dask-cuda 23.12.0a231027__tar.gz → 24.2.0a3__tar.gz

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/MANIFEST.in

@@ -1 +1,2 @@
 include dask_cuda/_version.py
+include dask_cuda/VERSION

{dask-cuda-23.12.0a231027/dask_cuda.egg-info → dask-cuda-24.2.0a3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: dask-cuda
-Version: 23.12.0a231027
+Version: 24.2.0a3
 Summary: Utilities for Dask and CUDA interactions
 Author: NVIDIA Corporation
 License: Apache-2.0
@@ -17,12 +17,11 @@ Classifier: Programming Language :: Python :: 3.10
 Requires-Python: >=3.9
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: dask==2023.9.2
-Requires-Dist: distributed==2023.9.2
 Requires-Dist: pynvml<11.5,>=11.0.0
 Requires-Dist: numpy>=1.21
 Requires-Dist: numba>=0.57
 Requires-Dist: pandas<1.6.0dev0,>=1.3
+Requires-Dist: rapids-dask-dependency==23.12.*,>=0.0.0a0
 Requires-Dist: zict>=2.0.0
 Provides-Extra: docs
 Requires-Dist: numpydoc>=1.1.0; extra == "docs"

dask-cuda-24.2.0a3/dask_cuda/VERSION

@@ -0,0 +1 @@
+24.02.00a3

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/__init__.py

@@ -11,6 +11,7 @@ import dask.dataframe.shuffle
 import dask.dataframe.multi
 import dask.bag.core
 
+from ._version import __git_commit__, __version__
 from .cuda_worker import CUDAWorker
 from .explicit_comms.dataframe.shuffle import (
     get_rearrange_by_column_wrapper,
@@ -19,9 +20,6 @@ from .explicit_comms.dataframe.shuffle import (
 from .local_cuda_cluster import LocalCUDACluster
 from .proxify_device_objects import proxify_decorator, unproxify_decorator
 
-__version__ = "23.12.00"
-
-from . import compat
 
 # Monkey patching Dask to make use of explicit-comms when `DASK_EXPLICIT_COMMS=True`
 dask.dataframe.shuffle.rearrange_by_column = get_rearrange_by_column_wrapper(

dask-cuda-24.2.0a3/dask_cuda/_version.py

@@ -0,0 +1,20 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib.resources
+
+__version__ = (
+    importlib.resources.files("dask_cuda").joinpath("VERSION").read_text().strip()
+)
+__git_commit__ = ""

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/benchmarks/local_cudf_groupby.py

@@ -139,7 +139,7 @@ def pretty_print_results(args, address_to_index, p2p_bw, results):
             key="Device memory limit", value=f"{format_bytes(args.device_memory_limit)}"
         )
     print_key_value(key="RMM Pool", value=f"{not args.disable_rmm_pool}")
-    if args.protocol == "ucx":
+    if args.protocol in ["ucx", "ucxx"]:
         print_key_value(key="TCP", value=f"{args.enable_tcp_over_ucx}")
         print_key_value(key="InfiniBand", value=f"{args.enable_infiniband}")
         print_key_value(key="NVLink", value=f"{args.enable_nvlink}")

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/benchmarks/local_cudf_merge.py

@@ -217,7 +217,7 @@ def pretty_print_results(args, address_to_index, p2p_bw, results):
         )
     print_key_value(key="RMM Pool", value=f"{not args.disable_rmm_pool}")
     print_key_value(key="Frac-match", value=f"{args.frac_match}")
-    if args.protocol == "ucx":
+    if args.protocol in ["ucx", "ucxx"]:
         print_key_value(key="TCP", value=f"{args.enable_tcp_over_ucx}")
         print_key_value(key="InfiniBand", value=f"{args.enable_infiniband}")
         print_key_value(key="NVLink", value=f"{args.enable_nvlink}")

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/benchmarks/local_cudf_shuffle.py

@@ -146,7 +146,7 @@ def pretty_print_results(args, address_to_index, p2p_bw, results):
             key="Device memory limit", value=f"{format_bytes(args.device_memory_limit)}"
         )
     print_key_value(key="RMM Pool", value=f"{not args.disable_rmm_pool}")
-    if args.protocol == "ucx":
+    if args.protocol in ["ucx", "ucxx"]:
         print_key_value(key="TCP", value=f"{args.enable_tcp_over_ucx}")
         print_key_value(key="InfiniBand", value=f"{args.enable_infiniband}")
         print_key_value(key="NVLink", value=f"{args.enable_nvlink}")

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/benchmarks/local_cupy.py

@@ -193,7 +193,7 @@ def pretty_print_results(args, address_to_index, p2p_bw, results):
         )
     print_key_value(key="RMM Pool", value=f"{not args.disable_rmm_pool}")
     print_key_value(key="Protocol", value=f"{args.protocol}")
-    if args.protocol == "ucx":
+    if args.protocol in ["ucx", "ucxx"]:
         print_key_value(key="TCP", value=f"{args.enable_tcp_over_ucx}")
         print_key_value(key="InfiniBand", value=f"{args.enable_infiniband}")
         print_key_value(key="NVLink", value=f"{args.enable_nvlink}")

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/benchmarks/local_cupy_map_overlap.py

@@ -78,7 +78,7 @@ def pretty_print_results(args, address_to_index, p2p_bw, results):
         )
     print_key_value(key="RMM Pool", value=f"{not args.disable_rmm_pool}")
     print_key_value(key="Protocol", value=f"{args.protocol}")
-    if args.protocol == "ucx":
+    if args.protocol in ["ucx", "ucxx"]:
         print_key_value(key="TCP", value=f"{args.enable_tcp_over_ucx}")
         print_key_value(key="InfiniBand", value=f"{args.enable_infiniband}")
         print_key_value(key="NVLink", value=f"{args.enable_nvlink}")

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/benchmarks/utils.py

@@ -73,7 +73,7 @@ def parse_benchmark_args(description="Generic dask-cuda Benchmark", args_list=[]
     cluster_args.add_argument(
         "-p",
         "--protocol",
-        choices=["tcp", "ucx"],
+        choices=["tcp", "ucx", "ucxx"],
         default="tcp",
         type=str,
         help="The communication protocol to use.",

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/device_host_file.py

@@ -17,7 +17,7 @@ from distributed.protocol import (
     serialize_bytelist,
 )
 from distributed.sizeof import safe_sizeof
-from distributed.spill import CustomFile as KeyAsStringFile
+from distributed.spill import AnyKeyFile as KeyAsStringFile
 from distributed.utils import nbytes
 
 from .is_device_object import is_device_object

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/initialize.py

@@ -5,7 +5,6 @@ import click
 import numba.cuda
 
 import dask
-import distributed.comm.ucx
 from distributed.diagnostics.nvml import get_device_index_and_uuid, has_cuda_context
 
 from .utils import get_ucx_config
@@ -23,12 +22,21 @@ def _create_cuda_context_handler():
         numba.cuda.current_context()
 
 
-def _create_cuda_context():
+def _create_cuda_context(protocol="ucx"):
+    if protocol not in ["ucx", "ucxx"]:
+        return
     try:
         # Added here to ensure the parent `LocalCUDACluster` process creates the CUDA
         # context directly from the UCX module, thus avoiding a similar warning there.
         try:
-            distributed.comm.ucx.init_once()
+            if protocol == "ucx":
+                import distributed.comm.ucx
+
+                distributed.comm.ucx.init_once()
+            elif protocol == "ucxx":
+                import distributed_ucxx.ucxx
+
+                distributed_ucxx.ucxx.init_once()
         except ModuleNotFoundError:
             # UCX initialization has to be delegated to Distributed, it will take care
             # of setting correct environment variables and importing `ucp` after that.
@@ -39,20 +47,35 @@ def _create_cuda_context():
             os.environ.get("CUDA_VISIBLE_DEVICES", "0").split(",")[0]
         )
         ctx = has_cuda_context()
-        if (
-            ctx.has_context
-            and not distributed.comm.ucx.cuda_context_created.has_context
-        ):
-            distributed.comm.ucx._warn_existing_cuda_context(ctx, os.getpid())
+        if protocol == "ucx":
+            if (
+                ctx.has_context
+                and not distributed.comm.ucx.cuda_context_created.has_context
+            ):
+                distributed.comm.ucx._warn_existing_cuda_context(ctx, os.getpid())
+        elif protocol == "ucxx":
+            if (
+                ctx.has_context
+                and not distributed_ucxx.ucxx.cuda_context_created.has_context
+            ):
+                distributed_ucxx.ucxx._warn_existing_cuda_context(ctx, os.getpid())
 
         _create_cuda_context_handler()
 
-        if not distributed.comm.ucx.cuda_context_created.has_context:
-            ctx = has_cuda_context()
-            if ctx.has_context and ctx.device_info != cuda_visible_device:
-                distributed.comm.ucx._warn_cuda_context_wrong_device(
-                    cuda_visible_device, ctx.device_info, os.getpid()
-                )
+        if protocol == "ucx":
+            if not distributed.comm.ucx.cuda_context_created.has_context:
+                ctx = has_cuda_context()
+                if ctx.has_context and ctx.device_info != cuda_visible_device:
+                    distributed.comm.ucx._warn_cuda_context_wrong_device(
+                        cuda_visible_device, ctx.device_info, os.getpid()
+                    )
+        elif protocol == "ucxx":
+            if not distributed_ucxx.ucxx.cuda_context_created.has_context:
+                ctx = has_cuda_context()
+                if ctx.has_context and ctx.device_info != cuda_visible_device:
+                    distributed_ucxx.ucxx._warn_cuda_context_wrong_device(
+                        cuda_visible_device, ctx.device_info, os.getpid()
+                    )
 
     except Exception:
         logger.error("Unable to start CUDA Context", exc_info=True)
@@ -64,6 +87,7 @@ def initialize(
     enable_infiniband=None,
     enable_nvlink=None,
     enable_rdmacm=None,
+    protocol="ucx",
 ):
     """Create CUDA context and initialize UCX-Py, depending on user parameters.
 
@@ -118,7 +142,7 @@ def initialize(
     dask.config.set({"distributed.comm.ucx": ucx_config})
 
     if create_cuda_context:
-        _create_cuda_context()
+        _create_cuda_context(protocol=protocol)
 
 
 @click.command()
@@ -127,6 +151,12 @@ def initialize(
     default=False,
     help="Create CUDA context",
 )
+@click.option(
+    "--protocol",
+    default=None,
+    type=str,
+    help="Communication protocol, such as: 'tcp', 'tls', 'ucx' or 'ucxx'.",
+)
 @click.option(
     "--enable-tcp-over-ucx/--disable-tcp-over-ucx",
     default=False,
@@ -150,10 +180,11 @@ def initialize(
 def dask_setup(
     service,
     create_cuda_context,
+    protocol,
     enable_tcp_over_ucx,
     enable_infiniband,
     enable_nvlink,
     enable_rdmacm,
 ):
     if create_cuda_context:
-        _create_cuda_context()
+        _create_cuda_context(protocol=protocol)

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/local_cuda_cluster.py

@@ -319,8 +319,11 @@ class LocalCUDACluster(LocalCluster):
         if enable_tcp_over_ucx or enable_infiniband or enable_nvlink:
             if protocol is None:
                 protocol = "ucx"
-            elif protocol != "ucx":
-                raise TypeError("Enabling InfiniBand or NVLink requires protocol='ucx'")
+            elif protocol not in ["ucx", "ucxx"]:
+                raise TypeError(
+                    "Enabling InfiniBand or NVLink requires protocol='ucx' or "
+                    "protocol='ucxx'"
+                )
 
         self.host = kwargs.get("host", None)
 
@@ -371,7 +374,7 @@ class LocalCUDACluster(LocalCluster):
         ) + ["dask_cuda.initialize"]
         self.new_spec["options"]["preload_argv"] = self.new_spec["options"].get(
             "preload_argv", []
-        ) + ["--create-cuda-context"]
+        ) + ["--create-cuda-context", "--protocol", protocol]
 
         self.cuda_visible_devices = CUDA_VISIBLE_DEVICES
         self.scale(n_workers)

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/tests/test_dgx.py

@@ -73,10 +73,13 @@ def test_default():
     assert not p.exitcode
 
 
-def _test_tcp_over_ucx():
-    ucp = pytest.importorskip("ucp")
+def _test_tcp_over_ucx(protocol):
+    if protocol == "ucx":
+        ucp = pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        ucp = pytest.importorskip("ucxx")
 
-    with LocalCUDACluster(enable_tcp_over_ucx=True) as cluster:
+    with LocalCUDACluster(protocol=protocol, enable_tcp_over_ucx=True) as cluster:
         with Client(cluster) as client:
             res = da.from_array(numpy.arange(10000), chunks=(1000,))
             res = res.sum().compute()
@@ -93,10 +96,17 @@ def _test_tcp_over_ucx():
             assert all(client.run(check_ucx_options).values())
 
 
-def test_tcp_over_ucx():
-    ucp = pytest.importorskip("ucp")  # NOQA: F841
+@pytest.mark.parametrize(
+    "protocol",
+    ["ucx", "ucxx"],
+)
+def test_tcp_over_ucx(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
 
-    p = mp.Process(target=_test_tcp_over_ucx)
+    p = mp.Process(target=_test_tcp_over_ucx, args=(protocol,))
     p.start()
     p.join()
     assert not p.exitcode
@@ -117,9 +127,22 @@ def test_tcp_only():
     assert not p.exitcode
 
 
-def _test_ucx_infiniband_nvlink(enable_infiniband, enable_nvlink, enable_rdmacm):
+def _test_ucx_infiniband_nvlink(
+    skip_queue, protocol, enable_infiniband, enable_nvlink, enable_rdmacm
+):
     cupy = pytest.importorskip("cupy")
-    ucp = pytest.importorskip("ucp")
+    if protocol == "ucx":
+        ucp = pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        ucp = pytest.importorskip("ucxx")
+
+    if enable_infiniband and not any(
+        [at.startswith("rc") for at in ucp.get_active_transports()]
+    ):
+        skip_queue.put("No support available for 'rc' transport in UCX")
+        return
+    else:
+        skip_queue.put("ok")
 
     if enable_infiniband is None and enable_nvlink is None and enable_rdmacm is None:
         enable_tcp_over_ucx = None
@@ -135,6 +158,7 @@ def _test_ucx_infiniband_nvlink(enable_infiniband, enable_nvlink, enable_rdmacm)
             cm_tls_priority = ["tcp"]
 
     initialize(
+        protocol=protocol,
         enable_tcp_over_ucx=enable_tcp_over_ucx,
         enable_infiniband=enable_infiniband,
         enable_nvlink=enable_nvlink,
@@ -142,6 +166,7 @@ def _test_ucx_infiniband_nvlink(enable_infiniband, enable_nvlink, enable_rdmacm)
     )
 
     with LocalCUDACluster(
+        protocol=protocol,
         interface="ib0",
         enable_tcp_over_ucx=enable_tcp_over_ucx,
         enable_infiniband=enable_infiniband,
@@ -171,6 +196,7 @@ def _test_ucx_infiniband_nvlink(enable_infiniband, enable_nvlink, enable_rdmacm)
             assert all(client.run(check_ucx_options).values())
 
 
+@pytest.mark.parametrize("protocol", ["ucx", "ucxx"])
 @pytest.mark.parametrize(
     "params",
     [
@@ -185,16 +211,19 @@ def _test_ucx_infiniband_nvlink(enable_infiniband, enable_nvlink, enable_rdmacm)
     _get_dgx_version() == DGXVersion.DGX_A100,
     reason="Automatic InfiniBand device detection Unsupported for %s" % _get_dgx_name(),
 )
-def test_ucx_infiniband_nvlink(params):
-    ucp = pytest.importorskip("ucp")  # NOQA: F841
+def test_ucx_infiniband_nvlink(protocol, params):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
 
-    if params["enable_infiniband"]:
-        if not any([at.startswith("rc") for at in ucp.get_active_transports()]):
-            pytest.skip("No support available for 'rc' transport in UCX")
+    skip_queue = mp.Queue()
 
     p = mp.Process(
         target=_test_ucx_infiniband_nvlink,
         args=(
+            skip_queue,
+            protocol,
             params["enable_infiniband"],
             params["enable_nvlink"],
             params["enable_rdmacm"],
@@ -203,9 +232,8 @@ def test_ucx_infiniband_nvlink(params):
     p.start()
     p.join()
 
-    # Starting a new cluster on the same pytest process after an rdmacm cluster
-    # has been used may cause UCX-Py to complain about being already initialized.
-    if params["enable_rdmacm"] is True:
-        ucp.reset()
+    skip_msg = skip_queue.get()
+    if skip_msg != "ok":
+        pytest.skip(skip_msg)
 
     assert not p.exitcode

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/tests/test_explicit_comms.py

@@ -44,7 +44,7 @@ def _test_local_cluster(protocol):
             assert sum(c.run(my_rank, 0)) == sum(range(4))
 
 
-@pytest.mark.parametrize("protocol", ["tcp", "ucx"])
+@pytest.mark.parametrize("protocol", ["tcp", "ucx", "ucxx"])
 def test_local_cluster(protocol):
     p = mp.Process(target=_test_local_cluster, args=(protocol,))
     p.start()
@@ -160,7 +160,7 @@ def _test_dataframe_shuffle(backend, protocol, n_workers, _partitions):
 
 @pytest.mark.parametrize("nworkers", [1, 2, 3])
 @pytest.mark.parametrize("backend", ["pandas", "cudf"])
-@pytest.mark.parametrize("protocol", ["tcp", "ucx"])
+@pytest.mark.parametrize("protocol", ["tcp", "ucx", "ucxx"])
 @pytest.mark.parametrize("_partitions", [True, False])
 def test_dataframe_shuffle(backend, protocol, nworkers, _partitions):
     if backend == "cudf":
@@ -256,7 +256,7 @@ def _test_dataframe_shuffle_merge(backend, protocol, n_workers):
 
 @pytest.mark.parametrize("nworkers", [1, 2, 4])
 @pytest.mark.parametrize("backend", ["pandas", "cudf"])
-@pytest.mark.parametrize("protocol", ["tcp", "ucx"])
+@pytest.mark.parametrize("protocol", ["tcp", "ucx", "ucxx"])
 def test_dataframe_shuffle_merge(backend, protocol, nworkers):
     if backend == "cudf":
         pytest.importorskip("cudf")
@@ -293,7 +293,7 @@ def _test_jit_unspill(protocol):
             assert_eq(got, expected)
 
 
-@pytest.mark.parametrize("protocol", ["tcp", "ucx"])
+@pytest.mark.parametrize("protocol", ["tcp", "ucx", "ucxx"])
 def test_jit_unspill(protocol):
     pytest.importorskip("cudf")
 

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/tests/test_from_array.py

@@ -5,12 +5,16 @@ from distributed import Client
 
 from dask_cuda import LocalCUDACluster
 
-pytest.importorskip("ucp")
 cupy = pytest.importorskip("cupy")
 
 
-@pytest.mark.parametrize("protocol", ["ucx", "tcp"])
+@pytest.mark.parametrize("protocol", ["ucx", "ucxx", "tcp"])
 def test_ucx_from_array(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
+
     N = 10_000
     with LocalCUDACluster(protocol=protocol) as cluster:
         with Client(cluster):

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/tests/test_initialize.py

@@ -13,7 +13,6 @@ from dask_cuda.utils import get_ucx_config
 from dask_cuda.utils_test import IncreasedCloseTimeoutNanny
 
 mp = mp.get_context("spawn")  # type: ignore
-ucp = pytest.importorskip("ucp")
 
 # Notice, all of the following tests is executed in a new process such
 # that UCX options of the different tests doesn't conflict.
@@ -21,11 +20,16 @@ ucp = pytest.importorskip("ucp")
 # of UCX before retrieving the current config.
 
 
-def _test_initialize_ucx_tcp():
+def _test_initialize_ucx_tcp(protocol):
+    if protocol == "ucx":
+        ucp = pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        ucp = pytest.importorskip("ucxx")
+
     kwargs = {"enable_tcp_over_ucx": True}
-    initialize(**kwargs)
+    initialize(protocol=protocol, **kwargs)
     with LocalCluster(
-        protocol="ucx",
+        protocol=protocol,
         dashboard_address=None,
         n_workers=1,
         threads_per_worker=1,
@@ -50,18 +54,29 @@ def _test_initialize_ucx_tcp():
             assert all(client.run(check_ucx_options).values())
 
 
-def test_initialize_ucx_tcp():
-    p = mp.Process(target=_test_initialize_ucx_tcp)
+@pytest.mark.parametrize("protocol", ["ucx", "ucxx"])
+def test_initialize_ucx_tcp(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
+
+    p = mp.Process(target=_test_initialize_ucx_tcp, args=(protocol,))
     p.start()
     p.join()
     assert not p.exitcode
 
 
-def _test_initialize_ucx_nvlink():
+def _test_initialize_ucx_nvlink(protocol):
+    if protocol == "ucx":
+        ucp = pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        ucp = pytest.importorskip("ucxx")
+
     kwargs = {"enable_nvlink": True}
-    initialize(**kwargs)
+    initialize(protocol=protocol, **kwargs)
     with LocalCluster(
-        protocol="ucx",
+        protocol=protocol,
         dashboard_address=None,
         n_workers=1,
         threads_per_worker=1,
@@ -87,18 +102,29 @@ def _test_initialize_ucx_nvlink():
             assert all(client.run(check_ucx_options).values())
 
 
-def test_initialize_ucx_nvlink():
-    p = mp.Process(target=_test_initialize_ucx_nvlink)
+@pytest.mark.parametrize("protocol", ["ucx", "ucxx"])
+def test_initialize_ucx_nvlink(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
+
+    p = mp.Process(target=_test_initialize_ucx_nvlink, args=(protocol,))
     p.start()
     p.join()
     assert not p.exitcode
 
 
-def _test_initialize_ucx_infiniband():
+def _test_initialize_ucx_infiniband(protocol):
+    if protocol == "ucx":
+        ucp = pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        ucp = pytest.importorskip("ucxx")
+
     kwargs = {"enable_infiniband": True}
-    initialize(**kwargs)
+    initialize(protocol=protocol, **kwargs)
     with LocalCluster(
-        protocol="ucx",
+        protocol=protocol,
         dashboard_address=None,
         n_workers=1,
         threads_per_worker=1,
@@ -127,17 +153,28 @@ def _test_initialize_ucx_infiniband():
 @pytest.mark.skipif(
     "ib0" not in psutil.net_if_addrs(), reason="Infiniband interface ib0 not found"
 )
-def test_initialize_ucx_infiniband():
-    p = mp.Process(target=_test_initialize_ucx_infiniband)
+@pytest.mark.parametrize("protocol", ["ucx", "ucxx"])
+def test_initialize_ucx_infiniband(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
+
+    p = mp.Process(target=_test_initialize_ucx_infiniband, args=(protocol,))
     p.start()
     p.join()
     assert not p.exitcode
 
 
-def _test_initialize_ucx_all():
-    initialize()
+def _test_initialize_ucx_all(protocol):
+    if protocol == "ucx":
+        ucp = pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        ucp = pytest.importorskip("ucxx")
+
+    initialize(protocol=protocol)
     with LocalCluster(
-        protocol="ucx",
+        protocol=protocol,
         dashboard_address=None,
         n_workers=1,
         threads_per_worker=1,
@@ -166,8 +203,14 @@ def _test_initialize_ucx_all():
             assert all(client.run(check_ucx_options).values())
 
 
-def test_initialize_ucx_all():
-    p = mp.Process(target=_test_initialize_ucx_all)
+@pytest.mark.parametrize("protocol", ["ucx", "ucxx"])
+def test_initialize_ucx_all(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
+
+    p = mp.Process(target=_test_initialize_ucx_all, args=(protocol,))
     p.start()
     p.join()
     assert not p.exitcode

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/tests/test_local_cuda_cluster.py

@@ -87,23 +87,38 @@ async def test_with_subset_of_cuda_visible_devices():
                 }
 
 
+@pytest.mark.parametrize(
+    "protocol",
+    ["ucx", "ucxx"],
+)
 @gen_test(timeout=20)
-async def test_ucx_protocol():
-    pytest.importorskip("ucp")
+async def test_ucx_protocol(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
 
     async with LocalCUDACluster(
-        protocol="ucx", asynchronous=True, data=dict
+        protocol=protocol, asynchronous=True, data=dict
     ) as cluster:
         assert all(
-            ws.address.startswith("ucx://") for ws in cluster.scheduler.workers.values()
+            ws.address.startswith(f"{protocol}://")
+            for ws in cluster.scheduler.workers.values()
         )
 
 
+@pytest.mark.parametrize(
+    "protocol",
+    ["ucx", "ucxx"],
+)
 @gen_test(timeout=20)
-async def test_explicit_ucx_with_protocol_none():
-    pytest.importorskip("ucp")
+async def test_explicit_ucx_with_protocol_none(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
 
-    initialize(enable_tcp_over_ucx=True)
+    initialize(protocol=protocol, enable_tcp_over_ucx=True)
     async with LocalCUDACluster(
         protocol=None, enable_tcp_over_ucx=True, asynchronous=True, data=dict
     ) as cluster:
@@ -113,11 +128,18 @@ async def test_explicit_ucx_with_protocol_none():
 
 
 @pytest.mark.filterwarnings("ignore:Exception ignored in")
+@pytest.mark.parametrize(
+    "protocol",
+    ["ucx", "ucxx"],
+)
 @gen_test(timeout=20)
-async def test_ucx_protocol_type_error():
-    pytest.importorskip("ucp")
+async def test_ucx_protocol_type_error(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
 
-    initialize(enable_tcp_over_ucx=True)
+    initialize(protocol=protocol, enable_tcp_over_ucx=True)
     with pytest.raises(TypeError):
         async with LocalCUDACluster(
             protocol="tcp", enable_tcp_over_ucx=True, asynchronous=True, data=dict
@@ -337,6 +359,7 @@ async def test_pre_import():
 
 
 # Intentionally not using @gen_test to skip cleanup checks
+@pytest.mark.xfail(reason="https://github.com/rapidsai/dask-cuda/issues/1265")
 def test_pre_import_not_found():
     async def _test_pre_import_not_found():
         with raises_with_cause(RuntimeError, None, ImportError, None):
@@ -477,20 +500,30 @@ async def test_worker_fraction_limits():
             )
 
 
-def test_print_cluster_config(capsys):
+@pytest.mark.parametrize(
+    "protocol",
+    ["ucx", "ucxx"],
+)
+def test_print_cluster_config(capsys, protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
+
     pytest.importorskip("rich")
     with LocalCUDACluster(
-        n_workers=1, device_memory_limit="1B", jit_unspill=True, protocol="ucx"
+        n_workers=1, device_memory_limit="1B", jit_unspill=True, protocol=protocol
     ) as cluster:
         with Client(cluster) as client:
             print_cluster_config(client)
             captured = capsys.readouterr()
             assert "Dask Cluster Configuration" in captured.out
-            assert "ucx" in captured.out
+            assert protocol in captured.out
             assert "1 B" in captured.out
             assert "[plugin]" in captured.out
 
 
+@pytest.mark.xfail(reason="https://github.com/rapidsai/dask-cuda/issues/1265")
 def test_death_timeout_raises():
     with pytest.raises(asyncio.exceptions.TimeoutError):
         with LocalCUDACluster(

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/tests/test_proxy.py

@@ -400,10 +400,14 @@ class _PxyObjTest(proxy_object.ProxyObject):
 
 
 @pytest.mark.parametrize("send_serializers", [None, ("dask", "pickle"), ("cuda",)])
-@pytest.mark.parametrize("protocol", ["tcp", "ucx"])
+@pytest.mark.parametrize("protocol", ["tcp", "ucx", "ucxx"])
 @gen_test(timeout=120)
 async def test_communicating_proxy_objects(protocol, send_serializers):
     """Testing serialization of cuDF dataframe when communicating"""
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
     cudf = pytest.importorskip("cudf")
 
     def task(x):
@@ -412,7 +416,7 @@ async def test_communicating_proxy_objects(protocol, send_serializers):
         serializers_used = x._pxy_get().serializer
 
         # Check that `x` is serialized with the expected serializers
-        if protocol == "ucx":
+        if protocol in ["ucx", "ucxx"]:
             if send_serializers is None:
                 assert serializers_used == "cuda"
             else:
@@ -443,11 +447,15 @@ async def test_communicating_proxy_objects(protocol, send_serializers):
             await client.submit(task, df)
 
 
-@pytest.mark.parametrize("protocol", ["tcp", "ucx"])
+@pytest.mark.parametrize("protocol", ["tcp", "ucx", "ucxx"])
 @pytest.mark.parametrize("shared_fs", [True, False])
 @gen_test(timeout=20)
 async def test_communicating_disk_objects(protocol, shared_fs):
     """Testing disk serialization of cuDF dataframe when communicating"""
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
     cudf = pytest.importorskip("cudf")
     ProxifyHostFile._spill_to_disk.shared_filesystem = shared_fs
 

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/tests/test_utils.py

@@ -79,11 +79,18 @@ def test_get_device_total_memory():
             assert total_mem > 0
 
 
-def test_get_preload_options_default():
-    pytest.importorskip("ucp")
+@pytest.mark.parametrize(
+    "protocol",
+    ["ucx", "ucxx"],
+)
+def test_get_preload_options_default(protocol):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
 
     opts = get_preload_options(
-        protocol="ucx",
+        protocol=protocol,
         create_cuda_context=True,
     )
 
@@ -93,14 +100,21 @@ def test_get_preload_options_default():
     assert opts["preload_argv"] == ["--create-cuda-context"]
 
 
+@pytest.mark.parametrize(
+    "protocol",
+    ["ucx", "ucxx"],
+)
 @pytest.mark.parametrize("enable_tcp", [True, False])
 @pytest.mark.parametrize("enable_infiniband", [True, False])
 @pytest.mark.parametrize("enable_nvlink", [True, False])
-def test_get_preload_options(enable_tcp, enable_infiniband, enable_nvlink):
-    pytest.importorskip("ucp")
+def test_get_preload_options(protocol, enable_tcp, enable_infiniband, enable_nvlink):
+    if protocol == "ucx":
+        pytest.importorskip("ucp")
+    elif protocol == "ucxx":
+        pytest.importorskip("ucxx")
 
     opts = get_preload_options(
-        protocol="ucx",
+        protocol=protocol,
         create_cuda_context=True,
         enable_tcp_over_ucx=enable_tcp,
         enable_infiniband=enable_infiniband,

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda/utils.py

@@ -287,7 +287,7 @@ def get_preload_options(
     if create_cuda_context:
         preload_options["preload_argv"].append("--create-cuda-context")
 
-    if protocol == "ucx":
+    if protocol in ["ucx", "ucxx"]:
         initialize_ucx_argv = []
         if enable_tcp_over_ucx:
             initialize_ucx_argv.append("--enable-tcp-over-ucx")
@@ -625,6 +625,10 @@ def get_worker_config(dask_worker):
         import ucp
 
         ret["ucx-transports"] = ucp.get_active_transports()
+    elif scheme == "ucxx":
+        import ucxx
+
+        ret["ucx-transports"] = ucxx.get_active_transports()
 
     # comm timeouts
     ret["distributed.comm.timeouts"] = dask.config.get("distributed.comm.timeouts")

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3/dask_cuda.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: dask-cuda
-Version: 23.12.0a231027
+Version: 24.2.0a3
 Summary: Utilities for Dask and CUDA interactions
 Author: NVIDIA Corporation
 License: Apache-2.0
@@ -17,12 +17,11 @@ Classifier: Programming Language :: Python :: 3.10
 Requires-Python: >=3.9
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: dask==2023.9.2
-Requires-Dist: distributed==2023.9.2
 Requires-Dist: pynvml<11.5,>=11.0.0
 Requires-Dist: numpy>=1.21
 Requires-Dist: numba>=0.57
 Requires-Dist: pandas<1.6.0dev0,>=1.3
+Requires-Dist: rapids-dask-dependency==23.12.*,>=0.0.0a0
 Requires-Dist: zict>=2.0.0
 Provides-Extra: docs
 Requires-Dist: numpydoc>=1.1.0; extra == "docs"

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda.egg-info/SOURCES.txt

@@ -3,9 +3,10 @@ MANIFEST.in
 README.md
 pyproject.toml
 setup.py
+dask_cuda/VERSION
 dask_cuda/__init__.py
+dask_cuda/_version.py
 dask_cuda/cli.py
-dask_cuda/compat.py
 dask_cuda/cuda_worker.py
 dask_cuda/device_host_file.py
 dask_cuda/disk_io.py

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/dask_cuda.egg-info/requires.txt

@@ -1,9 +1,8 @@
-dask==2023.9.2
-distributed==2023.9.2
 pynvml<11.5,>=11.0.0
 numpy>=1.21
 numba>=0.57
 pandas<1.6.0dev0,>=1.3
+rapids-dask-dependency==23.12.*,>=0.0.0a0
 zict>=2.0.0
 
 [docs]

{dask-cuda-23.12.0a231027 → dask-cuda-24.2.0a3}/pyproject.toml

@@ -7,7 +7,7 @@ requires = [
 
 [project]
 name = "dask-cuda"
-version = "23.12.00a231027"
+dynamic = ["version"]
 description = "Utilities for Dask and CUDA interactions"
 readme = { file = "README.md", content-type = "text/markdown" }
 authors = [
@@ -16,12 +16,11 @@ authors = [
 license = { text = "Apache-2.0" }
 requires-python = ">=3.9"
 dependencies = [
-    "dask ==2023.9.2",
-    "distributed ==2023.9.2",
     "pynvml >=11.0.0,<11.5",
     "numpy >=1.21",
     "numba >=0.57",
     "pandas >=1.3,<1.6.0dev0",
+    "rapids-dask-dependency==23.12.*,>=0.0.0a0",
     "zict >=2.0.0",
 ]
 classifiers = [
@@ -123,6 +122,9 @@ filterwarnings = [
 [tool.setuptools]
 license-files = ["LICENSE"]
 
+[tool.setuptools.dynamic]
+version = {file = "dask_cuda/VERSION"}
+
 [tool.setuptools.packages.find]
 exclude = [
     "docs",

dask-cuda-23.12.0a231027/dask_cuda/compat.py

@@ -1,118 +0,0 @@
-import pickle
-
-import msgpack
-from packaging.version import Version
-
-import dask
-import distributed
-import distributed.comm.utils
-import distributed.protocol
-from distributed.comm.utils import OFFLOAD_THRESHOLD, nbytes, offload
-from distributed.protocol.core import (
-    Serialized,
-    decompress,
-    logger,
-    merge_and_deserialize,
-    msgpack_decode_default,
-    msgpack_opts,
-)
-
-if Version(distributed.__version__) >= Version("2023.8.1"):
-    # Monkey-patch protocol.core.loads (and its users)
-    async def from_frames(
-        frames, deserialize=True, deserializers=None, allow_offload=True
-    ):
-        """
-        Unserialize a list of Distributed protocol frames.
-        """
-        size = False
-
-        def _from_frames():
-            try:
-                # Patched code
-                return loads(
-                    frames, deserialize=deserialize, deserializers=deserializers
-                )
-                # end patched code
-            except EOFError:
-                if size > 1000:
-                    datastr = "[too large to display]"
-                else:
-                    datastr = frames
-                # Aid diagnosing
-                logger.error("truncated data stream (%d bytes): %s", size, datastr)
-                raise
-
-        if allow_offload and deserialize and OFFLOAD_THRESHOLD:
-            size = sum(map(nbytes, frames))
-        if (
-            allow_offload
-            and deserialize
-            and OFFLOAD_THRESHOLD
-            and size > OFFLOAD_THRESHOLD
-        ):
-            res = await offload(_from_frames)
-        else:
-            res = _from_frames()
-
-        return res
-
-    def loads(frames, deserialize=True, deserializers=None):
-        """Transform bytestream back into Python value"""
-
-        allow_pickle = dask.config.get("distributed.scheduler.pickle")
-
-        try:
-
-            def _decode_default(obj):
-                offset = obj.get("__Serialized__", 0)
-                if offset > 0:
-                    sub_header = msgpack.loads(
-                        frames[offset],
-                        object_hook=msgpack_decode_default,
-                        use_list=False,
-                        **msgpack_opts,
-                    )
-                    offset += 1
-                    sub_frames = frames[offset : offset + sub_header["num-sub-frames"]]
-                    if deserialize:
-                        if "compression" in sub_header:
-                            sub_frames = decompress(sub_header, sub_frames)
-                        return merge_and_deserialize(
-                            sub_header, sub_frames, deserializers=deserializers
-                        )
-                    else:
-                        return Serialized(sub_header, sub_frames)
-
-                offset = obj.get("__Pickled__", 0)
-                if offset > 0:
-                    sub_header = msgpack.loads(frames[offset])
-                    offset += 1
-                    sub_frames = frames[offset : offset + sub_header["num-sub-frames"]]
-                    # Patched code
-                    if "compression" in sub_header:
-                        sub_frames = decompress(sub_header, sub_frames)
-                    # end patched code
-                    if allow_pickle:
-                        return pickle.loads(
-                            sub_header["pickled-obj"], buffers=sub_frames
-                        )
-                    else:
-                        raise ValueError(
-                            "Unpickle on the Scheduler isn't allowed, "
-                            "set `distributed.scheduler.pickle=true`"
-                        )
-
-                return msgpack_decode_default(obj)
-
-            return msgpack.loads(
-                frames[0], object_hook=_decode_default, use_list=False, **msgpack_opts
-            )
-
-        except Exception:
-            logger.critical("Failed to deserialize", exc_info=True)
-            raise
-
-    distributed.protocol.loads = loads
-    distributed.protocol.core.loads = loads
-    distributed.comm.utils.from_frames = from_frames