dask-cuda 24.4.0__py3-none-any.whl → 24.8.2__py3-none-any.whl

dask_cuda/VERSION

@@ -1 +1 @@
-24.04.00
+24.08.02

dask_cuda/__init__.py

@@ -20,6 +20,18 @@ from .local_cuda_cluster import LocalCUDACluster
 from .proxify_device_objects import proxify_decorator, unproxify_decorator
 
 
+if dask.config.get("dataframe.query-planning", None) is not False and dask.config.get(
+    "explicit-comms", False
+):
+    raise NotImplementedError(
+        "The 'explicit-comms' config is not yet supported when "
+        "query-planning is enabled in dask. Please use the shuffle "
+        "API directly, or use the legacy dask-dataframe API "
+        "(set the 'dataframe.query-planning' config to `False`"
+        "before importing `dask.dataframe`).",
+    )
+
+
 # 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.dataframe.shuffle.rearrange_by_column

dask_cuda/_version.py

@@ -15,6 +15,16 @@
 import importlib.resources
 
 __version__ = (
-    importlib.resources.files("dask_cuda").joinpath("VERSION").read_text().strip()
+    importlib.resources.files(__package__).joinpath("VERSION").read_text().strip()
 )
-__git_commit__ = "7ed67359e2264393e97ad57fc39829c3a69aa1f3"
+try:
+    __git_commit__ = (
+        importlib.resources.files(__package__)
+        .joinpath("GIT_COMMIT")
+        .read_text()
+        .strip()
+    )
+except FileNotFoundError:
+    __git_commit__ = ""
+
+__all__ = ["__git_commit__", "__version__"]

dask_cuda/benchmarks/common.py

@@ -117,16 +117,18 @@ def run(client: Client, args: Namespace, config: Config):
     wait_for_cluster(client, shutdown_on_failure=True)
     assert len(client.scheduler_info()["workers"]) > 0
     setup_memory_pools(
-        client,
-        args.type == "gpu",
-        args.rmm_pool_size,
-        args.disable_rmm_pool,
-        args.enable_rmm_async,
-        args.enable_rmm_managed,
-        args.rmm_release_threshold,
-        args.rmm_log_directory,
-        args.enable_rmm_statistics,
-        args.enable_rmm_track_allocations,
+        client=client,
+        is_gpu=args.type == "gpu",
+        disable_rmm=args.disable_rmm,
+        disable_rmm_pool=args.disable_rmm_pool,
+        pool_size=args.rmm_pool_size,
+        maximum_pool_size=args.rmm_maximum_pool_size,
+        rmm_async=args.enable_rmm_async,
+        rmm_managed=args.enable_rmm_managed,
+        release_threshold=args.rmm_release_threshold,
+        log_directory=args.rmm_log_directory,
+        statistics=args.enable_rmm_statistics,
+        rmm_track_allocations=args.enable_rmm_track_allocations,
     )
     address_to_index, results, message_data = gather_bench_results(client, args, config)
     p2p_bw = peer_to_peer_bandwidths(message_data, address_to_index)

dask_cuda/benchmarks/local_cudf_merge.py

@@ -7,8 +7,7 @@ import numpy as np
 import pandas as pd
 
 import dask
-from dask.base import tokenize
-from dask.dataframe.core import new_dd_object
+import dask.dataframe as dd
 from dask.distributed import performance_report, wait
 from dask.utils import format_bytes, parse_bytes
 
@@ -25,12 +24,20 @@ from dask_cuda.benchmarks.utils import (
 # <https://gist.github.com/rjzamora/0ffc35c19b5180ab04bbf7c793c45955>
 
 
-def generate_chunk(i_chunk, local_size, num_chunks, chunk_type, frac_match, gpu):
+# Set default shuffle method to "tasks"
+if dask.config.get("dataframe.shuffle.method", None) is None:
+    dask.config.set({"dataframe.shuffle.method": "tasks"})
+
+
+def generate_chunk(input):
+    i_chunk, local_size, num_chunks, chunk_type, frac_match, gpu = input
+
     # Setting a seed that triggers max amount of comm in the two-GPU case.
     if gpu:
         import cupy as xp
 
         import cudf as xdf
+        import dask_cudf  # noqa: F401
     else:
         import numpy as xp
         import pandas as xdf
@@ -105,25 +112,25 @@ def get_random_ddf(chunk_size, num_chunks, frac_match, chunk_type, args):
 
     parts = [chunk_size for _ in range(num_chunks)]
     device_type = True if args.type == "gpu" else False
-    meta = generate_chunk(0, 4, 1, chunk_type, None, device_type)
+    meta = generate_chunk((0, 4, 1, chunk_type, None, device_type))
     divisions = [None] * (len(parts) + 1)
 
-    name = "generate-data-" + tokenize(chunk_size, num_chunks, frac_match, chunk_type)
-
-    graph = {
-        (name, i): (
-            generate_chunk,
-            i,
-            part,
-            len(parts),
-            chunk_type,
-            frac_match,
-            device_type,
-        )
-        for i, part in enumerate(parts)
-    }
-
-    ddf = new_dd_object(graph, name, meta, divisions)
+    ddf = dd.from_map(
+        generate_chunk,
+        [
+            (
+                i,
+                part,
+                len(parts),
+                chunk_type,
+                frac_match,
+                device_type,
+            )
+            for i, part in enumerate(parts)
+        ],
+        meta=meta,
+        divisions=divisions,
+    )
 
     if chunk_type == "build":
         if not args.no_shuffle:

dask_cuda/benchmarks/local_cudf_shuffle.py

@@ -8,8 +8,6 @@ import pandas as pd
 
 import dask
 import dask.dataframe
-from dask.dataframe.core import new_dd_object
-from dask.dataframe.shuffle import shuffle
 from dask.distributed import Client, performance_report, wait
 from dask.utils import format_bytes, parse_bytes
 
@@ -33,7 +31,7 @@ except ImportError:
 
 
 def shuffle_dask(df, args):
-    result = shuffle(df, index="data", shuffle="tasks", ignore_index=args.ignore_index)
+    result = df.shuffle("data", shuffle_method="tasks", ignore_index=args.ignore_index)
     if args.backend == "dask-noop":
         result = as_noop(result)
     t1 = perf_counter()
@@ -94,18 +92,24 @@ def create_data(
         )
 
     # Create partition based to the specified partition distribution
-    dsk = {}
+    futures = []
     for i, part_size in enumerate(dist):
         for _ in range(part_size):
             # We use `client.submit` to control placement of the partition.
-            dsk[(name, len(dsk))] = client.submit(
-                create_df, chunksize, args.type, workers=[workers[i]], pure=False
+            futures.append(
+                client.submit(
+                    create_df, chunksize, args.type, workers=[workers[i]], pure=False
+                )
             )
-    wait(dsk.values())
+    wait(futures)
 
     df_meta = create_df(0, args.type)
-    divs = [None] * (len(dsk) + 1)
-    ret = new_dd_object(dsk, name, df_meta, divs).persist()
+    divs = [None] * (len(futures) + 1)
+    ret = dask.dataframe.from_delayed(
+        futures,
+        meta=df_meta,
+        divisions=divs,
+    ).persist()
     wait(ret)
 
     data_processed = args.in_parts * args.partition_size
@@ -254,7 +258,9 @@ def parse_args():
     ]
 
     return parse_benchmark_args(
-        description="Distributed shuffle (dask/cudf) benchmark", args_list=special_args
+        description="Distributed shuffle (dask/cudf) benchmark",
+        args_list=special_args,
+        check_explicit_comms=False,
     )
 
 

dask_cuda/benchmarks/utils.py

@@ -11,11 +11,13 @@ from typing import Any, Callable, Mapping, NamedTuple, Optional, Tuple
 import numpy as np
 import pandas as pd
 
+from dask import config
 from dask.distributed import Client, SSHCluster
 from dask.utils import format_bytes, format_time, parse_bytes
 from distributed.comm.addressing import get_address_host
 
 from dask_cuda.local_cuda_cluster import LocalCUDACluster
+from dask_cuda.utils import parse_device_memory_limit
 
 
 def as_noop(dsk):
@@ -47,7 +49,11 @@ def as_noop(dsk):
         raise RuntimeError("Requested noop computation but dask-noop not installed.")
 
 
-def parse_benchmark_args(description="Generic dask-cuda Benchmark", args_list=[]):
+def parse_benchmark_args(
+    description="Generic dask-cuda Benchmark",
+    args_list=[],
+    check_explicit_comms=True,
+):
     parser = argparse.ArgumentParser(description=description)
     worker_args = parser.add_argument_group(description="Worker configuration")
     worker_args.add_argument(
@@ -88,15 +94,41 @@ def parse_benchmark_args(description="Generic dask-cuda Benchmark", args_list=[]
         "'forkserver' can be used to avoid issues with fork not being allowed "
         "after the networking stack has been initialised.",
     )
+    cluster_args.add_argument(
+        "--disable-rmm",
+        action="store_true",
+        help="Disable RMM.",
+    )
+    cluster_args.add_argument(
+        "--disable-rmm-pool",
+        action="store_true",
+        help="Uses RMM for allocations but without a memory pool.",
+    )
     cluster_args.add_argument(
         "--rmm-pool-size",
         default=None,
         type=parse_bytes,
         help="The size of the RMM memory pool. Can be an integer (bytes) or a string "
-        "(like '4GB' or '5000M'). By default, 1/2 of the total GPU memory is used.",
+        "(like '4GB' or '5000M'). By default, 1/2 of the total GPU memory is used."
+        ""
+        ".. note::"
+        "    This size is a per-worker configuration, and not cluster-wide.",
     )
     cluster_args.add_argument(
-        "--disable-rmm-pool", action="store_true", help="Disable the RMM memory pool"
+        "--rmm-maximum-pool-size",
+        default=None,
+        help="When ``--rmm-pool-size`` is specified, this argument indicates the "
+        "maximum pool size.  Can be an integer (bytes), or a string (like '4GB' or "
+        "'5000M'). By default, the total available memory on the GPU is used. "
+        "``rmm_pool_size`` must be specified to use RMM pool and to set the maximum "
+        "pool size."
+        ""
+        ".. note::"
+        "    When paired with `--enable-rmm-async` the maximum size cannot be "
+        "    guaranteed due to fragmentation."
+        ""
+        ".. note::"
+        "    This size is a per-worker configuration, and not cluster-wide.",
     )
     cluster_args.add_argument(
         "--enable-rmm-managed",
@@ -317,6 +349,24 @@ def parse_benchmark_args(description="Generic dask-cuda Benchmark", args_list=[]
     if args.multi_node and len(args.hosts.split(",")) < 2:
         raise ValueError("--multi-node requires at least 2 hosts")
 
+    # Raise error early if "explicit-comms" is not allowed
+    if (
+        check_explicit_comms
+        and args.backend == "explicit-comms"
+        and config.get(
+            "dataframe.query-planning",
+            None,
+        )
+        is not False
+    ):
+        raise NotImplementedError(
+            "The 'explicit-comms' config is not yet supported when "
+            "query-planning is enabled in dask. Please use the legacy "
+            "dask-dataframe API by setting the following environment "
+            "variable before executing:",
+            "    DASK_DATAFRAME__QUERY_PLANNING=False",
+        )
+
     return args
 
 
@@ -384,10 +434,29 @@ def get_worker_device():
         return -1
 
 
+def setup_rmm_resources(statistics=False, rmm_track_allocations=False):
+    import cupy
+
+    import rmm
+    from rmm.allocators.cupy import rmm_cupy_allocator
+
+    cupy.cuda.set_allocator(rmm_cupy_allocator)
+    if statistics:
+        rmm.mr.set_current_device_resource(
+            rmm.mr.StatisticsResourceAdaptor(rmm.mr.get_current_device_resource())
+        )
+    if rmm_track_allocations:
+        rmm.mr.set_current_device_resource(
+            rmm.mr.TrackingResourceAdaptor(rmm.mr.get_current_device_resource())
+        )
+
+
 def setup_memory_pool(
     dask_worker=None,
+    disable_rmm=None,
+    disable_rmm_pool=None,
     pool_size=None,
-    disable_pool=False,
+    maximum_pool_size=None,
     rmm_async=False,
     rmm_managed=False,
     release_threshold=None,
@@ -395,45 +464,66 @@ def setup_memory_pool(
     statistics=False,
     rmm_track_allocations=False,
 ):
-    import cupy
-
     import rmm
-    from rmm.allocators.cupy import rmm_cupy_allocator
 
     from dask_cuda.utils import get_rmm_log_file_name
 
     logging = log_directory is not None
 
-    if rmm_async:
-        rmm.mr.set_current_device_resource(
-            rmm.mr.CudaAsyncMemoryResource(
-                initial_pool_size=pool_size, release_threshold=release_threshold
-            )
-        )
-    else:
-        rmm.reinitialize(
-            pool_allocator=not disable_pool,
-            managed_memory=rmm_managed,
-            initial_pool_size=pool_size,
-            logging=logging,
-            log_file_name=get_rmm_log_file_name(dask_worker, logging, log_directory),
-        )
-    cupy.cuda.set_allocator(rmm_cupy_allocator)
-    if statistics:
-        rmm.mr.set_current_device_resource(
-            rmm.mr.StatisticsResourceAdaptor(rmm.mr.get_current_device_resource())
+    if pool_size is not None:
+        pool_size = parse_device_memory_limit(pool_size, alignment_size=256)
+
+    if maximum_pool_size is not None:
+        maximum_pool_size = parse_device_memory_limit(
+            maximum_pool_size, alignment_size=256
         )
-    if rmm_track_allocations:
-        rmm.mr.set_current_device_resource(
-            rmm.mr.TrackingResourceAdaptor(rmm.mr.get_current_device_resource())
+
+    if release_threshold is not None:
+        release_threshold = parse_device_memory_limit(
+            release_threshold, alignment_size=256
         )
 
+    if not disable_rmm:
+        if rmm_async:
+            mr = rmm.mr.CudaAsyncMemoryResource(
+                initial_pool_size=pool_size,
+                release_threshold=release_threshold,
+            )
+
+            if maximum_pool_size is not None:
+                mr = rmm.mr.LimitingResourceAdaptor(
+                    mr, allocation_limit=maximum_pool_size
+                )
+
+            rmm.mr.set_current_device_resource(mr)
+
+            setup_rmm_resources(
+                statistics=statistics, rmm_track_allocations=rmm_track_allocations
+            )
+        else:
+            rmm.reinitialize(
+                pool_allocator=not disable_rmm_pool,
+                managed_memory=rmm_managed,
+                initial_pool_size=pool_size,
+                maximum_pool_size=maximum_pool_size,
+                logging=logging,
+                log_file_name=get_rmm_log_file_name(
+                    dask_worker, logging, log_directory
+                ),
+            )
+
+            setup_rmm_resources(
+                statistics=statistics, rmm_track_allocations=rmm_track_allocations
+            )
+
 
 def setup_memory_pools(
     client,
     is_gpu,
+    disable_rmm,
+    disable_rmm_pool,
     pool_size,
-    disable_pool,
+    maximum_pool_size,
     rmm_async,
     rmm_managed,
     release_threshold,
@@ -445,8 +535,10 @@ def setup_memory_pools(
         return
     client.run(
         setup_memory_pool,
+        disable_rmm=disable_rmm,
+        disable_rmm_pool=disable_rmm_pool,
         pool_size=pool_size,
-        disable_pool=disable_pool,
+        maximum_pool_size=maximum_pool_size,
         rmm_async=rmm_async,
         rmm_managed=rmm_managed,
         release_threshold=release_threshold,
@@ -459,7 +551,9 @@ def setup_memory_pools(
     client.run_on_scheduler(
         setup_memory_pool,
         pool_size=1e9,
-        disable_pool=disable_pool,
+        disable_rmm=disable_rmm,
+        disable_rmm_pool=disable_rmm_pool,
+        maximum_pool_size=maximum_pool_size,
         rmm_async=rmm_async,
         rmm_managed=rmm_managed,
         release_threshold=release_threshold,

dask_cuda/cli.py

@@ -101,6 +101,20 @@ def cuda():
     total device memory), string (like ``"5GB"`` or ``"5000M"``), or ``"auto"`` or 0 to
     disable spilling to host (i.e. allow full device memory usage).""",
 )
+@click.option(
+    "--enable-cudf-spill/--disable-cudf-spill",
+    default=False,
+    show_default=True,
+    help="""Enable automatic cuDF spilling. WARNING: This should NOT be used with
+    JIT-Unspill.""",
+)
+@click.option(
+    "--cudf-spill-stats",
+    type=int,
+    default=0,
+    help="""Set the cuDF spilling statistics level. This option has no effect if
+    `--enable-cudf-spill` is not specified.""",
+)
 @click.option(
     "--rmm-pool-size",
     default=None,
@@ -120,6 +134,10 @@ def cuda():
     memory on the GPU is used. ``rmm_pool_size`` must be specified to use RMM pool and
     to set the maximum pool size.
 
+    .. note::
+        When paired with `--enable-rmm-async` the maximum size cannot be guaranteed due
+        to fragmentation.
+
     .. note::
         This size is a per-worker configuration, and not cluster-wide.""",
 )
@@ -326,6 +344,8 @@ def worker(
     name,
     memory_limit,
     device_memory_limit,
+    enable_cudf_spill,
+    cudf_spill_stats,
     rmm_pool_size,
     rmm_maximum_pool_size,
     rmm_managed_memory,
@@ -398,6 +418,8 @@ def worker(
             name,
             memory_limit,
             device_memory_limit,
+            enable_cudf_spill,
+            cudf_spill_stats,
             rmm_pool_size,
             rmm_maximum_pool_size,
             rmm_managed_memory,

dask_cuda/cuda_worker.py

@@ -20,7 +20,7 @@ from distributed.worker_memory import parse_memory_limit
 
 from .device_host_file import DeviceHostFile
 from .initialize import initialize
-from .plugins import CPUAffinity, PreImport, RMMSetup
+from .plugins import CPUAffinity, CUDFSetup, PreImport, RMMSetup
 from .proxify_host_file import ProxifyHostFile
 from .utils import (
     cuda_visible_devices,
@@ -41,6 +41,8 @@ class CUDAWorker(Server):
         name=None,
         memory_limit="auto",
         device_memory_limit="auto",
+        enable_cudf_spill=False,
+        cudf_spill_stats=0,
         rmm_pool_size=None,
         rmm_maximum_pool_size=None,
         rmm_managed_memory=False,
@@ -166,6 +168,12 @@ class CUDAWorker(Server):
         if device_memory_limit is None and memory_limit is None:
             data = lambda _: {}
         elif jit_unspill:
+            if enable_cudf_spill:
+                warnings.warn(
+                    "Enabling cuDF spilling and JIT-Unspill together is not "
+                    "safe, consider disabling JIT-Unspill."
+                )
+
             data = lambda i: (
                 ProxifyHostFile,
                 {
@@ -217,6 +225,7 @@ class CUDAWorker(Server):
                         track_allocations=rmm_track_allocations,
                     ),
                     PreImport(pre_import),
+                    CUDFSetup(spill=enable_cudf_spill, spill_stats=cudf_spill_stats),
                 },
                 name=name if nprocs == 1 or name is None else str(name) + "-" + str(i),
                 local_directory=local_directory,

dask_cuda/explicit_comms/dataframe/shuffle.py

@@ -8,13 +8,18 @@ from math import ceil
 from operator import getitem
 from typing import Any, Callable, Dict, List, Optional, Set, TypeVar
 
+import numpy as np
+import pandas as pd
+
 import dask
 import dask.config
 import dask.dataframe
+import dask.dataframe as dd
 import dask.utils
 import distributed.worker
 from dask.base import tokenize
-from dask.dataframe.core import DataFrame, Series, _concat as dd_concat, new_dd_object
+from dask.dataframe import DataFrame, Series
+from dask.dataframe.core import _concat as dd_concat
 from dask.dataframe.shuffle import group_split_dispatch, hash_object_dispatch
 from distributed import wait
 from distributed.protocol import nested_deserialize, to_serialize
@@ -153,9 +158,16 @@ def compute_map_index(
     if column_names[0] == "_partitions":
         ind = df[column_names[0]]
     else:
-        ind = hash_object_dispatch(
-            df[column_names] if column_names else df, index=False
-        )
+        # Need to cast numerical dtypes to be consistent
+        # with `dask.dataframe.shuffle.partitioning_index`
+        dtypes = {}
+        index = df[column_names] if column_names else df
+        for col, dtype in index.dtypes.items():
+            if pd.api.types.is_numeric_dtype(dtype):
+                dtypes[col] = np.float64
+        if dtypes:
+            index = index.astype(dtypes, errors="ignore")
+        ind = hash_object_dispatch(index, index=False)
     return ind % npartitions
 
 
@@ -185,15 +197,8 @@ def partition_dataframe(
     partitions
         Dict of dataframe-partitions, mapping partition-ID to dataframe
     """
-    if column_names[0] != "_partitions" and hasattr(df, "partition_by_hash"):
-        return dict(
-            zip(
-                range(npartitions),
-                df.partition_by_hash(
-                    column_names, npartitions, keep_index=not ignore_index
-                ),
-            )
-        )
+    # TODO: Use `partition_by_hash` if/when dtype-casting is added
+    # (See: https://github.com/rapidsai/cudf/issues/16221)
     map_index = compute_map_index(df, column_names, npartitions)
     return group_split_dispatch(df, map_index, npartitions, ignore_index=ignore_index)
 
@@ -468,18 +473,19 @@ def shuffle(
         npartitions = df.npartitions
 
     # Step (a):
-    df = df.persist()  # Make sure optimizations are apply on the existing graph
+    df = df.persist()  # Make sure optimizations are applied on the existing graph
     wait([df])  # Make sure all keys has been materialized on workers
+    persisted_keys = [f.key for f in c.client.futures_of(df)]
     name = (
         "explicit-comms-shuffle-"
-        f"{tokenize(df, column_names, npartitions, ignore_index)}"
+        f"{tokenize(df, column_names, npartitions, ignore_index, batchsize)}"
     )
     df_meta: DataFrame = df._meta
 
     # Stage all keys of `df` on the workers and cancel them, which makes it possible
     # for the shuffle to free memory as the partitions of `df` are consumed.
     # See CommsContext.stage_keys() for a description of staging.
-    rank_to_inkeys = c.stage_keys(name=name, keys=df.__dask_keys__())
+    rank_to_inkeys = c.stage_keys(name=name, keys=persisted_keys)
     c.client.cancel(df)
 
     # Get batchsize
@@ -526,23 +532,27 @@ def shuffle(
     # TODO: can we do this without using `submit()` to avoid the overhead
     #       of creating a Future for each dataframe partition?
 
-    dsk = {}
+    _futures = {}
     for rank in ranks:
         for part_id in rank_to_out_part_ids[rank]:
-            dsk[(name, part_id)] = c.client.submit(
+            _futures[part_id] = c.client.submit(
                 getitem,
                 shuffle_result[rank],
                 part_id,
                 workers=[c.worker_addresses[rank]],
             )
 
+    # Make sure partitions are properly ordered
+    futures = [_futures.pop(i) for i in range(npartitions)]
+
     # Create a distributed Dataframe from all the pieces
-    divs = [None] * (len(dsk) + 1)
-    ret = new_dd_object(dsk, name, df_meta, divs).persist()
+    divs = [None] * (len(futures) + 1)
+    kwargs = {"meta": df_meta, "divisions": divs, "prefix": "explicit-comms-shuffle"}
+    ret = dd.from_delayed(futures, **kwargs).persist()
     wait([ret])
 
     # Release all temporary dataframes
-    for fut in [*shuffle_result.values(), *dsk.values()]:
+    for fut in [*shuffle_result.values(), *futures]:
         fut.release()
     return ret
 

dask_cuda/local_cuda_cluster.py

@@ -10,7 +10,7 @@ from distributed.worker_memory import parse_memory_limit
 
 from .device_host_file import DeviceHostFile
 from .initialize import initialize
-from .plugins import CPUAffinity, PreImport, RMMSetup
+from .plugins import CPUAffinity, CUDFSetup, PreImport, RMMSetup
 from .proxify_host_file import ProxifyHostFile
 from .utils import (
     cuda_visible_devices,
@@ -73,6 +73,14 @@ class LocalCUDACluster(LocalCluster):
         starts spilling to host memory. Can be an integer (bytes), float (fraction of
         total device memory), string (like ``"5GB"`` or ``"5000M"``), or ``"auto"``, 0,
         or ``None`` to disable spilling to host (i.e. allow full device memory usage).
+    enable_cudf_spill : bool, default False
+        Enable automatic cuDF spilling.
+
+        .. warning::
+            This should NOT be used together with JIT-Unspill.
+    cudf_spill_stats : int, default 0
+        Set the cuDF spilling statistics level. This option has no effect if
+        ``enable_cudf_spill=False``.
     local_directory : str or None, default None
         Path on local machine to store temporary files. Can be a string (like
         ``"path/to/files"``) or ``None`` to fall back on the value of
@@ -114,6 +122,10 @@ class LocalCUDACluster(LocalCluster):
         memory on the GPU is used. ``rmm_pool_size`` must be specified to use RMM pool
         and to set the maximum pool size.
 
+        .. note::
+            When paired with `--enable-rmm-async` the maximum size cannot be guaranteed
+            due to fragmentation.
+
         .. note::
             This size is a per-worker configuration, and not cluster-wide.
     rmm_managed_memory : bool, default False
@@ -205,6 +217,8 @@ class LocalCUDACluster(LocalCluster):
         threads_per_worker=1,
         memory_limit="auto",
         device_memory_limit=0.8,
+        enable_cudf_spill=False,
+        cudf_spill_stats=0,
         data=None,
         local_directory=None,
         shared_filesystem=None,
@@ -255,6 +269,8 @@ class LocalCUDACluster(LocalCluster):
         self.device_memory_limit = parse_device_memory_limit(
             device_memory_limit, device_index=nvml_device_index(0, CUDA_VISIBLE_DEVICES)
         )
+        self.enable_cudf_spill = enable_cudf_spill
+        self.cudf_spill_stats = cudf_spill_stats
 
         self.rmm_pool_size = rmm_pool_size
         self.rmm_maximum_pool_size = rmm_maximum_pool_size
@@ -298,6 +314,12 @@ class LocalCUDACluster(LocalCluster):
             if device_memory_limit is None and memory_limit is None:
                 data = {}
             elif jit_unspill:
+                if enable_cudf_spill:
+                    warnings.warn(
+                        "Enabling cuDF spilling and JIT-Unspill together is not "
+                        "safe, consider disabling JIT-Unspill."
+                    )
+
                 data = (
                     ProxifyHostFile,
                     {
@@ -410,6 +432,7 @@ class LocalCUDACluster(LocalCluster):
                         track_allocations=self.rmm_track_allocations,
                     ),
                     PreImport(self.pre_import),
+                    CUDFSetup(self.enable_cudf_spill, self.cudf_spill_stats),
                 },
             }
         )

dask_cuda/plugins.py

@@ -14,6 +14,21 @@ class CPUAffinity(WorkerPlugin):
         os.sched_setaffinity(0, self.cores)
 
 
+class CUDFSetup(WorkerPlugin):
+    def __init__(self, spill, spill_stats):
+        self.spill = spill
+        self.spill_stats = spill_stats
+
+    def setup(self, worker=None):
+        try:
+            import cudf
+
+            cudf.set_option("spill", self.spill)
+            cudf.set_option("spill_stats", self.spill_stats)
+        except ImportError:
+            pass
+
+
 class RMMSetup(WorkerPlugin):
     def __init__(
         self,

dask_cuda/tests/test_cudf_builtin_spilling.py

@@ -20,7 +20,7 @@ from cudf.core.buffer.spill_manager import (  # noqa: E402
     get_global_manager,
     set_global_manager,
 )
-from cudf.testing._utils import assert_eq  # noqa: E402
+from cudf.testing import assert_eq  # noqa: E402
 
 if get_global_manager() is not None:
     pytest.skip(

dask_cuda/tests/test_dask_cuda_worker.py

@@ -231,6 +231,64 @@ def test_rmm_logging(loop):  # noqa: F811
                     assert v is rmm.mr.LoggingResourceAdaptor
 
 
+def test_cudf_spill_disabled(loop):  # noqa: F811
+    cudf = pytest.importorskip("cudf")
+    with popen(["dask", "scheduler", "--port", "9369", "--no-dashboard"]):
+        with popen(
+            [
+                "dask",
+                "cuda",
+                "worker",
+                "127.0.0.1:9369",
+                "--host",
+                "127.0.0.1",
+                "--no-dashboard",
+            ]
+        ):
+            with Client("127.0.0.1:9369", loop=loop) as client:
+                assert wait_workers(client, n_gpus=get_n_gpus())
+
+                cudf_spill = client.run(
+                    cudf.get_option,
+                    "spill",
+                )
+                for v in cudf_spill.values():
+                    assert v is False
+
+                cudf_spill_stats = client.run(cudf.get_option, "spill_stats")
+                for v in cudf_spill_stats.values():
+                    assert v == 0
+
+
+def test_cudf_spill(loop):  # noqa: F811
+    cudf = pytest.importorskip("cudf")
+    with popen(["dask", "scheduler", "--port", "9369", "--no-dashboard"]):
+        with popen(
+            [
+                "dask",
+                "cuda",
+                "worker",
+                "127.0.0.1:9369",
+                "--host",
+                "127.0.0.1",
+                "--no-dashboard",
+                "--enable-cudf-spill",
+                "--cudf-spill-stats",
+                "2",
+            ]
+        ):
+            with Client("127.0.0.1:9369", loop=loop) as client:
+                assert wait_workers(client, n_gpus=get_n_gpus())
+
+                cudf_spill = client.run(cudf.get_option, "spill")
+                for v in cudf_spill.values():
+                    assert v is True
+
+                cudf_spill_stats = client.run(cudf.get_option, "spill_stats")
+                for v in cudf_spill_stats.values():
+                    assert v == 2
+
+
 @patch.dict(os.environ, {"CUDA_VISIBLE_DEVICES": "0"})
 def test_dashboard_address(loop):  # noqa: F811
     with popen(["dask", "scheduler", "--port", "9369", "--no-dashboard"]):

dask_cuda/tests/test_dgx.py

@@ -15,6 +15,10 @@ mp = mp.get_context("spawn")  # type: ignore
 psutil = pytest.importorskip("psutil")
 
 
+def _is_ucx_116(ucp):
+    return ucp.get_ucx_version()[:2] == (1, 16)
+
+
 class DGXVersion(Enum):
     DGX_1 = auto()
     DGX_2 = auto()
@@ -102,9 +106,11 @@ def _test_tcp_over_ucx(protocol):
 )
 def test_tcp_over_ucx(protocol):
     if protocol == "ucx":
-        pytest.importorskip("ucp")
+        ucp = pytest.importorskip("ucp")
     elif protocol == "ucxx":
-        pytest.importorskip("ucxx")
+        ucp = pytest.importorskip("ucxx")
+    if _is_ucx_116(ucp):
+        pytest.skip("https://github.com/rapidsai/ucx-py/issues/1037")
 
     p = mp.Process(target=_test_tcp_over_ucx, args=(protocol,))
     p.start()
@@ -217,9 +223,11 @@ def _test_ucx_infiniband_nvlink(
 )
 def test_ucx_infiniband_nvlink(protocol, params):
     if protocol == "ucx":
-        pytest.importorskip("ucp")
+        ucp = pytest.importorskip("ucp")
     elif protocol == "ucxx":
-        pytest.importorskip("ucxx")
+        ucp = pytest.importorskip("ucxx")
+    if _is_ucx_116(ucp) and params["enable_infiniband"] is False:
+        pytest.skip("https://github.com/rapidsai/ucx-py/issues/1037")
 
     skip_queue = mp.Queue()
 

dask_cuda/tests/test_explicit_comms.py

@@ -25,6 +25,22 @@ from dask_cuda.utils_test import IncreasedCloseTimeoutNanny
 mp = mp.get_context("spawn")  # type: ignore
 ucp = pytest.importorskip("ucp")
 
+QUERY_PLANNING_ON = dask.config.get("dataframe.query-planning", None) is not False
+
+# Skip these tests when dask-expr is active (for now)
+query_planning_skip = pytest.mark.skipif(
+    QUERY_PLANNING_ON,
+    reason=(
+        "The 'explicit-comms' config is not supported "
+        "when query planning is enabled."
+    ),
+)
+
+# Set default shuffle method to "tasks"
+if dask.config.get("dataframe.shuffle.method", None) is None:
+    dask.config.set({"dataframe.shuffle.method": "tasks"})
+
+
 # Notice, all of the following tests is executed in a new process such
 # that UCX options of the different tests doesn't conflict.
 
@@ -82,6 +98,7 @@ def _test_dataframe_merge_empty_partitions(nrows, npartitions):
                     pd.testing.assert_frame_equal(got, expected)
 
 
+@query_planning_skip
 def test_dataframe_merge_empty_partitions():
     # Notice, we use more partitions than rows
     p = mp.Process(target=_test_dataframe_merge_empty_partitions, args=(2, 4))
@@ -92,7 +109,14 @@ def test_dataframe_merge_empty_partitions():
 
 def check_partitions(df, npartitions):
     """Check that all values in `df` hashes to the same"""
-    hashes = partitioning_index(df, npartitions)
+    dtypes = {}
+    for col, dtype in df.dtypes.items():
+        if pd.api.types.is_numeric_dtype(dtype):
+            dtypes[col] = np.float64
+    if not dtypes:
+        dtypes = None
+
+    hashes = partitioning_index(df, npartitions, cast_dtype=dtypes)
     if len(hashes) > 0:
         return len(hashes.unique()) == 1
     else:
@@ -111,11 +135,10 @@ def _test_dataframe_shuffle(backend, protocol, n_workers, _partitions):
         worker_class=IncreasedCloseTimeoutNanny,
         processes=True,
     ) as cluster:
-        with Client(cluster) as client:
-            all_workers = list(client.get_worker_logs().keys())
+        with Client(cluster):
             comms.default_comms()
             np.random.seed(42)
-            df = pd.DataFrame({"key": np.random.random(100)})
+            df = pd.DataFrame({"key": np.random.randint(0, high=100, size=100)})
             if backend == "cudf":
                 df = cudf.DataFrame.from_pandas(df)
 
@@ -124,15 +147,13 @@ def _test_dataframe_shuffle(backend, protocol, n_workers, _partitions):
 
             for input_nparts in range(1, 5):
                 for output_nparts in range(1, 5):
-                    ddf = dd.from_pandas(df.copy(), npartitions=input_nparts).persist(
-                        workers=all_workers
-                    )
+                    ddf1 = dd.from_pandas(df.copy(), npartitions=input_nparts)
                     # To reduce test runtime, we change the batchsizes here instead
                     # of using a test parameter.
                     for batchsize in (-1, 1, 2):
                         with dask.config.set(explicit_comms_batchsize=batchsize):
                             ddf = explicit_comms_shuffle(
-                                ddf,
+                                ddf1,
                                 ["_partitions"] if _partitions else ["key"],
                                 npartitions=output_nparts,
                                 batchsize=batchsize,
@@ -160,6 +181,32 @@ def _test_dataframe_shuffle(backend, protocol, n_workers, _partitions):
                                 got = ddf.compute().sort_values("key")
                                 assert_eq(got, expected)
 
+                                # Check that partitioning is consistent with "tasks"
+                                ddf_tasks = ddf1.shuffle(
+                                    ["key"],
+                                    npartitions=output_nparts,
+                                    shuffle_method="tasks",
+                                )
+                                for i in range(output_nparts):
+                                    expected_partition = ddf_tasks.partitions[
+                                        i
+                                    ].compute()["key"]
+                                    actual_partition = ddf.partitions[i].compute()[
+                                        "key"
+                                    ]
+                                    if backend == "cudf":
+                                        expected_partition = (
+                                            expected_partition.values_host
+                                        )
+                                        actual_partition = actual_partition.values_host
+                                    else:
+                                        expected_partition = expected_partition.values
+                                        actual_partition = actual_partition.values
+                                    assert all(
+                                        np.sort(expected_partition)
+                                        == np.sort(actual_partition)
+                                    )
+
 
 @pytest.mark.parametrize("nworkers", [1, 2, 3])
 @pytest.mark.parametrize("backend", ["pandas", "cudf"])
@@ -220,6 +267,7 @@ def _test_dask_use_explicit_comms(in_cluster):
         check_shuffle()
 
 
+@query_planning_skip
 @pytest.mark.parametrize("in_cluster", [True, False])
 def test_dask_use_explicit_comms(in_cluster):
     def _timeout(process, function, timeout):
@@ -282,6 +330,7 @@ def _test_dataframe_shuffle_merge(backend, protocol, n_workers):
             assert_eq(got, expected)
 
 
+@query_planning_skip
 @pytest.mark.parametrize("nworkers", [1, 2, 4])
 @pytest.mark.parametrize("backend", ["pandas", "cudf"])
 @pytest.mark.parametrize("protocol", ["tcp", "ucx", "ucxx"])

dask_cuda/tests/test_local_cuda_cluster.py

@@ -500,6 +500,54 @@ async def test_worker_fraction_limits():
             )
 
 
+@gen_test(timeout=20)
+async def test_cudf_spill_disabled():
+    cudf = pytest.importorskip("cudf")
+
+    async with LocalCUDACluster(
+        asynchronous=True,
+    ) as cluster:
+        async with Client(cluster, asynchronous=True) as client:
+            cudf_spill = await client.run(
+                cudf.get_option,
+                "spill",
+            )
+            for v in cudf_spill.values():
+                assert v is False
+
+            cudf_spill_stats = await client.run(
+                cudf.get_option,
+                "spill_stats",
+            )
+            for v in cudf_spill_stats.values():
+                assert v == 0
+
+
+@gen_test(timeout=20)
+async def test_cudf_spill():
+    cudf = pytest.importorskip("cudf")
+
+    async with LocalCUDACluster(
+        enable_cudf_spill=True,
+        cudf_spill_stats=2,
+        asynchronous=True,
+    ) as cluster:
+        async with Client(cluster, asynchronous=True) as client:
+            cudf_spill = await client.run(
+                cudf.get_option,
+                "spill",
+            )
+            for v in cudf_spill.values():
+                assert v is True
+
+            cudf_spill_stats = await client.run(
+                cudf.get_option,
+                "spill_stats",
+            )
+            for v in cudf_spill_stats.values():
+                assert v == 2
+
+
 @pytest.mark.parametrize(
     "protocol",
     ["ucx", "ucxx"],

dask_cuda/tests/test_proxy.py

@@ -537,10 +537,10 @@ def test_from_cudf_of_proxy_object():
     assert has_parallel_type(df)
 
     ddf = dask_cudf.from_cudf(df, npartitions=1)
-    assert has_parallel_type(ddf)
+    assert has_parallel_type(ddf._meta)
 
     # Notice, the output is a dask-cudf dataframe and not a proxy object
-    assert type(ddf) is dask_cudf.core.DataFrame
+    assert type(ddf._meta) is cudf.DataFrame
 
 
 def test_proxy_object_parquet(tmp_path):

dask_cuda/tests/test_version.py

@@ -0,0 +1,12 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.
+
+import dask_cuda
+
+
+def test_version_constants_are_populated():
+    # __git_commit__ will only be non-empty in a built distribution
+    assert isinstance(dask_cuda.__git_commit__, str)
+
+    # __version__ should always be non-empty
+    assert isinstance(dask_cuda.__version__, str)
+    assert len(dask_cuda.__version__) > 0

{dask_cuda-24.4.0.dist-info → dask_cuda-24.8.2.dist-info}/METADATA

@@ -1,9 +1,9 @@
 Metadata-Version: 2.1
 Name: dask-cuda
-Version: 24.4.0
+Version: 24.8.2
 Summary: Utilities for Dask and CUDA interactions
 Author: NVIDIA Corporation
-License: Apache-2.0
+License: Apache 2.0
 Project-URL: Homepage, https://github.com/rapidsai/dask-cuda
 Project-URL: Documentation, https://docs.rapids.ai/api/dask-cuda/stable/
 Project-URL: Source, https://github.com/rapidsai/dask-cuda
@@ -18,25 +18,25 @@ Classifier: Programming Language :: Python :: 3.11
 Requires-Python: >=3.9
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: click >=8.1
-Requires-Dist: numba >=0.57
-Requires-Dist: numpy <2.0a0,>=1.23
-Requires-Dist: pandas >=1.3
-Requires-Dist: pynvml <11.5,>=11.0.0
-Requires-Dist: rapids-dask-dependency ==24.4.*
-Requires-Dist: zict >=2.0.0
+Requires-Dist: click>=8.1
+Requires-Dist: numba>=0.57
+Requires-Dist: numpy<2.0a0,>=1.23
+Requires-Dist: pandas>=1.3
+Requires-Dist: pynvml<11.5,>=11.0.0
+Requires-Dist: rapids-dask-dependency==24.8.*
+Requires-Dist: zict>=2.0.0
 Provides-Extra: docs
-Requires-Dist: numpydoc >=1.1.0 ; extra == 'docs'
-Requires-Dist: sphinx ; extra == 'docs'
-Requires-Dist: sphinx-click >=2.7.1 ; extra == 'docs'
-Requires-Dist: sphinx-rtd-theme >=0.5.1 ; extra == 'docs'
+Requires-Dist: numpydoc>=1.1.0; extra == "docs"
+Requires-Dist: sphinx; extra == "docs"
+Requires-Dist: sphinx-click>=2.7.1; extra == "docs"
+Requires-Dist: sphinx-rtd-theme>=0.5.1; extra == "docs"
 Provides-Extra: test
-Requires-Dist: cudf ==24.4.* ; extra == 'test'
-Requires-Dist: dask-cudf ==24.4.* ; extra == 'test'
-Requires-Dist: kvikio ==24.4.* ; extra == 'test'
-Requires-Dist: pytest ; extra == 'test'
-Requires-Dist: pytest-cov ; extra == 'test'
-Requires-Dist: ucx-py ==0.37.* ; extra == 'test'
+Requires-Dist: cudf==24.8.*; extra == "test"
+Requires-Dist: dask-cudf==24.8.*; extra == "test"
+Requires-Dist: kvikio==24.8.*; extra == "test"
+Requires-Dist: pytest; extra == "test"
+Requires-Dist: pytest-cov; extra == "test"
+Requires-Dist: ucx-py==0.39.*; extra == "test"
 
 Dask CUDA
 =========

{dask_cuda-24.4.0.dist-info → dask_cuda-24.8.2.dist-info}/RECORD

@@ -1,16 +1,16 @@
-dask_cuda/VERSION,sha256=cS6_wRwcl4ZhhE4gk3mZbC_4dD-r28YRdEQeaKdKu1U,9
-dask_cuda/__init__.py,sha256=Vt42yCT1WhjZgehiNLDLw2uvfxjqLThD8uKDMyKQYsw,1454
-dask_cuda/_version.py,sha256=OcuA1fDRuukb6OC8LzTwSO2aJ3ZtB1N2pm08Nrga8y8,778
-dask_cuda/cli.py,sha256=XNRH0bu-6jzRoyWJB5qSWuzePJSh3z_5Ng6rDCnz7lg,15970
-dask_cuda/cuda_worker.py,sha256=bIu-ESeIpJG_WaTYrv0z9z5juJ1qR5i_5Ng3CN1WK8s,8579
+dask_cuda/VERSION,sha256=5YtjwV2EoD7E5Ed4K-PvnU0eEtdkkn33JHuNFDy8oKA,8
+dask_cuda/__init__.py,sha256=JLDWev7vI_dPusLgRdOwXBz-xfhlX_hc-DzmLtrEYO0,1918
+dask_cuda/_version.py,sha256=cHDO9AzNtxkCVhwYu7hL3H7RPAkQnxpKBjElOst3rkI,964
+dask_cuda/cli.py,sha256=Y3aObfAyMwOIo0oVz3-NC2InGLShOpeINwW5ROTF2s8,16616
+dask_cuda/cuda_worker.py,sha256=uqyoDKsSe7sKN3StMVyz_971rj0Sjpmwfv7Bj083Wss,8959
 dask_cuda/device_host_file.py,sha256=yS31LGtt9VFAG78uBBlTDr7HGIng2XymV1OxXIuEMtM,10272
 dask_cuda/disk_io.py,sha256=urSLKiPvJvYmKCzDPOUDCYuLI3r1RUiyVh3UZGRoF_Y,6626
 dask_cuda/get_device_memory_objects.py,sha256=R3U2cq4fJZPgtsUKyIguy9161p3Q99oxmcCmTcg6BtQ,4075
 dask_cuda/initialize.py,sha256=Gjcxs_c8DTafgsHe5-2mw4lJdOmbFJJAZVOnxA8lTjM,6462
 dask_cuda/is_device_object.py,sha256=CnajvbQiX0FzFzwft0MqK1OPomx3ZGDnDxT56wNjixw,1046
 dask_cuda/is_spillable_object.py,sha256=CddGmg0tuSpXh2m_TJSY6GRpnl1WRHt1CRcdWgHPzWA,1457
-dask_cuda/local_cuda_cluster.py,sha256=hoEiEfJqAQrRS7N632VatSl1245GiWMT5B77Wc-i5C0,17928
-dask_cuda/plugins.py,sha256=cnHsdrXx7PBPmrzHX6YEkCH5byCsUk8LE2FeTeu8ZLU,4259
+dask_cuda/local_cuda_cluster.py,sha256=jgXjd6OvEDfQ3iXU8hV_UfULa13GZsli0SGC2PIouZk,18882
+dask_cuda/plugins.py,sha256=DCf7PnIBu_VNjFfrFeb1zCNuEnCaX9oz4Umn76t02Mc,4630
 dask_cuda/proxify_device_objects.py,sha256=99CD7LOE79YiQGJ12sYl_XImVhJXpFR4vG5utdkjTQo,8108
 dask_cuda/proxify_host_file.py,sha256=Wf5CFCC1JN5zmfvND3ls0M5FL01Y8VhHrk0xV3UQ9kk,30850
 dask_cuda/proxy_object.py,sha256=bZq92kjgFB-ad_luSAFT_RItV3nssmiEk4OOSp34laU,29812
@@ -18,36 +18,37 @@ dask_cuda/utils.py,sha256=RWlLK2cPHaCuNNhr8bW8etBeGklwREQJOafQbTydStk,25121
 dask_cuda/utils_test.py,sha256=WNMR0gic2tuP3pgygcR9g52NfyX8iGMOan6juXhpkCE,1694
 dask_cuda/worker_spec.py,sha256=7-Uq_e5q2SkTlsmctMcYLCa9_3RiiVHZLIN7ctfaFmE,4376
 dask_cuda/benchmarks/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-dask_cuda/benchmarks/common.py,sha256=sEIFnRZS6wbyKCQyB4fDclYLc2YqC0PolurR5qzuRxw,6393
+dask_cuda/benchmarks/common.py,sha256=2MnDdQjvHfGaUWDgiTcTGI_EeKPmVBEwoWfsJUNpOjU,6613
 dask_cuda/benchmarks/local_cudf_groupby.py,sha256=T9lA9nb4Wzu46AH--SJEVCeCm3650J7slapdNR_08FU,8904
-dask_cuda/benchmarks/local_cudf_merge.py,sha256=POjxoPx4zY1TjG2S_anElL6rDtC5Jhn3nF4HABlnwZg,12447
-dask_cuda/benchmarks/local_cudf_shuffle.py,sha256=M-Lp3O3q8uyY50imQqMKZYwkAmyR0NApjx2ipGxDkXw,8608
+dask_cuda/benchmarks/local_cudf_merge.py,sha256=AsuVnMA3H93sJwjjgi4KaIdYKnnX1OeRMPiXizrwHGk,12577
+dask_cuda/benchmarks/local_cudf_shuffle.py,sha256=2xWJZf3gwDNimXKZN2ivtU3OE_qec1KNOhgL4_AGQZU,8655
 dask_cuda/benchmarks/local_cupy.py,sha256=aUKIYfeR7c77K4kKk697Rxo8tG8kFabQ9jQEVGr-oTs,10762
 dask_cuda/benchmarks/local_cupy_map_overlap.py,sha256=_texYmam1K_XbzIvURltui5KRsISGFNylXiGUtgRIz0,6442
-dask_cuda/benchmarks/utils.py,sha256=baL5zK6VS6Mw_M4x9zJe8vMLUd2SZd1lS78JrL-h6oo,26896
+dask_cuda/benchmarks/utils.py,sha256=4k8KnJPOczKDQNBPRWlaGsU2zdEA09BDGgklUXggwMU,30008
 dask_cuda/explicit_comms/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 dask_cuda/explicit_comms/comms.py,sha256=Su6PuNo68IyS-AwoqU4S9TmqWsLvUdNa0jot2hx8jQQ,10400
 dask_cuda/explicit_comms/dataframe/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-dask_cuda/explicit_comms/dataframe/shuffle.py,sha256=YferHNWKsMea8tele-ynPVr_6RAZNZIR-VzK_uFuEQU,20131
-dask_cuda/tests/test_cudf_builtin_spilling.py,sha256=u3kW91YRLdHFycvpGfSQKrEucu5khMJ1k4sjmddO490,4910
-dask_cuda/tests/test_dask_cuda_worker.py,sha256=gViHaMCSfB6ip125OEi9D0nfKC-qBXRoHz6BRodEdb4,17729
+dask_cuda/explicit_comms/dataframe/shuffle.py,sha256=4xfhfbTGa36YPs_ex1_fFhzfGMYJq-QkS5q0RwgeHh8,20645
+dask_cuda/tests/test_cudf_builtin_spilling.py,sha256=qVN9J0Hdv66A9COFArLIdRriyyxEKpS3lEZGHbVHaq8,4903
+dask_cuda/tests/test_dask_cuda_worker.py,sha256=o5g0_t-2M_2lfPeOPTS4NVF4rnQF0ZWAZekXw2h0xPc,19610
 dask_cuda/tests/test_device_host_file.py,sha256=79ssUISo1YhsW_7HdwqPfsH2LRzS2bi5BjPym1Sdgqw,5882
-dask_cuda/tests/test_dgx.py,sha256=Oh2vwL_CdUzSVQQoiIu6SPwXGRtmXwaW_Hh3ipXPUOc,7162
-dask_cuda/tests/test_explicit_comms.py,sha256=I4lSW-NQ0E08baEoG7cY4Ix3blGb1Auz88q2BNd1cPA,13136
+dask_cuda/tests/test_dgx.py,sha256=BPCF4ZvhrVKkT43OOFHdijuo-M34vW3V18C8rRH1HXg,7489
+dask_cuda/tests/test_explicit_comms.py,sha256=Pa5vVx63qWtScnVJuS31WESXIt2FPyTJVFO-0OUbbmU,15276
 dask_cuda/tests/test_from_array.py,sha256=okT1B6UqHmLxoy0uER0Ylm3UyOmi5BAXwJpTuTAw44I,601
 dask_cuda/tests/test_gds.py,sha256=6jf0HPTHAIG8Mp_FC4Ai4zpn-U1K7yk0fSXg8He8-r8,1513
 dask_cuda/tests/test_initialize.py,sha256=Rba59ZbljEm1yyN94_sWZPEE_f7hWln95aiBVc49pmY,6960
-dask_cuda/tests/test_local_cuda_cluster.py,sha256=G3kR-4o-vCqWWfSuQLFKVEK0F243FaDSgRlDTUll5aU,18376
+dask_cuda/tests/test_local_cuda_cluster.py,sha256=Lc9QncyGwBwhaZPGBfreXJf3ZC9Zd8SjDc2fpeQ-BT0,19710
 dask_cuda/tests/test_proxify_host_file.py,sha256=Yiv0sDcUoWw0d2oiPeHGoHqqSSM4lfQ4rChCiaxb6EU,18994
-dask_cuda/tests/test_proxy.py,sha256=6iicSYYT2BGo1iKUQ7jM00mCjC4gtfwwxFXfGwH3QHc,23807
+dask_cuda/tests/test_proxy.py,sha256=OnGnPkl5ksCb-3hpEKG2z1OfPK9DbnOCtBHOjcUUjhg,23809
 dask_cuda/tests/test_spill.py,sha256=xN9PbVERBYMuZxvscSO0mAM22loq9WT3ltZVBFxlmM4,10239
 dask_cuda/tests/test_utils.py,sha256=JRIwXfemc3lWSzLJX0VcvR1_0wB4yeoOTsw7kB6z6pU,9176
+dask_cuda/tests/test_version.py,sha256=vK2HjlRLX0nxwvRsYxBqhoZryBNZklzA-vdnyuWDxVg,365
 dask_cuda/tests/test_worker_spec.py,sha256=Bvu85vkqm6ZDAYPXKMJlI2pm9Uc5tiYKNtO4goXSw-I,2399
 examples/ucx/client_initialize.py,sha256=YN3AXHF8btcMd6NicKKhKR9SXouAsK1foJhFspbOn70,1262
 examples/ucx/local_cuda_cluster.py,sha256=7xVY3EhwhkY2L4VZin_BiMCbrjhirDNChoC86KiETNc,1983
-dask_cuda-24.4.0.dist-info/LICENSE,sha256=MjI3I-EgxfEvZlgjk82rgiFsZqSDXHFETd2QJ89UwDA,11348
-dask_cuda-24.4.0.dist-info/METADATA,sha256=mjy5YntsqIKdldcPEj-jjR70Aphe-B0LXqjcZjVFS9U,2570
-dask_cuda-24.4.0.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-dask_cuda-24.4.0.dist-info/entry_points.txt,sha256=UcRaKVEpywtxc6pF1VnfMB0UK4sJg7a8_NdZF67laPM,136
-dask_cuda-24.4.0.dist-info/top_level.txt,sha256=3kKxJxeM108fuYc_lwwlklP7YBU9IEmdmRAouzi397o,33
-dask_cuda-24.4.0.dist-info/RECORD,,
+dask_cuda-24.8.2.dist-info/LICENSE,sha256=MjI3I-EgxfEvZlgjk82rgiFsZqSDXHFETd2QJ89UwDA,11348
+dask_cuda-24.8.2.dist-info/METADATA,sha256=6iMwPI8cWrEYDYz73vm8pw-LkVeEgTQzymJgRxj32VQ,2546
+dask_cuda-24.8.2.dist-info/WHEEL,sha256=R0nc6qTxuoLk7ShA2_Y-UWkN8ZdfDBG2B6Eqpz2WXbs,91
+dask_cuda-24.8.2.dist-info/entry_points.txt,sha256=UcRaKVEpywtxc6pF1VnfMB0UK4sJg7a8_NdZF67laPM,136
+dask_cuda-24.8.2.dist-info/top_level.txt,sha256=3kKxJxeM108fuYc_lwwlklP7YBU9IEmdmRAouzi397o,33
+dask_cuda-24.8.2.dist-info/RECORD,,

{dask_cuda-24.4.0.dist-info → dask_cuda-24.8.2.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: bdist_wheel (0.43.0)
+Generator: setuptools (72.1.0)
 Root-Is-Purelib: true
 Tag: py3-none-any