- Shipped
- September 28, 2026 at 2:09 PM UTC
- Author
- Kamo
- Commit
- f4d979d
llama-server's host-RAM prompt cache defaults to 8192 MiB (--cache-ram), the container's whole limit, so it grew until the kernel killed ollama: on 2026-09-28 it held 14 prompts in 3 548 MiB on a ~4.4 GiB model and the 15th save was OOMKilled (8 restarts in 9.5 h, a 502 to every call in flight). LLAMA_ARG_CACHE_RAM=1024 keeps about four prompts (agent steps share ~340 tokens of prefix, so nothing is lost); LLAMA_ARG_CTX_CHECKPOINTS=8 bounds the per-slot checkpoints (2-3 per prompt at ~55 MiB) that the default 32 left at 1.7 GiB. Ollama starts llama-server with neither flag, and the env reaches it.