sgl-project/sglang · error · TypeError

MiniMaxH3DiTModel.forward received unexpected kwargs: {unexp

Error message

MiniMaxH3DiTModel.forward received unexpected kwargs: {unexpected}; supported kwargs: {sorted(_FORWARD_SUPPORTED_KWARGS)}

What it means

MiniMaxH3DiTModel.forward enforces a strict keyword contract: any kwarg not in _FORWARD_SUPPORTED_KWARGS raises TypeError. This catches typos and stale call sites early instead of silently ignoring parameters via **kwargs.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/minimax_h3.py:2373

                audio_row_ids,
                audio_embed.to(_BF16_DTYPE),
            )

        t_emb = self._time_embedding(unique_timesteps)
        return embeddings, t_emb

    def forward(self, **kwargs: Any) -> tuple[torch.Tensor, torch.Tensor]:
        """Packed inference forward.

        Keyword names follow the checkpoint's serving contract.
        Returns `(video_logits, audio_logits)` from rows selected by
        `img_pos_for_infer_output_info` and `audio_pos_info`, with condition
        rows zeroed by update masks.
        """
        # Strict keyword contract: refuse any kwarg forward does not consume.
        unexpected = sorted(set(kwargs) - _FORWARD_SUPPORTED_KWARGS)
        if unexpected:
            raise TypeError(
                "MiniMaxH3DiTModel.forward received unexpected kwargs: "
                f"{unexpected}; supported kwargs: "
                f"{sorted(_FORWARD_SUPPORTED_KWARGS)}"
            )

        x = _required_kwarg(kwargs, "x")
        audio_x = _required_kwarg(kwargs, "audio_x")
        img_position_ids = _required_kwarg(kwargs, "img_position_ids")
        unique_timesteps = _required_kwarg(kwargs, "unique_timesteps")
        inverse_indices = (
            _required_kwarg(kwargs, "inverse_indices").view(-1).to(torch.long)
        )
        update_mask = _required_kwarg(kwargs, "update_mask")
        subblock_sparse_query_block_mask = kwargs.get(
            "subblock_sparse_query_block_mask"
        )
        block_token_tags = kwargs.get("block_token_tags")
        token_tags = kwargs.get("token_tags")

View on GitHub (pinned to 0132848349)

Solutions

Check the error message: it lists the supported kwargs; remove/rename the unexpected ones
Align pipeline code with the current forward signature (see _FORWARD_SUPPORTED_KWARGS in the source)
Pin versions of the multimodal_gen runtime and model code to matching commits

Example fix

// before
model(x=x, img_pos_for_infer=img_pos, ...)
// after
model(x=x, img_pos_for_infer_output_info=img_pos, ...)

Defensive patterns

Strategy: validation

Validate before calling

from sglang.multimodal_gen.runtime.models.dits.minimax_h3 import _FORWARD_SUPPORTED_KWARGS
bad = set(my_kwargs) - _FORWARD_SUPPORTED_KWARGS
assert not bad, f"unsupported kwargs: {bad}"

Try / catch

try:
    model.forward(**kwargs)
except TypeError as e:
    if 'unexpected kwargs' in str(e):
        log_and_align_signature(e)  # message lists supported kwargs
    raise

Prevention

Update pipeline and model code together (same commit/PR)
Write a smoke test that calls forward with the pipeline's exact kwargs

When it happens

Trigger: Calling forward with a misspelled or deprecated kwarg (e.g. img_pos_for_infer instead of img_pos_for_infer_output_info), or a caller written against an older/newer signature.

Common situations: Version drift between the pipeline code and the DiT model signature; copy-pasted kwargs from a different DiT implementation; renamed flags after refactors.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/84929cc23cd4aeef. Report an issue: GitHub.