Stability-AI/generative-models · error · ValueError

unknown merge strategy {self.merge_strategy}

Error message

unknown merge strategy {self.merge_strategy}

What it means

VideoTransformerBlock validates merge_strategy in __init__ and only 'fixed' (buffer) and 'learned' (parameter) are accepted; any other string raises ValueError at construction time.

Source

Thrown at sgm/modules/autoencoding/temporal_ae.py:52

            dims=3,
            use_scale_shift_norm=False,
            use_conv=False,
            up=False,
            down=False,
            kernel_size=video_kernel_size,
            use_checkpoint=False,
            skip_t_emb=True,
        )

        self.merge_strategy = merge_strategy
        if self.merge_strategy == "fixed":
            self.register_buffer("mix_factor", torch.Tensor([alpha]))
        elif self.merge_strategy == "learned":
            self.register_parameter(
                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
            )
        else:
            raise ValueError(f"unknown merge strategy {self.merge_strategy}")

    def get_alpha(self, bs):
        if self.merge_strategy == "fixed":
            return self.mix_factor
        elif self.merge_strategy == "learned":
            return torch.sigmoid(self.mix_factor)
        else:
            raise NotImplementedError()

    def forward(self, x, temb, skip_video=False, timesteps=None):
        if timesteps is None:
            timesteps = self.timesteps

        b, c, h, w = x.shape

        x = super().forward(x, temb)

        if not skip_video:

View on GitHub (pinned to e8cd657656)

Solutions

1. Set merge_strategy to 'fixed' or 'learned' in the block config
2. Fix the typo in the config key value
3. Check the class __init__ signature for the default ('learned') and rely on it

Example fix

// before
VideoTransformerBlock(in_channels=320, merge_strategy="average")
// after
VideoTransformerBlock(in_channels=320, merge_strategy="learned")

VALID = {"fixed", "learned"}
assert merge_strategy in VALID, f"merge_strategy must be one of {VALID}, got {merge_strategy!r}"

def is_valid_merge_strategy(s) -> bool:
    return s in ("fixed", "learned")

try:
    block = VideoTransformerBlock(in_channels=320, merge_strategy=cfg.strategy)
except ValueError as e:
    logger.error("bad merge_strategy in config, falling back to 'learned'")
    block = VideoTransformerBlock(in_channels=320, merge_strategy="learned")

When it happens

Trigger: Instantiating VideoTransformerBlock with merge_strategy set to anything other than 'fixed' or 'learned', e.g. a config typo like 'fixed_with_decay' or 'average'.

Common situations: YAML/JSON config typos for temporal autoencoder attention blocks; copying older config keys into a newer code version that renamed allowed values.

Understand the failure class

Background: Invalid enum value errors: "Unknown type", "Invalid scope", "must be one of" — when a string is not on the library's allowed list — this error's family across 23 libraries.

Related errors

• unknown merge strategy {merge_strategy}
• Unknown loss type {self.loss_type}
• provide num_res_blocks either as an int (globally constant)
• Order {order} too high for step {i}
• Decay must be between 0 and 1

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/0dd0f73eb0a9c46d. Report an issue: GitHub.