google-research/timesfm · error · RuntimeError

Model is not compiled. Please call compile() first.

Error message

Model is not compiled. Please call compile() first.

What it means

`forecast()` requires the model to have been compiled (`compile()` sets `compiled_decode`); since `compiled_decode` is None, forecasting cannot proceed. TimesFM compiles the JAX decode kernel up front for fast, shape-specialized decoding, so forecasting without compilation is unsupported. The library raises RuntimeError to enforce the required call order: load checkpoint -> compile -> forecast.

Source

Thrown at src/timesfm/timesfm_2p5/timesfm_2p5_base.py:160

  forecast_config: ForecastConfig | None = None
  compiled_decode: Callable[..., Any] | None = None
  global_batch_size: int = 0

  def load_checkpoint(self, path: str):
    """Loads a TimesFM model from a checkpoint."""
    raise NotImplementedError()

  def compile(self, forecast_config: ForecastConfig | None = None):
    """Compiles the TimesFM model for fast decoding."""
    raise NotImplementedError()

  def forecast(
    self, horizon: int, inputs: list[np.ndarray]
  ) -> tuple[np.ndarray, np.ndarray]:
    """Forecasts the time series."""
    if self.compiled_decode is None:
      raise RuntimeError("Model is not compiled. Please call compile() first.")

    assert self.global_batch_size > 0
    assert self.forecast_config is not None

    context = self.forecast_config.max_context
    num_inputs = len(inputs)
    if (w := num_inputs % self.global_batch_size) != 0:
      inputs += [np.array([0.0] * 3)] * (self.global_batch_size - w)

    output_points = []
    output_quantiles = []
    values = []
    masks = []
    idx = 0
    for each_input in inputs:
      value = linear_interpolation(strip_leading_nans(np.array(each_input)))
      if (w := len(value)) >= context:
        value = value[-context:]

View on GitHub (pinned to 331c6d33cb)

Solutions

1. Call `model.compile(forecast_config)` before the first `forecast()` call
2. If the process restarted, re-create the model, re-run compile, then forecast
3. If you intended covariates, call `forecast_with_covariates` after compile with return_backcast=True

Example fix

// before
model = TimesFMFlax(...); model.load_checkpoint(path)
point, quantiles = model.forecast(96, inputs)
// after
model = TimesFMFlax(...); model.load_checkpoint(path)
model.compile(forecast_config=ForecastConfig(max_context=512, max_horizon=96, return_backcast=False))
point, quantiles = model.forecast(96, inputs)

if getattr(model, 'compiled_decode', True) is None:
    model.compile(forecast_config=ForecastConfig(max_context=512, max_horizon=horizon))

def is_compiled(model) -> bool:
    return getattr(model, 'compiled_decode', None) is not None

try:
    point, quantiles = model.forecast(horizon, inputs)
except RuntimeError as e:
    if 'not compiled' in str(e):
        model.compile(forecast_config=ForecastConfig(max_context=512, max_horizon=horizon))
        point, quantiles = model.forecast(horizon, inputs)
    else:
        raise

When it happens

Trigger: Calling `model.forecast(horizon, inputs)` (directly or via `forecast_with_covariates`) before ever calling `compile()`; constructing the model, loading a checkpoint, and immediately forecasting; re-instantiating the model in a new process and forgetting to recompile (compiled state is not persisted in the checkpoint).

Common situations: Notebook cells executed out of order after a kernel restart; a script that only loads the checkpoint; code paths that skip compile because an older TimesFM version forecasted without compilation.

Related errors

• For XReg, `return_backcast` must be set to True in the forec
• At least one of dynamic_numerical_covariates, dynamic_catego
• Unsupported mode: {xreg_mode}
• Forecast horizon length inferred from the dynamic covariates
• Context + horizon must be less than the context limit. {fc.m

AI-assisted analysis of google-research/timesfm@331c6d33cb (2026-08-29). Data as JSON: /api/errors/8bb5c0d28c618cee. Report an issue: GitHub.