1Container start → load model → serve requests → container stop → 所有 RAM / 磁碟 / port 自動回收

1brew services start → load model on demand → serve → ??? → 你忘記 stop
2                                              → RAM / 磁碟一直被佔
3                                              → 下次重開機才釋放

1brew services start ollama          → daemon 啟動、沒 model 載入、RAM 占用 ~200 MB
2                                     port 11434 listening
3ollama run gemma3:4b "hello"        → 把 model 載入 RAM (~4-5 GB)
4                                     立刻 generate response
5                                     model 留在 RAM
6(idle 5 分鐘、無新 request)         → Ollama 自動 unload model
7                                     RAM 釋放、daemon 仍跑著
8ollama run gemma3:4b "next"         → 重新 load model（~5-10 秒）、generate
9brew services stop ollama           → daemon 結束、port 釋放

 1# 看當前 loaded models
 2ollama ps
 3# NAME         ID              SIZE      PROCESSOR    UNTIL
 4# gemma3:4b    a2af6cc3eb7f    5.5 GB    100% Metal   4 minutes from now
 5
 6# 啟動時調 keep_alive（持續佔 RAM 直到 ollama 重啟）
 7OLLAMA_KEEP_ALIVE=-1 brew services restart ollama
 8
 9# 啟動時讓 model 用完立即 unload
10OLLAMA_KEEP_ALIVE=0 brew services restart ollama

1# 把 idle 的 model 立刻從 RAM 卸掉、但 daemon 仍跑
2curl -s http://localhost:11434/api/generate \
3  -d '{"model": "gemma3:4b", "keep_alive": 0}'
4
5# 或關掉整個 daemon
6brew services stop ollama

 1python main.py                      → ComfyUI server start、port 8188 listening
 2                                     RAM ~3 GB（Python venv + 框架）
 3第一次 Queue Prompt (用 SDXL)        → 載入 sd_xl_base_1.0.safetensors (~6 GB)
 4                                     RAM 跳到 ~9-10 GB
 5                                     generate 完成、model 留在 RAM
 6連續多張生成                          → 維持 ~9-10 GB、沒 unload
 7idle 1 小時                          → 仍 ~9-10 GB（沒 timer）
 8切到 ControlNet workflow             → 多載 ControlNet model (~2 GB)、ComfyUI 自動 swap
 9                                     RAM 暫升、SD 部分可能被 evict 到 disk
10Ctrl+C / pkill                       → process 結束、RAM 完全釋放

 1# 找 PID
 2ps aux | grep "ComfyUI/main.py" | grep -v grep
 3
 4# 優雅關（讓它 cleanup）
 5pkill -INT -f "ComfyUI/main.py"
 6
 7# 強制 kill（如果上面沒反應、最多等 5 秒再強制）
 8pkill -KILL -f "ComfyUI/main.py"
 9
10# 確認 port 釋放
11lsof -i :8188 | head -3

 1# 1. 確認當前狀態（記下要還回去多少 RAM）
 2vm_stat | head -3
 3df -h ~ | tail -1
 4ollama ps
 5ps aux | grep -E "ollama|ComfyUI|llama-server" | grep -v grep
 6
 7# 2. 釋放當前載入的 LLM models（Ollama）
 8brew services stop ollama
 9# 或保留 daemon、只 unload model：
10# curl -s http://localhost:11434/api/generate -d '{"model": "<your model>", "keep_alive": 0}'
11
12# 3. 結束 ComfyUI / 其他 GUI server
13pkill -INT -f "ComfyUI/main.py" 2>/dev/null
14pkill -INT -f "llama-server" 2>/dev/null
15sleep 5
16# 強制（如果上面沒清乾淨）
17pkill -KILL -f "ComfyUI/main.py" 2>/dev/null
18pkill -KILL -f "llama-server" 2>/dev/null
19
20# 4. 驗證所有 port 釋放
21lsof -i :11434 -i :1234 -i :8080 -i :8188 -i :8000 2>&1 | head
22
23# 5. 確認釋放量
24vm_stat | head -3
25# free pages 該明顯增加

 1# Ollama models 總占用
 2du -sh ~/.ollama/models/blobs
 3# 4.1G    /Users/tarragon/.ollama/models/blobs
 4
 5# 逐 model 看大小
 6ollama list
 7# NAME                       ID              SIZE      MODIFIED
 8# gemma4:e4b                 c6eb396dbd59    9.6 GB    Less than a second ago
 9# nomic-embed-text:latest    0a109f422b47    274 MB    3 hours ago
10
11# ComfyUI checkpoints 累積
12du -sh ~/.ollama ~/Projects/ComfyUI/models 2>/dev/null
13# 4.2G    /Users/tarragon/.ollama
14# 7.0G    /Users/tarragon/Projects/ComfyUI/models

 1# 刪掉很久沒用的 model
 2ollama rm <model-tag>
 3
 4# 一次清掉所有 Ollama models（保留 daemon）
 5ollama list | tail -n +2 | awk '{print $1}' | xargs -I {} ollama rm {}
 6
 7# 看 ComfyUI checkpoints 哪些可清
 8ls -lh ~/Projects/ComfyUI/models/checkpoints/
 9
10# 手動刪不要的 .safetensors（小心、不能 undo）
11rm ~/Projects/ComfyUI/models/checkpoints/<old-model>.safetensors

 1# 找誰占
 2lsof -i :11434
 3# COMMAND  PID  USER   ...   NAME
 4# ollama   xxx  ...    ...   TCP localhost:11434 (LISTEN)
 5
 6# 看是不是 zombie process
 7ps aux | grep $(lsof -ti :11434 | head -1)
 8
 9# 清掉
10kill -9 $(lsof -ti :11434)
11
12# 或重啟 service（會自動清舊 instance）
13brew services restart ollama

 1# 健康檢查
 2curl -s http://localhost:11434/api/version
 3
 4# 沒回應、看 service 狀態
 5brew services list | grep ollama
 6
 7# 沒在跑、重啟
 8brew services start ollama
 9
10# 看 log
11tail -50 /opt/homebrew/var/log/ollama.log

1# 看 stdout / stderr log
2tail -30 /tmp/comfyui.log  # 如果啟動時 redirect 到 log
3
4# 看是不是 GPU / Metal stuck（極少見、但 SDXL 大量並發可能踩到）
5# 解法：kill + 重啟
6pkill -9 -f "ComfyUI/main.py"

 1# Baseline
 2vm_stat | grep "Pages free"
 3# Pages free:                              1090076.   ← ~17 GB free
 4
 5# 啟動 Ollama + load 4B model
 6brew services start ollama
 7ollama run gemma3:4b "hello"
 8ollama ps
 9# NAME       SIZE     PROCESSOR    UNTIL
10# gemma3:4b  5.5 GB   100% Metal   4 minutes from now
11
12vm_stat | grep "Pages free"
13# Pages free:                               750000.   ← 跌 ~5 GB（model 載入）
14
15# 額外啟動 ComfyUI + load SDXL
16nohup python main.py > /tmp/comfyui.log 2>&1 &
17# 在 GUI 上 Queue Prompt 跑一次 SDXL generation
18vm_stat | grep "Pages free"
19# Pages free:                               280000.   ← 再跌 ~7.5 GB（SDXL 載入 + Python venv）
20
21# kill 全部
22brew services stop ollama
23pkill -9 -f "ComfyUI/main.py"
24sleep 3
25vm_stat | grep "Pages free"
26# Pages free:                              1090000.   ← 回到 baseline

 1# 加進 ~/.zshrc
 2llm-cleanup() {
 3  echo "[*] Stopping Ollama..."
 4  brew services stop ollama 2>/dev/null
 5
 6  echo "[*] Killing ComfyUI..."
 7  pkill -INT -f "ComfyUI/main.py" 2>/dev/null
 8  sleep 3
 9  pkill -KILL -f "ComfyUI/main.py" 2>/dev/null
10
11  echo "[*] Killing other model servers..."
12  pkill -KILL -f "llama-server" 2>/dev/null
13  pkill -KILL -f "lm-studio-server" 2>/dev/null
14
15  echo "[*] Verifying ports..."
16  for p in 11434 1234 8080 8188 8000; do
17    lsof -i :$p 2>/dev/null | head -2
18  done
19
20  echo "[*] Free RAM:"
21  vm_stat | grep "Pages free"
22}

1Found 71 markdown files under content/llm
2  [10/71] 86 chunks in 4.5s
3  [20/71] 181 chunks in 8.6s
4  ...
5  [70/71] 461 chunks in 22.2s
6Wrote 463 records to scripts/rag-demo/index.pkl (22.3s)

1Loaded 463 chunks from scripts/rag-demo/index.pkl
2=== Retrieved chunks ===
3  0.870  llm/knowledge-cards/transformer.md#chunk2
4  ...
5（LLM 生成 response）

1# 看當前 RAM 占用
2ollama ps
3# NAME                       SIZE      UNTIL
4# nomic-embed-text:latest    274 MB    4 minutes from now
5# gemma3:4b                  5.5 GB    4 minutes from now

1# 看當前所有 MCP server
2ps aux | grep blog_mcp_server | grep -v grep
3
4# 如果 client crash 留下 zombie：
5pkill -f "blog_mcp_server.py"

1User 在 Claude Desktop 打字
2  ↓
3Claude Desktop (~200 MB)
4  ↓ MCP stdio
5blog_mcp_server.py (~50 MB)
6  ↓ HTTP /api/embeddings + /v1/chat/completions
7Ollama daemon (~200 MB)
8  ↓ load
9nomic-embed-text 模型 (~274 MB) + 主 chat model (~6 GB)

1# 看 Ollama 載了哪些 model
2ollama ps
3
4# 看所有 LLM-related process
5ps aux | grep -E "ollama|comfyui|mcp" | grep -v grep | awk '{print $4, $11, $12, $13}' | sort -rn
6
7# 系統整體
8vm_stat | head -3

1# Ollama models 累積
2du -sh ~/.ollama/models
3
4# RAG index 累積（多個 corpus）
5du -sh scripts/rag-demo/index*.pkl 2>/dev/null
6
7# ComfyUI checkpoints / VAE / LoRA / etc
8du -sh ~/Projects/ComfyUI/models/*

1# 一鍵 audit 所有 LLM service
2for p in 11434 1234 8080 8188 8000; do
3  echo "=== port $p ==="
4  lsof -i :$p 2>/dev/null | head -2
5done
6
7# 找 zombie subprocess（沒 parent 的 mcp server）
8ps aux | grep "mcp_server" | grep -v grep

1# 每月跑一次的 cleanup
2llm-rag-cleanup() {
3  echo "[*] Old indexes (>30 days):"
4  find scripts -name 'index*.pkl' -mtime +30 -ls
5  echo "[*] Ollama models (review):"
6  ollama list
7  echo "[*] Python caches:"
8  find ~/Projects -name __pycache__ -type d | head -10
9}

 1# 1. RAG ingest 階段 RAM 量
 2ollama ps  # 先看 baseline
 3python3 scripts/rag-demo/ingest.py &
 4INGEST_PID=$!
 5ollama ps  # 看 embedding model 載入後
 6vm_stat | head -3
 7wait $INGEST_PID
 8
 9# 2. RAG query 階段 RAM 量
10ollama ps  # 看 idle 後 unload
11python3 scripts/rag-demo/query.py --show-retrieved "test query"
12ollama ps  # 看 chat model 載入
13
14# 3. MCP server 階段 process / RAM
15python3 scripts/mcp-demo/test_client.py &
16CLIENT_PID=$!
17sleep 2
18ps aux | grep blog_mcp_server | grep -v grep
19wait $CLIENT_PID
20
21# 4. 完成釋放
22ollama list | tail -n +2 | awk '{print $1}' | xargs -I {} \
23  curl -s http://localhost:11434/api/generate -d "{\"model\":\"{}\",\"keep_alive\":0}"