Unblock CI for 25.04 (#6519)

This PR addresses two issues that are currently blocking the cuML CI on
the 25.04 release branch:

1. Out-of-memory (OOM) errors occurring in SVM tests on CUDA 11.8:
- Several SVM-related tests, particularly `test_svc_methods`, are
failing with OOM errors and segmentation faults
- This only surfaces with CUDA 11.8 and is likely due to memory
allocation patterns
- As a temporary workaround, we skip these tests on CUDA 11.8 while the
root cause is investigated

2. XGBoost test dependency compatibility:
- XGBoost 3.0.0 has a known issue that manifests when using older NVIDIA
drivers with recent CUDA toolkit versions
(dmlc/xgboost#11397)
- To maintain stability, we constrain the XGBoost test dependency to
versions < 3.0.0
- This ensures consistent test behavior across different driver/toolkit
combinations

We expect to remove the constraint on the xgboost version once the issue
is resolved in a future xgboost release.

We expect to be able to address the SVM test issue by reducing its
memory footprint (see #6514),
however here we are taking a more conservative approach to ensure that
the CI pipeline is stable.

The remaining failing CI job is _optional_, the issue is going to be
addressed on branch-25.06.

Author: csadorf

conda/environments/all_cuda-118_arch-aarch64.yaml

@@ -77,5 +77,5 @@ dependencies:
 - sysroot_linux-aarch64==2.28
 - treelite==4.4.1
 - umap-learn==0.5.6
-- xgboost>=2.1.0
+- xgboost>=2.1.0,<3.0.0
 name: all_cuda-118_arch-aarch64

conda/environments/all_cuda-118_arch-x86_64.yaml

@@ -77,5 +77,5 @@ dependencies:
 - sysroot_linux-64==2.28
 - treelite==4.4.1
 - umap-learn==0.5.6
-- xgboost>=2.1.0
+- xgboost>=2.1.0,<3.0.0
 name: all_cuda-118_arch-x86_64

conda/environments/all_cuda-128_arch-aarch64.yaml

@@ -73,5 +73,5 @@ dependencies:
 - sysroot_linux-aarch64==2.28
 - treelite==4.4.1
 - umap-learn==0.5.6
-- xgboost>=2.1.0
+- xgboost>=2.1.0,<3.0.0
 name: all_cuda-128_arch-aarch64

conda/environments/all_cuda-128_arch-x86_64.yaml

@@ -73,5 +73,5 @@ dependencies:
 - sysroot_linux-64==2.28
 - treelite==4.4.1
 - umap-learn==0.5.6
-- xgboost>=2.1.0
+- xgboost>=2.1.0,<3.0.0
 name: all_cuda-128_arch-x86_64

dependencies.yaml

@@ -493,7 +493,7 @@ dependencies:
           - pytest-xdist
           - seaborn
           - *scikit_learn
-          - &xgboost xgboost>=2.1.0
+          - &xgboost xgboost>=2.1.0,<3.0.0
           - statsmodels
           - umap-learn==0.5.6
           - pynndescent

python/cuml/cuml/tests/test_device_selection.py

@@ -72,6 +72,7 @@
 from cuml.internals.safe_imports import cpu_only_import
 
 np = cpu_only_import("numpy")
+cp = gpu_only_import("cupy")
 pd = cpu_only_import("pandas")
 cudf = gpu_only_import("cudf")
 
@@ -1121,6 +1122,14 @@ def test_svc_methods(
     class_type,
     probability,
 ):
+    # Skip on CUDA 11.8 due to segfaults in decision_function
+    # See: https://github.com/rapidsai/cuml/issues/6480
+    if 11080 <= cp.cuda.runtime.runtimeGetVersion() < 11090:
+        pytest.skip(
+            "Skipping test_svc_methods on CUDA 11.8 due to segfaults in "
+            "decision_function (#6480)"
+        )
+
     if class_type == "single_class":
         X_train = X_train_class
         y_train = y_train_class

python/cuml/pyproject.toml

@@ -142,7 +142,7 @@ test = [
     "seaborn",
     "statsmodels",
     "umap-learn==0.5.6",
-    "xgboost>=2.1.0",
+    "xgboost>=2.1.0,<3.0.0",
 ] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`.
 
 [project.urls]