sgl-project/sglang · error · ValueError

Cannot duplicate `image` of batch size {latent_condition.sha

Error message

Cannot duplicate `image` of batch size {latent_condition.shape[0]} to {batch_size} text prompts.

What it means

When batching image-to-image/conditioning, QwenImage can broadcast init latents only when latent_condition divides evenly into batch_size. If batch_size is not a multiple of latent_condition.shape[0] (e.g. 2 condition latents with 3 text prompts), the repeat count batch_size // latent.shape[0] would drop or misalign latents, so postprocess_image_latent raises.

Source

Thrown at python/sglang/multimodal_gen/configs/pipeline_configs/qwen_image.py:569

    def preprocess_condition_image(
        self, image, target_width, target_height, _vae_image_processor
    ):
        return resize(image, target_height, target_width, resize_mode="default"), (
            target_width,
            target_height,
        )

    def postprocess_image_latent(self, latent_condition, batch):
        batch_size = batch.batch_size
        if batch_size > latent_condition.shape[0]:
            if batch_size % latent_condition.shape[0] == 0:
                # expand init_latents for batch_size
                additional_image_per_prompt = batch_size // latent_condition.shape[0]
                image_latents = latent_condition.repeat(
                    additional_image_per_prompt, 1, 1, 1
                )
            else:
                raise ValueError(
                    f"Cannot duplicate `image` of batch size {latent_condition.shape[0]} to {batch_size} text prompts."
                )
        else:
            image_latents = latent_condition
        image_latent_height, image_latent_width = image_latents.shape[3:]
        num_channels_latents = self.dit_config.arch_config.in_channels // 4
        image_latents = _pack_latents(
            image_latents,
            batch_size,
            num_channels_latents,
            image_latent_height,
            image_latent_width,
        )

        return image_latents

    def prepare_pos_cond_kwargs(self, batch, device, rotary_emb, dtype):
        return self._prepare_edit_cond_kwargs(

View on GitHub (pinned to 0132848349)

Solutions

1. Use one image for all prompts (latent batch 1 divides anything)
2. Make len(image) == len(prompt) so the ratio is exactly 1
3. Ensure effective batch_size (including CFG factor) is a multiple of the number of condition images

Example fix

# before
pipe(prompt=["p1", "p2", "p3"], image=[img1, img2])  # 3 % 2 != 0

# after
pipe(prompt=["p1", "p2", "p3", "p4"], image=[img1, img2])  # 4 % 2 == 0

n_img, n_prompt = len(image), len(prompt)
assert n_prompt % n_img == 0, (
    f"num prompts ({n_prompt}) must be a multiple of num images ({n_img})")

When it happens

Trigger: Calling image editing with image=[img1, img2] but 3 (or any non-multiple) prompt strings; latent_condition batch dim from multiple input images while num_prompts % num_images != 0.

Common situations: Batched img2img requests where the number of input images doesn't divide the number of prompts; per-prompt negative examples (CFG doubling) making effective batch_size odd relative to the image count.

Related errors

• The number of initial states is expected to be equal to the
• Mismatched batch sizes: mixed_qkv.shape[0]={B}, a.shape[0]={
• Mismatched batch sizes: mixed_qkv.shape[0]={B}, a.shape[0]={
• `write_pos` must have shape {(batch,)}.
• QwenImageEditPlus expects either one shared condition image

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/e45f59032297071d. Report an issue: GitHub.