feat(llm): swap frank-large to halogen-flash-server, retire frank-npu

frank-large moves from the kingjones30/ROCmFPX llama.cpp build to
peonist-ai/halogen-flash-server 0.5.8 serving Qwen3.8-Flash-Next (w4b
.hgn 115.55 GiB + quality overlay). Upstream measures this engine at
roughly 4x the prefill throughput of the llama.cpp family on the same
gfx1151 silicon, and prefill dominates any prompt with real context.

Shape: a rootless Quadlet container (dankclaw/frank-large/quadlet/),
engine on container loopback :8730 (never published), OpenAI front-end
:8731 published as host 7770, HALOGEN_MODEL_ID=frank-large so no client
renamed anything. Context 262144, KV pool 262144 positions (~7.2 GiB,
conservative for a shared 121 GiB host), 4 slots, host reserve 24,
vision tower on, weights pinned, all other flags at vendor defaults.
Weights are fetched by HALOGEN_DOWNLOAD on first start into a writable
/home/dank/models/halogen.

The llama.cpp era (unit, env, serve script) is staged in .AGENTTRASH/
rather than deleted; the STRIX payload and the ROCmFPX build stay on
disk as the rollback target. deploy_llm_large becomes a user-quadlet
handler: it retires the old ROOT unit (moving its checkout symlink into
the home trash, never rm) and gates the first start on the download
rather than reporting a failure for a fetch that is working exactly as
designed.

frank-npu is retired in the same change. The kernel cmdline moves from
iommu=pt to amd_iommu=off for the 13-16% of prefill upstream measured,
and that takes the XDNA2 NPU with it: without PASID/SVA the driver
cannot open /dev/accel/accel0 (err=-19). Its five dependent routes are
repointed (Hermes vision -> frank-large, Hermes title generation ->
frank-small, harness vision -> frank-large, tiny -> frank-small, task ->
frank-large:medium) and its provider blocks are removed from both the
harness config and Hermes's live config. Audio input leaves the fleet
with no replacement; image input survives on halogen's vision tower.

Docs: new halogen-flash-server runbook, new IOMMU decision record, and
frank-npu rewritten as a retirement record. AGENTS.md topology, runtime
sections, benchmarking policy (frank-large is a user unit now) and the
harness model-config section are corrected.

.AGENTTRASH/frank-large.env.llama.cpp

@@ -0,0 +1,126 @@
+# Tunables for frank-large (kingjones777/Qwen3.8-Flash-Next-ROCmFP4-
+# STRIX-GGUF on the kingjones30/ROCmFPX llama.cpp runtime, Dankclaw,
+# Strix Halo gfx1151, unified memory).
+#
+# Runtime: kingjones30/ROCmFPX (fork of charlie12345/ROCmFPX with the
+# qwen4exp-on-rocmfpx-d3ca537 patch applied — the model card's pinned
+# runtime: ROCmFP4 tensor types AND the qwen4exp architecture in one
+# tree; stock llama.cpp and charlie's fork alone cannot load this GGUF).
+# Built per the card recipe:
+#   cmake -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1151 -DGGML_NATIVE=ON -DCMAKE_BUILD_TYPE=Release
+# at /home/dank/GitPackages/kingjones-ROCmFPX/build.
+# Replaced the agentionai ROCmFP4-FAST imatrix deployment (charlie's
+# build-qwen38 runtime) on 2026-08-30.
+#
+# Model: Qwen3.8-Flash-Next base (qwen4exp MoE: 512 experts / 10 used,
+# 48 layers = 12 attention + 36 SSM, QSA indexer, hyper-connections, PLE
+# at layer 1), Q4_0_ROCMFP4_STRIX (file_type 105), 113.47 GiB single
+# file. STRIX is the top quality tier: PLE table at Q8_0 (50.66 GiB),
+# token embeddings + lm head at Q6_K. The PLE table NEVER enters GPU
+# memory — it streams from the file through the OS page cache. Two hard
+# card constraints:
+#  1. mmap MUST stay enabled (default) — --no-mmap forces the 50.66 GiB
+#     table into anonymous memory and the cgroup killer takes the
+#     process with nothing in the server log (only dmesg).
+#  2. Do NOT pin the PLE tensor to CPU with -ot — measured gen fell
+#     ~23 -> 13.4 tok/s; the kernel streams it better.
+# NOTE: the card's own build_attn_qsa still calls ggml_fill on the
+# F16-cast kq mask (the path that aborted our charlie build with
+# --flash-attn on before scripts/patches/qwen4exp-qsa-f32-mask.patch).
+# If a clean kingjones-tree build aborts graph build with
+# GGML_ASSERT(a-type == GGML_TYPE_F32), port that fix (F32 base mask +
+# cast at the attention call) into this tree and rebuild.
+#
+# WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE. The ROCm HIP runtime reads those bare env vars (see the 2026-08-13 TEMP
+# incident). llama.cpp reads --host/--port/--temp as CLI flags, never as
+# env vars.
+
+# kingjones30/ROCmFPX runtime binary (card recipe build).
+LLAMA_BIN=/home/dank/GitPackages/kingjones-ROCmFPX/build/bin/llama-server
+
+# Model package from
+# https://huggingface.co/kingjones777/Qwen3.8-Flash-Next-ROCmFP4-STRIX-GGUF
+# Layout: single unsharded GGUF, text-only (the repo ships no mmproj;
+# the vision tower below is borrowed from the agentionai FAST package —
+# same base model, f16 — kept only if it loads clean, drop
+# MMPROJ_PATH if it does not).
+MODEL_DIR=/home/dank/models/kingjones-Qwen3.8-Flash-Next-ROCmFP4-STRIX
+MODEL_FILE=Qwen3.8-Flash-Next-Q4_0-ROCmFP4-STRIX.gguf
+MMPROJ_PATH=/home/dank/models/agentionai-Qwen3.8-Flash-Next-ROCmFP4-FAST/mmproj/mmproj-Qwen3.8-Flash-Next-f16.gguf
+
+# Qwen-Sharp chat template (tracked in the repo, staged on the host). The
+# Sharp template drives thinking on by default (medium effort); the server
+# emits reasoning in DeepSeek format so clients receive reasoning_content
+# separately instead of literal thinking tags in content.
+CHAT_TEMPLATE_FILE=/home/dank/serverAppConfigs/scripts/chat-template.jinja
+CHAT_TEMPLATE_KWARGS='{"enable_thinking":true,"preserve_thinking":true,"preserve_reasoning":true}'
+
+# Alias + network.
+ALIAS=frank-large
+BIND_ADDR=0.0.0.0
+LISTEN_PORT=7770
+
+# Serving: 401408-token context, TWO slots (parallel 2), batch 4096 /
+# ubatch 512, 16+16 threads, f16 KV, 16 context checkpoints, 8 GiB
+# prompt cache. --fit off: auto-fit segfaults on this model's
+# hyper-connection/indexer config; -ngl 999 puts every layer on the
+# SVM-backed GPU pool. ubatch 512 matches the card's re-measurement
+# (-ub 1024/2048/4096 showed no effect; -t 32 vs 16 is noise).
+#
+# CTX_SIZE=401408 / PARALLEL=2: carried over from the FAST deployment
+# where it loaded clean (~74 s, health 200). The card measures the full
+# 262144 window at ~71.7 GiB GTT at load on the same 128 GB board; QSA
+# caps KV growth so the extra window is cheap. KV footprint scales
+# with the 12 attention layers only. Verify load on first STRIX boot.
+CTX_SIZE=401408
+PARALLEL=2
+
+BATCH_SIZE=4096
+UBATCH_SIZE=512
+NUM_THREADS=16
+THREAD_BATCH=16
+CACHE_TYPE_K=f16
+CACHE_TYPE_V=f16
+CTX_CHECKPOINTS=16
+CACHE_RAM=8192
+SLOT_SAVE_PATH=/home/dank/serverAppConfigs/scripts/slot-state
+SPLIT_MODE=none
+FIT=off
+NGL=999
+FLASH_ATTN=on
+
+# Sampling (card instruct table, non-thinking).
+TEMPERATURE=0.7
+TOP_K=20
+TOP_P=0.8
+MIN_P=0.0
+PRESENCE_PENALTY=1.5
+
+# Reasoning on (Sharp template drives it; preserve_thinking + preserve_reasoning
+# keep the reasoning trail in the conversation). DeepSeek format projects the
+# reasoning into the API's reasoning_content field instead of literal thinking
+# tags.
+REASONING=on
+REASONING_FORMAT=deepseek
+
+# Strix Halo unified memory + native gfx1151.
+HSA_OVERRIDE_GFX_VERSION=11.5.1
+
+VERBOSITY=4
+
+
+# Speculative decoding: disabled. The card's Usage block ships no spec
+# flags. The Flash-Next MTP heads do not load on this loader (requires
+# per-layer blk.N.hc_attn_norm; every head carries only blk.48 —
+# verified 3x on build-qwen38). DRAFT_MODEL_* point at the retired
+# agentionai FAST head and are unused while SPEC_TYPE=none.
+DRAFT_MODEL_DIR=/home/dank/models/mtp-Qwen3.8-Flash-Next-FAST
+DRAFT_MODEL_FILE=Qwen3.8-Flash-Next-MTP-ROCmFP4-FAST.gguf
+SPEC_TYPE=none
+SPEC_DRAFT_N_MAX=3
+SPEC_DRAFT_N_MIN=0
+SPEC_DRAFT_P_MIN=0.75
+SPEC_DRAFT_NGL=999
+SPEC_DRAFT_DEVICE=ROCm0
+SPEC_DRAFT_TYPE_K=f16
+SPEC_DRAFT_TYPE_V=f16