The ik_llama fast q8_KV cache type crashes at model init on this
QWEN3MOE/MTP + multimodal arch at commit b166e269 with a ggml_view size
assert (ggml.c:5427) right after the output buffer is created, looping
the service. Empirically isolated by toggling cache type: q8_0 and f16
load and serve (smoke-tested correct output), q8_KV asserts. Switch to
q8_0 + Hadamard (K/V), which still gives the 8-bit cache + Hadamard
quality win. KB updated with the crash lesson.
--split-mode graph requires >= 2 devices in ik_llama.cpp and errors hard
at model load on a single RTX 3090 ("It is not possible to use split mode
'graph' with less than 2 devices"), looping the service. Set SPLIT_MODE
=none (the default), which works with the --override-tensor CPU expert
offload. KB updated with the two-device requirement so we do not assume
graph degrades gracefully on one GPU.
Switch frank-default-small on Dankserver from upstream llama.cpp (b10121)
to ikawrakow/ik_llama.cpp (b166e269, built with the staged pip cu13
toolkit). The model (Ornith-1.5-35B-A3B CRACK Q6_K) is unchanged and the
old binary is kept for rollback.
Runtime changes:
- KV cache quantized to q8_KV with --k-cache-hadamard/--v-cache-hadamard
- --split-mode graph (single-GPU no-op, retained per operator pref)
- migrate ngram speculation to canonical
--spec-type ngram-mod:n_min=48,n_max=64,ngram_size_n=24
(legacy --spec-ngram-mod-* are rejected in ik main)
- N_GPU_LAYERS 999 (ik -ngl rejects the literal "all"); experts stay
pinned to CPU via --override-tensor
Build: GGML_CUDA_FA_ALL_QUANTS=ON + GGML_IQK_FA_ALL_QUANTS=ON mandatory
for the quantized KV cache; recipe + reasoning baked into
docs/LLM/ikllama-ornith.md. Tests and both indices updated.
Fan out subagents to compare mainline llama.cpp, charlie12345/ROCmFPX,
and ik_llama.cpp, then add their findings to the project knowledge base:
- docs/Research/Llama Cpp Runtimes.md (positioning + quantization + server fit)
- docs/Research/Llama Cpp Code Diff.md (code-level quant diff, table, recommendations)
- docs/Research/SGLang.md (what SGLang is, comparison, our-boxes fit)
- docs/Research/index.md (Research progressive-disclosure index)
Cross-linked from docs/index.md (root) and docs/LLM/index.md.
AGENTS.md:
- Add Documentation Authoring section: docs/ is an OKF bundle; point doc authors
at the canonical spec in the system-config knowledge base
(~/.pi/agent/docs/references/okf-spec.md), not the orphaned
docs/references/okf-spec.md; consolidate read-index/one-concept-per-file/
progressive-disclosure/citation rules.
- Map AGENTS.md to docs_completeness.bats in the focused pre-commit hook (it was
last committed before the hook existed and had no mapping).
frank-default-large now serves the kingjones Qwen3.8-27B-ROCmFPX-Q8_0-AGENT
model with a standalone MTP draft head (Q4_0), f16 K + TurboQuant V KV cache,
and the Sharp chat template. Context is maxed to 262144 (one slot); reasoning
is on with preserve_thinking + preserve_reasoning in the chat kwargs.
- frank-default-large.env: Q8_0 AGENT model paths, CTX_SIZE=262144, REASONING=on,
REASONING_FORMAT=none, CHAT_TEMPLATE_KWARGS {enable_thinking, preserve_thinking,
preserve_reasoning}
- serve-frank-large.sh: standalone draft-mtp profile (already switched off the
Kairic/PromptForge path)
- download-q8agent.sh: reproducible HF fetch of the three payload files
- docs/LLM/kingjones-q8agent-rocmfpx.md: Q8_0 AGENT runbook
- AGENTS.md: topology + ACTIVE runtime section + TTM/GTT load note
- qwen3.8-strix-optimizations, Dankclaw Improvements: TTM pages_limit now 114 GiB
(applied 2026-08-25), full 262k KV + Q8 payload fit in one slot
- frank_large_config.bats: assert 262144 context, reasoning + preserve_* enabled
The TTM/GTT was raised to 114 GiB (29,883,536 pages) to hold the full 262k KV.
SearXNG queries all ~50 upstream engines from one Gluetun VPN exit IP, so
rate-sensitive engines (Google, DuckDuckGo) answer with reCAPTCHA/429 and
SearXNG suspends the offending engine for hours-to-days, showing up as
"unavailable". Add docs/Dankclaw/SearXNG Providers.md covering: the
suspension model, the two failure modes (CAPTCHA vs selector drift), engine
config (settings.yml engines: list — duplicate google, timeout,
retry_on_http_error, JSON feed), request tuning (outgoing block, UA rotation,
suspended_times — correcting the old rate_limit/useragents blog myths),
routing/proxies (per-engine proxies, dedicated 2nd Gluetun, dedicated IP), and
verification (/json, /config, check.searx.space). Crosslink into SearXNG.md
and index.md.
Replace the retired kingjones ROCmFP4 STRIX MTP payload on Dankclaw with
cafonz/Qwen3.8-27B-ROCmI4-MTP-GGUF (eemin namespace, public), served by the
charlie12345/ROCmFPX llama.cpp build (ROCmI4 W4A4, c49ebdbd) on the Strix Halo
iGPU. Q4_0_ROCMI4 with an embedded NextN/MTP head (no separate draft file);
strict-Qwen MTP n16/p0.60/backend sampling on ROCm0. Binary
/home/dank/rocmi4-qwen38/ROCmFPX/build-rocmi4-w4a4/bin/llama-server; model dir
/home/dank/models/eemin-Qwen3.8-27B-ROCMI4-MTP. Updated serve script + env,
ubatch bench, large-config bats test, docs, AGENTS.md.
- frank-default-small (Dankserver): Ornith-1.5-35B-A3B-CRACK Q8_0, bundled
native draft-mtp (n_max 4), MoE expert FFN tensors pinned to CPU via an
explicit --override-tensor \.ffn_(up|down|gate|gate_up)_(ch|)exps=CPU
(replaces --n-cpu-moe), Qwen-Sharp chat template.
- frank-default-large (Dankclaw): kingjones Qwen3.8-27B-Uncensored-Q6_0
ROCmFP4 AGENT dense GGUF, standalone mtp-Qwen3.8-27B-Q4_0 draft head
(--model-draft, draft-mtp n_max 4), Qwen-Sharp chat template. Drop the
retired PromptForge .pfs sidecar lanes.
- Track scripts/chat-template.jinja (v22.4.0 Sharp) and apply via
--chat-template-file on both services.
- Docs: new kingjones-strix-mtp and dealign-ornith-crack runbook pages;
update index, deployment, and both system specs; models.json names.
Smoke-tested both models on the real hardware on scratch ports before
commit: model loads, MTP draft attaches, Sharp template applies, chat
completion returns.
Fanned out 5 read-only doc-auditor subagents (one per service group) that
fact-checked the knowledge base against the live system + repo config,
researched official setup docs online, and verified progress. Applied:
- All 34 broken links fixed (real pages / external URLs / correct relative
paths); 0 broken links remain. No .AGENTTRASH links (the 4 mentions are
accurate descriptive text, not links).
- Fact-check fixes: Readarr retired (May 2026), Radarr v4->v6, Calibre
9.2.1 live, Calibre-Web 0.6.25 live, Copyparty v1.20.21, LLM index
ncmoe 38 + 786432 ctx, frank-default-small active (not frank-pp-worker),
Hermes/SearXNG/Grafana hostnames, Loki/Alloy local images, Observability
diagram (Promtail->Alloy), VLLM retired note, DFlash status, benchmark
live-drift notes, Homepage test count 116.
- Homepage.md: rewritten to the settled design (single site, user + /admin/
pages, server-side status.json via a 3-min systemd-timer checker, terminal
TUI theme). Added mermaid diagrams: site structure, status states, checker
flow. Recorded the classification (user vs admin; *arr excluded; LLM admin;
backends status-only) and the documented future shift to client-side probing.
- Homepage Design Grilling.md (new): the decision log — settled Q1-Q5 plus 38
open questions (Q6-Q43) across 10 categories, each with a recommendation.
- docs/index.md: link the grilling page (satisfies the index-links-every-page test).
- report/homepage-grilling-2026-08-22.html: self-contained terminal-TUI progress
report (settled decisions, verified stack baseline, overnight processing, work
threads, the open questions, and how to continue).
Baseline recorded: Dankclaw fully up (10/10); the Dankserver media stack is down
(only Calibre-Web + the LLM are up; no .env on the target) — a separate thread.
No homepage code built yet; the design is pending the operator's answers to Q6-Q43.
- docs/index.md: LLM Fleet section (all 34 pages linked), merge note,
See Also now internal
- docs/LLM/index.md: retitled, dangling [[Dankserver]]/[[Dankclaw]]
links fixed, sibling section replaced with internal links
- docs/LLM/deployment.md: rewritten for the unified deploy flow,
manual-update paths point at the merged checkout, incident history
kept intact
- Deployment Flow / Dankclaw Stack / Gitea: stale vllmScripts
references and retired-topology notes updated to the 2026-08-22
state
- Historical docs (frank-redo, PP deployment, FastFlowLM) keep their
dated runbook paths as-is (records of retired topologies)
Add a hermes-dashboard compose service that runs 'hermes dashboard' on
0.0.0.0:9119 with HTTP basic auth. The username + scrypt password hash
live in the git-ignored /home/dank/hermes-data config (plaintext never
stored). A separate container from the Signal gateway, since the gateway
entrypoint does not start the dashboard. Both share the data volume and
are boot-persistent. Docs page + index link added (docs-first contract).
- README: fix broken 'Also See' placeholder URL -> live Gitea LAN URL
- index.md: expand vllmScripts See Also with the endpoints it serves
- OpenWebUI.md: :7770 backend is the llama.cpp PP head (split to the
3090), managed by vllmScripts; note the FLM NPU :52625 alternative;
fix the broken [[llama.cpp Server]] wikilink
- Gitea.md: hosted repositories table (both projects)
- Dankclaw Stack.md: host-level LLM endpoints in the service inventory
- AGENTS.md: runtime note — LLM endpoints live in vllmScripts, not here
Pre-existing failures (untracked ComfyUI WIP): tests 6, 8, 9, 10.
Docs seams green: ok 19 (compose service docs), ok 20 (index links).
Live-verified on Dankclaw 2026-08-14:
- official rocm/comfyui image is built PYTORCH_ROCM_ARCH=gfx942;gfx950
(datacenter only) -> torch.cuda.is_available() False on gfx1151
- AMD gfx1151-tagged PyTorch build (rocm/vllm:...gfx1151...pytorch_2.9.1)
-> avail True, Radeon 8060S, cap (11,5); needs /dev/kfd + /dev/dri
- torch.backends.vulkan is compiled out of standard wheels; no ComfyUI
--vulkan flag exists; Vulkan is a dead end for ComfyUI
- Lemonade Server does image gen via wrapped stable-diffusion.cpp, whose
HIP kernels include gfx1151 -> works where ComfyUI's torch cannot
New pages: ComfyUI ROCm and Vulkan, Lemonade and sd.cpp; ComfyUI.md
qualified; compose stays --cpu until a gfx1151-enabled torch image is built
Frank