sgl-project/sglang · error · ValueError

Unknown action normalization method {method!r}.

Error message

Unknown action normalization method {method!r}.

What it means

normalize_action/denormalize_action support a fixed set of normalization schemes (e.g. 'meanstd'/'minmax'/'quantile' as implemented). Any other method string raises ValueError, guarding against silently applying the wrong normalization to action inputs/outputs.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/cosmos3_action.py:198

    }


def normalize_action(
    action: torch.Tensor, method: str, stats: dict[str, torch.Tensor]
) -> torch.Tensor:
    if method == "quantile":
        q01, q99 = stats["q01"].to(action), stats["q99"].to(action)
        return (2.0 * (action - q01) / (q99 - q01).clamp(min=1e-8) - 1.0).clamp(
            -1.0, 1.0
        )
    if method == "meanstd":
        return (action - stats["mean"].to(action)) / stats["std"].to(action).clamp(
            min=1e-8
        )
    if method == "minmax":
        lo, hi = stats["min"].to(action), stats["max"].to(action)
        return (2.0 * (action - lo) / (hi - lo).clamp(min=1e-8) - 1.0).clamp(-1.0, 1.0)
    raise ValueError(f"Unknown action normalization method {method!r}.")


def denormalize_action(
    action: torch.Tensor, method: str, stats: dict[str, torch.Tensor]
) -> torch.Tensor:
    if method == "quantile":
        q01, q99 = stats["q01"].to(action), stats["q99"].to(action)
        return (action + 1.0) / 2.0 * (q99 - q01) + q01
    if method == "meanstd":
        return action * stats["std"].to(action) + stats["mean"].to(action)
    if method == "minmax":
        lo, hi = stats["min"].to(action), stats["max"].to(action)
        return (action + 1.0) / 2.0 * (hi - lo) + lo
    raise ValueError(f"Unknown action normalization method {method!r}.")

View on GitHub (pinned to 0132848349)

Solutions

1. Use one of the implemented method names — check the if-chains at the top of normalize_action/denormalize_action in cosmos3_action.py
2. Fix the config/request field that carries the normalization method to the canonical spelling
3. If a new scheme is genuinely needed, implement it in both normalize_action and denormalize_action and add it to validation docs

Example fix

# before
method = "zscore"
# after
method = "meanstd"

method = method.lower().strip()
assert method in {"meanstd", "minmax", "quantile"}, f"unknown normalization method {method!r}"  # verify against normalize_action's implemented branches

def is_supported_normalization(method: str) -> bool:
    return method.lower().strip() in {"meanstd", "minmax", "quantile"}

When it happens

Trigger: Passing method='gaussian', 'zscore', '' or any unlisted string to normalize_action (called from _prepare_action_latents with the value from the request or pipeline config).

Common situations: Config uses a different spelling than the code ('z-score' vs 'meanstd'); new dataset using a normalization scheme not yet implemented; copy-paste from another repo's config format.

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

• Cosmos3 action endpoint supports action_mode='policy' or 'in
• condition_video_keep must be 'first' or 'last', got {keep!r}
• rollout_sde_type must be one of {_VALID_ROLLOUT_SDE_TYPES},
• Cosmos3 action input accepts one image field; use a list or
• Cosmos3 observation image arrays must use uint8 dtype

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/89a31a9dce1559ee. Report an issue: GitHub.