serverAppConfigs/AGENTS.md

AGENTS.md — serverAppConfigs Agent Rules

Overview

This repository is the single source of truth for the two homelab servers. It tracks both the Podman container stacks and the LLM inference fleet. One repo, one checkout per host, one deploy script.

• Dankserver (192.168.0.235) — media/library stack (RomM, Copyparty, Calibre-Web, the *arr stack, Jellyfin) as bare containers, plus the frank-default-small LLM service.
• Dankclaw (192.168.0.143) — AI/search/development/observability stack (SearXNG via Gluetun, ComfyUI, OpenWebUI, Hermes, Gitea, Grafana, Prometheus, Loki, Alloy, Pi-hole) plus the frank-default-large LLM service. Caddy on Dankclaw reverse-proxies both hosts.

2026-08-22 merge: the former vllmScripts repository (LLM fleet scripts, units, env tunables, and runbook) is merged into this repo. Its full git history is preserved (two-parent import commit under the retired llm/ prefix, reorganized in follow-up commits). The retired target checkouts (/home/dank/vLLM on Dankserver, /home/dank/vllmScripts on Dankclaw) are archived. All services now run from /home/dank/serverAppConfigs on each host.

LLM fleet topology (2026-08-30)

Target	Runtime	Service	Model	Port
Dankserver	FreeToken v0.1.2	frank-default-small	batsclamp Huihui-Qwen3.6-35B-A3B-abliterated-FP8 (block-fp8 e4m3 128x128, ~35 GB safetensors), experts host-RAM LRU, 262144 ctx / 4 concurrent, reasoning on	7770
Dankclaw	kingjones30/ROCmFPX llama.cpp (card-pinned fork)	frank-default-large	kingjones777 Qwen3.8-Flash-Next-ROCmFP4-STRIX (113.47 GiB single GGUF, Q4_0_ROCMFP4_STRIX, PLE streams via mmap), 401408 ctx / 2 slots, fa on, no spec decode, reasoning on	7770
Dankclaw	FastFlowLM (FLM) v1.0.1 (flm serve, dpkg fastflowlm) on the XDNA2 NPU	frank-default-sightsound	Gemma4-E4B-IT-NPU2 pack (vision + audio towers); whole local registry served on one port, per-request routing by exact FLM tag; iommu=pt boot arg mandatory (SVA bind needs enabled IOMMU)	7771

History: the 2026-08-17 redo retired the pipeline-parallel split and the FastFlowLM NPU endpoint. The 2026-08-18 swap retired the vLLM service on Dankclaw. On 2026-08-21 a DFlash2 Vulkan deployment briefly replaced CIRU, then was reverted the same day for speed. On 2026-08-22 the service briefly swapped to the Kairic Edge IU4 runtime, then was reverted to v3 the same day (operator decision). On 2026-08-25 the service switched to the kingjones Qwen3.8-27B-Uncensored-Q6_0-ROCmFPX_AGENT model (standalone MTP draft, 262144 context, reasoning on). On 2026-09-01 both services were swapped: Dankserver switched from ik_llama.cpp Ornith CRACK Q8_0 to stock llama.cpp Tiel-Coder-35B (qwen35moe hybrid MoE), and Dankclaw switched from the Q6_0 AGENT model to the Flash-Next 56B MoE (qwen4exp, PR #27742 merged into charlie's ROCmFPX fork, --fit off, 128k context). On 2026-09-02 Dankserver moved to the FreeToken runtime serving the official Qwen3.6-35B-A3B-FP8. On 2026-08-30 both services swapped again: Dankclaw to the kingjones STRIX model on the card's pinned kingjones30/ROCmFPX runtime, Dankserver to the Huihui abliterated FP8 checkpoint on FreeToken. On 2026-08-30 the NPU endpoint returned as frank-default-sightsound — FLM v1.0.1 on XDNA2 serving sight+sound (see the Runtime & Troubleshooting section and its runbook). Runbooks: serverAppConfigs-docs/LLM/index.md. Retired scripts live in old/ (reference only) and .AGENTTRASH/.

Repository Layout

• dankserver/ — compose.yaml + Caddyfile + .env.example for the media stack
• dankclaw/ — compose.yaml + Caddyfile + .env.example for the AI stack
• shared/ — homepage, Prometheus, Loki, Alloy, SearXNG configs
• scripts/ — deploy.sh, deploy-plan.py, and deploy-remote.sh (service-scoped deployer), LLM serve scripts, systemd units, env tunables, and bench harnesses
• old/ — retired PP/NPU-era LLM scripts (reference only)
• serverAppConfigs-docs/ — knowledge base (Obsidian vault). serverAppConfigs-docs/LLM/ is the LLM fleet
• tests/ — bats test suite

Documentation Authoring

The serverAppConfigs-docs/ vault is the project knowledge base (Obsidian, wikilinks) and is an OKF (Open Knowledge Format) knowledge bundle. Follow the OKF spec for every doc you create or edit.

• Canonical spec: ~/.pi/agent/docs/references/okf-spec.md — the OKF v0.1 spec in the system-config knowledge base at ~/.pi/agent/docs. The repo's own serverAppConfigs-docs/references/okf-spec.md is deprecated/removed; use the system one. The spec index and maintenance guide are ~/.pi/agent/docs/references/index.md and ~/.pi/agent/docs/references/okf-maintenance.md.
• Read before you touch: open the relevant index.md first (serverAppConfigs-docs/index.md → serverAppConfigs-docs/LLM/index.md). Read the whole target file before editing it.
• One concept per file, with YAML frontmatter that has a non-empty type: field (reference / guide / configuration / research) plus description, tags, and timestamp. Split broad topics; keep pages ~1–500 lines.
• Progressive disclosure: an index.md at each level links its children. Every serverAppConfigs-docs/**/*.md (except index.md) must be linked from serverAppConfigs-docs/index.md with [[Title]] — the docs-completeness test checks this.
• Cross-link siblings as [[Title]] wikilinks; cite external claims under a # Citations section.
• Always update docs when a task changes behavior, then run ./tests/run.sh (the docs-completeness seam is in there) before commit.

Git Workflow

All edits flow through git. Do not edit files on the remote machine directly.

Protocol

1. Source the config from the target machine — fetch the live compose/Caddyfile/serve script from the server where it runs
2. Place it in this repo — commit the fetched version (redact secrets into .env)
3. Edit here — make all changes locally in the working tree
4. Run the tests — ./tests/run.sh must pass
5. Commit and push — push to origin (Gitea on Dankclaw)
6. Pull on the target — git pull on the remote machine
7. Deploy — scripts/deploy.sh pulls, plans, and dispatches only the changed service(s), then checks the selected service health

Deployment

Use scripts/deploy.sh to deploy committed changes to a target. The planner maps changed paths to individual services; the remote runner restarts only the selected service and records its successful revision outside the checkout.

./scripts/deploy.sh dankclaw
./scripts/deploy.sh dankserver
./scripts/deploy.sh dankclaw --service searxng
./scripts/deploy.sh dankserver --service frank-default-small
./scripts/deploy.sh dankclaw --dry-run

A documentation-only change is a no-op. The LLM services are selected only by their own unit/env/serve files or the shared chat template, so unrelated container changes cannot interrupt inference. frank-default-npu is relink-only when explicitly selected.

Set FRANK_SUDO_PW for headless runs on a host without passwordless sudo (remote sudo reads it from stdin via sudo -S). The value is never stored in this repo.

graph LR
    A[Edit locally] --> B[Tests pass]
    B --> C[Commit]
    C --> D[Push to origin]
    D --> E[deploy.sh computes service plan]
    E --> F[Target pulls with ff-only]
    F --> G[deploy-remote.sh dispatches selected service]
    G --> H[Per-service health check]

Boot Persistence & Self-Healing (MANDATORY)

Every Podman container governed by this project that is successfully deployed and proven working MUST have a self-healing, boot-persistent Quadlet unit. A container that runs only from compose.yaml (or a manual podman run) dies silently at reboot: rootless Podman does not auto-start bare containers, and restart: unless-stopped does not survive the explicit stop that shutdown performs. Do not report such a service as "up" without its unit.

Each container's unit must provide:

• Boot persistence — [Install] WantedBy=default.target, installed to ~/.config/containers/systemd/<name>.container (generating <name>.service) and enabled in the user manager (systemctl --user enable <name>.service). Linger must be on for the deploy user (loginctl show-user dank -p Linger → Linger=yes) so user units start at boot with no login session.
• Self-healing — [Service] Restart=always (or Restart=on-failure) with a short RestartSec, so a crash or OOM brings the container back without operator action. Add a quadlet health check where the service has one.
• Repo provenance — the .container file lives in this repo (dankclaw/<svc>/quadlet/ or dankserver/<svc>/quadlet/), is installed by the service-scoped deployer, and is covered by the quadlet test seam. The host-side unit is generated state; the repo file is the source of truth.

Verification before a service is declared deployed: systemctl --user is-enabled <name>.service → enabled, systemctl --user is-active <name>.service → active, and the container answers its health check. A running container with no unit is a gap: write the unit, test it, deploy it.

Incident that established this rule (2026-08-26, Dankclaw reboot): Caddy ran for weeks as a bare compose container and never came back after reboot; the Pi-hole unit existed but crash-looped on a sysctl conflict (a later-sorted drop-in reset net.ipv4.ip_unprivileged_port_start to 80), taking LAN DNS down with it.

Testing

Tests are mandatory before commit. bats must be installed on the development machine.

./tests/run.sh

Test seams: compose validity, env consistency (no leaked secrets), Caddyfile validity, quadlet units, docs completeness, script syntax (all of scripts/, including the LLM serve scripts), LLM unit integrity (units point at files in this repo), LLM secret hygiene, deployment flow, and service-scoped deployment planning/dispatch. See serverAppConfigs-docs/Architecture/Testing Policy.md.

Secret Rules

• Never commit real secrets. The live Gitea database password and the live sudo password are test-enforced as leak checks.
• .env files are git-ignored. Commit only .env.example with placeholders. The LLM env tunables in scripts/ are tracked by exception (!scripts/*.env). They hold paths and numbers, never credentials.
• Do not document passwords in this repo, including sudo passwords. SSH keys are the auth mechanism. Headless sudo passes the password via the FRANK_SUDO_PW environment variable. The value never enters the repo.

Deletion & Trash Policy (MANDATORY)

Never run rm to delete files. The command deletes permanently. Move every item you delete into trash instead. The trash keeps the deletion recoverable and reviewable.

1. In-repo files: move the item to .AGENTTRASH/ beside its directory.
2. Home and other paths: create a dated staging directory ~/.AGENTTRASH/YYYYMMDD-topic and move each item there.
3. List the staged items and their sizes for the user.
4. Wait for explicit user permission before you empty the trash.

Rules:

• rm, rm -rf, and git clean -f are forbidden without explicit user permission.
• git * --no-verify is forbidden.
• A moved item stays in the trash until the user approves its purge.
• The trash is a temporary hold, not a long-term archive.

Benchmarking Policy (MANDATORY)

STOP the production service BEFORE running any benchmark on a target. These servers serve real users (about 3 people). Running a bench while the service is up starves the GPU and corrupts results. Restart the service when the bench finishes.

ALWAYS bench using llama-bench for llama.cpp or vllm bench for vllm runtimes. DO NOT ROLL YOUR OWN BENCH UNLESS YOU CANNOT TEST THE DESIRED PARAMETER WITH THE OFFICIAL BENCHES. ASK THE USER BEFORE DEVIATING FROM THIS DIRECTIVE.

1. Stop the service before any benchmark: sudo systemctl stop frank-default-large (Dankclaw) or sudo systemctl stop frank-default-small (Dankserver). Then mask it to prevent the self-healing unit from restarting: sudo systemctl mask frank-default-large (Dankclaw). CRITICAL WARNING: Do NOT use systemctl disable on these service units. On these hosts, systemctl disable doesn't just remove symlinks — it deletes the unit file itself from /etc/systemd/system/. You'd have to manually copy the unit file back from this repo, run daemon-reload, then enable and start. Mask (systemctl mask) creates a symlink to /dev/null instead, which is safe and reversible.
2. Run the bench (llama-bench for llama.cpp / vllm bench for vLLM). Scratch ports are not used for tests or benches on the LLM hosts — a test process cannot share the single GPU with the running service anyway (the model + KV fill the card), so changes are validated directly against the service itself.
3. Restart the service when the bench completes: unmask, then start: sudo systemctl unmask frank-default-large && sudo systemctl start frank-default-large (Dankclaw).
4. Verify health before reporting: curl http://192.168.0.235:7770/health or curl http://192.168.0.143:7770/health.

Pi Model Config (~/.pi/agent/models.json)

The Pi agent reaches the fleet through /home/dank/.pi/agent/models.json. This file lives outside the repository. Each provider is one host endpoint. Each model id is the served alias. The two aliases are frank-default-small (Dankserver :7770) and frank-default-large (Dankclaw :7770).

The model id values are stable. The file /home/dank/.pi/agent/settings.json references them. Two resolution rules apply (verified in @earendil-works/pi-coding-agent model-resolver.js):

• The default model resolves as a defaultProvider + defaultModel pair. The defaultProvider must name the provider that serves the defaultModel.
• A subagent override uses a bare model id. A bare id resolves only when it is unique across all providers.

When you rename a service or move a model to a new host, update models.json. Update the provider key, the model id, contextWindow, and input. If the default model moved hosts, also update defaultProvider in settings.json. Set contextWindow to the per-slot context (both hosts 262144). The full schema is in the agent knowledge base page ~/.pi/agent/docs/pi/models.md.

Runtime & Troubleshooting

frank-default-sightsound — FLM NPU sight+sound endpoint (Dankclaw, ACTIVE)

The third inference endpoint on Dankclaw. The flm serve runtime (dpkg package fastflowlm, v1.0.1, binary /usr/bin/flm) serves the whole local model registry (36 tags at first deploy) on a single port — 7771 — and routes chat completions per request by the exact FLM registry tag in the model field. Unknown tags do NOT 404: /v1/chat/completions answers HTTP 200 with {"error":"Failed to load <tag> model!"} and no choices (status-only clients silently get nothing), so clients must send exact registry tags, never service aliases. Primary model = Gemma4 E4B IT NPU2 pack (vision + audio towers); the uncensored candidate (qwythos:9b) is text-only and fails the modality requirement, so this endpoint trades "uncensored" for "sight + sound". The NPU is a single-lane device — generations serialize FIFO; treat this as a background/triage endpoint, not an interactive multi-user chat.

Hard requirements verified at first deploy (2026-08-30):

• IOMMU must be enabled. With iommu=off on the kernel cmdline every open of /dev/accel/accel0 fails with err=-19: the XDNA2 driver's SVA bind needs PASID/SVA, which requires enabled IOMMU hardware. Fix applied: iommu=pt in GRUB_CMDLINE_LINUX_DEFAULT + reboot (passthrough keeps the iGPU default domain identity-mapped — zero DMA-translation overhead for frank-default-large — while enabling IOMMU + PASID; PASID domains are paged regardless of the default-domain type). Never set iommu=off on this host again; it silently kills the NPU.
• Health probe = GET /v1/version (FLM has no /health route).
• Unit: User=dank, LimitMEMLOCK=infinity (FLM requires unlimited memlock), Restart=on-failure, RestartSec=10. Deploy via scripts/deploy.sh dankclaw --service frank-default-sightsound.
• Verified live at first deploy: image probe (red PNG as image_url data-URI → correct color answer, vision tower consumed the pixels) and an audio control-pair probe (tone WAV via input_audio consumed ~22 audio-tower tokens vs a text-only control).

Runbook: serverAppConfigs-docs/LLM/frank-default-sightsound.md.

kingjones Qwen3.8-Flash-Next STRIX on kingjones30/ROCmFPX (Dankclaw, ACTIVE)

The large service serves kingjones777/Qwen3.8-Flash-Next-ROCmFP4-STRIX-GGUF (public, 113.47 GiB single file, Q4_0_ROCMFP4_STRIX = file_type 105) from /home/dank/models/kingjones-Qwen3.8-Flash-Next-ROCmFP4-STRIX/ with the card's pinned runtime: the kingjones30/ROCmFPX fork (charlie12345/ROCmFPX + qwen4exp-on-rocmfpx-d3ca537 patch — ROCmFP4 types and qwen4exp arch in one tree), built at /home/dank/GitPackages/kingjones-ROCmFPX/build with -DGGML_HIP=ON -DGPU_TARGETS=gfx1151 -DGGML_NATIVE=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_WEBUI=OFF (WebUI provisioning needs npm/HF-bucket and aborts the link at 98% without the OFF flag). Args: our validated profile — 401408 context / 2 slots, b 4096 / ub 512 (ub 512 matches the card's re-measurement), t 16 / tb 16, f16 KV, --fit off, --flash-attn on (works UNPATCHED on this model — the QSA F16 ggml_fill assert that needed scripts/patches/qwen4exp-qsa-f32-mask.patch is specific to the agentionai FAST layer set; if a future build aborts with GGML_ASSERT(a-type == GGML_TYPE_F32), port that fix), reasoning on (Sharp template, deepseek format), no speculative decoding (card ships no spec flags; Flash-Next MTP heads do not load on this loader family). Two hard card constraints: mmap must stay on (--no-mmap forces the 50.66 GiB Q8_0 PLE table into anonymous memory → silent cgroup kill, only dmesg evidence; llama.cpp's "use_mmap=1" log line is correct here — mmap pages ARE page cache) and never pin the PLE tensor with -ot (23 → 13.4 t/s). The 51.2B-param PLE table never enters GPU memory (~20 bytes/token-file per 500 decoded). On-host file sha256 verified == HF LFS oid c6770d74...49fa. Env scripts/frank-default-large.env, serve scripts/serve-frank-large.sh. Runbook: serverAppConfigs-docs/LLM/kingjones-strix-rocmfp4.md.

Serving status (2026-08-30): swapped from the agentionai FAST deployment. Validated live before the systemd handoff: clean load, health 200, reasoning completions, and an operator session's ~111k token prefill at 213–232 t/s (card rung: 196 at 131k). Note llama.cpp clamps n_ctx_train 262144 → per-slot 131072 with parallel 2.

agentionai ROCmFP4-FAST imatrix llama.cpp (Dankclaw, prior, 2026-08-29 to 2026-08-30)

Historical — rollback target. agentionai/Qwen3.8-Flash-Next-ROCmFP4-FAST-imatrix-GGUF (87.06 GiB, Q4_0_ROCMFP4_FAST imatrix, v2/ joined-PLE layout only — the root -ple16 per-head split aborts this loader family) on the charlie build-qwen38 tree (/home/dank/GitPackages/ROCmFPX, PR #27742 merged) at 32768/1, ub 4096. That tree REQUIRES scripts/patches/qwen4exp-qsa-f32-mask.patch for --flash-attn on (F16-kq_mask ggml_fill abort otherwise). Flash-Next MTP heads do not load on any current loader (require full per-layer blk.<il>.hc_attn_norm.weight; heads carry only blk.48 — verified 3x). Binary and payload retained. Runbook: serverAppConfigs-docs/LLM/agentionai-fast-imatrix.md.

Huihui-Qwen3.6-35B-A3B-abliterated-FP8 on FreeToken (Dankserver, ACTIVE)

The small service serves batsclamp/Huihui-Qwen3.6-35B-A3B-abliterated-FP8 (public, ~35 GB, 26 safetensors shards) from /home/dank/models/huihui-qwen36-35b-a3b-abliterated-fp8/ with FreeToken v0.1.2 (/home/dank/freetoken/.venv, ft serve). Native modelopt block-fp8 (e4m3, dynamic, 128×128 weight_block_size, un-fused per-expert layout) — same format and loader path as the official Qwen3.6-35B-A3B-FP8 it replaced on 2026-08-30, so the entire memory/pool plan carries: experts host-RAM with 20% LRU GPU residency, paged KV (only the 10 full-attention layers page; 30 GDN layers use constant-size mamba slots), ~3.4 GiB VRAM headroom at boot, and the validated --memory-num-gpu-blocks-experts 2048 --memory-num-gpu-tokens-kv-cache 453005 --max-num-seqs 4 pool triple (never grow the expert pool). 262144 context, reasoning on (the checkpoint ships the froggeric fixed chat template — no override). FreeToken's API REQUIRES an explicit "model" field in chat requests (422 without it). Env scripts/frank-default-small.env, serve scripts/serve-frank-small.sh. Runbook: serverAppConfigs-docs/LLM/freetoken-frank-default-small.md.

Serving status (2026-08-30): swapped the model within the FreeToken runtime (official FP8 → Huihui abliterated FP8). ~40 tok/s verified on a 282-token completion post-deploy. Prior: official Qwen FP8 (2026-09-02, runtime swap), Tiel-Coder-35B on llama.cpp (2026-09-01), ik_llama.cpp Ornith CRACK (retired 2026-09-02).

charlie12345/ROCmFPX ROCmI4 W4A4 llama.cpp (Dankclaw, prior v3, 2026-08-18 to 2026-08-25)

The v4 runtime serves cafonz/Qwen3.8-27B-ROCmI4-MTP-GGUF (eemin namespace, public) from /home/dank/rocmi4-qwen38/ROCmFPX/build-rocmi4-w4a4/bin/llama-server (charlie12345/ROCmFPX commit c49ebdbd built with -DGGML_HIP_ROCMI4_W4A4=ON against system ROCm 7.2.4 / amdclang 22). The model is Q4_0_ROCMI4 with an embedded NextN/MTP head (no separate draft file); the gfx1151 native IU4 WMMA instruction batch-verifies proposals. Profile: strict Qwen MTP (--spec-mtp-strict-qwen, n_max 16, p_min 0.60, backend sampling, all layers on ROCm0), -t 16 -tb 32, batch 512 / ubatch 256 (ubatch swappable via the ubatch bench on 8k context), f16 KV, full 262144 context / 1 slot, TEMPERATURE 0.7 / TOP_P 0.95, reasoning on (deepseek, budget -1), --no-mmap. Two-file payload (~14.4 GiB) in /home/dank/models/eemin-Qwen3.8-27B-ROCMI4-MTP/. Full runbook: serverAppConfigs-docs/LLM/eemin-rocmi4-mtp.md.

Kairic Edge llama.cpp (Dankclaw, RETIRED experiment 2026-08-22)

The Kairic Edge llama-server serves jcbtc/Qwen3.8-27B-IU4-Kairic-Edge (pinned tag kairic-edge-qwen38-27b-v1). It was deployed and reverted to v3 the same day (operator decision). The runtime routes qualified prompt and verification shapes through the RDNA 3.5 native IU4 matrix instruction.

Release gap — CK config.h: the pinned CK commit ships only config.h.in. The release script and guide never generate third_party/composable_kernel/include/ck/config.h. Generate it via a CK CMake configure and copy it in before building (full runbook in serverAppConfigs-docs/LLM/kairic-edge.md). The file is untracked in CK git. A git clean removes it. Local rebuilds are byte-identical (d7625a0c...).

Target greedy argmax fast path: the pinned runner sets LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=1. The fast path rejects non-greedy requests (400) and the card's own chat sampling is non-greedy, so the production attempt ran it at 0. It still rejects reasoning budgets unless compatibility mode is used. Do not re-deploy Kairic without re-testing chat first.

Rollback target: the v3 model files remain in /home/dank/models/CIRU-Qwen3.8-27B-ActiveFPX-v3/ and the v3 binary in /home/dank/ciru-qwen38/ROCmFPX/build-promptforge/bin/llama-server. Restored to v3 on 2026-08-22 (commit c60156f).

DFlash2 llama.cpp (Dankclaw, retired 2026-08-21)

A DFlash2 Vulkan deployment briefly replaced CIRU on 2026-08-21, then was reverted the same day for speed. Its serve script, env, and the build worktree are archived in .AGENTTRASH/ and documented in serverAppConfigs-docs/LLM/dflash2-large-vulkan.md.

FastMTP patched llama.cpp (Dankserver, retired 2026-08-18)

Historical: the Qwen3.8-27B IQ4_XS FastMTP service was retired on 2026-08-18 (files in .AGENTTRASH/). The build notes are kept for reference. The HauhauCS FastMTP draft sidecar required a llama.cpp build from commit 4df29be4f4c3673f428170fda944a5b19f743bb8 plus the HauhauCS-FastMTP-llama.cpp.patch (from the HF repo). The build lives at /home/dank/llama.cpp-fastmtp/build on Dankserver. Loading the draft with a stock binary fails with expected 5120, 248320, got 5120, 32768. Do not rebuild from another commit without re-applying the patch.

REQUIRED build flags: -DGGML_CUDA_FA=ON -DGGML_CUDA_FA_ALL_QUANTS=ON. Without GGML_CUDA_FA_ALL_QUANTS the q8_0 KV cache + flash attention falls back to a slow path and prefill collapses ~12x (measured 2026-08-17: 95 t/s vs 1300+ t/s at an 8K prompt on the 3090). See the HauhauCS HF repo README for the base recipe.

Env var naming rule — HIP runtime collision

The ROCm HIP runtime reads the bare environment variables TEMP, HOST, and GPU_MAX_HEAP_SIZE. The serve scripts export the env file with set -a. A tunable named TEMP leaks into HIP. On 2026-08-13 TEMP=1.0 made HIP use the path 1.0 for its JIT kernel cache, then segfaulted every model load. Commit 227ce01 fixed it by renaming the sampling variable to TEMPERATURE.

Never name a tunable TEMP, HOST, or GPU_MAX_HEAP_SIZE. Check for collisions with this command:

strings /opt/rocm-*/lib/libamdhip64.so.* | grep -xE "TEMP|HOST|GPU_MAX_HEAP_SIZE"

Crash-loop diagnostic procedure

When a llama service segfaults during model load, follow these steps.

graph TD
    A[Service SEGVs at load] --> B[Read journal + kernel segfault line]
    B --> C{Segfault in libamdhip64?}
    C -- yes --> D[Bisect script vs raw binary]
    C -- no --> E[Check config, memory, model files]
    D --> F[Run gdb for backtrace]
    F --> G[Find root cause]

1. Read the service journal with journalctl -u <service> -n 100. Then read the kernel segfault line with journalctl -k. A segfault at address 8 in libamdhip64 means a NULL unref in the HIP runtime.
2. Run the serve script directly. Then run the raw binary with the same flags. A different outcome isolates the env file as the cause.
3. Run gdb for the backtrace with this command:

   gdb -batch -ex run -ex "bt 25" --args bash scripts/serve-frank-large.sh
   

   llama-server buffers its timed stdout. On SIGSEGV the log tail is lost. Trust gdb, not the log.
4. Kernel spam about amdgpu: SVM mapping failed is a symptom, not the root cause. Investigate other causes before assuming memory pressure.

Operational quirks

• Both hosts allow passwordless sudo for the deploy user ((ALL) NOPASSWD: ALL, verified 2026-08-22). Use plain sudo in committed scripts. Do not use echo password | sudo -S in committed scripts. The FRANK_SUDO_PW env var covers headless runs on a host without that.
• Never run pkill -f with a pattern that appears in your own command line. It kills your ssh session. Use pkill -x <exact-name>.
• Model load is slow: Dankserver loads the 30.6 GB Q6_K_P through RAM (--no-mmap, MoE experts on CPU) in a few minutes cold. Dankclaw loads the ~30 GiB Q8_0 AGENT payload (main + MTP draft + mmproj) a few minutes cold with a fresh KV cache; ~18 s with a warm page cache. Health returns 503 while loading. Do not kill the process. deploy.sh waits up to 20 minutes on both targets (see LLM_WAIT).
• A client abort and a server release look different in the journal. A client abort logs W srv stop: cancel task, id_task = N. A natural stop (EOS or limit) logs only slot release with truncated = 0. The server never preempts a running generation. When all slots are busy, new tasks defer into a queue. See the 2026-08-13 incident in serverAppConfigs-docs/LLM/deployment.md.
• ComfyUI on Dankclaw runs CPU-only (--cpu). ROCm wheels do not support gfx1151. Do not use CUDA images.
• SearXNG outbound traffic must stay on the Gluetun tunnel. Do not move its ports off the gluetun service.
• Retired hardware history: the Dankserver 35B cuBLAS crash and the DFlash premature-EOS bug are historical. See serverAppConfigs-docs/LLM/dankserver-vram-growth.md and serverAppConfigs-docs/LLM/qwen3.8-spec-matrix.md. The FastFlowLM NPU endpoint is retired. See serverAppConfigs-docs/LLM/FastFlowLM NPU Inference.md.

Dankclaw memory model

The GPU budget on Dankclaw is the rocm-smi figure: 64 GiB VRAM reported by rocm-smi (SVM-backed). The docs' older 96 GiB figure was the addressable SVM pool. The GPU node reports local_mem_size 0, so all allocations are GTT/SVM-backed. Trust rocm-smi and the runtime. Do not use free or dmidecode for the GPU budget. They show the CPU-side carve-out. This is intentional. Do not change it.

Why

Direct edits on the remote drift from version control. Future agents will not know what changed. Git is the single source of truth. The two workloads share two hosts, one Caddy, and one Gitea. One repo keeps them in one history.

User request → Fetch from remote → Commit here → Edit → Test → Commit → Push → Pull on remote → Dispatch changed service(s) → Health

Summary

Context	Action
Editing a compose/Caddyfile/serve script	Work through git — fetch, commit, edit, test, push, pull, deploy. Never edit on the remote directly.
Config exists only on remote	Pull it into the repo first — commit the fetched version (redacted) before making changes.
Unstaged local changes	Commit or discard — check git status before starting work. Ask the user if unsure.
Deploying a committed change	Run scripts/deploy.sh <target> — it pushes, pulls with fast-forward-only, plans changed paths, dispatches selected services, and checks their health. See serverAppConfigs-docs/LLM/deployment.md.
Benchmarking	Stop the service first, restart, verify health. See the Benchmarking Policy above.
Deleting a file	Move it to trash — .AGENTTRASH/ (in-repo) or ~/.AGENTTRASH/YYYYMMDD-topic. Never rm.
Tests exist	Run them — verify tests pass after every change. Do not commit failing tests.
Real secrets	Never commit — use host .env files. .env.example holds placeholders only. FRANK_SUDO_PW covers headless sudo.