[SiddhantGahankari]

Reference Issues/PRs

Fixes #134.

What does this implement/fix? Explain your changes.

Move the quantile capability check in SktimeExecutor.predict() before fit() and predict(). Requests for unsupported quantiles are rejected without computing a point forecast that would be discarded.

• Move the existing capability:pred_int guard immediately after request conversion, outside the point-forecast error handler.
• Preserve the existing RuntimeError message and HTTP 400 handling.
• Keep point-only requests and supported quantile requests on their existing inference paths.
• Update the method docstring to reflect validation before inference.

The change is limited to src/tserve/runtime/executors/sktime/executor.py.

Does your contribution introduce a new dependency? If yes, which one?

No.

PR checklist for new estimators

N/A.

Any other comments?

Manual verification passed using the current executor directly, without a running HTTP server:

• Run the reproduction in [BUG] SktimeExecutor runs fit/predict before rejecting unsupported quantile requests #134 and confirm the same rejection message with fit calls: 0 and predict calls: 0.
• Confirm a point-only request still succeeds (TrendForecaster returned sales forecasts of [18.0, 20.0]).
• Confirm a quantile-capable model still returns quantile predictions (NaiveForecaster(strategy="drift") returned all requested quantiles: 0.1, 0.5, and 0.9, for both forecast steps).

[read-the-docs-community]

Documentation build overview

📚 tserve | 🛠️ Build #35212295 | 📁 Comparing 53ff556 against latest (fa06b9b)

  🔍 Preview build  

No files changed.

[SiddhantGahankari]

@geetu040 Can you review this? If any changes required, please tell.

[geetu040]

Thanks for the PR, looks good.