Both the EasiiX Q8_0 and agentionai ROCmFP4-FAST MTP draft heads fail
with 'missing tensor blk.0.hc_attn_norm.weight' against the STRIX_LEAN
Uncensored model. The qwen4exp binary expects this hyper-connection
attn norm tensor which neither draft GGUF carries. Using ngram-mod
only until a compatible MTP head is found.
The EasiiX Q8_0 draft was incompatible with STRIX_LEAN (missing
blk.0.hc_attn_norm.weight). Switched to agentionai/Qwen3.8-Flash-Next-
MTP-ROCmFP4-FAST-GGUF (2.28 GiB, ROCmFP4-FAST quant matched to target).
n_max 3 per the HF card (n_max 4 measured worse). SPEC_TYPE=draft-mtp,ngram-mod.
The script runs ON Dankclaw but was calling ssh() to SSH back to
dankclaw for service management. Nested SSH fails (no key auth for
localhost). Replaced ssh() with local() using subprocess.run(shell=True)
for systemctl and sed commands.
The env file had the old STRIX paths (directory doesn't exist on disk).
Updated to the actual running STRIX_LEAN (Uncensored) variant:
- MODEL_DIR → kingjones-Qwen3.8-Flash-Next-Uncensored-ROCmFP4-STRIX_LEAN
- MODEL_FILE → 3-shard Q4_0 (00001-of-00003)
- MMPROJ_PATH → mmproj-Qwen3.8-Flash-Next-Uncensored-BF16.gguf
- serve-frank-large.sh: add --mmproj $MMPROJ_PATH
Reconciles the uncommitted local changes that were on Dankclaw.
serve-frank-large.sh: only pass --model-draft when SPEC_TYPE contains
'draft-mtp'. ngram-only configs (ngram-mod, ngram-simple, etc.) do not
need a draft model file — loading one wastes VRAM and may error.
bench-spec-decode.py: new benchmark script. Sweeps 4 configs (none,
ngram-mod, draft-mtp, draft-mtp+ngram-mod), generates 16k tokens per
config at temp 0 from a minimal prompt. Uses HTTP API (llama-bench
cannot target --spec-type). Follows benchmarking policy: stop+mask
the production service, bench, unmask+restart, health check.
podman's quadlet generator creates the default.target.wants symlink from
the unit's [Install] section at every user-manager start (is-enabled
reports 'generated' for all healthy quadlets, pihole included). An
explicit 'systemctl --user enable caddy.service' fails with 'transient or
generated' and aborted the deploy after the unit was already installed.
systemd-sysctl applies /etc/sysctl.conf AFTER all of sysctl.d/, so the
stale ip_unprivileged_port_start=80 line there beat every drop-in and kept
the runtime value at 80 after the first deploy pass. deploy_pihole now
comments out an active line in /etc/sysctl.conf (idempotent) and generalizes
the sysctl.d conflict scan to any file setting a value above 53.
Caddy left compose.yaml for a quadlet unit (caddy.service): joins
dankclaw_default for container-name upstreams, Pull=never for the local-only
image, Porkbun credentials via /home/dank/Containers/caddy/env written by
deploy from .env, bare-IP-catch-all health check. deploy_caddy installs,
enables, and starts the unit after retiring the bare container (state is in
host volumes).
Pi-hole: the repo now owns the sysctl drop-in
(dankclaw/pihole/sysctl/99-dnsmasq.conf, ip_unprivileged_port_start=53);
deploy_pihole installs it to /etc/sysctl.d/, stages conflicting later-sorted
drop-ins in the home trash, and applies 'sysctl --system' (the boot code
path). This fixes the 2026-08-26 crash loop (port-53 bind denied, restart
counter 1645) caused by a second drop-in resetting the value to 80.
Tests: new caddy_quadlet.bats seam; caddyfile/deployment_flow assertions
repointed from the removed compose service to the quadlet; pihole.bats pins
the sysctl drop-in and its deployment; test-changed.sh routes dankclaw/caddy/
and the pihole subtree to their seams. Docs: Caddy, Pi-hole, Dankclaw Stack,
work log.
--ctx-size is the total pool divided among --parallel slots, so with
PARALLEL=3 the 160k target must be a 491520 pool (163840/slot) to
actually deliver 160k context per slot. Env, serve default, bats updated.
Reconfigure to 3 parallel 160000-ctx slots (was 1 slot). KV cache is
now asymmetric: K at q8_0, V at q5_0, both with Hadamard. The 3 slots
triple the KV, so expert blocks 27-34 are pinned back to CPU (freed
~8.7 GiB) leaving output-adjacent 35-40 on CUDA. Q8_0 model,
split-mode none, spec disabled (ik limits spec to 1 slot). AGENTS.md:
drop the scratch-port rule (test/bench processes cannot share the
single GPU with the running service anyway). Env, serve script, bats,
docs updated.
Switch to CRACK Q8_0 model (37.8 GB, mmproj f16). Servt as a single
160000-ctx slot (PARALLEL=1); speculative decoding stays disabled. KV
cache is q6_0 + Hadamard. iq5_k is not a valid KV-cache type in ik_llama
(cache whitelist excludes iq5_*); q8_KV reproduces the ggml.c:5427 view
assert even at PARALLEL=1/no-spec, so q6_0 is the verified working 6-bit
choice. Env, serve script, tests, and KB updated. Deployment was left
down after the isolation test; this commit redeploys it with the
verified Q8_0/160k/q6_0 config.
Fused MoE (fused_moe / fused_up_gate / fused_mmad) is on by default in
ik_llama - no flag needed; only -no-fmoe disables it. Slide two more
expert blocks (28,29) from CPU onto the RTX 3090 (~1.26 GiB) for the
fused-MoE run, leaving ~1.6 GiB headroom. Update override regex, test,
and KB (including a note that the headroom is now tighter than the
2-3 GiB ideal).
--jinja: ik_llama only renders a custom --chat-template-file as real Jinja
when --jinja is passed (inputs.use_jinja ? apply_jinja : non-jinja
fallback). Without it the Qwen-Sharp template's thinking/image/tool
formatting is silently dropped. Add USE_JINJA=1 -> --jinja.
Experts: the RTX 3090 had ~5.4 GiB headroom; pull expert blocks 30,32,33,34
back to CUDA (each block ~630 MiB) from CPU for lower latency. Update the
OVERRIDE_EXPERT regex, its test, and the KB with the headroom-check note.
ik_llama limits speculative decoding to a single server slot (-np 1) and
hard-exits at load if a spec stage is configured with --parallel > 1:
"Speculative decoding is currently limited to a single server slot
(-np 1)." This service keeps PARALLEL=3 (one slot per user), so set
SPEC_TYPE=none to disable speculation. It is a low-parameter MoE (3B
active of 35B), so throughput is fine without it. MTP stays disabled.
KB and tests updated.
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.
The v6 restore commits dropped --chat-template-file and
--chat-template-kwargs while the artifact check still validated the
template file, so the Sharp template was checked but never applied.
Re-add both flags (with --jinja) so reasoning is emitted in DeepSeek
format and preserve_thinking/preserve_reasoning are honoured.
Reformat the script one-flag-per-line with long flag aliases
(--gpu-layers, --ctx-size, --ubatch-size, --cache-type-k/v) like
serve-frank-small, and update the 2 frank_large_config assertions
that grep the old short-flag forms (-ctk/-ctv, -ub).
- scripts/serve-frank-large.sh
- tests/frank_large_config.bats
The pre-commit hook runs scripts/test-changed.sh, which requires a
focused-test mapping for every staged path. The quadlets were moved into
dankclaw/<svc>/quadlet/ but the mapping still only matched top-level
dankclaw/*.container, so any nested quadlet or the pihole addn-hosts/env
data files fell through to the catch-all 'die' and blocked commit.
Add the nested-quadlet mappings (pihole-specific first, then the
dankclaw/*/quadlet/* fallback) plus the pihole data-file paths.
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