backend
backend
¶
Abstract base class for inference backends.
Classes¶
InferenceBackend
¶
Bases: ABC
Base class for all inference backends used in evaluation.
Attributes¶
framework_commit_value
property
¶
Default: empty (subclasses override). Used by runner.py error paths.
Methods:¶
generate
abstractmethod
¶
generate(
prompt: str,
*,
model: str,
system: str = "",
temperature: float = 0.0,
max_tokens: int = 2048,
) -> str
Generate a response and return just the text content.
generate_full
abstractmethod
¶
generate_full(
prompt: str,
*,
model: str,
system: str = "",
temperature: float = 0.0,
max_tokens: int = 2048,
) -> Dict[str, Any]
Generate a response and return full details.
Returns dict with keys: content, usage, model, latency_seconds, cost_usd.