Use RAPIDS 24.06 for the build (#109)

Author: seberg

.github/workflows/github-actions.yml

@@ -49,7 +49,8 @@ jobs:
           echo "  - python=3.10" >> environment-test-linux-cuda11.8-compilers-openmpi-ucx.yaml;
           mamba env create -n legate -f environment-test-linux-cuda11.8-compilers-openmpi-ucx.yaml;
           mamba activate legate;
-          mamba install -y build scikit-learn hypothesis 'pytest<8' notebook;
+          # Downgrade NumPy; as of 2024-07-01 it triggers a mypy bug
+          mamba install -y build scikit-learn hypothesis 'pytest<8' notebook 'numpy<2';
           pip install matplotlib seaborn xgboost
           export LEGATE_CORE_ARCH='arch-linux-py-relwithdebinfo';
           ./configure --with-python --cuda-arch 70 --with-cuda --build-type=relwithdebinfo --with-nccl --with-ucx && make && pip install .;

cmake/thirdparty/fetch_rapids.cmake

@@ -12,7 +12,7 @@
 # the License.
 
 # Use this variable to update RAPIDS and RAFT versions
-set(RAPIDS_VERSION "24.02")
+set(RAPIDS_VERSION "24.06")
 
 if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/RAFT_RAPIDS.cmake)
     file(DOWNLOAD https://raw.githubusercontent.com/rapidsai/rapids-cmake/branch-${RAPIDS_VERSION}/RAPIDS.cmake

legateboost/test/test_with_hypothesis.py

@@ -2,6 +2,7 @@
 from hypothesis import HealthCheck, Verbosity, assume, given, settings, strategies as st
 from sklearn.preprocessing import StandardScaler
 
+import cunumeric as cn
 import legateboost as lb
 from legate.core import TaskTarget, get_legate_runtime
 
@@ -25,11 +26,11 @@
 @st.composite
 def tree_strategy(draw):
     if get_legate_runtime().machine.count(TaskTarget.GPU) > 0:
-        max_depth = draw(st.integers(1, 12))
+        max_depth = draw(st.integers(1, 8))
     else:
         max_depth = draw(st.integers(1, 6))
     alpha = draw(st.floats(0.0, 1.0))
-    split_samples = draw(st.integers(1, 1000))
+    split_samples = draw(st.integers(1, 500))
     return lb.models.Tree(max_depth=max_depth, alpha=alpha, split_samples=split_samples)
 
 
@@ -143,6 +144,7 @@ def regression_dataset_strategy(draw):
     regression_param_strategy,
     regression_dataset_strategy(),
 )
+@cn.errstate(divide="raise", invalid="raise")
 def test_regressor(model_params, regression_params, regression_dataset):
     X, y, w = regression_dataset
     eval_result = {}
@@ -156,7 +158,7 @@ def test_regressor(model_params, regression_params, regression_dataset):
     )
     model.predict(X)
     loss = next(iter(eval_result["train"].values()))
-    assert non_increasing(loss, tol=1e-2)
+    assert non_increasing(loss, tol=1e-1)
     sanity_check_models(model)
 
 
@@ -238,12 +240,12 @@ def classification_dataset_strategy(draw):
     classification_param_strategy,
     classification_dataset_strategy(),
 )
+@cn.errstate(divide="raise", invalid="raise")
 def test_classifier(
     model_params: dict, classification_params: dict, classification_dataset: tuple
 ) -> None:
     X, y, w, name = classification_dataset
     eval_result = {}
-    model_params["n_estimators"] = 3
     model = lb.LBClassifier(**model_params, **classification_params).fit(
         X, y, sample_weight=w, eval_result=eval_result
     )
@@ -253,4 +255,4 @@ def test_classifier(
     loss = next(iter(eval_result["train"].values()))
     # multiclass models with higher learning rates don't always converge
     if len(model.classes_) == 2:
-        assert non_increasing(loss, 1e-2)
+        assert non_increasing(loss, 1e-1)

legateboost/utils.py

@@ -92,6 +92,7 @@ def preround(x: cn.ndarray) -> cn.ndarray:
     m = cn.sum(cn.abs(x))
     n = x.size
     delta = cn.floor(m / (one - two * n * eps))
+    delta = cn.maximum(delta, one)
     M = two ** cn.ceil(cn.log2(delta))
     return (x + M) - M