huggingface/transformers · error · ValueError

{generation_mode.name.replace('_', ' ').title()} requires `t

Error message

{generation_mode.name.replace('_', ' ').title()} requires `trust_remote_code=True` in your `generate` call, since it loads https://hf.co/{repo}.

What it means

Some generation modes were moved out of transformers into remote `custom_generate` repositories on the Hub. If the resolved mode maps to a Hub repo (contains '/'), generate warns that it moved and then requires `trust_remote_code=True` to authorize downloading and executing that repo's code; without it, it raises for safety.

Source

Thrown at src/transformers/generation/utils.py:2221

    def _get_deprecated_gen_repo(
        self,
        generation_mode: GenerationMode,
        trust_remote_code: bool,
        custom_generate: str | None = None,
    ) -> str | None:
        """
        Returns the Hub repo for a deprecated generation mode, if any.
        """
        if custom_generate is not None or "/" not in (repo := GENERATION_MODES_MAPPING[generation_mode]):
            return None

        logger.warning_once(
            f"{generation_mode.name.replace('_', ' ').title()} was moved to a `custom_generate` repo: https://hf.co/{repo}. "
            f"To prevent loss of backward compatibility, add `custom_generate='{repo}'` "
            "to your `generate` call before v4.62.0."
        )
        if not trust_remote_code:
            raise ValueError(
                f"{generation_mode.name.replace('_', ' ').title()} requires `trust_remote_code=True` in your `generate` call, "
                f"since it loads https://hf.co/{repo}."
            )
        return repo

    def _extract_generation_mode_kwargs(
        self,
        custom_generate,
        kwargs,
        synced_gpus,
        assistant_model,
        streamer,
    ) -> dict[str, Any]:
        """
        Extracts and returns the generation mode related keyword arguments from the provided kwargs.
        """
        generation_mode_kwargs = {
            "tokenizer": kwargs.pop("tokenizer", None),

View on GitHub (pinned to a597f97485)

Solutions

1. Pin explicit consent: add `custom_generate='<repo>'` (the repo named in the warning) to your generate call before v4.62.0.
2. Add `trust_remote_code=True` to the generate call to allow loading the remote repo.
3. If remote code is unacceptable, switch to a supported built-in generation mode (e.g. greedy/sampling/beam) and re-tune parameters.
4. Monitor the deprecation warning and migration docs so the transition to the required kwarg is deliberate.

Example fix

# before
out = model.generate(**inputs, penalty_alpha=0.6, top_k=4)  # moved mode, no consent -> ValueError

# after
out = model.generate(**inputs, penalty_alpha=0.6, top_k=4, trust_remote_code=True)
# or explicitly:
out = model.generate(**inputs, custom_generate='<repo-from-warning>', trust_remote_code=True)

# opt in deliberately before calling a moved mode
required_repo = GENERATION_MODES_MAPPING.get(resolved_mode)
if required_repo and "/" in required_repo:
    assert kwargs.get("trust_remote_code") is True or kwargs.get("custom_generate") == required_repo

try:
    out = model.generate(**inputs, **kwargs)
except ValueError as e:
    if "requires `trust_remote_code=True`" in str(e):
        out = model.generate(**inputs, **kwargs, trust_remote_code=True)
    else:
        raise

When it happens

Trigger: Parameterizing `generate` to a deprecated/moved mode (e.g. contrastive search-style settings mapped in `GENERATION_MODES_MAPPING` to a repo like 'hf-internal-testing/...') without `trust_remote_code=True` and without an explicit `custom_generate='<repo>'` kwarg.

Common situations: Upgrading to transformers versions where the mode was extracted (with a deprecation window before v4.62.0); old scripts relying on the built-in mode; security-hardened environments where trust_remote_code is disabled by policy.

Related errors

• Loading this model requires you to execute custom code conta
• {error_message} You can inspect the repository content at ht
• `crop` was called, but the current layer does not track past
• Once the sliding window size has been reached, `DynamicSlidi
• `crop` was called, but the current layer does not track past

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/b736621e14c31054. Report an issue: GitHub.