sgl-project/sglang · error · TypeError

Kimi-K3 image feature must be a torch.Tensor, got {type(item

Error message

Kimi-K3 image feature must be a torch.Tensor, got {type(item.feature)}

What it means

In materialize_item_features(), items without a deferred config must already hold a torch.Tensor in item.feature; any other type raises TypeError. This guards the local path where features are assumed precomputed before materialization.

Source

Thrown at python/sglang/srt/models/kimi_k3.py:3424

                        "Kimi-K3 cannot mix local preprocessed and deferred images"
                    )
                return materialize_multimodal_features(
                    [item.feature for item in selected_items],
                    device=device,
                    dtype=target_dtype,
                )

            deferred = [
                item.model_specific_data.get(DEFERRED_PREPROCESSING_KEY)
                for item in selected_items
            ]
            if any(config is not None for config in deferred):
                materialized = [None] * len(selected_items)
                deferred_by_backend = {}
                for index, (item, config) in enumerate(zip(selected_items, deferred)):
                    if config is None:
                        if not isinstance(item.feature, torch.Tensor):
                            raise TypeError(
                                "Kimi-K3 image feature must be a torch.Tensor, "
                                f"got {type(item.feature)}"
                            )
                        materialized[index] = item.feature
                    else:
                        deferred_by_backend.setdefault(config.backend, []).append(index)

                for backend, indices in deferred_by_backend.items():
                    group_items = [selected_items[index] for index in indices]
                    group_configs = [deferred[index] for index in indices]
                    # Map backend-group positions through the rank-local shard to global grid rows.
                    global_indices = [image_indices[index] for index in indices]
                    first_config = group_configs[0]
                    if backend == "gpu":
                        from sglang.srt.multimodal.processors.kimi_k25 import (
                            _gpu_preprocess_images,
                        )

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure local items have their .feature set to a torch.Tensor before calling get_image_feature()
2. Convert non-tensor features with torch.as_tensor(..., dtype=target_dtype) beforehand
3. If features are not computed, mark items as deferred with a proper backend config

Example fix

// before
item.feature = np_array  # then model.get_image_feature([item])

// after
item.feature = torch.from_numpy(np_array)
model.get_image_feature([item])

for i in items:
    if deferred_config(i) is None and not isinstance(i.feature, torch.Tensor):
        i.feature = torch.as_tensor(i.feature)

import torch
def features_are_tensors(items) -> bool:
    return all(isinstance(i.feature, torch.Tensor) for i in items if i.feature is not None)

try:
    model.get_image_feature(items)
except TypeError as e:
    if "torch.Tensor" in str(e):
        coerce_features_to_tensors(items)
        return model.get_image_feature(items)
    raise

When it happens

Trigger: A MultimodalDataItem with config None (not deferred) whose .feature is not a torch.Tensor — e.g. a numpy array, PIL image, raw pixel list, or None because preprocessing never ran.

Common situations: Skipping the local preprocessing step for some items; a pipeline change that stores raw pixels instead of tensor features; None feature from a failed earlier stage.

Related errors

• Unsupported image type: {type(image)}
• Input 'data' must be a torch.Tensor, but got {type(data)}
• Expected a PIL image, got {type(image)}
• Kimi-K3 encoder mode supports image input only
• Kimi-K3 expects one vision grid per MultimodalDataItem; spli

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/e5968949d86fda79. Report an issue: GitHub.