sgl-project/sglang · error · ValueError

Invalid latent H/W computed from batch.height/width: {batch.

Error message

Invalid latent H/W computed from batch.height/width: {batch.height=} {batch.width=} {self.vae_scale_factor=}

What it means

After dividing batch height/width by vae_scale_factor, the LTX-2 config checks that the resulting latent dimensions are positive. A zero or negative latent_height/latent_width means the request's resolution is smaller than one VAE latent cell. The error message includes batch.height, batch.width, and vae_scale_factor for diagnosis.

Source

Thrown at python/sglang/multimodal_gen/configs/pipeline_configs/ltx_2.py:345

        - This assumes `patch_size_t == 1` (no temporal patching).
        - Tokens are ordered as (frame, height, width) after packing.
        """
        if int(self.patch_size_t) != 1:
            raise ValueError(
                "LTX-2 SP time-sharding for packed token latents currently requires "
                f"{self.patch_size_t=}. (Expected 1)"
            )
        if int(seq_len) <= 0:
            raise ValueError(f"Expected {seq_len=} > 0 for packed token latents.")
        if int(self.vae_scale_factor) <= 0:
            raise ValueError(f"Invalid {self.vae_scale_factor=}. Must be > 0.")
        if int(self.patch_size) <= 0:
            raise ValueError(f"Invalid {self.patch_size=}. Must be > 0.")

        latent_height = int(batch.height) // int(self.vae_scale_factor)
        latent_width = int(batch.width) // int(self.vae_scale_factor)
        if latent_height <= 0 or latent_width <= 0:
            raise ValueError(
                "Invalid latent H/W computed from batch.height/width: "
                f"{batch.height=} {batch.width=} {self.vae_scale_factor=}"
            )
        if (latent_height % int(self.patch_size)) != 0 or (
            latent_width % int(self.patch_size)
        ) != 0:
            raise ValueError(
                "Invalid spatial patching for packed token latents. Expected latent H/W "
                "to be divisible by patch_size, got "
                f"{latent_height=} {latent_width=} {self.patch_size=}."
            )

        post_patch_h = latent_height // int(self.patch_size)
        post_patch_w = latent_width // int(self.patch_size)
        tokens_per_frame = int(post_patch_h) * int(post_patch_w)
        if tokens_per_frame <= 0:
            raise ValueError(
                f"Invalid tokens_per_frame={tokens_per_frame} from "

View on GitHub (pinned to 0132848349)

Solutions

1. Increase the request resolution so height and width are at least vae_scale_factor (ideally a multiple of vae_scale_factor * patch_size)
2. Check upstream code that computes batch.height/batch.width for truncation or unit confusion (pixels vs latent units)
3. Verify vae_scale_factor is not accidentally set too large for the intended resolution

Example fix

# before
batch = VideoBatch(height=64, width=64)  # with vae_scale_factor=32 -> latent 2x2 (ok) but height=24 -> 0

# after
batch = VideoBatch(height=768, width=1280)  # latent 96x160 with vae_scale_factor=8

lh = batch.height // config.vae_scale_factor; lw = batch.width // config.vae_scale_factor
assert lh > 0 and lw > 0, (batch.height, batch.width, config.vae_scale_factor)

When it happens

Trigger: Calling shard_latents_for_sp with a batch where height or width is smaller than vae_scale_factor (e.g. height=4 with vae_scale_factor=8 yields latent_height=0), or with non-positive height/width.

Common situations: Passing thumbnails or tiny test resolutions (e.g. 64x64 with a scale factor larger than the resolution), unit tests with dummy dimensions, or resolution/aspect-ratio math bugs upstream that produce tiny dims.

Related errors

• Invalid {self.vae_scale_factor=}. Must be > 0.
• Invalid {self.patch_size=}. Must be > 0.
• Invalid spatial patching for packed token latents. Expected
• Invalid tokens_per_frame={tokens_per_frame} from {latent_hei
• LTX-2 token latents seq_len={seq_len} is not divisible by to

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/2cbcbe38a4471b86. Report an issue: GitHub.