Record: llama.cpp rebuild on beast (build 10790) + Qwen3.8-Flash-Next (qwen38) CPU preset — 2026-09-02

llama.cpp rebuild (beast)

• Rebuilt and installed 2026-09-02 21:37–21:40: build 10790, commit 7cbe657e3 (previous: 10726 / 63f45bd2c). Full library set replaced (/usr/local/lib/libllama.so.0.3.0, libllama-server-impl.so, libggml-*.so.0.22.0, …) plus /usr/local/bin/llama-server.
• llama.service restarted 21:44 (MainPID 745958) to load the new build; healthy, 14 models registered, 0 kernel resets.
• qwen4exp (Qwen3.8-Flash-Next) support present: strings /usr/local/lib/libllama.so.0.3.0 | grep -c qwen4exp → 93. (Earlier “no support” readings in records/2026-09-02-01 / chat were from the 17 KB launcher binary and libggml-base only — the arch table lives in libllama.so, which was not grepped.)
• llama.cpp mainline context: PR ggml-org/llama.cpp #27742 “model: add Qwen3.8-Flash-Next (qwen4exp)” merged 2026-08-27; follow-up fix PR #27774 (“rotated-KV QSA fix”) still open — not in this build.
• Athena untouched (still build 10712 / commit f7807f27d, installed 2026-08-30) — revisit if athena should also gain qwen4exp.

Qwen3.8-Flash-Next CPU preset (qwen38) added to beast

Added to /usr/local/etc/config.ini (baseline: config.ini.bak-20260902-qwen38):

[batiai/Qwen3.8-Flash-Next-GGUF:Q4_K_M]
    alias = qwen38
    ngl = 0
    numa = distribute
    ctx-size = 32768
    temp = 0.6
    top-p = 0.95
    jinja = true

• Repo/quant choice: batiai/Qwen3.8-Flash-Next-GGUF:Q4_K_M (root-level, 119.2 GB over 3 splits) over unsloth/Qwen3.8-Flash-Next-GGUF UD-quants because unsloth keeps its files in per-quant subdirectories and llama.cpp’s --hf-repo resolution is unverified for subdir layouts; the batiai root layout mirrors every existing working preset.
• Model facts: MoE ~6B active; total reported 125–176B across sources; native 256K ctx; multimodal (mmproj files published); day-0 GGUFs from unsloth/batiai/AtomicChat/argyelan; MTP draft-head GGUFs exist for speculative decoding.
• Registered in /v1/models (status unloaded; beast now lists 14 models). Not yet downloaded/loaded: the first qwen38 request will download ~119 GB (239 GB free on /home), then load on CPU — that load is the final end-to-end qwen4exp proof, still pending.
• RAM note: ~119 GB resident when loaded — avoid co-loading with gpt-oss-120b (~59 GB) on the 256 GB box; the router keeps loaded workers resident.

Notes

• Service restarts are required after a llama.cpp install or config edit: the router snapshots both the binary and config.ini at llama.service start.
• If unsloth’s subdir layout is ever needed, confirm llama.cpp resolution first (cheap failure = no file matched; expensive success = starts the multi-GB download).