frank-large: cutover to vLLM TheRock gfx1151, MiMo-V2.6 BF16 + cross-model MTP

Engine: digest-pinned docker.io/kyuz0/vllm-therock-gfx1151 (vLLM 0.30,
TheRock ROCm 10, Triton GDN kernels); the serve script gates the image by
REPO DIGEST, the payload (config + index + exact shard count), the spec
draft (BOTH weight layouts: single model.safetensors for dflash, sharded
index for the Qwen3.5-9B MTP head), and whole /dev/dri, sweeps the bench
identity vllm-mimo, then exec podman wait so a container death trips
Restart=always. Shape: 4 x 262144 at util 0.45 (KV grows only on the 8
full-attention layers), MTP 2 draft tokens.

Measured 2026-10-01 window (frank-large stopped): base 13.1 t/s decode
(~94% of the LPDDR5x BF16 roof), DFlash-8 17.4, MTP-2 19.8 with 71%/44%
per-position acceptance - the cross-model head works because vLLM reads the
draft from a separate path while MiMo ships zero mtp.* tensors. Spec arms
do not preserve greedy bf16 text (batched-verify reduction order).

The seam was rewritten around the new laws (18 tests): unit self-healing,
digest gating, engine wiring, the measured spec table, the retired gufo
gates, the deployer MainPID check. frank_small port test follows the
PORT -> FRANK_LARGE_PUBLISH rename. Rollback: tag frank-large-gufo-q4kxl-20261001.

scripts/frank-large.service

@@ -1,20 +1,46 @@
+# frank-large.service — vLLM MiMo-V2.6-Distill-Qwen-9B (BF16) on the Strix
+# Halo gfx1151 iGPU, :7770. 2026-10-01 cutover from gufo (tag
+# frank-large-gufo-q4kxl-20261001 restores the Q4_K_XL + DFlash2 shape).
+#
+# WHY A ROOT UNIT: parity with frank-primary/frank-small — the unit is
+# relink_unit'd from this repo into /etc/systemd/system/, PID1 grants
+# LimitMEMLOCK=infinity (vLLM pins host buffers), and OOM protection lands
+# on the engine instead of the co-tenants. The container runs in dank's
+# ROOTLESS podman storage.
+#
+# The serve script drives `podman run` directly (image entrypoint is
+# overridden to `vllm serve`) and blocks on the container, so a container
+# death is a non-zero ExecStart. NOTE from frank-primary's experience:
+# vLLM's API server can drain and exit 0 after an engine fault, so the
+# restart policy MUST be always, not on-failure.
 [Unit]
-Description=gufo (gufo-org/gufo, ROCm gfx1151) Qwen3.8-27B UD-Q8_K_XL, DFlash2 spec decode, 262144 ctx / 1 session, root systemd unit (dankclaw, frank-large, Strix Halo iGPU)
+Description=frank-large — vLLM MiMo-V2.6-Distill-Qwen-9B (BF16) on the gfx1151 iGPU
 After=network-online.target
 Wants=network-online.target
+# Never let the start-rate limiter give up after a crash loop.
+StartLimitIntervalSec=0
 
 [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
-Restart=on-failure
+# systemd stop: podman wait alone cannot stop the container child; stop it
+# explicitly or the engine outlives the unit holding the GPU and port.
+ExecStop=/usr/bin/podman stop -t 120 frank-large-mimo
+# Boot cost: ~200-240 s warm (measured 2026-10-01 bench window); cold torch
+# compile is longer. Health stays 503 until the engine is up — do not kill.
+TimeoutStartSec=1800
+# ExecStop allows 120 s of drain; the stop timeout must exceed it (the
+# frank-primary 2026-10-01 lesson: a short TimeoutStopSec records 'Failed
+# with result timeout' on every stop).
+TimeoutStopSec=180
+Restart=always
 RestartSec=10
+# The engine holds 17.5 GiB of weights plus the KV pool in the GTT
+# aperture; it is the last process on this host the kernel should pick.
+OOMScoreAdjust=-500
 
 [Install]
 WantedBy=multi-user.target