__init__.py 2.4 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950
  1. """Nebius Token Factory provider profile."""
  2. from typing import Any
  3. from agent.reasoning_effort import NEBIUS_EFFORTS, clamp_effort
  4. from providers import register_provider
  5. from providers.base import ProviderProfile
  6. # Conservative allowlist of model families that expose reasoning effort.
  7. _REASONING_MARKERS = (
  8. "deepseek-r1", "deepseek-v4", "deepseek-reasoner", "gpt-oss", "glm-5", "kimi-k2", "minimax-m2", "qwen3",
  9. )
  10. class NebiusTokenFactoryProfile(ProviderProfile):
  11. """Nebius Token Factory - top-level reasoning_effort."""
  12. def build_api_kwargs_extras(
  13. self, *, reasoning_config: dict | None = None, model: str | None = None,
  14. supports_reasoning: bool = False, **context: Any,
  15. ) -> tuple[dict[str, Any], dict[str, Any]]:
  16. model_name = (model or "").strip().rsplit("/", 1)[-1].lower()
  17. if not supports_reasoning and not any(marker in model_name for marker in _REASONING_MARKERS):
  18. return {}, {}
  19. rc = reasoning_config if isinstance(reasoning_config, dict) else {}
  20. # Unset/blank effort defaults to medium (reasoning ON).
  21. effort = str(rc.get("effort", "medium") or "medium").strip().lower()
  22. if rc.get("enabled", True) is False or effort in {"none", "off", "disabled"}:
  23. return {}, {}
  24. # Canonical clamp: nearest weaker supported level, never escalate.
  25. return {}, {"reasoning_effort": clamp_effort(effort, NEBIUS_EFFORTS) or "medium"}
  26. nebius_token_factory = NebiusTokenFactoryProfile(
  27. name="nebius-token-factory",
  28. aliases=("nebius", "nebius-tokenfactory", "nebius-tf", "token-factory", "tokenfactory"),
  29. display_name="Nebius Token Factory", description="Nebius Token Factory — OpenAI-compatible inference",
  30. signup_url="https://tokenfactory.nebius.com/",
  31. env_vars=("NEBIUS_API_KEY", "NEBIUS_TOKEN_FACTORY_API_KEY", "NEBIUS_BASE_URL"),
  32. base_url="https://api.tokenfactory.nebius.com/v1",
  33. models_url="https://api.tokenfactory.nebius.com/v1/models?verbose=true", auth_type="api_key",
  34. default_aux_model="nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B",
  35. fallback_models=(
  36. "Qwen/Qwen3.5-397B-A17B-fast", "deepseek-ai/DeepSeek-V4-Pro", "zai-org/GLM-5.1", "moonshotai/Kimi-K2.5-fast",
  37. "MiniMaxAI/MiniMax-M2.5-fast", "deepseek-ai/DeepSeek-V3.2-fast", "NousResearch/Hermes-4-70B",
  38. "openai/gpt-oss-120b-fast", "meta-llama/Llama-3.3-70B-Instruct",
  39. ),
  40. )
  41. register_provider(nebius_token_factory)