sgl-project/sglang · error · ValueError

Cosmos3 AVAE dec_strides product must equal hop_size: produc

Error message

Cosmos3 AVAE dec_strides product must equal hop_size: product={stride_product}, hop_size={self.hop_size}.

What it means

The Cosmos3 audio VAE decodes latents back to audio frames, so the product of its decoder strides must exactly equal hop_size (frames per latent step, default prod(dec_strides)). If the config explicitly overrides hop_size inconsistently with dec_strides, decode would produce the wrong number of frames, so __init__ fails fast.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/vaes/cosmos3_avae.py:169

                default=2 if bool(config.get("stereo", True)) else 1,
            )
        )
        self.latent_channels = int(
            _cfg(config, "vocoder_input_dim", "io_channels", "latent_ch", default=64)
        )
        dec_strides = [
            int(s) for s in _cfg(config, "dec_strides", default=[2, 4, 5, 6, 8])
        ]
        self.hop_size = int(
            _cfg(
                config,
                "hop_size",
                default=math.prod(dec_strides) if dec_strides else 1920,
            )
        )
        stride_product = math.prod(dec_strides)
        if stride_product != self.hop_size:
            raise ValueError(
                "Cosmos3 AVAE dec_strides product must equal hop_size: "
                f"product={stride_product}, hop_size={self.hop_size}."
            )

        norm = str(_cfg(config, "normalization_type", default="none"))
        if bool(_cfg(config, "normalize_latents", default=False)) and norm == "none":
            norm = "tanh"
        self.normalization_type = norm
        self.tanh_input_scale = float(_cfg(config, "tanh_input_scale", default=1.5))
        self.tanh_output_scale = float(_cfg(config, "tanh_output_scale", default=3.5))
        self.tanh_clamp = float(_cfg(config, "tanh_clamp", default=0.995))

        self.decoder = OobleckDecoder(
            channels=int(_cfg(config, "dec_dim", default=320)),
            input_channels=self.latent_channels,
            audio_channels=self.audio_channels,
            upsampling_ratios=list(reversed(dec_strides)),
            channel_multiples=list(

View on GitHub (pinned to 0132848349)

Solutions

1. Remove the explicit hop_size override so it defaults to prod(dec_strides)
2. Or fix hop_size to equal math.prod(dec_strides) exactly
3. Verify the dec_strides list parsed correctly (non-empty, expected values) from the config file

Example fix

# before
config = {"dec_strides": [4, 4, 120], "hop_size": 960}
# after
config = {"dec_strides": [4, 4, 120]}  # hop_size defaults to 1920
# or config = {"dec_strides": [4, 4, 60], "hop_size": 960}

import math
stride_product = math.prod(dec_strides)
if hop_size is None:
    hop_size = stride_product
assert math.prod(dec_strides) == hop_size, "dec_strides must multiply to hop_size"

When it happens

Trigger: Loading a Cosmos3 AVAE config where hop_size (default 1920 when dec_strides is empty) disagrees with math.prod(dec_strides), e.g. setting hop_size=960 while dec_strides multiply to 1920, or providing dec_strides whose product differs from an inherited default hop_size.

Common situations: Editing a config to halve hop_size for shorter latents without adjusting strides; merging configs from different Cosmos3 checkpoints; checkpoint whose dec_strides list is parsed as empty (falling back to 1920) due to a schema/name change.

Related errors

• Either spatial_upsample or temporal_upsample must be True
• Invalid threshold_type for topk: {threshold_type}. Choose 'q
• Invalid threshold_type: {threshold_type}. Choose 'query_head
• SGLANG_DIFFUSION_ATTENTION_CONFIG is not set
• Cosmos3CausalAttention requires num_attention_heads divisibl

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5fd167d7551452c8. Report an issue: GitHub.