Class StatelessForwardConfig

rl_engine/executors/stateless_executor.py:23–49

Configuration for no-cache reference/reward model scoring.

@dataclass(frozen=True)
class StatelessForwardConfig:
    """Configuration for no-cache reference/reward model scoring."""

    mode: StatelessForwardMode = "both"
    use_cache: bool = False
    attention_backend: StatelessAttentionBackend = "flash_attention_2"
    reject_kv_cache_outputs: bool = True
    detach_outputs: bool = True
    return_token_scores: bool = False
    max_batch_size: Optional[int] = None
    temperature: float = 1.0
    output_dtype: torch.dtype = torch.float32

    def __post_init__(self) -> None:
        if self.mode not in {"reference", "reward", "both"}:
            raise ValueError("mode must be 'reference', 'reward', or 'both'")
        if self.use_cache:
            raise ValueError("StatelessForwardConfig.use_cache must be False")
        if self.attention_backend not in {"flash_attention_2", "sdpa", "eager", "model_default"}:
            raise ValueError(
                "attention_backend must be 'flash_attention_2', 'sdpa', 'eager', "
                "or 'model_default'"
            )
        if self.max_batch_size is not None and self.max_batch_size <= 0:
            raise ValueError("max_batch_size must be greater than zero")
        if self.temperature <= 0.0:
            raise ValueError("temperature must be greater than zero")