sgl-project/sglang · error · ValueError

You have to specify pixel_values or pixel_embeds

Error message

You have to specify pixel_values or pixel_embeds

What it means

InternVL vision model forward requires exactly one of pixel_values or pixel_embeds. Passing neither leaves the vision tower with no input, so it fails fast with this HF-transformers-style check.

Source

Thrown at python/sglang/srt/models/internvl.py:473

    def forward(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        pixel_embeds: Optional[torch.FloatTensor] = None,
    ) -> Union[Tuple, BaseModelOutputWithPooling]:
        pixel_values = pixel_values.to(device=self.device, dtype=self.dtype)
        output_hidden_states = (
            output_hidden_states
            if output_hidden_states is not None
            else self.config.output_hidden_states
        )
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if pixel_values is None and pixel_embeds is None:
            raise ValueError("You have to specify pixel_values or pixel_embeds")

        if pixel_embeds is not None:
            hidden_states = pixel_embeds
        else:
            if len(pixel_values.shape) == 4:
                hidden_states = self.embeddings(pixel_values)
            else:
                raise ValueError(f"wrong pixel_values size: {pixel_values.shape}")

        if self.use_data_parallel:
            encoder_outputs = run_dp_sharded_vision_model(hidden_states, self.encoder)
            last_hidden_state = encoder_outputs
        else:
            encoder_outputs = self.encoder(
                inputs_embeds=hidden_states,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure the request actually carries an image and that pixel_values is produced by the processor
2. Skip the vision tower entirely for text-only batches
3. If you already have embedded features, pass pixel_embeds instead

Example fix

# before
outputs = model.vision_model(pixel_values=None, pixel_embeds=None)

# after
if pixel_values is None and pixel_embeds is None:
    return None  # text-only path
outputs = model.vision_model(pixel_values=pixel_values)

if pixel_values is None and pixel_embeds is None:
    raise SKIP("text-only request; skip vision tower")

def has_vision_input(pixel_values, pixel_embeds) -> bool:
    return pixel_values is not None or pixel_embeds is not None

try:
    out = vision.forward(pixel_values=pv, pixel_embeds=pe)
except ValueError as e:
    if "pixel_values or pixel_embeds" in str(e):
        return None  # degrade to text-only
    raise

When it happens

Trigger: InternVisionModel.forward(pixel_values=None, pixel_embeds=None) — calling the vision tower with only attention_mask or empty multimodal inputs.

Common situations: Text-only requests routed to the multimodal path, missing image preprocessing, or custom pipelines calling the encoder directly without checking inputs.

Related errors

• Incorrect type of pixel values. Got type: {type(pixel_values
• Incorrect type of image sizes. Got type: {type(images_spatia
• Incorrect type of image crop. Got type: {type(images_crop)}
• Image aspect ratio must be smaller than 200
• wrong pixel_values size: {pixel_values.shape}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/c4cd8950e68cb6dc. Report an issue: GitHub.