vibehacker
Discuss
Nova Hale
1 day ago

ollama serve died mid-eval and left a zombie on port 11434

running mistral-small on an M2 Air for a cheap eval loop. hit ~4k tokens into a batch and the whole thing just... vanished. no crash log. Activity Monitor still shows ollama at 0% CPU hugging the port.

killed it, restarted, same port stuck for 10 minutes until I rebooted. sticky note of working num_ctx values did nothing this time.

anyone else getting silent deaths on Metal lately or am I the only one who still trusts local for CI-adjacent work?

5 comments

Join the discussion

Log in to comment.

  • Owen

    happened twice last week on my denver machine with llama.cpp + mistral. the process stays registered but Metal already ate the context. i stopped trusting ollama serve for overnight batches — wrap it and check /api/tags every 30s, restart if it hangs.

    also if you're mid-eval and it invents field names after a silent restart, treat the JSON like untrusted input. learned that the hard way.

  • Tara Nguyen

    this is why I cancelled my "run everything local" experiment after two weeks. $20/mo Claude API is boring but it doesn't zombie a port and waste a Friday night reboot.

    if the eval is CI-adjacent and you need sleep, cloud. keep local for the fun weekend tinkering where a silent death is just a story, not a missed ship.

    • Devon

      disagree on the $20/mo escape hatch for CI-adjacent. cloud evals drift between model versions overnight and you dont notice until monday. local zombies are annoying; silent quality drift is expensive.

      keep local, just stop treating ollama like a daemon you can ignore.

  • Chris Liu

    M2 Air + Metal + overnight batch is a known bad combo. I moved eval loops to a spare 3090 box running llama.cpp server — ~38 tok/s on mistral-small, and when it dies the systemd unit actually restarts.

    Ollama's 0% CPU zombie is the worst failure mode. At least a crash leaves a log. If you're still on laptop Metal for CI-adjacent work, wrap it or expect Friday reboots.

  • Remy

    same on my lyon box with dockerized ollama. i put a hard timeout of 90s around every /api/generate call now. if it hangs, kill -9 the container and bring it back. ugly but the silent port hold is worse.

    also dont trust Activity Monitor alone — check lsof -i :11434. sometimes the process is gone and the port is still stuck.

More like this

View all