diff --git a/ci/test_cuda_parallel_python.sh b/ci/test_cuda_parallel_python.sh
index d663db9457b..dbfbee3df22 100755
--- a/ci/test_cuda_parallel_python.sh
+++ b/ci/test_cuda_parallel_python.sh
@@ -20,4 +20,4 @@ python -m pip install "${CUDA_PARALLEL_WHEEL_PATH}[test]"
 
 # Run tests
 cd "/home/coder/cccl/python/cuda_parallel/tests/"
-python -m pytest -n auto -v
+python -m pytest -n 6 -v -m "not large"
diff --git a/python/cuda_parallel/pyproject.toml b/python/cuda_parallel/pyproject.toml
index 5234e5e6bb4..7c3da2d35cb 100644
--- a/python/cuda_parallel/pyproject.toml
+++ b/python/cuda_parallel/pyproject.toml
@@ -76,4 +76,7 @@ extend = "../../pyproject.toml"
 known-first-party = ["cuda.parallel"]
 
 [tool.pytest.ini_options]
-markers = ["no_verify_sass: skip SASS verification check"]
+markers = [
+  "no_verify_sass: skip SASS verification check",
+  "large: tests requiring large device memory allocations",
+]
diff --git a/python/cuda_parallel/tests/test_radix_sort.py b/python/cuda_parallel/tests/test_radix_sort.py
index 2262db5917d..a8a3706816e 100644
--- a/python/cuda_parallel/tests/test_radix_sort.py
+++ b/python/cuda_parallel/tests/test_radix_sort.py
@@ -12,6 +12,13 @@
 
 import cuda.parallel.experimental.algorithms as algorithms
 
+
+def get_mark(dt, log_size):
+    if log_size < 20:
+        return tuple()
+    return pytest.mark.large
+
+
 DTYPE_LIST = [
     np.uint8,
     np.uint16,
@@ -27,7 +34,11 @@
 
 PROBLEM_SIZES = [2, 10, 20]
 
-DTYPE_SIZE = [(dt, 2**log_size) for dt in DTYPE_LIST for log_size in PROBLEM_SIZES]
+DTYPE_SIZE = [
+    pytest.param(dt, 2**log_size, marks=get_mark(dt, log_size))
+    for dt in DTYPE_LIST
+    for log_size in PROBLEM_SIZES
+]
 
 
 def random_array(size, dtype, max_value=None) -> np.typing.NDArray:
@@ -251,7 +262,7 @@ def test_radix_sort_pairs_double_buffer(dtype, num_items):
 
 # These tests take longer to execute so we reduce the number of test cases
 DTYPE_SIZE_BIT_WINDOW = [
-    (dt, 2**log_size)
+    pytest.param(dt, 2**log_size, marks=get_mark(dt, log_size))
     for dt in [np.uint8, np.int16, np.uint32, np.int64, np.float64]
     for log_size in [2, 24]
 ]
diff --git a/python/cuda_parallel/tests/test_reduce.py b/python/cuda_parallel/tests/test_reduce.py
index 60536e134f9..c775d11116b 100644
--- a/python/cuda_parallel/tests/test_reduce.py
+++ b/python/cuda_parallel/tests/test_reduce.py
@@ -31,8 +31,14 @@ def type_to_problem_sizes(dtype):
         raise ValueError("Unsupported dtype")
 
 
+def get_mark(dt, log_size):
+    if log_size + np.log2(np.dtype(dt).itemsize) < 21:
+        return tuple()
+    return pytest.mark.large
+
+
 dtype_size_pairs = [
-    (dt, 2**log_size)
+    pytest.param(dt, 2**log_size, marks=get_mark(dt, log_size))
     for dt in [np.uint8, np.uint16, np.uint32, np.uint64]
     for log_size in type_to_problem_sizes(dt)
 ]
diff --git a/python/cuda_parallel/tests/test_unique_by_key.py b/python/cuda_parallel/tests/test_unique_by_key.py
index e55a64c414c..d258f7b6179 100644
--- a/python/cuda_parallel/tests/test_unique_by_key.py
+++ b/python/cuda_parallel/tests/test_unique_by_key.py
@@ -25,10 +25,19 @@
     np.float64,
 ]
 
+
+def get_mark(dt, log_size):
+    if log_size < 20:
+        return tuple()
+    return pytest.mark.large
+
+
 PROBLEM_SIZES = [2, 8, 16, 22]
 
 DTYPE_SIZE_PAIRS = [
-    (dt, 2**log_size) for dt in DTYPE_LIST for log_size in PROBLEM_SIZES
+    pytest.param(dt, 2**log_size, marks=get_mark(dt, log_size))
+    for dt in DTYPE_LIST
+    for log_size in PROBLEM_SIZES
 ]