diff options
| author | Henry J Webster <hwebs@hwebs.info> | 2026-08-04 00:02:54 -0500 |
|---|---|---|
| committer | Henry J. Webster <hwebs@hwebs.info> | 2026-08-04 16:05:20 -0500 |
| commit | df960b2ed92f9e45fc4984d19ab65c55074bb1d7 (patch) | |
| tree | 702ad999126a55f93122277c986a1bd6974a47be | |
| parent | 8f9fbf2f4a269eae143c9e279afc6608ce64b7d0 (diff) | |
kusanagi/ollama: double default context, add 1h keep-alive
The 12B model at 32k context uses only 7.7 GB of the 7900 XT's 20 GB, so
there is ample VRAM headroom for 64k with the q8_0 KV cache — agentic
coding tools are context-hungry and 32k is the practical bottleneck.
The keep-alive stops the default 5-minute idle unload from adding a
reload stall to every resumed session on a single-user workstation.
Update flake to get newer packages.
Assisted-by: Claude Code:claude-fable-5
| -rw-r--r-- | flake.lock | 6 | ||||
| -rw-r--r-- | machines/kusanagi/default.nix | 5 |
2 files changed, 7 insertions, 4 deletions
@@ -43,11 +43,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1784280462, - "narHash": "sha256-DtoqIqM7VkR6NxAkcLpMwmi02USwWb3JdmNGLyhthc0=", + "lastModified": 1785734586, + "narHash": "sha256-ODZkEK9Gy50yg6h98u7KkitZ3oc/uuTFK00bh1CRdNA=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "293d6abedf0478e681a4dfcfcb35b30fc796a32f", + "rev": "531670d871c0e29724a02f3cbcac170adc65b58c", "type": "github" }, "original": { diff --git a/machines/kusanagi/default.nix b/machines/kusanagi/default.nix index ff786f6..1c8ccb1 100644 --- a/machines/kusanagi/default.nix +++ b/machines/kusanagi/default.nix @@ -228,8 +228,11 @@ # KV-cache quantization (below) is a no-op without flash attention; ollama # falls back to f16 KV cache, inflating VRAM use and forcing CPU offload. OLLAMA_FLASH_ATTENTION = "1"; - OLLAMA_CONTEXT_LENGTH = "32768"; + OLLAMA_CONTEXT_LENGTH = "65536"; OLLAMA_KV_CACHE_TYPE = "q8_0"; + # Keep the model resident between agent turns; the default 5-minute + # keep-alive forces a multi-second reload after any pause. + OLLAMA_KEEP_ALIVE = "1h"; }; }; |