sgl-project/sglang · error · FileNotFoundError

Action normalization stats not found at

Error message

Action normalization stats not found at {stats_path}.

What it means

load_action_stats reads a JSON file of per-channel action normalization statistics (mean/std/min/max/q01/q99) required to normalize/denormalize actions. If the path does not exist it raises FileNotFoundError — the stats artifact was not shipped, mounted, or its path misconfigured.

Solutions

1. Verify the stats JSON exists at the configured path (ls the exact path from the error)
2. Point the config/server arg at the correct location of the model's action stats file
3. If the file was never generated, create it from your training data (per-channel mean/std/min/max/q01/q99) or export it from the model release
4. In containers, ensure the volume containing the stats file is mounted

Example fix

# before
--action-stats-path /models/cosmos3/stats.json  # missing
# after
--action-stats-path /models/cosmos3/action_stats.json  # actual filename in checkpoint

from pathlib import Path
assert Path(stats_path).is_file(), f"action stats file missing: {stats_path}"

def action_stats_available(stats_path: str) -> bool:
    return Path(stats_path).is_file()

try:
    stats = load_action_stats(stats_path)
except FileNotFoundError:
    raise RuntimeError(f"deploy blocked: mount or download action stats at {stats_path}") from None

When it happens

Trigger: Any action-conditioned Cosmos3 request: _prepare_action_latents/forward calls load_action_stats(stats_path) with a path (from pipeline config / server args) that doesn't exist on the scheduler's filesystem.

Common situations: Stats JSON not downloaded with the model checkpoint; wrong path after moving the model directory; container missing a volume mount that hosts the stats file; typo in the config key.

Related errors

• Cosmos3TokenizationStage requires a tokenizer; expected the…
• no safetensors files found in
• A GGUF encoder checkpoint cannot be combined with a second…
• A quantized checkpoint requires an in-tree native encoder…
• action_horizon must be a positive integer

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/8a2b850b955fb843. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/cosmos3_action.py:170

            "duration": f"{int(duration_seconds)}s",
            "fps": float(fps),
            "resolution": {"H": int(height), "W": int(width)},
            "aspect_ratio": canonical_aspect_ratio(int(width), int(height)),
        }
        prompts.append(json.dumps(prompt))

    if isinstance(description, (list, tuple)):
        return prompts
    return prompts[0]


def load_action_stats(
    stats_path: str, stats_key: str = "global"
) -> dict[str, torch.Tensor]:
    """Load per-channel action normalization stats from a JSON file."""
    path = Path(stats_path)
    if not path.exists():
        raise FileNotFoundError(
            f"Action normalization stats not found at {stats_path}."
        )
    raw = json.loads(path.read_text())
    if stats_key in raw:
        raw = raw[stats_key]
    return {
        k: torch.as_tensor(np.array(v, dtype=np.float32))
        for k, v in raw.items()
        if k in _STAT_KEYS
    }


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(

View on GitHub (pinned to 0132848349)