What causes "resize(mode={mode}) not implemented." in hpcaitech/Open-Sora?

Change mode to one of the supported values: 'nearest', 'area', 'bilinear', or 'bicubic'

hpcaitech/Open-Sora · error · NotImplementedError

resize(mode={mode}) not implemented.

Error message

resize(mode={mode}) not implemented.

What it means

The resize() helper in dc_ae's vo_ops only supports the bilinear/bicubic family (via antialiased interpolate) and 'nearest'/'area' modes. Any other mode string (e.g. 'linear', 'trilinear', 'nearest-exact', or a typo like 'neareast') reaches the else branch and raises NotImplementedError.

Source

Thrown at opensora/models/dc_ae/models/nn/vo_ops.py:231

def resize(
    x: torch.Tensor,
    size: Optional[Any] = None,
    scale_factor: Optional[list[float]] = None,
    mode: str = "bicubic",
    align_corners: Optional[bool] = False,
) -> torch.Tensor:
    if mode in {"bilinear", "bicubic"}:
        return F.interpolate(
            x,
            size=size,
            scale_factor=scale_factor,
            mode=mode,
            align_corners=align_corners,
        )
    elif mode in {"nearest", "area"}:
        return F.interpolate(x, size=size, scale_factor=scale_factor, mode=mode)
    else:
        raise NotImplementedError(f"resize(mode={mode}) not implemented.")


def build_kwargs_from_config(config: dict, target_func: Callable) -> dict[str, Any]:
    valid_keys = list(signature(target_func).parameters)
    kwargs = {}
    for key in config:
        if key in valid_keys:
            kwargs[key] = config[key]
    return kwargs


if __name__ == "__main__":
    test_chunked_interpolate()

View on GitHub (pinned to 7ad6a96a13)

Solutions

Change mode to one of the supported values: 'nearest', 'area', 'bilinear', or 'bicubic'
Check the config/dict that supplies the mode string for typos
If you truly need another mode, call F.interpolate directly instead of this wrapper

Example fix

// before
y = resize(x, scale_factor=2.0, mode="trilinear")
// after
y = resize(x, scale_factor=2.0, mode="nearest")

Defensive patterns

Strategy: validation

Validate before calling

SUPPORTED = {"bilinear", "bicubic", "nearest", "area"}
assert mode in SUPPORTED, f"unsupported resize mode {mode!r}; choose from {SUPPORTED}"

Type guard

def is_supported_resize_mode(mode: str) -> bool:
    return mode in {"bilinear", "bicubic", "nearest", "area"}

Try / catch

try:
    y = resize(x, mode=mode, ...)
except NotImplementedError:
    y = F.interpolate(x, mode="nearest", ...)  # explicit fallback

Prevention

Centralize resample mode strings in one constants module
Reject unknown mode strings at config-load time
Keep mode names lowercase and short-form

When it happens

Trigger: Calling resize(x, mode=...) with a mode not in {bilinear, bicubic, nearest, area} (and whatever antialias variants the earlier branches accept), typically from a model forward pass where the resample mode comes from a config.

Common situations: Copying a resample mode string from another library (e.g. torch.nn.Upsample's 'linear'/'trilinear' or diffusers' 'nearest-exact') into an opensora dc_ae config; typos in YAML configs.

Related errors

AI-assisted analysis of hpcaitech/Open-Sora@7ad6a96a13 (2026-08-28). Data as JSON: /api/errors/2caa7624299c7a60. Report an issue: GitHub.