Rootless podman cannot map the 'render' gid (990) into the container
namespace (dank's subgid range is 100000+), so --group-add render fails to
start the container. GPU access actually comes from the /dev/kfd + /dev/dri
passthrough, checked host-side (the run user is already in 'render' on the
host). So add only the 'video' group (which resolves) and drop 'render'.
Update the group test to assert video present and render ABSENT (with the
rootless rationale), per the approved rootless config.
Replace the CPU stock image with a container built from
Containerfile.comfyui on AMD's gfx1151-tagged vLLM base — the only torch
build verified to report torch.cuda.is_available()=True on the Strix Halo
APU (Radeon 8060S, capability (11,5)).
- Containerfile bakes the ComfyUI code in and installs its deps, but skips
the torch family (the base image already ships the gfx1151 ROCm
torch/torchvision/torchaudio; an unpinned 'torch' in requirements.txt would
clobber the gfx1151 kernels).
- compose comfyui service: build + image comfyui-rocm:gfx1151, passes
/dev/kfd + /dev/dri, adds video+render groups, drops --cpu. Models/data
persist under /opt/comfyui.
- Un-skip the four compose-wiring acceptance tests (the gate for this work).
The JSON rate cap (API_MAX=4/hour) is hardcoded in Python, not settable in
limiter.toml; pass_ip is the only bypass. Rootless podman strips the client IP
on published-port hits (the container sees a slirp source), so local subagents
must present a pass_ip source. Document the slirp IP-strip, the pass_ip
expansion, and the root-owned config file (one-off sudo cp to update).
Root cause of the 429s: the ip_limit rate limiter caps JSON (format!=html)
requests at API_MAX=4 per IP per hour, and that constant is hardcoded in
Python (searx/botdetection/ip_limit.py), not settable in limiter.toml. The
only bypass is pass_ip (filter_request returns before ip_limit runs).
Rootless podman strips the client IP on published-port hits, so the container
sees a slirp/bridge source (10.88.x/10.89.x), never the real client - hence
even a LAN-IP curl 429'd on /search. Expand pass_ip to loopback + the host's
private podman/slirp ranges (plus the existing LAN) so local subagents present
a trusted source and skip the rate limit entirely. All these ranges are
reachable only from the host or LAN, never the internet.
User request 2026-08-19 (second pass): parallel slots degraded
performance. PARALLEL=1, CTX_SIZE=262144 (native max, -c equals the
per-slot context with one slot). Concurrent requests queue on the
single slot; the server never preempts a running generation.
pi models.json contextWindow/maxTokens back to 262144 (outside repo).
Docs updated to match.
User request 2026-08-19: lower per-stream context from 262144 to 200000.
CTX_SIZE is the total; llama.cpp splits it across the 2 parallel slots.
f16 KV drops from ~32 GiB to ~24 GiB, adding OOM headroom.
Docs: sync the 2026-08-18 service swap (frank-default-large on dankclaw,
frank-default-small on dankserver) across AGENTS.md, README.md,
deployment.md, ciru-activefpx-promptforge.md, and index.md. The pi
models.json contextWindow/maxTokens for frank-default-large move
262144 -> 200000 (separate file, outside the repo).
Replace the dead dankserver open-webui (bare container, :3000, pointed at a
dead :9090 backend) with a boot-persistent quadlet on dankclaw:
- CPU image (:main); dankclaw is CPU-only, so the :cuda image is wasted.
- Isolated on slirp4netns; publishes 3005 (3000 is Gitea's HTTP port).
- Default LLM backend is the local vLLM on 52625 (frank-default-large).
- Data persists under /home/dank/Containers/open-webui/data.
- Caddy route now proxies to the published LAN port (the container is not on
Caddy's network).
Also corrects the model/port mapping in the docs: the live state (verified
2026-08-19) is frank-default-large on vLLM :52625 (dankclaw) and
frank-default-small on llama.cpp :7770 (dankserver) — the reverse of the
2026-08-17 note.
Quadlet units are generator-managed; systemctl enable fails on them
('transient or generated'). The [Install] WantedBy=default.target in each
.container file is itself the enablement (is-enabled reports 'generated'), and
with user linger on they start at boot. Drop the enable step; daemon-reload and
start are enough.
Device= in [Container] is a [Volume]-group key; the container device key is
AddDevice= (maps to podman --device). Update the /dev/net/tun device line and
the test seam.
The podman quadlet generator fails with "unsupported key 'CapAdd' in group
'Container'" and skips the whole unit. The correct key is AddCapability=
(maps to podman --cap-add). Update the gluetun unit and its test seam.
The quadlet .network unit is generator-managed and cannot be enabled; it is
pulled in automatically by the containers that set Wants=/After=searxng-network.
Enable only the three container units. The explicit start of the network is kept
(start works on generated units).
Update SearXNG, Gluetun, and Valkey to describe the quadlet units instead of
compose: searxng sets Network=gluetun.container, the units install to
~/.config/containers/systemd/, and scripts/deploy-searxng-quadlet.sh deploys
them. Add a Boot Persistence section to the Stack doc. Correct the Gluetun
tunnel check (the :8000 server 404s; verify tun0 instead) and note the pinned
10.89.2.0/24 subnet.
Remove gluetun/searxng/valkey from dankclaw/compose.yaml (they are now
quadlet units) so podman-compose up can no longer fight them for the netns and
port 8081. Add scripts/deploy-searxng-quadlet.sh: an idempotent installer that
copies the units to ~/.config/containers/systemd/, daemon-reloads, enables them
for boot, and starts them in dependency order. Safe to run on first migration
and on every later change.
Converts the podman-compose SearXNG stack to boot-persistent quadlet units so
it survives a reboot (the compose deployment did not). searxng sets
Network=gluetun.container, which makes it reuse gluetun's network namespace so
outbound search traffic exits the WireGuard tunnel; it publishes no port of its
own (the port lives on gluetun). gluetun and valkey join a pinned searxng
network (10.89.2.0/24, DNS-enabled) so searxng reaches valkey by name.
Secrets stay in /opt/searxng/.env (mode 0600), referenced via [Service]
EnvironmentFile= so ${VAR} expands. Adds a bats seam asserting the netns
wiring, port placement, boot install, and that no secret material is tracked.
The :8000 status server 404s on every path and the image has no curl, so the
prior healthchecks could not pass. Check /sys/class/net/tun0 instead: it exists
only while the WireGuard tunnel is connected, so the healthcheck reflects real
VPN state.
The gluetun image ships wget, not curl, and its status server answers 200 at
/ on port 8000 (not /health). The old curl-based healthcheck always failed, so
the container read 'unhealthy' even with a working tunnel. Use wget against /.
A direct hit on the published port has no X-Forwarded-For header, so the
bot-detection cannot resolve a client and returns 429. SearXNG must be
reached through Caddy (search.dankclaw.local), which sets X-Forwarded-For.
Add the podman bridge subnets to trusted_proxies so Caddy's source is
trusted for XFF extraction; note the CIDR is dynamic and should be pinned.
Verified agentic JSON search returns 26-31 relevant results via this path.
The host has no unqualified-search registries in registries.conf, so the
short names qmcgaw/gluetun and valkey/valkey did not resolve and the pull
hung. Qualify them with docker.io/ so the pull works on any host.
The compose only forwarded VPN_SERVICE_PROVIDER, VPN_TYPE and the private
key, so Gluetun could not form a tunnel. Forward the peer public key,
endpoint IP/port, addresses, allowed IPs and keepalive interval. Add
matching placeholders to .env.example (current Gluetun uses the
single-peer WIREGUARD_* model, not the old WIREGUARD_PEERS_* names).
The compose still runs ComfyUI with --cpu (the gfx1151 torch image is not
built yet), so the four compose-wiring ROCm tests fail. Skip them with a
pending note; the Containerfile test stays active. Re-enable when the GPU
migration is applied.
The trusted group lives in limiter.toml, not settings.yml. Add
shared/searxng/limiter.toml trusting 192.168.0.0/24 in pass_ip and
trusted_proxies, copy it into /opt/searxng/config on deploy, and correct
the SearXNG doc. New test seam: tests/searxng_limiter.bats.
Note that the Pi agent reaches the fleet via ~/.pi/agent/models.json
(outside the repo), that the model ids are stable and referenced by
settings.json, and the two resolution rules (default = provider+model pair;
subagents = unique bare id). Gives the per-slot contextWindow values.
llama.cpp splits -c across -np slots; to give each of the 2 slots the full
262144 native context, set CTX_SIZE=524288. KV ~32 GiB; fits the 96 GiB
unified pool. Update env default, serve-script fallback, and the deployment
doc line.
- Restore the pre-2026-08-17 Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
Q6_K_P + ngram-mod setup as frank-default-small on Dankserver (port 7770,
upstream llama.cpp b10121, MoE experts on CPU via --n-cpu-moe 36).
- Rename the CIRU Qwen3.8-27B ActiveFPX setup from frank-default-small to
frank-default-large (Dankclaw, port 52625); bump to CTX_SIZE=262144
(native max) and PARALLEL=2 (96 GiB unified pool).
- Retire the Qwen3.8-27B IQ4_XS FastMTP frank-default-large (Dankserver) to
.AGENTTRASH.
- deploy.sh: SERVICE[dankserver]=frank-default-small,
SERVICE[dankclaw]=frank-default-large; HEALTH_WAIT[dankserver]=600.
- CTX_SIZE 65536 -> 131072 (user request). KV at 128K is 8 GiB
(64 KiB/token f16); estimated ~43 GiB used of the 64 GiB SVM pool,
~21 GiB headroom.
- REASONING off -> on with REASONING_FORMAT=deepseek, so <think>
reasoning lands in message.reasoning_content (OpenAI-compatible)
instead of being left unparsed in content.
- Dankclaw System Specs: correct RAM to 64 GiB (61 visible), GPU
budget to the rocm-smi 64 GiB SVM figure, and replace the vLLM
relevance section with the CIRU ROCmFPX runtime facts. Note the
retired vLLM venvs in .AGENTTRASH. Refresh the disk figures.
- deployment.md: update the dankclaw target row (CIRU runtime).
Retire the vLLM service (frank-default-small, port 52625) on Dankclaw
and serve jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge through the CIRU
ROCmFPX llama.cpp runtime instead.
- Build llama-server from ciru-ai/ROCmFPX tag
qwen3.8-activefpx-promptforge-v2.3 (commit 103af8b41) against the
private TheRock ROCm 7.15 SDK in the retired ling3-ciru venv. The
model card requires ROCm 7.15 for the advertised Strix Halo
performance; the system ROCm 7.2.4 selects slower kernels.
CK pinned at fdf4bb7fcc984811cef48ce817d89aac064b984a; generated
ck/config.h + ck/version.h copied into the CK source tree so the
ggml-hip build can include them. LLAMA_BUILD_WEBUI=OFF (HF bucket
llama-ui returns 404; server is consumed via the OpenAI API).
- New scripts/serve-frank-small.sh + frank-default-small.env: 64K ctx,
one slot, f16 KV, context checkpoints, prompt cache, native depth-4
MTP on ROCm0, PromptForge DualView sidecars (FFN + GDN), temp 0.7
top-p 0.95, reasoning off. LD_LIBRARY_PATH puts the 7.15 SDK libs
before any system ROCm.
- Update the systemd unit description (same service name and port).
- deploy.sh: update the dankclaw target comment.
- Docs: new docs/ciru-activefpx-promptforge.md with the build runbook;
update docs/index.md and AGENTS.md topology.
- The vLLM venvs (vllm-venv, vllm-vk-venv) were moved to
/home/dank/.AGENTTRASH/ on the target; the vllm-vulkan plugin and
docs were removed from the repo in an earlier commit.
The box was upgraded from 32 GB to 64 GB. Two vLLM sizing pages
still stated 30 GiB physical RAM; rocm-smi reports 64 GiB and free
shows 61 GiB visible. Update the hardware tables and the ROCm
overcommit guidance to match.
Also removes the vllm-vulkan backend page, its index links, and the
frank-vk-compat plugin from the working tree. The vLLM service on
Dankclaw is being retired in favor of the CIRU llama.cpp runtime.
The Avesed template only shows a generic format; the Qwen3.5-MoE
model kept emitting <function=get_weather(city=...)> (args baked into
the name). Add a WORKED EXAMPLE block derived from the actual tool
schema (name + parameter keys) to the system prompt.
The Avesed tokenizer_config.json ships no chat_template, so vLLM used its
built-in template, which does not render the <tool_call> instructions;
the model then emitted malformed XML (arguments baked into the function
name) and the qwen3_xml parser could not extract clean calls.
Clients send tool_choice=auto; without these flags vLLM rejects with
HTTP 400. The Avesed chat template renders the Qwen3 XML tool format,
so qwen3_xml is the matching parser.
- frank-default-large: GGML_CUDA_FA_ALL_QUANTS=ON is mandatory in the
FastMTP build; without it q8_0 KV + FA falls back to a slow path and
prefill collapses ~12x (95 -> 1312 t/s at 8K, measured)
- frank-default-small: FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE + pure-python
flash_attn (triton_amd) + stub; without it the vision encoder warmup
OOMs via torch SDPA
Frank