docs: initialize wiki with session report and Phase 01 summary

Variet Agent
2026-04-11 18:16:53 +09:00
parent 8fa9c549b1
commit d63b1ab84c
2 changed files with 124 additions and 0 deletions

57
Phase-01-LLM-Tuning.md Normal file

@@ -0,0 +1,57 @@
# Phase 01 — LLM Tuning
## 5 모델 확정 성능 (v3 재튜닝 포함)
| # | 모델 | 역할 | 측정 t/s | 컨텍스트 | 비고 |
|:-:|------|------|:--------:|:-------:|------|
| 1 | Gemma 4 26B-A4B Q4_K_M | fast | 71.89 | 256K | Vision GPU (~1s) |
| 2 | Qwen 3.5 35B-A3B Q4_K_M | **balanced (primary)** | 64.16 | 256K | Vision CPU (6.4s), thinking mode |
| 3 | Gemma 4 31B Dense Q4_K_M | deep-coder | 16.0 | 192K | 코딩 특화 |
| 4 | Qwen 3.5 27B Dense Q4_K_M | deep-logic | 16.7 | 256K | 깊은 추론 |
| 5 | Qwen 3.5 122B-A10B Q4_K_M | ultra | 8.95 | 256K | GPU 1 단독 |
## Hardware Asymmetry
```
GPU 0 PCIe 3.0 x4 (3.94 GB/s) ← bottleneck
GPU 1 PCIe 4.0 x16 (31.5 GB/s)
Total VRAM: 24 GB
```
GPU 0의 PCIe x4 제약이 이중 GPU 추론 속도 상한을 결정합니다.
## v3 재튜닝 변경사항 (2026-04-11)
### balanced (Qwen 3.5 35B-A3B)
- `-ub 128 → 256` (prefill +78%: 649 → 1,157 t/s)
- `-ts 0.5,0.5 → 0.48,0.52` (PP 활성화)
- `--mmproj models/mmproj-F16.gguf` + `--no-mmproj-offload` (비전 CPU)
- `--mlock/--poll/--prio/-t/-tb` 제거
### fast (Gemma 4 26B-A4B)
- `cache-type: f16 → q8_0` (VRAM 2.5 GB 절약)
- `-ts 0.43,0.57` (13/17 layer split)
- `--mmproj models/gemma-4-26B-mmproj-F16.gguf` (Vision GPU)
- `--mlock/--poll/--prio/-t/-tb` 제거
## 아키텍처 발견
### Qwen 3.5 35B-A3B (Gated Delta Net)
- 40 레이어 중 **10개만 full attention** (매 4번째)
- 나머지 30개는 SSM (Mamba-like) 레이어
- KV 캐시 실제 크기: 1.4 GB (q4_0, 256K) — 초기 추정(5GB) 대비 작음
### Gemma 4 26B-A4B
- 30 레이어 중 **5개만 full attention** (매 6번째)
- 나머지 25개는 SWA (Sliding Window, 1,536 cells)
- 128 experts / 8 active (MoE)
## Reference
- [docs/v3_balanced_retuning_log.md](../../docs/v3_balanced_retuning_log.md)
- [docs/v3_fast_retuning_log.md](../../docs/v3_fast_retuning_log.md)
- [.planning/phases/01-llm-tuning/VERIFICATION.md](../../.planning/phases/01-llm-tuning/VERIFICATION.md)
## Status
**Phase 01**: ✅ COMPLETE (2026-04-07 초기, 2026-04-11 재검증)

67
Session-2026-04-11.md Normal file

@@ -0,0 +1,67 @@
# Session 2026-04-11 — Phase 01 Retuning & Hermes Agent v0.8.0
**Type**: Maintenance
**Focus**: LLM tuning re-verification, Hermes Agent update, primary model switch
---
## Summary
Phase 01 (LLM Tuning) 재검증 세션과 Hermes Agent 주요 업데이트를 병행하여 진행. 두 LLM 역할 최적화 파라미터 갱신 및 Qwen 3.5 35B를 메인 모델로 승격, Hermes Agent v0.7.x → v0.8.0 업데이트 완료.
## Key Outcomes
### LLM Engine
- **balanced** (Qwen 3.5 35B-A3B): 61.62 → **64.16 t/s** (+4.1%)
- **fast** (Gemma 4 26B-A4B): 74.65 → **71.89 t/s** (Vision GPU 추가 trade-off)
- Default role: `fast``balanced`
- 두 역할 모두 `--mlock/--poll/--prio/-t/-tb` 제거
### Qwen Primary Promotion
- 속도 차이 1.25 t/s (negligible)
- 35B > 26B 품질 우위
- Thinking mode + 한국어/코딩 강점
- Vision CPU 오프로드 수용 (6.4s/image)
### Hermes Agent
- 버전: `fff237e1``e902e55b` (340 commits, v0.8.0)
- 로컬 8개 패치 자동 병합 (0 conflict)
- 설정: `custom/qwen3.5-35b-a3b` + `DISCORD_HOME_CHANNEL` 수동 설정
## Experiments (Not Adopted)
- **Speculative Decoding** (E2B draft): +14% gen vs -31% cold start → 채택 안 함
- **llama.cpp b8757**: Gemma 4 9% 회귀 → b8660 유지
## Files Changed
- `config/engine_models.json`
- `docs/v3_balanced_retuning_log.md` (new)
- `docs/v3_fast_retuning_log.md` (new)
- `.planning/reports/20260411-session-report.md` (new)
- `.planning/phases/01-llm-tuning/VERIFICATION.md` (updated)
- `.planning/STATE.md` (updated)
- `.planning/HANDOFF.json` (updated)
- `agents/hermes-agent` (submodule bump)
- `scripts/bench_short.py`, `bench_long.py`, `test_ts_ratios.py` (new utilities)
## Git History
```
e02626f chore(session): pause work — Qwen promoted to primary + Hermes v0.8.0
0dee779 refactor(phase-01): v3 retune fast & balanced roles
```
## Hardware Constraints Documented
- GPU 0: PCIe 3.0 x4 (3.94 GB/s) — bottleneck
- GPU 1: PCIe 4.0 x16 (31.5 GB/s)
- Total VRAM: 24 GB (2x RTX 3060 12GB)
## Next Session
- Run `/gsd-resume-work` to reload state from `HANDOFF.json`
- Options:
- Start Hermes Agent (`run_hermes_agent.bat`)
- Resume Phase 05 (VS Code Extension Packaging)
- New feature development