sgl-project/sglang · error · ValueError

Invalid grid metadata for kimi image tokens: {vals} (expecte

Error message

Invalid grid metadata for kimi image tokens: {vals} (expected [t,h,w] or [h,w])

What it means

After flattening grid_thw, the Kimi helper requires at least 2 values ([h,w] or [t,h,w] with more) to derive patch dimensions. Fewer than 2 values means the grid is malformed (empty or a single scalar), so it cannot compute the image token count and raises ValueError.

Source

Thrown at python/sglang/srt/multimodal/processors/kimi_common.py:79

        merge_h, merge_w = self.hf_config.vision_config.merge_kernel_size

        if isinstance(grid_thw, torch.Tensor):
            vals = grid_thw.flatten().tolist()
        elif isinstance(grid_thw, np.ndarray):
            vals = grid_thw.reshape(-1).tolist()
        elif isinstance(grid_thw, (list, tuple)):
            vals = list(np.array(grid_thw).reshape(-1).tolist())
        else:
            raise TypeError(
                f"Unsupported grid type for kimi image tokens: {type(grid_thw)}"
            )

        if len(vals) >= 3:
            _t, h, w = vals[-3], vals[-2], vals[-1]
        elif len(vals) == 2:
            _t, h, w = 1, vals[0], vals[1]
        else:
            raise ValueError(
                f"Invalid grid metadata for kimi image tokens: {vals} "
                "(expected [t,h,w] or [h,w])"
            )

        h, w = int(h), int(w)
        return (h * w) // (merge_h * merge_w)

    def _build_kimi_mm_data_from_grids(
        self, prompt, embeddings, **kwargs
    ) -> MultimodalProcessorOutput:
        image_token_id = kwargs.get("image_token_id", 0)
        img_grid_thw = kwargs.get("img_grid_thw", None)

        if not isinstance(prompt, list):
            prompt = self._tokenizer.encode(prompt)

        image_token_counts = [
            self._num_image_tokens_from_grid(grid) for grid in img_grid_thw

View on GitHub (pinned to 0132848349)

Solutions

1. Validate every image's grid has t*h*w >= 2 entries before calling
2. Check upstream vision-encoder output for dropped/failed images (corrupt files, zero-byte uploads)
3. Pad or regenerate missing grid entries from the processor's image_sizes instead of forwarding empty grids

Example fix

// before
build_from_grids(prompt, grids=[torch.tensor([])])  # one image, empty grid

// after
assert all(g.numel() >= 2 for g in grids), f"malformed grid: {grids}"
build_from_grids(prompt, grids=grids)

assert all(np.asarray(g).size >= 2 for g in grids), "each grid needs [t,h,w] or [h,w]"

When it happens

Trigger: Passing an empty tensor/array/list as grid_thw, or a grid with a single element, to _num_image_tokens_from_grid via _build_kimi_mm_data_from_grids or get_mm_data.

Common situations: Empty per-image grid lists because the vision encoder returned no output for a corrupt/blank image, or a data pipeline that dropped grid entries during batching/serialization.

Related errors

• Grid dim ({_mm_grid_attrs[modality]}) not found in {mm_input
• Invalid Kimi image grid metadata: {values}; expected [h, w]
• Unsupported grid type for kimi image tokens: {type(grid_thw)
• The number of image placeholders exceeds img_grid_thw entrie
• The number of image placeholders does not match img_grid_thw

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/afbef697acf26a54. Report an issue: GitHub.