
 | Image | `comfyui-rocm:gfx1151` — built from `dankclaw/Containerfile.comfyui` |
 | Base image | `docker.io/rocm/vllm:rocm7.12.0_gfx1151_ubuntu24.04_py3.12_pytorch_2.9.1_vllm_0.16.0` |
+| Upstream latest | 0.37.0 (2026-09-21) — ships **native Qwen-Image 2.1** support (CORE-423, `comfy/ldm/qwen_image21/`). The image's custom GGUF patches may be unnecessary on the next rebuild; check before re-applying. |
 | Port | 8188 (host) → 8188 (container) |
 | Backend | ROCm GPU (gfx1151 / Radeon 8060S). No `--cpu` flag. |
 | Data | `/home/dank/comfyui-data/{models,input,output,custom_nodes}` → `/opt/ComfyUI/...` |
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "Gluetun is a lightweight VPN client container."
+tags:
+  - dankclaw
+  - gluetun
+timestamp: 2026-08-28
+---
+
 # Gluetun

 Gluetun is a lightweight VPN client container. It connects to a VPN provider (WireGuard or OpenVPN) and shares that tunnel with other containers that attach to its network namespace.
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "The Hermes web dashboard is a browser UI for managing the agent."
+tags:
+  - dankclaw
+  - hermes
+timestamp: 2026-09-16
+---
+
 # Hermes-Dashboard

 The Hermes web dashboard is a browser UI for managing the agent. It shows config, API keys, and sessions. It is a management tool. It is not a chat window. Chat happens over Signal or the CLI. See [[Hermes]].
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "Hermes is the self-improving AI agent from Nous Research."
+tags:
+  - dankclaw
+  - hermes
+timestamp: 2026-09-15
+---
+
 # Hermes

 Hermes is the self-improving AI agent from Nous Research. It runs in a terminal, Telegram, Discord, Slack, WhatsApp, or Signal, and keeps a persistent learning loop: it creates skills from experience, searches past conversations, and builds a model of the user across sessions.
@@ -20,6 +29,7 @@ is not durable. The tracked, sanitized configuration baseline is
 | Item | Value |
 |------|-------|
 | Image | built from `NousResearch/hermes-agent` (official Dockerfile in repo) |
+| Upstream latest | agent v0.21.4 / CalVer `v2026.9.21` (2026-09-21); hermes-webui v0.52.113 (2026-08-14). Pin the CalVer tag, not `:latest`. |
 | License | MIT |
 | Model backends | Nous Portal, OpenRouter, OpenAI, or any custom endpoint |
 | Interfaces | TUI, Telegram, Discord, Slack, WhatsApp, Signal |
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "OpenWebUI is a self-hosted chat interface for LLMs."
+tags:
+  - dankclaw
+  - openwebui
+timestamp: 2026-09-21
+---
+
 # OpenWebUI

 OpenWebUI is a self-hosted chat interface for LLMs. It connects to OpenAI-compatible endpoints (vLLM, llama.cpp, Ollama) and adds RAG, plugins, tool calling, and multi-user support.
@@ -12,32 +21,110 @@ https://github.com/open-webui/open-webui
 | Item | Value |
 |------|-------|
 | Image | `ghcr.io/open-webui/open-webui:main` (verified from the live quadlet unit) |
+| Upstream latest | 0.11.4 (2026-09-21) — security/access-control advisory; `:main` is the rolling image, not a stable pin. Imagegen endpoint is `POST /api/v1/images/generations` (the `comfyui` engine is still upstream). |
 | Port | 3005 (host) → 8080 (container). Port 3000 is Gitea's HTTP, so it is not usable. |
 | Data volume | `/home/dank/Containers/open-webui/data` → `/app/backend/data` |
-| Default backend | `frank-large` on Dankclaw `:7770` (ROCmFPX llama.cpp fork — see the LLM knowledge base below) — reached via host LAN IP `192.168.0.143:7770` |
+| Backends | Three fleet endpoints: `frank-large` (host.containers.internal:7770), `frank-medium` (192.168.0.235:7770), `frank-small` (host.containers.internal:7771) — see [The LLM backends](#the-llm-backends) |
 | Network | Pinned bridge `systemd-open-webui` (`10.89.3.0/24`, gateway `10.89.3.1`) — see [History](#history) |
 | Persistence | Quadlet unit `open-webui.service`; `[Install] WantedBy=default.target` |
 | Deploy | `scripts/deploy-open-webui-quadlet.sh dankclaw` |

 ## The LLM backends

-The fleet runs two OpenAI-compatible endpoints. This mapping is the **live**
-state (verified 2026-09-03). Both services listen on `:7770`, distinguished by
-host; `frank-large` moved from `:52625` to `:7770` on 2026-08-22.
+OpenWebUI serves the three fleet LLM endpoints. The model picker is built from
+the `model` table in `webui.db` (one row per model), NOT from live connection
+discovery — so the picker shows exactly the models seeded there.

-| Endpoint | Model name | Host | Capability |
-|----------|-----------|------|------------|
-| `:7770` (ROCmFPX llama.cpp fork) | `frank-large` | Dankclaw | completion |
-| `:7770` (ROCmFPX llama.cpp fork) | `frank-medium` | Dankserver | completion |
+| Picker model | Connection URL | Runtime | Host |
+|--------------|----------------|---------|------|
+| `frank-large` | `http://host.containers.internal:7770/v1` | unsloth llama.cpp (Strix Halo iGPU, 200k ctx) | Dankclaw |
+| `frank-medium` | `http://192.168.0.235:7770/v1` | vllm-radiance TP=2 (FP8, DFlash2, 262k ctx) | Dankserver |
+| `frank-small` | `http://host.containers.internal:7771/v1` | llama.cpp (RTX 3090, Gemma4-12B, 262k ctx) | Dankclaw |

-The default connection points at the on-host endpoint (`frank-large`).
-To add the other backend, open **Admin → Connections** in the UI and point an
-OpenAI connection at `http://192.168.0.235:7770/v1` (key `none`).
+Two gotchas make the wiring non-obvious:
+
+- **llama.cpp `/v1/models` is Ollama-shaped.** llama.cpp returns
+  `{"models":[{"name":...}]}`; OpenWebUI's `extract_data()` only reads the
+  OpenAI `{"data":[{"id":...}]}` shape, so connection discovery silently misses
+  the llama.cpp endpoints (frank-large, frank-small). The fix is to set
+  `model_ids` in each connection's `api_config` — when non-empty, OpenWebUI
+  builds the model list directly and never fetches `/v1/models`. frank-medium
+  (vLLM) serves the OpenAI shape and would discover fine, but uses `model_ids`
+  too for consistency.
+- **The host's own services are reached via `host.containers.internal`.** The
+  container sits on a netavark bridge; it can hairpin to a *different* host's
+  LAN IP (dankserver `192.168.0.235`) but not to its *own* host's non-published
+  ports via the host LAN IP (`192.168.0.143:7770/7771` is refused).
+  `host.containers.internal` (169.254.1.2) resolves to the host on the bridge
+  and reaches both llama.cpp services.
+
+The Ollama connection is disabled (`ollama.enable=false`).
+
+> All three endpoints are bare-metal systemd services managed by this
+> repository. Model swaps, context sizing, and spec tuning happen there. See
+> the [LLM knowledge base](../LLM/index.md).
+
+## Web search
+
+Chat web search is served by [[SearXNG]] on this host, wired in the quadlet
+(`ENABLE_WEB_SEARCH`, `WEB_SEARCH_ENGINE=searxng`, `SEARXNG_QUERY_URL`).
+
+- **Access path:** the backend calls `http://192.168.0.143:8081/search?format=json`
+  directly — gluetun's published port, NOT the `search.franksdank.cloud` vhost.
+  The Anubis→Authentik guard chain is a browser mechanism (proof-of-work +
+  session flow); a plain HTTP client cannot solve it, and the internal port
+  bypasses Caddy entirely.
+- **Limiter:** SearXNG caps non-HTML (JSON) requests at 4/IP/hour (`API_MAX`,
+  a hardcoded Python constant — no config knob). The `pass_ip` list is checked
+  FIRST and short-circuits the whole botdetection chain, and OpenWebUI's bridge
+  egress (`10.89.x`) is in `pass_ip`, so the cap never applies. Verified: an
+  8-request burst from the container returned 8×200.
+- **Config storage:** OpenWebUI keeps web-search settings per-key in its SQLite
+  config table (`webui.db`, `web.search.*`). The quadlet env vars seed that
+  table on first start (fill-missing only); afterwards the DB is authoritative
+  and Admin-UI edits persist there.
+
+## SSO (authentik)
+
+Login is authentik via the OIDC authorization-code flow (wired 2026-09-24).
+The vhost stays exempt from the Anubis guard chain: the OAuth dance is
+browser-driven and only touches the two exempt vhosts (`ai.` and `auth.`).
+
+- **authentik objects** (created via the admin API, 2026-09-24): application
+  `open-webui` (slug `open-webui`) with the OAuth2 provider `Open WebUI`
+  (confidential client, `authorization_code` + `refresh_token`, Strict
+  redirect URI `https://ai.franksdank.cloud/oauth/oidc/callback`, implicit
+  consent flow). The client ID is in the quadlet; the client secret lives in
+  the git-ignored host file `/home/dank/Containers/open-webui/env` (mode 600,
+  loaded via `EnvironmentFile=`).
+- **Roles:** the scope mapping `Open WebUI roles` (scope `roles`) reads the
+  application entitlements `Open WebUI Users` → `user` and `Open WebUI
+  Admins` → `admin`. The entitlements are bound to the groups
+  `openwebui_users` / `openwebui_admins`. `akadmin` (frank8465@protonmail.com)
+  is in `openwebui_admins`; `qa-e2e-admin` is in `openwebui_users`.
+- **Account merge:** `OAUTH_MERGE_ACCOUNTS_BY_EMAIL=true` folds an SSO login
+  into the existing local account with the same email — the pre-SSO admin
+  (frank8465@protonmail.com) keeps its data. `ENABLE_OAUTH_SIGNUP=true` lets
+  other `openwebui_users` create accounts on first login.
+- **IdP reachability:** the container is on its own bridge (10.89.3.0/24) and
+  cannot reach `authentik-server:9000` by name. It reaches the IdP over the
+  public URL `https://auth.franksdank.cloud` — Caddy publishes `0.0.0.0:443`
+  (netavark DNATs published ports from other containers) and Pi-hole resolves
+  the name to the host LAN IP. The server-side code exchange rides that path.
+- **Config authority:** the OAuth env vars in the quadlet seed the per-key
+  config DB on first run (fill-missing only); on the live data dir the Admin
+  Panel is authoritative. To rotate the client secret: create a new token in
+  authentik, update the host env file, and set the new value in the Admin
+  Panel (or the DB), then restart the container.
+- **SSO-only (2026-09-24):** after the OIDC round-trip was verified, the local
+  username/password path was retired. `ENABLE_LOGIN_FORM=false` hides the form
+  and `ENABLE_PASSWORD_AUTH=false` refuses password logins, so the only way in
+  is the authentik OIDC flow. The `ui.enable_login_form` config-DB override was
+  flipped to `false` in the same change — the per-key DB wins over env, so both
+  must agree, and the DB value is the one to change on a live data dir. To
+  restore the password form: set `ui.enable_login_form` back to `true` in the
+  Admin Panel (or DB) and drop the two env vars from the quadlet.

-> Both endpoints are bare-metal systemd services managed by this repository
-(the former `vllmScripts` repo, merged in 2026-08-22). Model swaps, context
-sizing, and spec tuning happen there. See the [LLM knowledge
-base](../LLM/index.md).

 ## Boot persistence (Quadlet)

@@ -71,12 +158,18 @@ The quadlet carries a built-in healthcheck (30s interval, 3 retries, 300s start
 - The image tracks upstream `main`; pin a digest in the quadlet unit if reproducible rollback matters more than tracking. (The old "CPU-only host" framing is stale: this APU runs the gfx1151 ROCm build — see [[ComfyUI]].)
 - The default admin account is created on first login; secure it immediately.
 - If the backend is loading (3–8 minute model load), OpenWebUI shows connection errors until the model is ready.
- The LLM is reached via the host LAN IP `192.168.0.143:7770`, not the bridge gateway: the gateway (`10.89.3.1`) lives in the rootless netns, while llama-server runs in the host netns, so the gateway refuses. `host.containers.internal` is slirp4netns-only and does not resolve on a bridge either. To reach the Dankserver backend, use the LAN address `192.168.0.235:7770`.
+- The on-host LLM services (frank-large `:7770`, frank-small `:7771`) are reached via `host.containers.internal`, NOT the host LAN IP: the container's netavark bridge cannot hairpin to its own host's non-published ports via `192.168.0.143` (connection refused), and the bridge gateway (`10.89.3.1`) lives in the rootless netns while llama-server runs in the host netns, so it refuses too. `host.containers.internal` (169.254.1.2) resolves to the host on the bridge and works — verified end-to-end for both endpoints. The Dankserver backend (frank-medium) is a different host, reached via its LAN IP `192.168.0.235:7770`.

 ## History

 - **2026-08-22 — slirp4netns → pinned bridge.** The unit ran on `Network=slirp4netns`. The slirp4netns helper process died (no OOM, no kill entry, no crash log found), and the container kept running with only `lo` in its netns: published port 3005 accepted TCP and reset every request, so chat served empty responses for ~2 days. Nothing logged an error. Moving to a kernel bridge removes the per-container helper entirely, and the healthcheck's interface check makes the zombie state visible if it ever returns.

+- **2026-09-21 — model picker rebuilt for the 3-model fleet.** The picker carried 39 stale rows (37 dead Ollama models + 2 `frank-default-*`) and was missing `frank-medium`. Rebuilt the `model` table to exactly `frank-large`/`frank-medium`/`frank-small` and set `model_ids` on each connection (llama.cpp's Ollama-shaped `/v1/models` is invisible to OpenWebUI's discovery). Pointed the two on-host endpoints at `host.containers.internal` — the bridge refuses the host LAN IP for non-published ports. All three verified end-to-end via `/openai/chat/completions`.
+
+- **2026-09-24 — SSO via authentik.** OIDC authorization-code flow wired per the [SSO (authentik)](#sso-authentik) section: application/provider/entitlements created in authentik, client secret in the host env file, roles from the `roles` scope claim. E2E coverage in `tests/e2e/specs/18-openwebui-sso.spec.ts`.
+
+- **2026-09-24 — flipped to SSO-only.** The OIDC round-trip verified in a real browser (login view renders no password form, only "Continue with authentik", and the flow lands on a signed-in session), so the local username/password path was retired: `ENABLE_LOGIN_FORM=false` + `ENABLE_PASSWORD_AUTH=false` in the quadlet and `ui.enable_login_form` set to `false` in the config DB (the DB wins over env, so both were changed together). `qa-e2e-admin` was demoted out of the `authentik Admins` group (it keeps `openwebui_users` / `homepage-admins` for app access).
+
 ## See Also

 - [[Dankclaw Stack]]
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "PostgreSQL is the database backend for Gitea."
+tags:
+  - dankclaw
+  - postgres
+timestamp: 2026-09-03
+---
+
 # Postgres

 PostgreSQL is the database backend for Gitea. It runs on the same podman network as Gitea and holds the repository metadata, users, and issue data.
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "How Google and other upstream search engines get blocked behind the Gluetun"
+tags:
+  - dankclaw
+  - searxng
+timestamp: 2026-08-28
+---
+
 # SearXNG Providers

 How Google and other upstream search engines get blocked behind the Gluetun
@@ -354,8 +363,15 @@ misleading).
   disable pagination for rate-sensitive engines.
 3. **Promote by evidence** — raise `weight` only for engines that return useful
   results consistently; weight is not a health check.
-4. **Avoid credentials** — do not add API-backed engines unless the provider is
-   genuinely free, anonymous, and compatible with our privacy requirements.
+4. **Avoid credentials, with one carve-out** — do not add API-backed engines
+   unless the provider is genuinely free, anonymous, and compatible with our
+   privacy requirements. Carve-out (2026-09-26): a paid, user-ordered API
+   engine (Kagi) is acceptable when the blast radius is bounded — the
+   `kagi` engine runs with `paging: false` and weight 2, so it spends exactly
+   one API credit per aggregated query, and its key never enters git
+   (deploy-time rendering from `/opt/searxng/.env`; see [[SearXNG]]). It is
+   also immune to the shared-VPN-exit-IP blocking this page is about: it
+   authenticates, so reputation of the exit IP is irrelevant.
 5. **Verify first** — `/json` + `?engines=google` and the searx-checker to see
   exactly which engines are down and why. Confirm the running SearXNG version
   (check whether `parsers`/`rate_limit` exist in the pinned build).
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "SearXNG is a privacy-respecting metasearch engine."
+tags:
+  - dankclaw
+  - searxng
+timestamp: 2026-09-03
+---
+
 # SearXNG

 SearXNG is a privacy-respecting metasearch engine. It aggregates results from many upstream search engines into one result set and serves both an HTML UI and a JSON API for agents.
@@ -51,8 +60,8 @@ Deploy with `scripts/deploy-searxng-quadlet.sh dankclaw`. It copies the units, d
 ```yaml
 use_default_settings: true
 server:
-  secret_key: "${SEARXNG_SECRET}"   # from .env
-  base_url: "${SEARXNG_BASE_URL}"  # resolves to http://192.168.0.143:8081/
+  secret_key: "${SEARXNG_SECRET}"   # rendered at deploy by scripts/searxng-render-settings.py
+  base_url: "${SEARXNG_BASE_URL}"   # same renderer; SearXNG does NOT expand env vars itself
  limiter: true
  method: "POST"

@@ -81,8 +90,16 @@ engines:
  - name: duckduckgo
    disabled: false
    weight: 2
-    paging: false
-    timeout: 5.0
+  - name: kagi
+    engine: kagi
+    categories: [general, web]
+    shortcut: kg
+    api_key: "${KAGI_API_KEY}"      # rendered at deploy from /opt/searxng/.env
+    kagi_categ: search
+    disabled: false
+    weight: 2
+    paging: false                  # one API credit per aggregated query
+    timeout: 8.0
  - name: google cse
    disabled: true
  - name: brave
@@ -159,10 +176,39 @@ root cause, the suspend-model table, the `engines:` config, and the
  LAN by construction (no public route to the LAN IP). No second agent-only
  instance is needed; a duplicate would double upstream request volume from the
  same rate-sensitive VPN exit IP for zero benefit.
- The config file is root-owned (uid of the rootless podman namespace). Update it with a one-off `sudo cp`, then run `systemctl --user restart searxng.service` to reload.
+- The config file is root-owned (uid of the rootless podman namespace). The
+  official path is `deploy.sh --service searxng`: it copies the repo settings
+  and renders the `${VAR}` placeholders via `scripts/searxng-render-settings.py`
+  from `/opt/searxng/.env`. A hotfix can still hand-copy with `sudo cp`, but
+  remember the renderer: an unrendered copy silently serves the literal
+  placeholders (before 2026-09-26 the live `secret_key` literally was
+  `"${SEARXNG_SECRET}"` and nobody noticed).
 - `base_url` must match the public address or proxy rewrites break.
 - Keep the valkey cache container on the same podman network as searxng.

+
+## Kagi (paid API engine)
+
+Added 2026-09-26 (user-ordered). The stock SearXNG image ships a native
+`kagi` engine module (`searx/engines/kagi.py`); it authenticates with
+`Authorization: Bearer <key>` against `POST https://kagi.com/api/v1/search`.
+
+- **Enabled**: `weight: 2` alongside bing/duckduckgo, categories
+  `[general, web]`, shortcut `kg`, `kagi_categ: search`.
+- **Cost control**: `paging: false` — exactly one API credit per aggregated
+  query; `timeout: 8.0` so a slow paid query is not wasted (scrapers use 5.0).
+- **Key handling**: the repo file keeps `api_key: "${KAGI_API_KEY}"`;
+  `deploy-remote.sh` bakes the real key into the live settings.yml via
+  `scripts/searxng-render-settings.py`, reading `/opt/searxng/.env`
+  (dank-owned, mode 600). The key never enters git and never appears in any
+  process argv. SearXNG itself does NOT expand env vars in settings.yml —
+  without the renderer the engine fails `setup()` and is dropped at load.
+- **Immune to the VPN-exit-IP problem**: Kagi authenticates, so it does not
+  care about the shared Gluetun exit IP reputation (see [[SearXNG Providers]]).
+- **Verified live 2026-09-26**: direct API probe through the gluetun netns
+  (200, `data.search[]` shape), then a live aggregated query returned
+  29/29 results tagged `kagi`, zero unresponsive engines.
+
 ## See Also

 - [[Dankclaw Stack]]
@@ -1,3 +1,12 @@
+---
+type: reference
+description: "Seerr is the unified request manager for the *arr stack (the maintained `seerr-team/seerr` successor to the archived Overseerr/Jellyseerr line)."
+tags:
+  - dankclaw
+  - seerr
+timestamp: 2026-09-02
+---
+
 # Seerr

 Seerr is the unified request manager for the *arr stack (the maintained `seerr-team/seerr` successor to the archived Overseerr/Jellyseerr line). Frank's household uses it to request movies/TV/music; Seerr routes approvals to Sonarr/Radarr/Lidarr/Readarr on FrankBook-Pro and pushes downloads to qBittorrent.
--- a/Show More
+++ b/Show More