__init__.py 1.4 KB

12345678910111213141516171819202122232425262728293031323334
  1. """NVIDIA NIM provider profile."""
  2. from typing import Any
  3. from providers import register_provider
  4. from providers.base import ProviderProfile
  5. class NvidiaProviderProfile(ProviderProfile):
  6. """NVIDIA NIM accepts a stricter ToolMessage schema than most OpenAI-compatible APIs."""
  7. @staticmethod
  8. def _needs_strip(msg: Any) -> bool:
  9. return isinstance(msg, dict) and msg.get("role") == "tool" and ("name" in msg or "tool_name" in msg)
  10. def prepare_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
  11. """Copy-on-write: only tool messages that lose a field are copied
  12. (no deep copy of large tool outputs); untouched input returned as-is."""
  13. if not any(self._needs_strip(msg) for msg in messages):
  14. return messages
  15. return [
  16. {k: v for k, v in msg.items() if k not in ("name", "tool_name")} if self._needs_strip(msg) else msg
  17. for msg in messages
  18. ]
  19. nvidia = NvidiaProviderProfile(
  20. name="nvidia", aliases=("nvidia-nim", "nim", "build-nvidia", "nemotron"), env_vars=("NVIDIA_API_KEY",), display_name="NVIDIA NIM",
  21. description="NVIDIA NIM — accelerated inference", signup_url="https://build.nvidia.com/",
  22. fallback_models=("nvidia/llama-3.1-nemotron-70b-instruct", "nvidia/llama-3.3-70b-instruct"),
  23. base_url="https://integrate.api.nvidia.com/v1", default_max_tokens=16384,
  24. )
  25. register_provider(nvidia)