Skip to content

Environment Variables

Environment variables configure process-level defaults and library paths. Values are read during startup or module import, so changes take effect after a restart. Worker-side settings also need to be available on the relevant Ray nodes.

Variable Default Purpose
AEGAEON_WEB_DIR Unset Optional external UI asset directory
AEGAEON_LOG_FILE output.log Service log path
AEGAEON_TTFT_SLO 10 TTFT target in seconds
AEGAEON_TPOT_SLO 0.1 TPOT target in seconds
AEGAEON_DEFAULT_MODEL_REPO modelscope Model source: modelscope or huggingface
AEGAEON_PROFILE_ROOT Bundled profiles, then repository profiles/ Model/device performance profiles
AEGAEON_SHM_DIR /dev/shm Shared-memory cache directory
AEGAEON_RAY_ADDRESS RAY_ADDRESS / auto Ray cluster address; CLI overrides it
AEGAEON_RAY_TEMP_DIR Unset Local Ray temporary directory
AEGAEON_ARENA_FRACTION 0.875 QuickLoader GPU arena fraction
CUDA_VISIBLE_DEVICES Process environment Visible GPU set
AEGAEON_MAX_REQUESTS_PER_SECOND Unset Per-model requests/second, averaged over 60 seconds; includes the pending request
AEGAEON_MAX_TOKENS_PER_SECOND Unset Per-model generated tokens/second, averaged over 60 seconds; checked at request admission
AEGAEON_FOUNDRY_HOOK_PATH Unset Absolute path to libcuda_hook.so
AEGAEON_CUDA_COMPAT_PATH Probed paths CUDA compatibility library directory
LD_LIBRARY_PATH Process environment Dynamic-library search path
HTTP_PROXY, HTTPS_PROXY, NO_PROXY Process environment Download/network proxy configuration

Graph settings belong in YAML configuration; startup commands are in the CLI reference.