sgl-project/sglang · error · ValueError

Step3-VL image item has num_patches > 0 but no patch_pixel_v

Error message

Step3-VL image item has num_patches > 0 but no patch_pixel_values.

What it means

Step3-VL-10B mirror of the patch-feature guard: an image with num_patch > 0 must come with patch pixel features; when item_patch_features is None the loop raises ValueError because patch features are required to assemble the output.

Source

Thrown at python/sglang/srt/models/step3_vl_10b.py:557

        patch_offset = 0
        for thumb_count, num_patches_list, patch_count in item_metadata:
            item_thumb_features = all_thumb_features[
                thumb_offset : thumb_offset + thumb_count
            ]
            thumb_offset += thumb_count
            item_patch_features = (
                all_patch_features[patch_offset : patch_offset + patch_count]
                if patch_count > 0
                else None
            )
            patch_offset += patch_count

            cur_patch_idx = 0
            for i, num_patch in enumerate(num_patches_list):
                cur_feature = []
                if num_patch > 0:
                    if item_patch_features is None:
                        raise ValueError(
                            "Step3-VL image item has num_patches > 0 but no patch_pixel_values."
                        )
                    patch_slice = item_patch_features[
                        cur_patch_idx : cur_patch_idx + num_patch
                    ]
                    cur_feature.append(patch_slice.view(-1, patch_slice.shape[-1]))
                cur_feature.append(
                    item_thumb_features[i].view(-1, item_thumb_features.shape[-1])
                )
                cur_patch_idx += num_patch
                merged_image_features.append(
                    torch.cat(cur_feature) if len(cur_feature) > 1 else cur_feature[0]
                )
        return self._flatten_embeddings(merged_image_features)

    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
        pattern = MultiModalityDataPaddingPatternMultimodalTokens()
        return pattern.pad_input_tokens(input_ids, mm_inputs)

View on GitHub (pinned to 0132848349)

Solutions

1. Provide patch_pixel_values/patch features for every image with num_patches > 0
2. Zero out num_patches when patch inputs are intentionally omitted
3. Keep processor and model code from the same release

Example fix

# before
item.model_specific_data = {"num_patches": [2, 1]}
# after
item.model_specific_data = {"num_patches": [2, 1], "patch_pixel_values": patches}

assert item_patch_features is not None or all(n == 0 for n in num_patches_list)

When it happens

Trigger: num_patches_list containing positive counts while patch pixel features are missing for that item in step3_vl_10b.get_image_feature (python/sglang/srt/models/step3_vl_10b.py:557).

Common situations: Preprocessor emitting patch counts but not patch pixels across versions; trimming patch inputs manually; DP/EP pipelines dropping the patch tensor.

Related errors

• Step3-VL image item has num_patches > 0 but no patch_pixel_v
• Step3-VL image item is missing num_patches.
• cos/sin shape does not cover image tokens and head_dim
• Unsupported image type: {type(image)}
• QwenImageEditPlus expects either one shared condition image

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/23565ccace9703d4. Report an issue: GitHub.