Configurable ollama keep_alive

Closes #1859

default/config.yaml

@@ -98,6 +98,13 @@ mistral:
   # Enables prefilling of the reply with the last assistant message in the prompt
   # CAUTION: The prefix is echoed into the completion. You may want to use regex to trim it out.
   enablePrefix: false
+# -- OLLAMA API CONFIGURATION --
+ollama:
+  # Controls how long the model will stay loaded into memory following the request
+  # * -1: Keep the model loaded indefinitely
+  # * 0: Unload the model immediately after the request
+  # * 5m: Keep the model loaded for 5 minutes after the request. Accepts duration strings (e.g. 5h30m40s)
+  keepAlive: -1
 # -- SERVER PLUGIN CONFIGURATION --
 enableServerPlugins: false
 # User session timeout *in seconds* (defaults to 24 hours).

src/endpoints/backends/text-completions.js

@@ -5,7 +5,7 @@ const Readable = require('stream').Readable;
 
 const { jsonParser } = require('../../express-common');
 const { TEXTGEN_TYPES, TOGETHERAI_KEYS, OLLAMA_KEYS, INFERMATICAI_KEYS, OPENROUTER_KEYS, VLLM_KEYS, DREAMGEN_KEYS, FEATHERLESS_KEYS } = require('../../constants');
-const { forwardFetchResponse, trimV1 } = require('../../util');
+const { forwardFetchResponse, trimV1, getConfigValue } = require('../../util');
 const { setAdditionalHeaders } = require('../../additional-headers');
 
 const router = express.Router();
@@ -325,11 +325,12 @@ router.post('/generate', jsonParser, async function (request, response) {
         }
 
         if (request.body.api_type === TEXTGEN_TYPES.OLLAMA) {
+            const keepAlive = getConfigValue('ollama.keepAlive', -1);
             args.body = JSON.stringify({
                 model: request.body.model,
                 prompt: request.body.prompt,
                 stream: request.body.stream ?? false,
-                keep_alive: -1,
+                keep_alive: keepAlive,
                 raw: true,
                 options: _.pickBy(request.body, (_, key) => OLLAMA_KEYS.includes(key)),
             });