docs(frank-large): unsloth Qwen3.8-Flash-Next runbook; retire the CIRU Orca runbook

New runbook frank-large-unsloth-qwen38-flash-next.md (runtime, payload, deployed config, the explicit -md, the 114 GiB GTT memory model, client contract, rollback). frank-large-ciru-orca.md marked RETIRED (rollback target, payload and runtime intact on disk). AGENTS.md topology row, frank-large section, harness note, runbook index, and the GPU-budget lesson (GTT aperture ~114 GiB, not the stale 64 GiB rocm-smi figure — the CIRU Orca held 92.6 GiB on the same device). LLM index header, mermaid, active list, and the vault root index updated. Work-log entry.
2026-09-21 12:30:35 -07:00
parent b149d02187
commit 7010f5b754
6 changed files with 307 additions and 63 deletions
@@ -12,7 +12,7 @@ merge absorbed the former `vllmScripts` repo, history preserved).
  Calibre-Web, Jellyfin, Plex) plus the `frank-medium` 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` (CIRU v4.4.1 llama.cpp on
+  Alloy, Pi-hole) plus TWO LLM services: `frank-large` (unsloth llama.cpp prebuilt on
  the gfx1151 iGPU) and `frank-small` (llama.cpp on the RTX 3090). Caddy on
  Dankclaw reverse-proxies both hosts.
 - **FrankBook-Pro** (192.168.0.112) — macOS mini running the *arr stack
@@ -25,7 +25,7 @@ merge absorbed the former `vllmScripts` repo, history preserved).
 | Target | Service | Runtime | Model | Port |
 |--------|---------|---------|-------|------|
 | Dankserver | `frank-medium` | vllm-radiance 0.9.3 (rootless Quadlet, image `localhost/radiance-vllm:0.9.3`, gfx1201) | orcarouter Qwen3.8-27B-Uncensored-FP8 (FP8 E4M3 block-wise, 30.9 GB / 7 shards) with the z-lab DFlash2 external drafter (depth 7, live mean accept length 4.24/7 — A/B tested and beats MTP self-speculation on decode: 100.0 vs 85.9 t/s), 2x Radeon AI PRO R9700 in ONE tensor-parallel process (TP=2 over the R4D one-shot allreduce — no RCCL collective; both cards on CPU root ports at Gen5 x8 since the 2026-09-20 ProArt B850-Creator swap), 262144 ctx over up to 8 conversations (auto-sized 384,238-token KV pool = 1.47x the window), fp8 KV, prefix caching, vision | 7770 |
-| Dankclaw | `frank-large` | CIRU v4.4.1 llama.cpp (QSA window hotfix, root systemd unit, gfx1151 HIP build) | jcbtc Qwen3.8-Flash-CIRU-STRIX-Orca on the gfx1151 iGPU (Strix Halo LPDDR), 262144 ctx single stream (1 slot), MTP depth 4, PLE 4096 MiB, text only | 7770 |
+| Dankclaw | `frank-large` | unsloth llama.cpp b11030 prebuilt (ROCm gfx1151, root systemd unit) | unsloth Qwen3.8-Flash-Next UD-IQ4_XS on the gfx1151 iGPU (Strix Halo LPDDR, 114 GiB GTT), 200000 ctx single stream (1 slot), MTP depth 2 via the explicit `-md` shared-Q8_0 head, text only | 7770 |
 | Dankclaw | `frank-small` | mainline llama.cpp master 37b53fd (root systemd unit) | HauhauCS Gemma4-12B-QAT-Uncensored-Balanced Q4_K_M (BF16 mmproj + MTP assistant sidecar, `draft-mtp,ngram-mod`) on the RTX 3090, 1048576 pool / 4 slots of 262,144, q8_0 KV, the fleet's fastest vision endpoint | 7771 |

 Retired endpoints (`frank-npu`, and every prior runtime) are in
@@ -265,7 +265,7 @@ a role pointing at a provider the file does not define fails at call time. As
 of 2026-09-20 both `frank-small` and `frank-medium` accept images
 (`input: [text, image]`) — frank-medium's checkpoint serves vision through the
 Gemma4 merge path with no separate mmproj, verified on held-out synthetic
-images; `frank-large` remains text-only because its CIRU deployment loads no
+images; `frank-large` remains text-only because its unsloth deployment loads no
 vision projector (the F16 mmproj is on disk, one env flip away). Note also that
 the reasoning channel differs by runtime: llama.cpp endpoints emit
 `reasoning_content`, vLLM emits `reasoning`.
@@ -273,40 +273,45 @@ There is **no audio**
 input anywhere in the fleet: that capability left with the FLM NPU endpoint on
 2026-09-11.

-### frank-large — CIRU v4.4.1 Orca (Dankclaw, ACTIVE)
+### frank-large — unsloth llama.cpp Qwen3.8-Flash-Next UD-IQ4_XS (Dankclaw, ACTIVE)

-CIRU v4.4.1 (the QSA window hotfix) as a **root systemd unit** (`scripts/
-frank-large.service`, `User=dank`), serving `jcbtc/Qwen3.8-Flash-CIRU-STRIX-
-Orca` (79.4 GiB main GGUF + a 4.1 GiB MTP head + a 52.4 GiB PLE payload, all
-sha256-verified) on port 7770 from the gfx1151 iGPU. One stream, the model's
-FULL 262,144 trained window (`PARALLEL_SLOTS=1`; the MTP drafter needs exactly
-one slot in this runtime), MTP depth 4, PLE cache 4096 MiB, reasoning extracted
-to `reasoning_content`, and **no vision projector** (the F16 mmproj is on disk,
-`ENABLE_VISION=0` — one env flip to enable). It replaced halogen-flash-server,
-which had replaced ik_llama.cpp.
+The unsloth llama.cpp prebuilt `app-b11030-mix-5ff778e` (ROCm gfx1151,
+`/home/dank/GitPackages/unsloth-llama-b11030/`, bundles its own ROCm runtime —
+no host SDK gate, no venv) as a **root systemd unit** (`scripts/
+frank-large.service`, `User=dank`), serving
+`unsloth/Qwen3.8-Flash-Next-GGUF:UD-IQ4_XS` (87.3 GiB main GGUF, three HF
+split parts concatenated, + a 2.6 GiB MTP head, all sha256-verified) on port
+7770 from the gfx1151 iGPU. One stream, 200,000 ctx (`PARALLEL_SLOTS=1`,
+operator's requested shape; the trained window is 262,144), MTP depth 2
+(`SPEC_DRAFT_N_MAX=2`, the model card's default), reasoning extracted to
+`reasoning_content`, and **no vision projector** (the F16 mmproj is on disk,
+`ENABLE_VISION=0` — one env flip to enable). It replaced the CIRU v4.4.1
+Orca runtime, which had replaced halogen-flash-server.

-The runtime is a **private host build** (`/home/dank/GitPackages/ciru-runtime-
-v4.4.1-qsa-hotfix/`), not a container image: a gfx1151 HIP build of the CIRU
-llama.cpp fork with its own bundled ROCm SDK in a venv (`.venv-rocm/`), and the
-binary resolves the SDK libs via RUNPATH. The launcher (`scripts/ciru/
-run-orca-server.sh` → `run-server.sh`) has a **runtime gate**: it exits 2 unless
-`${repo_root}/runtime/hip/lib/libamdhip64.so` and `runtime/rocr/lib/
-libhsa-runtime64.so` exist. The serve script (`scripts/serve-frank-large.sh`)
-satisfies that gate idempotently by `ln -sfn`-ing both to the venv's unversioned
-devel symlinks, then `exec`s the launcher with `--chat-template-kwargs
-'{"preserve_thinking":true}'` (Qwen3.8's template needs this to carry thinking
-blocks across turns; a request-level value still wins).
+The runtime is the first with an MTP graph for the qwen4exp architecture
+(stock ggml-org mainline has none), so the unsloth prebuilt is the
+llama.cpp-compatible choice. The MTP head is loaded via an **explicit
+`-md`**: sidecar auto-discovery does not search the `mtp/` subfolder, and
+without it `--spec-type draft-mtp` silently runs at the main model's speed.
+The serve script (`scripts/serve-frank-large.sh`) validates binary → main
+GGUF → MTP head and exits with a pointer message if any is missing, then
+`exec`s `llama-server` with `--jinja --alias frank-large --chat-template-
+kwargs '{"preserve_thinking":true}'` (Qwen3.8's template needs this to
+carry thinking blocks across turns; a request-level value still wins).

-The serve script validates launcher → built binary → venv runtime → main GGUF →
-MTP head → PLE payload + manifest + scale, and exits with a pointer message if
-any is missing. `PLE_CACHE_MIB=4096` is the vendor production-profile default;
-the 52.4 GiB PLE is mmap'd and paged, not resident.
+The iGPU addresses **114 GiB** (the GTT aperture over the shared LPDDR,
+`gtt_total` in `/sys/class/drm/card1/device/`; `rocm-smi`'s 2 GiB "dedicated"
+is not the budget). The 87.3 GiB model is mmap'd in with `-ngl 999`; budget
+at 200k ≈ 100 GiB (weights + MTP head + KV + buffers), leaving ~14 GiB of
+aperture headroom. The n-gram embedding (PLE) is inside the GGUF — no
+external PLE payload or pager, unlike the CIRU Orca packaging.

-The halogen Quadlet (`dankclaw/frank-large/quadlet/`) was retired to
-`.AGENTTRASH/` — one owner per endpoint, now the scripts/ trio.
+Fetch: `scripts/download-frank-large-unsloth.sh` (resumable, each part
+verified against its LFS sha256 fetched live from the Hub tree API, then the
+assembled GGUF against `SHA256SUMS`; `--verify-only` re-checks).

-Runbook: `serverAppConfigs-docs/LLM/frank-large-ciru-orca.md`.
-Rollback: [[halogen-flash-server]].
+Runbook: `serverAppConfigs-docs/LLM/frank-large-unsloth-qwen38-flash-next.md`.
+Rollback: [[frank-large-ciru-orca]] (payload and runtime intact on disk).

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

@@ -530,9 +535,13 @@ 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**

@@ -743,7 +752,7 @@ Active:

 | Service | Runbook |
 |---------|---------|
-| frank-large (CIRU v4.4.1 Orca, Dankclaw) | `serverAppConfigs-docs/LLM/frank-large-ciru-orca.md` |
+| frank-large (unsloth llama.cpp Qwen3.8-Flash-Next UD-IQ4_XS, Dankclaw) | `serverAppConfigs-docs/LLM/frank-large-unsloth-qwen38-flash-next.md` |
 | frank-small (llama.cpp Gemma4-12B QAT, Dankclaw) | `serverAppConfigs-docs/LLM/frank-small-gemma4-12b.md` |
 | frank-medium (radiance TP=2 + MTP, Dankserver) | `serverAppConfigs-docs/LLM/frank-medium-radiance-mxfp4.md` |
 | frank-medium model swap (toolchain + process) | `serverAppConfigs-docs/LLM/frank-medium-model-swap.md` |
@@ -757,6 +766,7 @@ Retired (rollback targets and history):
 | kingjones STRIX Qwen3.8-Flash-Next ROCmFP4 | Dankclaw | 2026-08-30 → 09-11 | `kingjones-strix-rocmfp4.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` |
 | 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` |
 | 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` |
 | agentionai ROCmFP4-FAST imatrix | Dankclaw | 2026-08-29 → 08-30 | `agentionai-fast-imatrix.md` |
@@ -1,7 +1,7 @@
 ---
 type: reference
 title: CIRU v4.4.1 Orca on Dankclaw (frank-large)
-description: The active frank-large runbook — CIRU v4.4.1 (QSA window hotfix) serving jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca on the Dankclaw Strix Halo iGPU: the private host build, the runtime gate, the payload, the deployed config, MTP and PLE, the client contract, and the failure modes worth knowing.
+description: The retired frank-large runbook (2026-09-19 to 2026-09-21, rollback target) — CIRU v4.4.1 (QSA window hotfix) serving jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca on the Dankclaw Strix Halo iGPU: the private host build, the runtime gate, the payload, the deployed config, MTP and PLE, the client contract, and the failure modes worth knowing.
 tags:
  - llm
  - ciru
@@ -17,11 +17,13 @@ timestamp: 2026-09-19

 # CIRU v4.4.1 Orca on Dankclaw (frank-large)

-> **ACTIVE since 2026-09-19.** It replaced halogen-flash-server (see
-> [[halogen-flash-server]], the current rollback), which had replaced
-> ik_llama.cpp (see [[frank-large-ik-uncensored]]). Deployed as a **root
-> systemd unit** (`scripts/frank-large.service`, `User=dank`), port 7770, on the
-> same gfx1151 iGPU the halogen quadlet used. The RTX 3090 it does not touch
+> **RETIRED 2026-09-21** (replaced by
+> [[frank-large-unsloth-qwen38-flash-next]]; **rollback target**). It was
+> ACTIVE since 2026-09-19 and replaced halogen-flash-server (see
+> [[halogen-flash-server]]), which had replaced ik_llama.cpp (see
+> [[frank-large-ik-uncensored]]). It was deployed as a **root systemd
+> unit** (`scripts/frank-large.service`, `User=dank`), port 7770, on the
+> same gfx1151 iGPU the halogen quadlet used. The RTX 3090 it did not touch
 > belongs to [[frank-small-gemma4-12b]].

 CIRU is a llama.cpp fork tuned for the Qwen3.8-Flash-Next ("qwen4exp")
@@ -0,0 +1,182 @@
+---
+type: reference
+title: unsloth llama.cpp Qwen3.8-Flash-Next UD-IQ4_XS on Dankclaw (frank-large)
+description: The active frank-large runbook — the unsloth llama.cpp prebuilt (b11030-mix-5ff778e, ROCm gfx1151) serving unsloth/Qwen3.8-Flash-Next-GGUF:UD-IQ4_XS on the Dankclaw Strix Halo iGPU: the prebuilt runtime, the 200k/one-stream shape, the explicit -md MTP wiring, the payload, the client contract, and the failure modes worth knowing.
+tags:
+  - llm
+  - unsloth
+  - dankclaw
+  - frank-large
+  - strix-halo
+  - gfx1151
+  - qwen4exp
+  - rocm
+  - mtp
+timestamp: 2026-09-21
+---
+
+# unsloth llama.cpp Qwen3.8-Flash-Next UD-IQ4_XS on Dankclaw (frank-large)
+
+> **ACTIVE since 2026-09-21.** It replaced the CIRU v4.4.1 Orca runtime
+> (see [[frank-large-ciru-orca]], the current rollback target), which had
+> replaced halogen-flash-server (see [[halogen-flash-server]]). Deployed as a
+> **root systemd unit** (`scripts/frank-large.service`, `User=dank`), port
+> 7770, on the same gfx1151 iGPU. The RTX 3090 it does not touch belongs to
+> [[frank-small-gemma4-12b]].
+
+The model is `unsloth/Qwen3.8-Flash-Next-GGUF` at **UD-IQ4_XS** (unsloth
+dynamic quant): the 125B-param / 6B-active Qwen3.8-Flash-Next MoE with its
+51B n-gram embedding (PLE) folded into the single GGUF, 48 layers of
+12 × (3 × (Gated DeltaNet → MoE) → 1 × (QSA → MoE)), native 262,144-token
+window, vision encoder. The runtime is the **unsloth prebuilt**
+`app-b11030-mix-5ff778e-linux-x64-rocm-gfx1151` — the first runtime with an
+MTP graph for the qwen4exp architecture (stock ggml-org mainline has no MTP
+graph, no cross-model tensor borrowing, and no `--spec-type draft-mtp` for
+this model).
+
+## Why this swap
+
+- The operator asked for llama.cpp + the unsloth UD-IQ4_XS quant at
+  **one conversation, 200k context** (2026-09-21).
+- UD-IQ4_XS (87.3 GiB) is a standard llama.cpp quant — no private runtime,
+  no PLE sidecar payload, no QSA hotfix: one GGUF, one MTP head, done.
+- The unsloth prebuilt carries the qwen4exp MTP graph (1.3–1.7x at
+  concurrency 1, per the model card), which the CIRU fork also had but which
+  stock mainline does not.
+
+## Shape
+
+One root unit, one process, one published port:
+
+| half | port | reachable |
+|---|---|---|
+| llama-server (OpenAI front-end + engine, single process) | 7770 | published as host **7770** |
+
+## Runtime
+
+- **Prebuilt**: `app-b11030-mix-5ff778e-linux-x64-rocm-gfx1151.tar.gz` from
+  the [unslothai/llama.cpp releases](https://github.com/unslothai/llama.cpp/releases)
+  (tag `b11030-mix-5ff778e` or newer; the MTP graph landed in
+  `b10715-mix-86bd2d3`). 0.4 GB, bundles its own ROCm runtime
+  (`libamdhip64.so.7.16`), so there is **no host SDK to gate on and no venv
+  to symlink** — unlike the CIRU predecessor's runtime-root gate.
+- **Installed at** `/home/dank/GitPackages/unsloth-llama-b11030/`
+  (extracted tarball; `llama-server` at the top level).
+- Verified on-host (2026-09-21): `--version` → `0.4.1-dev (build 11030,
+  commit 6ba30d05b)`; `--list-devices` → `ROCm0: AMD Radeon 8060S Graphics
+  (116732 MiB)`; a Gemma4-12B smoke load served a clean completion on the
+  iGPU before the cutover.
+
+## Model
+
+- **Checkpoint**: `unsloth/Qwen3.8-Flash-Next-GGUF` → `UD-IQ4_XS` (the
+  `:UD-IQ4_XS` tag is a directory in the repo, not a submodel).
+- **Transport**: three llama.cpp split parts (HF caps uploads at 50 GB):
+  `00001` (10,946,624 B) + `00002` (49,835,229,856 B) + `00003`
+  (43,836,407,744 B) = **93,682,584,224 B** concatenated in numeric order.
+- **Assembled at** `/home/dank/models/Qwen3.8-Flash-Next-UD-IQ4_XS/`:
+  - `Qwen3.8-Flash-Next-UD-IQ4_XS.gguf` — main model (87.3 GiB)
+  - `mtp/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf` — MTP head (2.6 GiB)
+  - `vision/mmproj-F16.gguf` — projector (904 MB, not loaded while
+    `ENABLE_VISION=0`)
+  - `parts/` + `SHA256SUMS` — transport parts and the assembled digest
+- **Fetch**: `scripts/download-frank-large-unsloth.sh` — resumable
+  (`curl -fL -C -`), each part verified against its **LFS sha256 fetched
+  live from the Hub tree API** (not a file that could drift), then the
+  assembled GGUF against `SHA256SUMS`. `--verify-only` re-checks without
+  downloading.
+- The n-gram embedding (PLE) is **inside the GGUF** as ordinary tensors —
+  there is no external PLE payload and no pager, unlike the CIRU Orca
+  packaging (79.4 GiB main + 52.4 GiB paged PLE).
+
+## Deployed config
+
+From `scripts/frank-large.env` (the flag list is owned by
+`scripts/serve-frank-large.sh`):
+
+| knob | value | note |
+|---|---|---|
+| `MODEL_DIR` | `/home/dank/models/Qwen3.8-Flash-Next-UD-IQ4_XS` | serve script refuses to start without the GGUF and the MTP head |
+| `RUNTIME_DIR` | `/home/dank/GitPackages/unsloth-llama-b11030` | prebuilt tarball, extracted |
+| `HOST` / `PORT` | `0.0.0.0` / `7770` | same address the CIRU predecessor served |
+| `CONTEXT_SIZE` | `200000` | operator's requested shape (2026-09-21); trained window is 262144 |
+| `PARALLEL_SLOTS` | `1` | one conversation |
+| `ENABLE_MTP` | `1` | the shared-Q8_0 head speculates |
+| `SPEC_DRAFT_N_MAX` | `2` | the model card's default ("higher drafts more but each guess is accepted less often") |
+| `ENABLE_VISION` | `0` | projector on disk, one env flip away |
+
+Serve flags: `-m <GGUF> -md <MTP head> --spec-type draft-mtp
+--spec-draft-n-max 2 -c 200000 -np 1 -ngl 999 --jinja --alias frank-large
+--host 0.0.0.0 --port 7770 --chat-template-kwargs
+'{"preserve_thinking":true}'`.
+
+### The explicit `-md` (do not drop it)
+
+Sidecar auto-discovery does **not** search the `mtp/` subfolder. Without
+`-md`, `--spec-type draft-mtp` finds no head and the server runs at the main
+model's speed **with no error saying why**. The serve script gates on the
+head's presence and passes it explicitly.
+
+### Memory model
+
+The iGPU (Radeon 8060S, gfx1151) addresses **116,732 MiB (~114 GiB)** — the
+GTT aperture over the shared LPDDR (`gtt_total=122402963456` in
+`/sys/class/drm/card1/device/`, confirmed by `llama-server --list-devices`).
+The 87.3 GiB model is mmap'd into that aperture with
+`-ngl 999`, the same regime the CIRU Orca predecessor ran in (92.6 GiB
+occupied at rest). Budget at 200k: weights 87.3 + MTP head 2.6 + KV ~5
+(QSA 12 layers × 2 KV heads × 256 dim × f16 × 200k) + compute buffers ~5
+≈ **100 GiB**, leaving ~14 GiB of aperture headroom. The host's 121 GiB of
+LPDDR is shared with the CPU and the ~28 co-tenant containers; the 64 GiB
+swap file is the backstop.
+
+## Client contract
+- OpenAI-compatible chat completions at `:7770/v1/chat/completions`;
+  `/health` 200 when serving.
+- **Reasoning** is extracted to `reasoning_content` (llama.cpp's field —
+  this is a llama.cpp runtime, not vLLM's `reasoning`).
+- **Tools** work via `--jinja` (the model's own template). The Qwen3.8
+  template gates `reasoning_effort` to `{xhigh, medium, low}` — the harness
+  folds its levels onto that set in `models.yml` (`reasoningEffortMap`).
+  `preserve_thinking` is on by default (request-level
+  `chat_template_kwargs` still wins).
+- **Vision**: off (`ENABLE_VISION=0`); the F16 projector is on disk. The
+  fleet's image path is frank-small.
+- **Audio**: none (no audio in the fleet since the FLM NPU endpoint retired
+  2026-09-11).
+
+## Troubleshooting
+
+- **`missing runtime` / `missing target model` / `missing MTP head`** from
+  the serve script: the unit exits with a pointer instead of restart-looping.
+  Re-run the fetch script (`scripts/download-frank-large-unsloth.sh`) or
+  re-extract the prebuilt.
+- **Health 503 for minutes after (re)start**: the 87 GiB load is slow cold
+  (page-cache miss). Do not kill the process; `deploy.sh` waits up to 20
+  minutes (`LLM_WAIT`).
+- **Serving at the main model's speed with no spec metrics**: the MTP head
+  is not attached — check the process cmdline for `-md` (see above).
+- **`Address already in use` on 7770**: a stray manual run of the same model
+  holds the port. The deploy handler sweeps `pkill -f
+  'Qwen3.8-Flash-Next-UD-IQ4_XS'` (model path, not binary name — frank-small
+  is also a `llama-server`) and gates on `wait_port_free 7770`.
+- **Host memory pressure**: this endpoint is the largest LPDDR consumer on
+  the box. If the co-tenants start thrashing, the levers are KV quant
+  (`--cache-type-k/v q8_0` halves the ~5 GiB KV) and `SPEC_DRAFT_N_MAX`
+  (smaller draft state), not the weights.
+
+## Rollback
+
+The CIRU Orca payload and runtime are intact on disk:
+`/home/dank/models/Qwen3.8-Flash-CIRU-STRIX-Orca/` (main GGUF, MTP head, PLE
+payload) and `/home/dank/GitPackages/ciru-runtime-v4.4.1-qsa-hotfix/`.
+Rollback is one commit restoring the CIRU `serve-frank-large.sh` +
+`frank-large.env` + the deploy handler's `LLM_LARGE_*` paths (the
+predecessor's version is in git history), then `deploy.sh dankclaw
+--service frank-large`. See [[frank-large-ciru-orca]].
+
+## References
+
+- [unsloth/Qwen3.8-Flash-Next-GGUF](https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF) — model card, MTP README, imatrix
+- [unslothai/llama.cpp releases](https://github.com/unslothai/llama.cpp/releases) — the prebuilt runtime (tag `b11030-mix-5ff778e`)
+- [MTP PR #28243](https://github.com/ggml-org/llama.cpp/pull/28243) — the qwen4exp MTP graph that the prebuilt carries
@@ -1,11 +1,11 @@
 # LLM Fleet — Knowledge Base

-Documentation for the LLM inference fleet on the two hosts (2026-09-20):
+Documentation for the LLM inference fleet on the two hosts (2026-09-21):
 **vllm-radiance on Dankserver** serving frank-medium as ONE tensor-parallel
-process across both Radeon AI PRO R9700 cards (just1moremodel
-Qwen3.8-27B-Uncensored MXFP4-awq + the z-lab DFlash2-FP8 drafter, TP=2 over the
-R4D one-shot allreduce), and TWO endpoints on Dankclaw — the CIRU v4.4.1
-runtime (frank-large, jcbtc Qwen3.8-Flash-CIRU-STRIX-Orca w4b on the gfx1151
+process across both Radeon AI PRO R9700 cards (orcarouter
+Qwen3.8-27B-Uncensored-FP8 + the z-lab DFlash2 drafter, TP=2 over the
+R4D one-shot allreduce), and TWO endpoints on Dankclaw — the unsloth
+llama.cpp prebuilt (frank-large, Qwen3.8-Flash-Next UD-IQ4_XS on the gfx1151
 iGPU, port 7770) and mainline llama.cpp (frank-small, HauhauCS Gemma4-12B QAT
 Q4_K_M with MTP and vision on the RTX 3090, port 7771).

@@ -23,18 +23,18 @@ graph LR
 gfx1201, 32 GiB each]
        B2[vllm-radiance 0.9.3, TP=2
 frank-medium, port 7770]
-        C[Qwen3.8-27B MXFP4-awq + DFlash2-FP8
+        C[Qwen3.8-27B FP8 + DFlash2 drafter
 ONE process on both cards, 262144 ctx, up to 8 seqs]
        A --> B2 --> C
    end

    subgraph Dankclaw[AMD Strix Halo - Ubuntu 26.04]
        E2[RTX 3090 (PCIe x4 dock) +
-gfx1151 iGPU 64 GiB unified memory]
-        E[CIRU v4.4.1 runtime
+gfx1151 iGPU 114 GiB GTT]
+        E[unsloth llama.cpp b11030 (ROCm gfx1151)
 frank-large, port 7770]
-        F[jcbtc Qwen3.8-Flash-CIRU-STRIX-Orca w4b
-262144 ctx / 1 slot, MTP depth 4 + PLE]
+        F[Qwen3.8-Flash-Next UD-IQ4_XS
+200000 ctx / 1 slot, MTP depth 2]
        G[mainline llama.cpp 37b53fd
 frank-small, port 7771]
        H[Gemma4-12B QAT Q4_K_M
@@ -96,11 +96,18 @@ frank-small, port 7771]
  (bs1/bs3/bs4/bs6, 8k–700k KV) on Dankserver RTX 3090
 - [[deployment]] — deploy and update process for both servers
 - [[frank-redo-2026-08-17]] — the 2026-08-17 topology change
- [[frank-large-ciru-orca]] — ACTIVE frank-large runbook since 2026-09-19
-  (root systemd unit, port 7770): the CIRU v4.4.1 runtime (private host build,
-  QSA hotfix) serving jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca w4b on the gfx1151
-  iGPU — one 262,144-token conversation, MTP depth 4, 4096 MiB PLE cache,
-  runtime-root gate, client contract, troubleshooting
+- [[frank-large-unsloth-qwen38-flash-next]] — ACTIVE frank-large runbook
+  since 2026-09-21 (root systemd unit, port 7770): the unsloth llama.cpp
+  prebuilt (b11030-mix-5ff778e, ROCm gfx1151) serving
+  unsloth/Qwen3.8-Flash-Next-GGUF:UD-IQ4_XS on the gfx1151 iGPU — one
+  200,000-token conversation, MTP depth 2 via the explicit -md shared-Q8_0
+  head, the 114 GiB GTT memory model, client contract, troubleshooting
+- [[frank-large-ciru-orca]] — RETIRED 2026-09-21 (replaced by
+  [[frank-large-unsloth-qwen38-flash-next]]; rollback target): the CIRU
+  v4.4.1 runtime (private host build, QSA hotfix) serving
+  jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca w4b on the gfx1151 iGPU — one
+  262,144-token conversation, MTP depth 4, 4096 MiB PLE cache, runtime-root
+  gate, client contract, troubleshooting (2026-09-19 → 09-21)
 - [[halogen-flash-server]] — RETIRED 2026-09-19 (replaced by
  [[frank-large-ciru-orca]]; rollback target): peonist-ai halogen-flash-server
  0.5.8 serving peonist-ai/halogen-qwen3.8-flash-next (Qwen3.8-Flash-Next w4b
@@ -161,7 +168,7 @@ frank-small, port 7771]
  (deployed and reverted 2026-08-22)
 - [[Dankserver System Specs]] — the 2x R9700 box (frank-medium)
 - [[Dankclaw System Specs]] — the Strix Halo gfx1151 box (frank-large on the
-  CIRU v4.4.1 runtime since 2026-09-19, frank-small on the RTX 3090)
+  unsloth llama.cpp prebuilt since 2026-09-21, frank-small on the RTX 3090)
 - [[frank-npu]] — RETIRED 2026-09-11: the FLM NPU sight+sound endpoint
  on Dankclaw (Gemma4 E4B vision+audio pack, XDNA2, port 7771), kept as
  the retirement record — the IOMMU/SVA err=-19 root cause and where
@@ -147,11 +147,18 @@ The LLM inference fleet on the two hosts. Full index:
  agentionai/Qwen3.8-Flash-Next-ROCmFP4-FAST-imatrix (87.06 GiB,
  Q4_0_ROCMFP4_FAST) + own MTP draft, build-qwen38 (retired 2026-08-30;
  rollback record)
- [[frank-large-ciru-orca]] — the ACTIVE frank-large runbook since 2026-09-19
-  (root systemd unit, port 7770): the CIRU v4.4.1 runtime (private host build,
-  QSA hotfix) serving jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca w4b on the gfx1151
-  iGPU — one 262,144-token conversation, MTP depth 4, 4096 MiB PLE cache,
-  runtime-root gate, client contract
+- [[frank-large-unsloth-qwen38-flash-next]] — the ACTIVE frank-large runbook
+  since 2026-09-21 (root systemd unit, port 7770): the unsloth llama.cpp
+  prebuilt (b11030-mix-5ff778e, ROCm gfx1151) serving
+  unsloth/Qwen3.8-Flash-Next-GGUF:UD-IQ4_XS on the gfx1151 iGPU — one
+  200,000-token conversation, MTP depth 2 via the explicit -md shared-Q8_0
+  head, the 114 GiB GTT memory model, client contract
+- [[frank-large-ciru-orca]] — RETIRED 2026-09-21 (replaced by
+  [[frank-large-unsloth-qwen38-flash-next]]; rollback target, 2026-09-19 →
+  09-21): the CIRU v4.4.1 runtime (private host build, QSA hotfix) serving
+  jcbtc/Qwen3.8-Flash-CIRU-STRIX-Orca w4b on the gfx1151 iGPU — one
+  262,144-token conversation, MTP depth 4, 4096 MiB PLE cache, runtime-root
+  gate, client contract
 - [[halogen-flash-server]] — RETIRED 2026-09-19 (replaced by
  [[frank-large-ciru-orca]]; rollback target, 2026-09-16 → 09-19):
  peonist-ai/halogen-flash-server 0.5.8 serving peonist-ai/halogen-qwen3.8-
@@ -11,6 +11,42 @@ timestamp: 2026-09-14

 # Bundle Change Log

+## 2026-09-21 — frank-large: CIRU Orca → unsloth llama.cpp Qwen3.8-Flash-Next UD-IQ4_XS
+
+- **The change.** At the operator's request, frank-large (dankclaw :7770)
+  moves from the CIRU v4.4.1 Orca runtime (jcbtc
+  Qwen3.8-Flash-CIRU-STRIX-Orca, 262144 ctx) to the unsloth llama.cpp
+  prebuilt (b11030-mix-5ff778e, ROCm gfx1151, bundles its own ROCm 7.16)
+  serving `unsloth/Qwen3.8-Flash-Next-GGUF:UD-IQ4_XS` (87.3 GiB, three HF
+  split parts concatenated), one conversation at 200,000 ctx, MTP depth 2.
+  The fleet contract is unchanged: port 7770, alias `frank-large`,
+  text-only.
+- **Why the unsloth prebuilt.** The qwen4exp MTP graph exists only in the
+  unsloth fork (stock mainline has no MTP for this architecture); the
+  prebuilt bundles its own ROCm, so there is no host SDK gate or venv the
+  way the CIRU build had one.
+- **The explicit `-md`.** Sidecar auto-discovery does not search the `mtp/`
+  subfolder — without an explicit `-md <head>`, `--spec-type draft-mtp`
+  silently runs at the main model's speed. The serve script gates on the
+  head's presence and passes it.
+- **Memory model.** The iGPU's budget is the GTT aperture: 116,732 MiB
+  (~114 GiB), `gtt_total=122402963456` — not the old "64 GiB rocm-smi"
+  lesson (stale: the CIRU Orca held 92.6 GiB on the same device). 200k
+  budget ≈ 100 GiB, ~14 GiB of headroom.
+- **Verified before cutover.** Gemma4-12B smoke load on the iGPU via the
+  prebuilt: clean completion with reasoning extraction (the first probe
+  returned empty content because the thinking block consumed the 32-token
+  budget — a budget artifact, not a decode failure). CIRU restored for the
+  ~3-hour model fetch (health 200 in ~40 s).
+- **Docs.** New runbook `frank-large-unsloth-qwen38-flash-next.md`;
+  `frank-large-ciru-orca.md` marked RETIRED (rollback target, payload and
+  runtime intact on disk); AGENTS.md topology/section/runbook index and the
+  GPU-budget lesson updated; LLM index header + mermaid + active list
+  updated.
+- **Files.** `scripts/serve-frank-large.sh`, `scripts/frank-large.env`,
+  `scripts/frank-large.service`, `scripts/download-frank-large-unsloth.sh`,
+  `scripts/deploy-remote.sh`, `tests/frank_large_unit.bats`.
+
 ## 2026-09-21 — node-dankclaw Prometheus target fixed (netavark host-bind quirk)

 - **The gap.** A deep audit found the `node-dankclaw` scrape target red while