Files

T

Frank 49e8dfff66 docs: the vault and AGENTS.md describe two Dankclaw endpoints

halogen-flash-server is ACTIVE again with its 2026-09-16 measured ledger (budget
33.5 GiB against a need of 27.2 — no clamp — and 14.7 GiB left for everything
else), ik_llama.cpp is retired with its measurements kept as the record of why,
and the new frank-small runbook covers the -mtp file, the 64 KiB/token KV
arithmetic, the MTP contract, and the reasoning_effort trap that cost me every
subagent on the endpoint.

AGENTS.md: topology, Active Services, runbook index, and three new llama.cpp
lessons (MTP is a property of the file; --jinja is what `tools` needs; a model's
template can reject the harness's thinking level). Three pre-existing defects
found while editing it: a truncated sentence mid-file, a harness path that does
not exist on either server, and the benchmarking stop/mask recipe still calling
frank-large a root unit.

Truthfulness notes worth keeping: the co-tenant stall NOTE prints again, so the
claim that it "no longer prints" is gone and the three observed block counts are
recorded instead of a guessed threshold; and the Observability page says out
loud why frank-large has no scrape job rather than leaving a silent absence.

2026-09-16 14:52:51 -07:00

7.2 KiB

Raw Blame History

type, description, tags, timestamp

type

description

frank-large — ik_llama.cpp, Qwen3.8-Flash-Next-Uncensored i1-Q4_K_M

RETIRED 2026-09-16, after one day. Replaced by halogen-flash-server, which was the runtime before it as well — the fleet went ik → halogen → back. Its unit, env, serve script and the Qwen-Sharp template are archived in the repo under scripts/.AGENTTRASH/*.ik-llama.cpp and *.ik-llama-qwen-sharp, and the payload, build and chat template stay on disk. The measurements below are not a stale plan: they are the complete account of what this model costs on this hardware, and they are why the fleet went back. It served mradermacher/Qwen3.8-Flash-Next-Uncensored (111 GiB GGUF + mmproj-f16) at 192.168.0.143:7770 while it ran.

Shape

Runtime: ik_llama.cpp master d85a1d7, built at /home/dank/GitPackages/ik_llama.cpp/build with -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=86 -DGGML_AVX512=ON -DGGML_AVX512_VBMI=ON -DGGML_AVX512_VNNI=ON -DGGML_AVX512_BF16=ON -DLLAMA_CURL=OFF.
Unit: root systemd unit /etc/systemd/system/frank-large.service symlinked to scripts/frank-large.service in the repo; enabled at boot; Restart=always. Serve script scripts/serve-frank-large.sh + env scripts/frank-large.env.
Topology: 3090 (PCIe x4 in the M.2 dock, link downgraded from x16) carries dense/attention/KV (~11.5 GiB); ALL MoE expert tensors (~84 GiB) and the ~33 GiB PLE n-gram table stay in LPDDR via CPU buffers.
Client contract: greedy default (--temp 0.0), reasoning on (deepseek format, unlimited budget), vision via mmproj, slot-state persistence under scripts/slot-state, 262144 ctx, 1 slot.

Measured performance envelope (2026-09-15/16 campaign, all single-instance verified)

config	decode tok/s	prefill tok/s	verdict
shipped (T16, `-ot exps=CPU`, NO_PINNED_WEIGHTS)	3.8-4.0	~193 at 1.8k / 191 at 32k	shipped
T8 / T4 / T2 / T1	4.18 / 4.20 / 4.26 / 3.97	small-prompt only	decode flat across threads
`-ncmoe 42` (6 layers' experts on GPU)	4.54	32 (small prompt)	VRAM 22.0/24.6 GiB — fragile
`-ncmoe 48 -thp`	4.60	18	110 GiB anon host buffer — co-tenant killer, never ship
`--prefetch-experts`	4.03	—	no effect (page cache already resident)
`--defer-experts --defer-ple -thp`	crashes	—	experts under `-ngl 999` still target CUDA0 → cudaMalloc OOM
mainline llama.cpp `37b53fd` (`--n-cpu-moe 48`)	1.01	34	ruled out — CPU MoE kernels 4x slower than ik's iqk

Large-prompt prefill ≈ 192 tok/s is the CPU AVX512-VNNI compute ceiling for CPU-resident experts (expert weight reads amortize across the batch; compute-bound). Halogen's "4x llama.cpp family" prefill claim is therefore unreachable with experts on CPU — the physics caps hybrid MoE prefill here.
Decode is gated by a serial per-token critical path: ~100 CPU↔GPU backend splits per token (graph splits = 100 in the log); the 3090 sits at 5% SM utilization during decode; thread count does not matter (T1 ≈ T16 ≈ 4 tok/s). The GPU sections are launch/sync-bound, not compute-bound.

Root cause chain (perf evidence)

perf at T16: ~53% cycles in ggml_graph_compute_thread + ~37% in ggml_compute_forward_mul_mat_id, of which ~98% is the pause instruction (thread-pool barrier spin). Actual vec_dot math: ~4%.
T1 = T16 decode (3.97 vs 3.8-4.0) ⇒ the barrier spin is threads waiting on a serial critical path, not parallel overhead.
GPU 95% idle during decode ⇒ the critical path is the GPU-section launch/sync overhead per split (x4 dock link, eager kernel launches).
ik's CUDA-graph capture would collapse those launches but aborts on qwen4exp: the per-layer hyper-connection broadcasts (ggml_repeat_4d, 4 sites in src/graphs/build_qwen4exp.cpp) resolve to a REPEAT node whose CUDA path does a capture-illegal operation (CUDA error: operation not permitted when stream is capturing → abort).
Memory placement is NOT the lever: mmap-off/anon/THP/deferred variants all land at 3.8-4.6 tok/s.

Why the 30 tok/s halogen bar is out of reach (honest record)

Halogen's engine held the whole 115.55 GiB w4b checkpoint resident in LPDDR (HALOGEN_FLASH_PIN_TRUNK=1) and served from one engine. This deployment must split 84 GiB of expert tensors across a 24 GB GPU + CPU, and every public runtime pays the split penalty above. Paths to >30 tok/s, in order of realism: (a) a runtime with capture-safe CUDA graphs + dedicated hc ops — mainline is building exactly this (ggml-org PR #28901 "qwen4exp: add hc ops"); re-evaluate mainline when its hc-ops + graph handling mature; (b) experts on GPU — impossible, 84 GiB > 24 GiB; (c) a smaller model.

Upstream issues tracking the slow TG path

ik_llama.cpp: #2374 (QSA opt POC), #2406 (single-threaded sigmoid), #2433 (cache-miss full reprocess), #2460 (PLE history correctness). MTP/speculative decoding is unavailable for this GGUF — mradermacher stripped the NextN head (verified: no blk.48.*/nextn tensors in the file).

Traps (learned the hard way 2026-09-15/16)

Pinned host buffers strand the host: a default-policy start page-locks ~108 GiB (CUDA_Host buffer), evicting co-tenants into swap and starving sshd until a power cycle. GGML_CUDA_NO_PINNED_WEIGHTS=1 + the serve-script guard exist for this. -ncmoe N with default NO_PINNED settings produces the same 108 GiB pinned buffer — never run it without the guard.
Any -ot/-ncmoe override disables mmap for the whole model load (create_tensors returns false → ml.use_mmap = false): the "mmap-backed CPU buffer" comment in the serve script is wrong on the mmap word; the buffer is anonymous. All placement variants perform alike anyway.
--defer-experts does not keep experts off the GPU under -ngl 999 — only PLE defers (32.78 GiB). With ngl 999 the load dies with cudaMalloc failed: out of memory at ~79.7 GiB on device 0.
The systemd unit self-heals through manual tests: pkill the server and the unit respawns it ~4 s later (Restart=always) — two instances, port races, phantom measurements. For manual experiments: sudo systemctl stop frank-large.service first (it is a root unit; plain systemctl --user lines in older docs are stale), and restore with sudo systemctl start frank-large.service.
Verify "HTTP server listening" in the CURRENT log and pgrep -xc llama-server == 1 before trusting any measurement.
FA must stay ON — FA-off crashes the loader on this arch.
Do not run builds concurrently with decode measurements (-j16 starves the service and corrupts both results).

Rollback

Stop and disable the root unit, trash it, restore the halogen quadlet from ~/.AGENTTRASH/20260915-ik-swap/halogen-quadlet to ~/.config/containers/systemd/, systemctl --user daemon-reload, start; the .hgn checkpoint and the kingjones STRIX llama.cpp payload remain on disk.

7.2 KiB Raw Blame History