Release the Ollama model after a run instead of holding it resident

[tmartin2113]

Problem

Ollama keeps a model resident after a request for OLLAMA_KEEP_ALIVE — five minutes by default, indefinitely when set to -1. MiroFish's models are large (granite4:small-h is ~19 GB), so on a shared Ollama instance a single run can keep every other model out of memory long after the work has finished.

On my box the report model was evicting the GPU-resident model used by everything else, and staying loaded until something forced it out.

Why the obvious fix doesn't work

Passing keep_alive through the OpenAI client does nothing. Ollama's OpenAI-compatible endpoint accepts the field and silently ignores it — no error, no warning — while the native API honours it. Verified on Ollama 0.32.1 with qwen3:0.6b, watching the UNTIL column of ollama ps:

Request	UNTIL
POST /v1/chat/completions {"keep_alive": "30s"}	Forever — ignored
POST /api/chat {"keep_alive": "30s"}	29 seconds from now

Worth knowing independently of this PR: a keep_alive added to extra_body would look correct in review and have no effect.

Change

LLMClient.unload() posts keep_alive=0 to the native /api/generate (what ollama stop <model> does), deriving the native root by stripping the /v1 suffix from the configured base URL.

It's called from a finally at the three places that drive the LLM, so a failed run frees the model too:

• report generation — api/report.py
• prepare_simulation — services/simulation_manager.py, which covers both the profile and config generators, they being phases 2 and 3 of one call
• standalone profile generation — api/simulation.py

Safety

• No-op for non-Ollama providers — guarded by the existing _is_ollama().
• Never raises. Failing to free memory must not fail the caller, so every path returns a bool and swallows exceptions.
• Opt-out: LLM_UNLOAD_AFTER_USE=0 restores the current behaviour exactly.

Testing

Against a live Ollama, exercising the committed source directly:

• a resident model with 9 minutes left on its keep-alive is released immediately, confirming the /v1 → native URL rewrite
• non-Ollama base URL → False, no request sent
• LLM_UNLOAD_AFTER_USE=0 → False, no request sent
• unreachable Ollama host → False, no exception escapes

Also run end to end on a real 2-section report and a generate-profiles call: model released after each, and generate-profiles still returned its profiles normally.

No tests are included since the repo has no suite to extend — happy to add one if you'd like a particular shape.

Note on the default

This defaults to on, which is a behaviour change: the next run reloads the model. For a single dedicated Ollama that's a small cost; for a shared one it's the whole point. Happy to flip it to opt-in if you'd rather not change existing behaviour — it's one word in the env lookup.

🤖 Generated with Claude Code