jingyaogong/minimind · error · ValueError

不支持的引擎类型: {engine_type}

Error message

不支持的引擎类型: {engine_type}

What it means

ValueError raised by the create_rollout_engine factory (trainer/rollout_engine.py) when engine_type is neither the literal 'torch' nor 'sglang'. It is deliberate input validation at the API boundary: only two backends exist, TorchRolloutEngine(policy_model, tokenizer, device, autocast_ctx) and SGLangRolloutEngine(sglang_base_url, sglang_model_path, sglang_shared_path). The f-string message interpolates the offending value, so the error text is whatever string was passed.

Source

Thrown at trainer/rollout_engine.py:224

# ===== 工厂函数 =====
def create_rollout_engine(
    engine_type: str = "torch",
    policy_model: torch.nn.Module = None,
    tokenizer = None,
    device: str = "cuda",
    autocast_ctx = None,
    sglang_base_url: str = None,
    sglang_model_path: str = None,
    sglang_shared_path: str = None,
) -> RolloutEngine:
    if engine_type == "torch":
        return TorchRolloutEngine(policy_model, tokenizer, device, autocast_ctx)
    elif engine_type == "sglang":
        return SGLangRolloutEngine(sglang_base_url, sglang_model_path, sglang_shared_path)
    else:
        raise ValueError(f"不支持的引擎类型: {engine_type}")

View on GitHub (pinned to 393e387e9a)

Solutions

1. Set engine_type to exactly 'torch' (in-process rollout with your policy model) or 'sglang' (external SGLang server rollout) — lowercase, no whitespace.
2. If you intended 'sglang', also supply sglang_base_url, sglang_model_path, sglang_shared_path; if 'torch', supply policy_model, tokenizer, device, autocast_ctx.
3. Normalize the value where it enters: engine_type = engine_type.strip().lower() before the factory call.
4. If you believe a third backend should exist, check the repo version — you may be on a branch that only implements two engines.

Example fix

# before
engine = create_rollout_engine(engine_type=config.get('engine'), ...)

# after
engine_type = (config.get('engine') or 'torch').strip().lower()
if engine_type not in ('torch', 'sglang'):
    raise ValueError(f"engine must be 'torch' or 'sglang', got {engine_type!r}")
engine = create_rollout_engine(engine_type=engine_type, ...)

SUPPORTED_ENGINES = ('torch', 'sglang')

engine_type = (engine_type or 'torch').strip().lower()
if engine_type not in SUPPORTED_ENGINES:
    raise ValueError(
        f"engine_type must be one of {SUPPORTED_ENGINES}, got {engine_type!r}"
    )
engine = create_rollout_engine(engine_type=engine_type, ...)

def is_supported_engine(value: str) -> bool:
    """Type guard for create_rollout_engine's engine_type argument."""
    return isinstance(value, str) and value.strip().lower() in ('torch', 'sglang')

When it happens

Trigger: Calling create_rollout_engine(engine_type=...) with any value outside {'torch','sglang'}: typos like 'Torch'/'torch '/'vllm', a config default that was never updated (e.g. engine_type: vllm in a YAML), or passing None because a CLI/config key was misspelled and the loader fell back to None. Note the check is case- and whitespace-sensitive with no normalization.

Common situations: Renaming or extending the trainer config with a new backend string (e.g. after adding a vLLM engine that is not merged); copying a config from an older/newer revision where the accepted names differ; reading engine_type from argparse with a wrong default; case mismatch from JSON/YAML ('SGLang' vs 'sglang').

AI-assisted analysis of jingyaogong/minimind@393e387e9a (2026-08-15). Data as JSON: /api/errors/001d6611640f25f6. Report an issue: GitHub.