hiyouga/LlamaFactory · error · ValueError

These `kt_config` values are derived from LLaMA-Factory argu

Error message

These `kt_config` values are derived from LLaMA-Factory arguments: {conflicts}.

What it means

Raised by _normalize_advanced_kt_config (model_args.py:563) when the user-supplied kt_config mapping contains any key in _KT_DERIVED_KEYS (the set of kt_weight_path, kt_lora_* , kt_train_mode, etc. shown above the __post_init__). Those settings are derived from the dedicated LlamaFactory kt_* arguments; supplying them inside kt_config too would create two competing sources of truth, so the conflicts are reported with their key names.

Source

Thrown at src/llamafactory/hparams/model_args.py:563

        return {"cpu": cpu_activation, "gpu": gpu_activation}

    @staticmethod
    def _get_accelerator_kt_config(training_args: Any) -> Any:
        accelerator_config = getattr(training_args, "accelerator_config", None)
        if isinstance(accelerator_config, dict):
            return accelerator_config.get("kt_config")
        return getattr(accelerator_config, "kt_config", None)

    def _normalize_advanced_kt_config(self, raw_config: Any) -> dict[str, Any]:
        if raw_config is None:
            return {}
        if not isinstance(raw_config, dict):
            raise TypeError("LLaMA-Factory `kt_config` must be a flat mapping.")

        config = dict(raw_config)
        conflicts = sorted(set(config) & self._KT_DERIVED_KEYS)
        if conflicts:
            raise ValueError(f"These `kt_config` values are derived from LLaMA-Factory arguments: {conflicts}.")
        return config

    def _get_advanced_kt_config(self, training_args: Any) -> dict[str, Any]:
        raw_config = getattr(training_args, "kt_config", None)
        accelerator_config = self._get_accelerator_kt_config(training_args)
        if raw_config is None:
            if accelerator_config is not None:
                raise ValueError(
                    "Put KTransformers settings in the LLaMA-Factory training YAML `kt_config`; "
                    "remove `kt_config` from the Accelerate config."
                )
            return {}
        if accelerator_config is not None and accelerator_config != raw_config:
            raise ValueError("LLaMA-Factory YAML and Accelerate config cannot define different KT settings.")
        return self._normalize_advanced_kt_config(raw_config)

    def configure_kt_checkpointing(self, training_args: Any) -> None:
        r"""Keep LLaMA-Factory as the single gradient-checkpointing entry point."""

View on GitHub (pinned to f28afaf635)

Solutions

1. Move the listed keys out of kt_config into the corresponding top-level kt_* arguments (e.g. kt_weight_path, kt_lora_rank)
2. Delete duplicates: keep kt_config only for genuinely advanced KTransformers keys not exposed as LlamaFactory arguments
3. Check the _KT_DERIVED_KEYS set in model_args.py for the authoritative blacklist

Example fix

# before
kt_config:
  kt_weight_path: /path/to/weights   # derived key

# after
kt_weight_path: /path/to/weights
kt_config:
  gen_config:
    temperature: 0.7

derived = {'kt_weight_path','kt_non_expert_weight_path','kt_lora_rank','kt_lora_expert_num','kt_train_mode','kt_use_lora_experts','kt_skip_expert_loading'}
conflicts = set(cfg.get('kt_config') or {}) & derived
assert not conflicts, f'move {sorted(conflicts)} to top-level arguments'

def no_derived_conflicts(kt_config: dict | None, derived: set[str]) -> bool:
    return not (set(kt_config or {}) & derived)

When it happens

Trigger: Adding kt_weight_path or kt_lora_rank inside the kt_config mapping while also (or instead) using the top-level kt_* arguments; porting a KTransformers recipe that uses those exact key names into kt_config.

Common situations: Copy-pasting KTransformers upstream configs whose keys collide with the derived set; users assuming kt_config accepts every KTransformers option (it accepts all EXCEPT the derived ones).

Related errors

• Put KTransformers settings in the LLaMA-Factory training YAM
• LLaMA-Factory YAML and Accelerate config cannot define diffe
• `kt_cpu_activation` must be `retain` or `recompute`.
• `kt_cpu_activation` is only valid when `use_kt: true`.
• `kt_cpu_activation: recompute` requires GPU gradient checkpo

AI-assisted analysis of hiyouga/LlamaFactory@f28afaf635 (2026-08-14). Data as JSON: /api/errors/b3db08ad68ead296. Report an issue: GitHub.