feat: add panel and multi-model CodSpeed benchmarks

.github/workflows/codspeed.yml

@@ -35,7 +35,7 @@ jobs:
         uses: CodSpeedHQ/action@v5
         with:
           mode: simulation
-          run: uv run pytest tests/benchmarks/test_core.py --codspeed -m benchmark -o addopts=
+          run: uv run pytest tests/benchmarks/test_core.py tests/benchmarks/test_orchestration.py --codspeed -m benchmark -o addopts=
 
   benchmarks-models:
     name: CodSpeed models

tests/benchmarks/conftest.py

@@ -12,6 +12,16 @@ def panel_df():
     )
 
 
+@pytest.fixture(scope="session")
+def large_panel_df():
+    return generate_series(
+        n_series=100,
+        freq="D",
+        min_length=500,
+        max_length=500,
+    )
+
+
 @pytest.fixture(scope="session")
 def chronos_bolt():
     from foundationforecast.models.chronos import Chronos

tests/benchmarks/test_core.py

@@ -12,9 +12,16 @@ def test_maybe_infer_freq(benchmark, panel_df):
     assert result == "D"
 
 
-def test_timeseries_dataset_from_df(benchmark, panel_df):
+@pytest.mark.parametrize(
+    "panel_df_fixture",
+    ["panel_df", "large_panel_df"],
+    ids=["small", "large"],
+)
+def test_timeseries_dataset_from_df(benchmark, panel_df_fixture, request):
+    df = request.getfixturevalue(panel_df_fixture)
+
     def build_dataset():
-        return TimeSeriesDataset.from_df(panel_df, batch_size=4)
+        return TimeSeriesDataset.from_df(df, batch_size=4)
 
     dataset = benchmark(build_dataset)
     assert len(dataset) > 0

tests/benchmarks/test_orchestration.py

@@ -0,0 +1,38 @@
+import pytest
+
+from foundationforecast import FoundationForecast
+from foundationforecast.core.forecaster import Forecaster
+from foundationforecast.core.utils import TimeSeriesDataset
+
+pytestmark = pytest.mark.benchmark
+
+
+class DatasetTouchingModel(Forecaster):
+    """Lightweight model that exercises dataset construction like real forecasters."""
+
+    batch_size = 32
+
+    def __init__(self, alias: str = "DatasetTouch"):
+        self.alias = alias
+
+    def forecast(self, df, h, freq=None, level=None, quantiles=None, **kwargs):
+        panel = kwargs.get("panel")
+        ds_kwargs = {"df": df, "batch_size": self.batch_size}
+        if panel is not None:
+            ds_kwargs["panel"] = panel
+        dataset = TimeSeriesDataset.from_df(**ds_kwargs)
+        fcst = dataset.make_future_dataframe(h=h, freq=freq or "D")
+        fcst[self.alias] = 1.0
+        return fcst
+
+
+@pytest.mark.parametrize("n_models", [3, 5], ids=["3-models", "5-models"])
+def test_foundation_forecast_multi_model(benchmark, large_panel_df, n_models):
+    models: list[Forecaster] = [
+        DatasetTouchingModel(alias=f"DatasetTouch{i}") for i in range(n_models)
+    ]
+    forecaster = FoundationForecast(models=models)
+    result = benchmark(forecaster.forecast, large_panel_df, h=12, freq="D")
+    assert len(result) == large_panel_df["unique_id"].nunique() * 12
+    for model in models:
+        assert model.alias in result.columns