num_ctx makes Ollama reload the model at a different context size (an
unload+load). Against a resident qwen3:32b at ctx 32768, our request asked for
8192 — forcing a 29GB reload on every call, which stalls indefinitely whenever
the backend's eviction path is unhealthy.
A/B against the gateway, identical requests otherwise:
stream, no num_ctx -> 200 in 0.31s
stream, num_ctx=8192 -> http 000, hangs 60s+
It was also counterproductive: it shrank the context the model already served.
Ryan Malloy
3891dd3566