Variet-Worker
e02626fda8
chore(session): pause work — Qwen promoted to primary + Hermes v0.8.0
...
Engine:
- default_role: fast → balanced (Qwen 3.5 35B primary)
- balanced: remove -t/-tb (no impact with -ngl 999)
Hermes Agent submodule:
- Update fff237e1 → e902e55b (v0.8.0, 340 commits merged)
- Local 8 file patches auto-merged (stash/pop, 0 conflicts)
- mmproj CPU offload, DISCORD_HOME_CHANNEL fix via external ~/.hermes/config.yaml
Decisions:
- Speculative decoding experiment rejected (+14% gen vs -31% cold start)
- llama.cpp kept at b8660 (b8757 has 9% Gemma 4 regression)
- Qwen superior for thinking/Korean/coding; speed diff negligible
Session handoff: .planning/HANDOFF.json
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com >
2026-04-11 18:14:41 +09:00
Variet-Worker
0dee779a73
refactor(phase-01): v3 retune fast & balanced roles
...
fast (Gemma 4 26B-A4B):
- Enable mmproj GPU loading (vision ~1s, 12x faster than CPU)
- KV f16 → q8_0 (save ~2.5 GB VRAM for mmproj)
- Tensor split 0.5,0.5 → 0.43,0.57 (13/17 layers)
- Remove --mlock/--poll/--prio/-t/-tb (no measurable impact)
- measured_tps 74.65 → 71.89 (trade 3.7% speed for vision)
balanced (Qwen 3.5 35B-A3B):
- Tensor split 0.5,0.5 → 0.48,0.52 (enables pipeline parallelism)
- Ubatch 128 → 256 (prefill +78%: 649 → 1,157 t/s)
- mmproj + --no-mmproj-offload (CPU vision, VRAM headroom)
- Remove useless flags same as fast
- measured_tps 61.62 → 64.16 (+4.1%)
Other:
- Document full retuning in docs/v3_{fast,balanced}_retuning_log.md
- Session report at .planning/reports/20260411-session-report.md
- Add bench utilities: bench_short/bench_long/test_ts_ratios
- Speculative decoding (E2B draft) experimented but rejected
(+14% gen vs -31% cold start + tokenizer mismatch + mmproj conflict)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com >
2026-04-11 14:55:27 +09:00
Variet-Worker
c111b3a9b0
feat: Variet Engine v1.0 + 5-model tuning complete
...
Phase 01 (LLM Tuning):
- Gemma4 26B: 74.65 t/s (fast)
- Qwen 35B: 61.62 t/s (balanced)
- Gemma4 31B: 16.0 t/s (deep-coder)
- Qwen 27B: 16.7 t/s (deep-logic)
- Qwen 122B: 8.95 t/s (ultra, GPU 1 only)
Phase 02 (API Engine):
- FastAPI reverse proxy on port 8000
- /engine/switch hot-swap with 503 protection
- config/engine_models.json as single source of truth
- Replaced 4 individual .bat files with unified engine
File cleanup:
- scripts/ 85 files -> 9 + _archive/
- Root .bat files -> _archive/
2026-04-07 18:08:58 +09:00
