sgl-project/sglang · error · ValueError

Unsupported RRDBNet conv_first input channels: {in_channels}

Error message

Unsupported RRDBNet conv_first input channels: {in_channels}

What it means

Raised in realesrgan_upscaler._build_net_from_state_dict when inferring scale from the RRDBNet checkpoint's conv_first weight shape: 3→scale 4, 12→scale 2, 48→scale 1. Any other input-channel count means the checkpoint is not a standard RRDBNet real-ESRGAN model.

Source

Thrown at python/sglang/multimodal_gen/runtime/postprocess/realesrgan_upscaler.py:226

# ---------------------------------------------------------------------------
# Architecture auto-detection
# ---------------------------------------------------------------------------


def _build_net_from_state_dict(state_dict: dict) -> nn.Module:
    """Detect architecture from checkpoint keys and return an unloaded network."""
    if "conv_first.weight" in state_dict:
        # RRDBNet (e.g., RealESRGAN_x4plus)
        num_feat = state_dict["conv_first.weight"].shape[0]
        in_channels = state_dict["conv_first.weight"].shape[1]
        if in_channels == 3:
            scale = 4
        elif in_channels == 12:
            scale = 2
        elif in_channels == 48:
            scale = 1
        else:
            raise ValueError(
                f"Unsupported RRDBNet conv_first input channels: {in_channels}"
            )
        num_block = sum(
            1
            for k in state_dict
            if k.startswith("body.") and k.endswith(".rdb1.conv1.weight")
        )
        num_grow_ch = state_dict["body.0.rdb1.conv1.weight"].shape[0]
        logger.info(
            "Detected RRDBNet: num_feat=%d, num_block=%d, num_grow_ch=%d, scale=%d",
            num_feat,
            num_block,
            num_grow_ch,
            scale,
        )
        return RRDBNet(
            num_in_ch=3,
            num_out_ch=3,

View on GitHub (pinned to 0132848349)

Solutions

1. Use a standard Real-ESRGAN RRDBNet or SRVGGNetCompact checkpoint (e.g. RealESRGAN_x4plus)
2. Point model_path at a local .pth or 'repo_id:filename' for a supported model

Example fix

// before
upscaler = RealESRGANUpscaler(model_path="some_srgan.pth")
// after
upscaler = RealESRGANUpscaler(model_path="ai-forever/RealESRGAN_x4:RealESRGAN_x4.pth")

sd = torch.load(pth, map_location="cpu", weights_only=True)
ch = sd.get("conv_first.weight").shape[1]
assert ch in (3, 12, 48), f"unsupported RRDB channels {ch}"

try:
    upscaler.upscale(frame)
except RuntimeError as e:
    if "not compatible" in str(e) or "conv_first" in str(e):
        raise ValueError("Use a standard Real-ESRGAN checkpoint") from e
    raise

When it happens

Trigger: Loading a .pth whose 'conv_first.weight' has an unexpected first dimension (e.g. 1, 6, or 64 channels), typically a different super-resolution architecture or a non-image model.

Common situations: Pointing model_path at a vanilla ESRGAN/SRGAN/Real-ESRGAN anime-video variant with different channel layout, or at an unrelated PyTorch file.

Related errors

• Real-ESRGAN weight file '{resolved_path}' is not compatible
• Failed to load Real-ESRGAN checkpoint from '{resolved_path}'
• Unsupported activation type: {act_type}
• All frames in a batch must have the same resolution
• RealESRGAN batch upscale did not produce all frames

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/e955c0d49ec0fd1c. Report an issue: GitHub.