Add CodSpeed continuous benchmarking

[codspeed-hq]

This sets up continuous performance measurement with CodSpeed for the library's core forecasting pipeline.

What was added

Benchmarks (benchmarks/) — 23 benchmarks run with pytest-codspeed. They deliberately avoid downloading or running model weights: the models used (benchmarks/_models.py, a constant forecaster and a seasonal-naive forecaster) are trivial, so what is measured is FoundationForecast's own data plumbing, which is what every model call pays for.

• test_bench_core_utils.py: input validation, frequency inference, ds/y dtype conversion, seasonality lookup, and both QuantileConverter conversions (levels <-> quantile columns).
• test_bench_dataset.py: TimeSeriesDataset.from_df (two batch sizes), batch iteration, make_future_dataframe, and the _DataProcessor tensor path (left padding/stacking of variable-length contexts, and missing-value imputation).
• test_bench_pipelines.py: FoundationForecast.forecast (single model, with prediction levels, four merged models, and on shuffled input that forces sorting and frequency inference), cross_validation with 1 and 4 windows, and detect_anomalies.

Fixtures in benchmarks/conftest.py generate deterministic panels (20 series x 200 rows, 200 series x 500 rows) with utilsforecast.data.generate_series, which is seeded by default so measurements are stable run to run.

pytest testpaths is still ["tests"], so the benchmarks do not run as part of the normal test suite; they are invoked explicitly by path.

CI (.github/workflows/codspeed.yaml) — runs on pushes to main, on pull requests, and via workflow_dispatch (used by CodSpeed for backtests). It mirrors the existing CI job: astral-sh/setup-uv@v6 with caching plus the same disk-cleanup step, since the dependency set (torch, gluonts and the model packages) is large. Benchmarks run through CodSpeedHQ/action@v5 in simulation mode on ubuntu-latest, authenticated with OIDC (id-token: write), so no token secret is required.

Other — pytest-codspeed added to the dev dependency group (lockfile updated), and a CodSpeed badge added to the README.

Verification

The benchmarks were run locally with codspeed run --mode simulation -- uv run pytest benchmarks/ --codspeed; all 23 benchmarks were measured successfully. ruff check, ruff format and mypy (same settings as the pre-commit hooks) pass on the new files, and uv run pytest tests/core still passes.

Next steps

• Merge to main so CodSpeed records a baseline; subsequent pull requests will then get automatic performance reports.
• If you later want to track the model inference paths themselves, those benchmarks would need cached weights and would be better suited to a separate, scheduled workflow.

[codspeed-hq]

Congrats! CodSpeed is installed 🎉

🆕 23 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• test_get_seasonality[D]: 514.5 µs
• test_get_seasonality[h]: 218.4 µs
• test_get_seasonality[MS]: 239.8 µs
• test_maybe_convert_col_to_datetime_from_string: 252.7 ms
• test_maybe_convert_col_to_float32: 22.1 ms
• test_maybe_infer_freq: 98.5 ms
• test_quantile_converter_from_level: 100.5 µs
• test_quantile_converter_level_to_quantiles: 63.2 ms
• test_quantile_converter_quantiles_to_level: 55 ms
• test_validate_input: 84.9 µs
• test_data_processor_impute_missing: 55.7 ms
• test_data_processor_left_pad_and_stack: 7 ms
• test_time_series_dataset_from_df[128]: 393.4 ms
• test_time_series_dataset_from_df[16]: 394.1 ms
• test_time_series_dataset_iteration: 112.3 µs
• test_time_series_dataset_make_future_dataframe: 14.5 ms
• test_cross_validation[1]: 186.8 ms
• test_cross_validation[4]: 676 ms
• test_detect_anomalies: 529.3 ms
• test_forecast_infer_freq_and_sort: 153.2 ms
• ...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

---