Coverage for src/lilbee/core/config/keys.py: 100%

1"""Config-field key sets shared by the settings boundary and the engine.""" 

2 

3from __future__ import annotations 

4 

5# API-key cfg field names: keep in sync with ``providers.sdk_backend.PROVIDER_KEYS``. 

6PROVIDER_API_KEYS: frozenset[str] = frozenset( 

7 { 

8 "llm_api_key", 

9 "openrouter_api_key", 

10 "gemini_api_key", 

11 "anthropic_api_key", 

12 "openai_api_key", 

13 "mistral_api_key", 

14 "deepseek_api_key", 

15 } 

16) 

17 

18# Settings whose value is baked into a loaded model and only takes effect 

19# after the worker reloads. 

20LOAD_AFFECTING_KEYS: frozenset[str] = frozenset( 

21 { 

22 "num_ctx", 

23 "num_ctx_max", 

24 "chat_n_ctx_target", 

25 "kv_cache_type", 

26 # Baked into the llama-server argv (--n-gpu-layers, --cpu-moe, --n-cpu-moe, 

27 # --flash-attn), so a change must rebuild the engine; the bind contract 

28 # (role/model + pin) would otherwise adopt a running engine's old offload, 

29 # cache, or flash-attention flags unchanged. 

30 "n_gpu_layers", 

31 "cpu_moe", 

32 "n_cpu_moe", 

33 "flash_attention", 

34 "chat_model", 

35 "embedding_model", 

36 "vision_model", 

37 "reranker_model", 

38 "reranker_type", 

39 # The embed/rerank window and the chat floor are sized from the chunk budget. 

40 "chunk_size", 

41 } 

42) 

43 

44# Writes here require reconstructing the services singleton. 

45PROVIDER_SWITCHING_KEYS: frozenset[str] = frozenset({"llm_provider"}) 

46 

47# Settings that change which physical devices an engine launches on, or which of 

48# them holds the model (--main-gpu). Part of the cross-process engine pin, so a 

49# process with a different placement binds its own engine rather than silently 

50# adopting the incumbent's GPUs. Kept out of LOAD_AFFECTING_KEYS so they do not 

51# change the settings reload path (they take effect on restart), only which 

52# engines are shareable. 

53PLACEMENT_PIN_KEYS: frozenset[str] = frozenset({"placement", "gpu_devices", "main_gpu"})