frank-primary: swap Dankserver to llama.cpp Qwen3.8-Flash-Next Q8_0

The 2026-09-26 same-day cutover that ended the Lucebox DS4.1 detour:
Dankserver's :7770 endpoint is now unsloth llama.cpp (the b11030-mix
prebuilt, ROCm gfx120X, bundled runtime) serving
unsloth/Qwen3.8-Flash-Next-GGUF at Q8_0 — 175.30 GiB across six sha256-
verified split parts — on both Radeon AI PRO R9700s, as a ROOT unit (the
frank-large shape: no container, no generated unit).

The deployed contract:

- Placement law: layer split -ts 1,1 with -ngl 999; the routed experts
  and the 51B-param n-gram (PLE) table are COMPUTED in host RAM via
  -ot exps=CPU,ngram_embeddings=CPU (the x4-pair measurement: streaming
  over PCIe loses at 3.7 t/s, computing in RAM does not care).
  per_layer_token_embd.weight=CPU is required by the shared MTP head's
  borrow path (a GPU-resident borrow crashes the draft load).
  mmap stays on; --cache-ram 16384.
- Spec chain draft-mtp,ngram-mod: the shared Q8_0 MTP head
  (mtp-...-shared-Q8_0.gguf, 2,786,568,256 B) at depth 2, with the
  draftless ngram-mod matcher chained behind it — the only legal way to
  layer two spec implementations in one process.
- Vision explicit: --mmproj names the BF16 projector (this build never
  autoloads; a server without it comes up healthy and text-only).
- Reasoning: --reasoning-format deepseek → reasoning_content; the
  Qwen3.8 template's {xhigh, medium, low} reasoning_effort gate folded
  by reasoningEffortMap in models.yml.
- One stream, the model's full 262,144 context; greedy server default
  with Qwen-recommended sampling riding compat.extraBody.

The deployer's deploy_llm_primary is rewritten for the root unit: it
gates every artifact at its declared byte count (stat -c %s, never wc -c
— which read the whole 377 GiB lucebox file for minutes on 2026-09-26),
retires the lucebox user quadlet to dated home trash, waits out port
7770, relinks/enables/restarts the root unit and waits up to 1800 s for
/health. The quadlet-era tree under dankserver/frank-primary/ is
archived (dankserver/.AGENTTRASH/20260926-llamacpp-cutover/ on this
box; git history everywhere else) and its planner mapping replaced by
the scripts/ trio.

Observability flips back to the llamacpp namespace: a frank-primary
Prometheus job at 192.168.0.235:7770 (cross-host LAN — the netavark
published-port quirk that forces host.containers.internal for the two
Dankclaw endpoints is same-host only), the LlmEndpointDown runbook
text, and the dashboard description for three llama.cpp endpoints.

Docs: the new runbook (frank-primary-llamacpp-qwen38-flash-next-q8.md),
RETIRED banners on the lucebox and radiance runbooks and their index
entries, the AGENTS.md fleet section rewritten for the root-unit shape,
and the homepage frank-medium card renamed to frank-primary with the
Q8_0 Flash-Next spec (docs.html #models and the modality note updated:
images on frank-primary and frank-small, audio on frank-small only).

models.yml: the dankserver provider is "llamacpp (dankserver)" serving
frank-primary, contextWindow 262144, maxTokens 65536 (compaction-gap
law: 183,501 + 65,536 <= 262,144), input [text, image].

Suite: 346 green (frank_primary_unit.bats rewritten for the root unit —
16 tests pinning the placement law, the shared-head requirement, the
explicit projector, HIP env hygiene (BIND_HOST, never bare HOST), and
the deployer's byte-count gates; the quadlet-era planner and
changed-tests mappings rewritten for the trio; the observability scrape
test asserts three llamacpp jobs again).

Rollback: the lucebox image (localhost/lucebox-rocm:gfx1201) and the
378 GiB payload stay on the host; the Qwen payload stays fetched, so a
re-swap in either direction is a redeploy, not a refetch.

dankserver/frank-primary/frank-primary-entrypoint.sh

@@ -1,90 +0,0 @@
-#!/usr/bin/env bash
-# frank-primary entrypoint: exec the native luce_server with the DeepSeek V4
-# launch shape for this host (2x R9700 gfx1201, 246 GiB RAM).
-#
-# The image's own entrypoint (server/scripts/entrypoint.sh) only forwards a
-# fixed Qwen/DFlash flag set and cannot express the DS4 flags, so this unit
-# bypasses it (PodmanArgs=--entrypoint) and execs the binary directly. The
-# C++ server reads LUCE_* environment variables straight from the process
-# environment, so no env plumbing is needed here.
-#
-# SOURCE BRANCH. The `deepseek41` architecture exists only on the
-# `feat/ds41-engine` branch of Luce-Org/lucebox: main carries `deepseek4`
-# alone and rejects this checkpoint with "unsupported model architecture
-# 'deepseek41'" (server/src/common/backend_plan.cpp). The image is built from
-# that branch — see serverAppConfigs-docs/LLM/frank-primary-lucebox-ds41.md.
-#
-# LAUNCH SHAPE (server/docs/DS41.md — the V4.1 page, not the V4 DS4.md one):
-#   --ds4-prefill exact   THE ONLY prefill this placement can run. Any
-#                         approximate mode (dense, sparse) makes
-#                         DeepSeek4Backend::requires_monolithic_model() true,
-#                         and a monolithic load asks ggml for the whole 187 GiB
-#                         expert set as ONE device-0 buffer — measured
-#                         2026-09-26: "allocating 191932.25 MiB on device 0:
-#                         cudaMalloc failed: out of memory", then
-#                         "monolithic HIP load required for dense prefill".
-#                         DS41.md's dense-prefill section measures that mode
-#                         where the experts fit (in-process owners); on a
-#                         32 GiB target it cannot. exact feeds the prompt token
-#                         by token, which is what the hybrid tier streams for.
-#   --chunk 1024          DS41.md's cold-expert chunk. The runtime caps the
-#                         chunk to what fits after headroom regardless.
-#   --max-ctx 131072      the V4.1 default window. The MLA compressed caches
-#                         are F16 by design and stay small; --cache-type-k/v
-#                         are ignored by the DeepSeek4 backend, so they are
-#                         not passed.
-#   --prefix-cache-slots 0 --prefill-cache-slots 0
-#                         the DS41.md bring-up recipe (the server default is
-#                         32 prefix slots). Lift them once capacity under load
-#                         has been measured, not before.
-#
-# EXPERT PLACEMENT. 377 GiB of weights against 64 GiB of VRAM: only the dense
-# weights, the compressed caches and a hot expert stack fit on hip:0, and the
-# remaining routed experts are streamed out of the mmapped page cache/NVMe by
-# the hybrid expert tier. LUCE_EXPERT_STREAM_DEVICE=1 moves the streamed-
-# expert slot cache onto the second R9700 — it defaults to the target device,
-# where it would compete with the hot stack. Both the owner stack and the
-# slot cache then auto-size to their device's free memory minus the 1.5 GiB
-# discrete headroom.
-#
-# No DSpark draft is published for DeepSeek-V4.1-Flash, so this is plain
-# autoregressive decode (no --draft, no LUCE_DS4_SPEC).
-set -euo pipefail
-
-MODEL="${FRANK_PRIMARY_MODEL:-/models/DeepSeek-V4.1-Flash/DeepSeek-V4.1-Flash-ROCMFP23.gguf}"
-BIN="${FRANK_PRIMARY_BIN:-/opt/lucebox-hub/server/build/luce_server}"
-
-# The second R9700 holds the streamed-expert slot cache and computes them.
-export LUCE_EXPERT_STREAM_DEVICE="${LUCE_EXPERT_STREAM_DEVICE:-1}"
-# The hot expert stack must be sized EXPLICITLY. Left alone the runtime
-# budgets gpu_free - kv - warm - safety (22.35 GiB here), which leaves only
-# its 0.50 GiB safety margin free while the fit check then demands the 1.50
-# GiB discrete headroom — its own auto-sizing can never pass on a 32 GiB
-# target: "device 0 memory: 30.97 of 31.86 GiB used, 0.89 GiB free (headroom
-# 1.50 GiB): does not fit". 18000 MB is the value DS41.md measured its R9700
-# rows at; it puts ~1500 experts (37/layer) resident and leaves device 0
-# ~5 GiB free, so the fit check passes with room for transient buffers.
-export LUCE_EXPERT_BUDGET_MB="${LUCE_EXPERT_BUDGET_MB:-18000}"
-# rocBLAS/BLAS dispatch and queue depth, as the fleet's other ROCm servers.
-export ROCBLAS_USE_HIPBLASLT="${ROCBLAS_USE_HIPBLASLT:-0}"
-export GPU_MAX_HW_QUEUES="${GPU_MAX_HW_QUEUES:-8}"
-
-if [[ ! -f "$MODEL" ]]; then
-    echo "frank-primary: model missing at $MODEL" >&2
-    exit 1
-fi
-if [[ ! -x "$BIN" ]]; then
-    echo "frank-primary: binary missing at $BIN (image build failed?)" >&2
-    exit 1
-fi
-
-exec "$BIN" "$MODEL" \
-    --host 0.0.0.0 \
-    --port 7770 \
-    --model-name frank-primary \
-    --target-device hip:0 \
-    --max-ctx 131072 \
-    --chunk 1024 \
-    --ds4-prefill exact \
-    --prefix-cache-slots 0 \
-    --prefill-cache-slots 0

dankserver/frank-primary/frank-primary-healthcheck.py

@@ -1,14 +0,0 @@
-#!/usr/bin/env python3
-"""frank-primary health probe: 200 on /health only once the engine serves.
-
-File-based (not an inline HealthCmd) because the podman 5.x v1 quadlet
-generator splits HealthCmd on spaces into an argv array.
-"""
-import sys
-import urllib.request
-
-try:
-    with urllib.request.urlopen("http://127.0.0.1:7770/health", timeout=10) as r:
-        sys.exit(0 if r.status == 200 else 1)
-except Exception:
-    sys.exit(1)

dankserver/frank-primary/lucebox-gfx1201.patch

@@ -1,40 +0,0 @@
-diff --git a/Dockerfile.rocm b/Dockerfile.rocm
-index 4ea7a1d7..0e22355b 100644
---- a/Dockerfile.rocm
-+++ b/Dockerfile.rocm
-@@ -59,18 +59,17 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
-         curl \
-         git \
-         git-lfs \
--        hipblas-dev \
--        hipblaslt-dev \
--        hipcub-dev \
-         libcurl4-openssl-dev \
-         ninja-build \
-         pkg-config \
-         python3 \
--        rocblas-dev \
--        rocprim-dev \
--        rocwmma-dev \
-     && rm -rf /var/lib/apt/lists/*
- 
-+# Ubuntu jammy ships cmake 3.22, which predates the block command that
-+# ROCm 7.14 hipblaslt-config.cmake requires (CMake >= 3.25). Install the
-+# official 3.31 binary; /usr/local/bin precedes /usr/bin on PATH.
-+RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.31.6/cmake-3.31.6-linux-x86_64.tar.gz     | tar -xz -C /usr/local --strip-components=1
-+
- WORKDIR /src
- 
- # COPY ordering mirrors the CUDA Dockerfile: C++ build inputs first so the
-@@ -152,11 +151,8 @@ ENV PATH=/opt/rocm/bin:${PATH}
- RUN apt-get update && apt-get install -y --no-install-recommends \
-         ca-certificates \
-         curl \
--        hipblas \
--        hipblaslt \
-         libgomp1 \
-         pciutils \
--        rocblas \
-     && rm -rf /var/lib/apt/lists/*
- 
- # uv (pinned) manages Python 3.12 + resolves the luce + pflash

dankserver/frank-primary/quadlet/frank-primary.container

@@ -1,90 +0,0 @@
-# Quadlet unit: frank-primary (rootless, boot-persistent).
-#
-# Install path on dankserver: ~/.config/containers/systemd/frank-primary.container
-# Unit name: frank-primary.service (container name: frank-primary)
-#
-# The dankserver LLM endpoint (:7770, alias frank-primary): the Lucebox
-# runtime (github.com/Luce-Org/lucebox, local image localhost/lucebox-rocm:
-# gfx1201 built from Dockerfile.rocm with LUCE_HIP_ARCHES=gfx1201 and
-# ROCM_VERSION=7.14.1-full to match the host driver's ROCm 7.x line)
-# serving DeepSeek-V4.1-Flash-ROCMFP23 (377.2 GiB GGUF, 40-layer MoE,
-# 384 routed experts + 1 shared per layer, top-6 routing, Engram tables at
-# layers 1 and 14) on the two Radeon AI PRO R9700s (gfx1201). The model
-# cannot be resident: dense weights, the compressed MLA caches and a hot
-# expert stack sit on hip:0, the streamed-expert slot cache on hip:1, and
-# every other routed expert is pulled out of the mmapped page cache (246 GiB
-# RAM holds most of the 179 GiB expert set; the rest reads from NVMe). Lazy
-# mmap page-in is the engine default — no flag.
-#
-# The image's entrypoint only forwards a fixed Qwen/DFlash flag set and
-# cannot express the DS4 flags, so the unit bypasses it (--entrypoint) and
-# runs the repo's frank-primary-entrypoint.sh, which execs the native
-# luce_server with the DS4 launch shape (see that script's header).
-#
-# SHAPE. Host network, :7770, served model name frank-primary, OpenAI /v1
-# (Chat Completions with tool metadata, Responses, Anthropic Messages,
-# Open WebUI metadata) — the fleet contract is unchanged from
-# frank-medium.
-#
-# DEVICES. /dev/kfd + /dev/dri through the amdgpu/KFD stack. Both R9700s are
-# visible (indices 0 and 1; the Raphael iGPU enumerates as index 2 and stays
-# out). They are NOT driven as a symmetric pair: the in-process expert-
-# parallel path (LUCE_DS4_MOE_TP_INPROC) is qualified for gfx1151+gfx1201
-# pairs, not gfx1201+gfx1201. Instead hip:0 owns the dense weights and the
-# hot expert stack while LUCE_EXPERT_STREAM_DEVICE=1 (entrypoint) gives hip:1
-# the streamed-expert slot cache, so each card carries one tier and neither
-# waits on the other.
-#
-# MOUNTS. /models (ro — the GGUF is mmapped, nothing writes), and the
-# entrypoint + healthcheck scripts straight out of this checkout
-# (deploy.sh git-pulls the checkout before dispatching).
-#
-# IMAGE. localhost/-aliased tag: this host's podman quadlet path cannot
-# resolve registry-prefixed refs from local storage at create time (same
-# as frank-medium). Pull=never keeps boots off the registry.
-#
-# CUTOVER. This unit REPLACES frank-medium on :7770. The deploy handler
-# (deploy_llm_primary in scripts/deploy-remote.sh) stops and masks the
-# radiance unit before starting this one; the radiance unit is retired to
-# the home trash after frank-primary is verified (trash policy, never rm).
-#
-# 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-primary — Lucebox DeepSeek-V4.1-Flash-ROCMFP23 on R9700 (gfx1201), experts in RAM
-After=network-online.target
-
-[Container]
-ContainerName=frank-primary
-Image=localhost/lucebox-rocm:gfx1201
-# Host network: luce_server binds 0.0.0.0:7770 directly; no PublishPort.
-AddDevice=/dev/kfd
-AddDevice=/dev/dri
-GroupAdd=keep-groups
-PodmanArgs=--privileged --ipc=host --security-opt seccomp=unconfined --cap-add SYS_PTRACE --network=host --entrypoint /usr/local/bin/frank-primary-entrypoint.sh
-# Both R9700s: hip:0 target, hip:1 the streamed-expert slot cache. Index 2,
-# the Raphael iGPU, stays out.
-Environment=HIP_VISIBLE_DEVICES=0,1
-Environment=ROCR_VISIBLE_DEVICES=0,1
-# The 377 GiB GGUF, mmapped; pages pull from disk/page cache on demand.
-Volume=/home/dank/models:/models:ro
-# Entrypoint + healthcheck from this checkout, mounted read-only.
-Volume=/home/dank/serverAppConfigs/dankserver/frank-primary/frank-primary-entrypoint.sh:/usr/local/bin/frank-primary-entrypoint.sh:ro
-Volume=/home/dank/serverAppConfigs/dankserver/frank-primary/frank-primary-healthcheck.py:/usr/local/bin/frank-primary-healthcheck.py:ro
-# 200 only once the engine serves; first start page-ins the dense weights
-# and builds the graphs, which takes a while. HealthCmd is split on spaces
-# into an argv array by the podman 5.x v1 generator, hence the file probe.
-HealthCmd=python3 /usr/local/bin/frank-primary-healthcheck.py
-HealthInterval=60s
-HealthTimeout=30s
-HealthRetries=3
-HealthStartPeriod=1800s
-Pull=never
-
-[Service]
-TimeoutStartSec=1800
-Restart=on-failure
-RestartSec=10
-
-[Install]
-WantedBy=default.target