hiyouga/LlamaFactory · error · ValueError

LLaMA-Factory YAML and Accelerate config cannot define diffe

Error message

LLaMA-Factory YAML and Accelerate config cannot define different KT settings.

What it means

Raised by _get_advanced_kt_config (model_args.py:577) when kt_config is defined in BOTH the LLaMA-Factory training YAML and the Accelerate config's accelerator_config.kt_config, and the two mappings are not equal. Divergent duplicates would make the effective KT settings depend on merge order, so LlamaFactory refuses to guess.

Source

Thrown at src/llamafactory/hparams/model_args.py:577

        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."""
        if self.use_unsloth or self.use_unsloth_gc:
            raise ValueError("KTransformers cannot be combined with Unsloth checkpoint wrapping.")
        if getattr(training_args, "gradient_checkpointing", False):
            raise ValueError(
                "KTransformers uses LLaMA-Factory's `disable_gradient_checkpointing`; "
                "remove `gradient_checkpointing: true`."
            )
        if getattr(training_args, "gradient_checkpointing_kwargs", None) is not None:
            raise ValueError("KTransformers supplies its checkpoint context; remove `gradient_checkpointing_kwargs`.")

        fsdp_config = getattr(training_args, "fsdp_config", None)
        if isinstance(fsdp_config, dict) and fsdp_config.get("activation_checkpointing"):
            raise ValueError("Disable FSDP activation checkpointing when using KTransformers.")
        if os.environ.get("FSDP_ACTIVATION_CHECKPOINTING", "false").lower() in {"1", "true", "yes"}:

View on GitHub (pinned to f28afaf635)

Solutions

1. Delete kt_config from the Accelerate config and keep only the training-YAML copy (preferred)
2. Or make the two mappings byte-for-byte equivalent in parsed value (mind types and quoting)
3. Add a pre-launch lint that fails if both files contain kt_config

Example fix

# before
# accelerate.yaml: kt_config: {gen_config: {temperature: 0.9}}
# train.yaml:    kt_config: {gen_config: {temperature: 0.7}}

# after
# accelerate.yaml: (kt_config removed)
# train.yaml:
kt_config:
  gen_config:
    temperature: 0.7

lf_kt = cfg.get('kt_config')
acc_kt = (yaml.safe_load(open('accelerate_config.yaml')).get('accelerator_config') or {}).get('kt_config')
assert not (acc_kt is not None and lf_kt is not None and acc_kt != lf_kt), 'kt_config defined twice and differs'

def kt_sources_agree(lf: dict | None, acc: dict | None) -> bool:
    return lf is None or acc is None or lf == acc

When it happens

Trigger: Editing kt_config in the training YAML but not updating the copy in the Accelerate file (or vice versa); a shared Accelerate base config pinned to an older KT block combined with an updated experiment YAML.

Common situations: Two-file setups drifting apart after iterative tuning; CI templates that regenerate one file; JSON vs YAML quoting differences making apparently-equal values unequal (e.g. string 'true' vs bool true).

Related errors

• Put KTransformers settings in the LLaMA-Factory training YAM
• These `kt_config` values are derived from LLaMA-Factory argu
• `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/21ff2af0ded7d1f5. Report an issue: GitHub.