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>
This commit is contained in:
Variet-Worker
@@ -1,32 +1,93 @@
|
||||
{
|
||||
"version": "1.0",
|
||||
"timestamp": "2026-04-09T21:54:17+09:00",
|
||||
"phase": "06",
|
||||
"phase_name": "install-and-evaluate-hermes-agent",
|
||||
"phase_dir": ".planning/phases/06-install-and-evaluate-hermes-agent",
|
||||
"timestamp": "2026-04-11T16:30:00+09:00",
|
||||
"session_type": "maintenance",
|
||||
"phase": "01",
|
||||
"phase_name": "llm-tuning",
|
||||
"phase_dir": ".planning/phases/01-llm-tuning",
|
||||
"plan": 1,
|
||||
"task": 5,
|
||||
"total_tasks": 5,
|
||||
"task": null,
|
||||
"total_tasks": null,
|
||||
"status": "paused",
|
||||
"completed_tasks": [
|
||||
{"id": 1, "name": "Hermes Agent Repository Clone & Config", "status": "done", "commit": "none"},
|
||||
{"id": 2, "name": "Variet Engine 로컬 연결 및 .env 세팅", "status": "done", "commit": "none"},
|
||||
{"id": 3, "name": "윈도우 플랫폼 버그 패치", "status": "done", "commit": "none"},
|
||||
{"id": 4, "name": "GSD Phase 06 완료 처리", "status": "done", "commit": "none"},
|
||||
{"id": 5, "name": "Gemma-4-A4B 무한루프 및 타임아웃 방어 옵션 주입", "status": "done", "commit": "none"}
|
||||
],
|
||||
"remaining_tasks": [],
|
||||
"blockers": [],
|
||||
"human_actions_pending": [
|
||||
{"action": "현재 실행중인 콘솔에서 run_hermes_agent.bat 를 재실행하여 파이썬 변경 사항 메모리 적재하기", "context": "방금 전 코드 패치가 적용되려면 반드시 새로 구동해야 합니다", "blocking": true}
|
||||
"summary": "Maintenance session — Phase 01 LLM tuning re-verification + Hermes Agent v0.8.0 update + Qwen promoted to primary role",
|
||||
"completed_work": [
|
||||
{"id": 1, "name": "Balanced role (Qwen 3.5 35B-A3B) v3 retuning", "status": "done", "commit": "0dee779"},
|
||||
{"id": 2, "name": "Fast role (Gemma 4 26B-A4B) v3 retuning + mmproj GPU", "status": "done", "commit": "0dee779"},
|
||||
{"id": 3, "name": "Speculative decoding experiment (E2B draft) + rejection decision", "status": "done", "commit": "0dee779"},
|
||||
{"id": 4, "name": "Benchmark utilities (bench_short/bench_long/test_ts_ratios)", "status": "done", "commit": "0dee779"},
|
||||
{"id": 5, "name": "E2B model files removed (3.8 GB disk recovery)", "status": "done", "commit": "0dee779"},
|
||||
{"id": 6, "name": "llama.cpp b8660 -> b8757 update + regression analysis", "status": "done", "commit": "pending"},
|
||||
{"id": 7, "name": "llama.cpp b8660 retained (Gemma 4 9% regression vs bug fixes)", "status": "done", "commit": "pending"},
|
||||
{"id": 8, "name": "Qwen 3.5 35B promoted to primary role (default_role=balanced)", "status": "done", "commit": "pending"},
|
||||
{"id": 9, "name": "Hermes Agent update fff237e1 -> e902e55b (v0.8.0, 340 commits)", "status": "done", "commit": "pending"},
|
||||
{"id": 10, "name": "Local Hermes patches preserved via stash/pop auto-merge (8 files)", "status": "done", "commit": "pending"},
|
||||
{"id": 11, "name": "Hermes config.yaml updated: Qwen default + DISCORD_HOME_CHANNEL fix", "status": "done", "commit": "pending"}
|
||||
],
|
||||
"decisions": [
|
||||
{"decision": "추론형(Reasoning) 모델의 SWA 관련 무한 루프 늪을 방지하기 위해 run_agent.py에 min_p(0.05)와 repeat_penalty(1.05) 직접 강제 주입", "rationale": "가장 확실하고 공식적인 로컬 커뮤니티 국룰 세팅으로, 온도를 높이는 것보다 환각이 덜 발생", "phase": "06"}
|
||||
{
|
||||
"decision": "Qwen 3.5 35B (balanced) promoted to primary role",
|
||||
"rationale": "Speed differential vs Gemma 4 26B is 1.25 t/s (negligible). Qwen larger (35B vs 26B), thinking mode, better Korean/coding, architecture mature. Vision CPU offload (6.4s/image) acceptable for Hermes Agent text-dominant workload.",
|
||||
"phase": "01"
|
||||
},
|
||||
{
|
||||
"decision": "llama.cpp kept at b8660 (rolled back from b8757)",
|
||||
"rationale": "b8757 has Gemma 4 correctness fixes but introduces ~9% speed regression on Gemma 4. b8757 binary retained in llama_bin_run.backup_b8660/ for rollback. Re-evaluate on next llama.cpp release.",
|
||||
"phase": "01"
|
||||
},
|
||||
{
|
||||
"decision": "Speculative Decoding rejected",
|
||||
"rationale": "+14% avg gen speedup vs -31% cold start penalty + tokenizer mismatch (E2B vs 26B) + mmproj conflict + 5+ additional flags complexity.",
|
||||
"phase": "01"
|
||||
},
|
||||
{
|
||||
"decision": "KV cache q8_0 for fast, q4_0 for balanced",
|
||||
"rationale": "fast needs mmproj GPU capacity (q8_0 saves 2.5 GB). balanced uses q4_0 for maximum VRAM headroom.",
|
||||
"phase": "01"
|
||||
},
|
||||
{
|
||||
"decision": "Removed --mlock/--poll/--prio/-t/-tb from all roles",
|
||||
"rationale": "Measured impact 0.04 t/s. Dedicated inference machine has no CPU contention.",
|
||||
"phase": "01"
|
||||
}
|
||||
],
|
||||
"performance_baselines": {
|
||||
"balanced_qwen3.5_35b": {
|
||||
"gen_short_tps": 64.16,
|
||||
"prefill_long_tps": 1157,
|
||||
"gen_long_tps": 62.00,
|
||||
"vram_free_mb": 3246,
|
||||
"mmproj_mode": "CPU",
|
||||
"vision_encode_sec": 6.4
|
||||
},
|
||||
"fast_gemma4_26b": {
|
||||
"gen_short_tps": 71.89,
|
||||
"prefill_long_tps": 1672,
|
||||
"gen_long_tps": 66.67,
|
||||
"vram_free_mb": 2073,
|
||||
"mmproj_mode": "GPU",
|
||||
"vision_encode_sec": 1.0
|
||||
}
|
||||
},
|
||||
"hardware_constraints": {
|
||||
"gpu0_pcie": "Gen3 x4 (3.94 GB/s)",
|
||||
"gpu1_pcie": "Gen4 x16 (31.5 GB/s)",
|
||||
"asymmetric_bottleneck": "GPU0 PCIe x4 limits inter-GPU transfers",
|
||||
"total_vram_mb": 24576
|
||||
},
|
||||
"blockers": [],
|
||||
"human_actions_pending": [
|
||||
{
|
||||
"action": "Run run_hermes_agent.bat to start Hermes Agent",
|
||||
"context": "All config ready. Qwen 3.5 35B loaded, DISCORD_HOME_CHANNEL set, v0.8.0 active.",
|
||||
"blocking": false
|
||||
}
|
||||
],
|
||||
"uncommitted_files": [
|
||||
"agents/hermes-agent/run_agent.py",
|
||||
"agents/hermes-agent/agent/auxiliary_client.py"
|
||||
"config/engine_models.json",
|
||||
"agents/hermes-agent (submodule)",
|
||||
"~/.hermes/config.yaml (outside repo)",
|
||||
"agents/hermes-agent/cli-config.yaml"
|
||||
],
|
||||
"next_action": "에이전트 재시작 후 봇과 정상 대화 검증하기",
|
||||
"context_notes": "모델의 30분 무한루프를 해결하고 휴식에 들어갑니다. 해결 완료 상태이며 사용자님의 로컬 256K A4B 모델은 성능에 문제가 없음을 확인했습니다."
|
||||
"next_action": "Commit remaining changes + git push + start Hermes Agent",
|
||||
"context_notes": "Maintenance session for Phase 01 and Hermes Agent (Phase 06 outcome). Phase 05 VS Code Extension Packaging still planning. Next real development session determined by user priority."
|
||||
}
|
||||
|
||||
Submodule agents/hermes-agent updated: fff237e111...e902e55b26
@@ -1,6 +1,6 @@
|
||||
{
|
||||
"version": "1.0",
|
||||
"default_role": "fast",
|
||||
"default_role": "balanced",
|
||||
"llama_server": {
|
||||
"path": "llama_bin_run/llama-server.exe",
|
||||
"internal_port": 8080,
|
||||
@@ -62,10 +62,6 @@
|
||||
"256",
|
||||
"-b",
|
||||
"512",
|
||||
"-t",
|
||||
"6",
|
||||
"-tb",
|
||||
"6",
|
||||
"-ts",
|
||||
"0.48,0.52"
|
||||
]
|
||||
|
||||
Reference in New Issue
Block a user