Cap runaway chat generations in router

2026-08-26 07:01:04 +02:00
parent 400de54ff9
commit d1eed44341
3 changed files with 59 additions and 0 deletions
@@ -567,6 +567,10 @@ services:
      PROFILE_CONTROL_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
      SWITCH_TIMEOUT: "600"
      REQUEST_TIMEOUT: "600"
+      # Last-resort guard for every OpenAI-compatible client. Without a
+      # request limit llama.cpp uses n_predict=-1 and a reasoning loop can
+      # consume the complete context before yielding visible output.
+      MAX_GENERATION_TOKENS: "8192"
      IMAGE_DIR: /data/images
      IMAGE_WORKER_URL: http://flux-worker:8086
      IMAGE_WORKER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"