google-research/timesfm · error · ValueError

Continuous quantile head is not supported for horizons > {se

Error message

Continuous quantile head is not supported for horizons > {self.model.os}.

What it means

The continuous quantile head (used when `use_continuous_quantile_head=True`) can only emit quantiles up to the model's trained output horizon `self.model.os`. If the compiled `max_horizon` exceeds `os`, quantile outputs beyond that are undefined, so compile() raises ValueError. This is a model-capability boundary, not a generic config error.

Source

Thrown at src/timesfm/timesfm_2p5/timesfm_2p5_flax.py:539

        new_context := math.ceil(fc.max_context / self.model.p) * self.model.p,
      )
      fc = dataclasses.replace(fc, max_context=new_context)
    if fc.max_horizon % self.model.o != 0:
      logging.info(
        "When compiling, max horizon needs to be multiple of the output patch"
        " size %d. Using max horizon = %d instead.",
        self.model.o,
        new_horizon := math.ceil(fc.max_horizon / self.model.o) * self.model.o,
      )
      fc = dataclasses.replace(fc, max_horizon=new_horizon)
    if fc.max_context + fc.max_horizon > self.model.config.context_limit:
      raise ValueError(
        "Context + horizon must be less than the context limit."
        f" {fc.max_context} + {fc.max_horizon} >"
        f" {self.model.config.context_limit}."
      )
    if fc.use_continuous_quantile_head and (fc.max_horizon > self.model.os):
      raise ValueError(
        f"Continuous quantile head is not supported for horizons > {self.model.os}."
      )

    self.forecast_config = fc
    self.model.compile(
      context=self.forecast_config.max_context,
      horizon=self.forecast_config.max_horizon,
      per_core_batch_size=fc.per_core_batch_size,
    )
    self.per_core_batch_size = self.forecast_config.per_core_batch_size
    self.num_devices = self.model.num_devices
    self.global_batch_size = (
      self.forecast_config.per_core_batch_size * self.model.num_devices
    )

    def compiled_decode_kernel(fc, horizon, inputs, masks):
      inputs = jnp.array(inputs, dtype=jnp.float32)
      masks = jnp.array(masks, dtype=jnp.bool)

View on GitHub (pinned to 331c6d33cb)

Solutions

1. Reduce `max_horizon` to at most `model.os` when using the continuous quantile head
2. Set `use_continuous_quantile_head=False` if you need longer horizons (quantiles come from the standard quantiles output instead)
3. Check `model.os` for your checkpoint before configuring

Example fix

// before
model.compile(forecast_config=ForecastConfig(max_context=512, max_horizon=256, use_continuous_quantile_head=True))
// after
model.compile(forecast_config=ForecastConfig(max_context=512, max_horizon=model.os, use_continuous_quantile_head=True))

if fc.use_continuous_quantile_head and fc.max_horizon > model.os:
    fc = dataclasses.replace(fc, max_horizon=model.os)  # or disable the quantile head
model.compile(forecast_config=fc)

When it happens

Trigger: Calling `model.compile()` with `ForecastConfig(use_continuous_quantile_head=True, max_horizon=os+1 or larger)`; requesting long-horizon continuous quantiles on a model variant whose quantile head supports only os steps.

Common situations: Enabling the quantile head to get smooth quantile forecasts but also needing a long horizon; copying a config that used a bigger model variant with a larger os; not rounding max_horizon down (note compile rounds max_horizon up to patch multiples before this check).

Related errors

• At least one of dynamic_numerical_covariates, dynamic_catego
• Forecast horizon length inferred from the dynamic covariates
• Context + horizon must be less than the context limit. {fc.m
• Horizon must be less than the max horizon. {horizon} > {fc.m
• train_dynamic_numerical_covariates and test_dynamic_numerica

AI-assisted analysis of google-research/timesfm@331c6d33cb (2026-08-29). Data as JSON: /api/errors/b33cf73fb194fec5. Report an issue: GitHub.