__init__.py 2.2 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243
  1. """LM Studio provider profile (local desktop app with an OpenAI-compatible server).
  2. Request quirk: reasoning goes out as top-level ``reasoning_effort``, gated on and clamped to
  3. the per-model ``allowed_options`` LM Studio publishes in ``/api/v1/models``. The agent probes
  4. and caches those options and passes them as ``lmstudio_reasoning_options``.
  5. Everything else stays where it lives today: endpoint, ``LM_API_KEY`` / ``LM_BASE_URL`` and the
  6. no-auth placeholder (``hermes_cli/auth.py`` registry row), the chat-model picker probe
  7. (``hermes_cli/models_local.py``). Hence ``base_url`` is left empty: a loopback ``base_url``
  8. would map every ``127.0.0.1`` / ``localhost`` endpoint to LM Studio in URL->provider inference
  9. and disable local-server detection for Ollama, llama.cpp and vLLM.
  10. """
  11. from typing import Any
  12. from agent.lmstudio_reasoning import resolve_lmstudio_effort
  13. from agent.reasoning_effort import generic_nested_reasoning
  14. from providers import register_provider
  15. from providers.base import ProviderProfile
  16. class LMStudioProfile(ProviderProfile):
  17. """LM Studio: top-level ``reasoning_effort`` from the model's published options."""
  18. def build_api_kwargs_extras(
  19. self, *, reasoning_config: dict | None = None, supports_reasoning: bool = False,
  20. lmstudio_reasoning_options: list[str] | None = None, **context: Any,
  21. ) -> tuple[dict[str, Any], dict[str, Any]]:
  22. """``supports_reasoning`` is true only when the model publishes a non-``off`` option;
  23. ``lmstudio_reasoning_options`` is the agent's cached probe of them. Auxiliary calls have no
  24. probe (``None``) and keep the generic nested ``reasoning`` they always sent."""
  25. if lmstudio_reasoning_options is None:
  26. return generic_nested_reasoning(reasoning_config), {}
  27. if not supports_reasoning:
  28. return {}, {}
  29. effort = resolve_lmstudio_effort(reasoning_config, lmstudio_reasoning_options)
  30. return {}, ({"reasoning_effort": effort} if effort is not None else {})
  31. register_provider(LMStudioProfile(
  32. name="lmstudio", display_name="LM Studio", supports_model_listing=False, supports_health_check=False,
  33. env_vars=("LM_API_KEY", "LM_BASE_URL"),
  34. ))