Align extract_partitioning_index logic with upstream shuffling (NVIDIA#60)

* update extract_partitioning_index with compat code

Signed-off-by: rjzamora <rzamora217@gmail.com>

* [Tutorials] Add a tutorial for PEFT data curation (NVIDIA#45)

This PR adds a new tutorial to demonstrate data curation for PEFT
use-cases.

Signed-off-by: Mehran Maghoumi <Maghoumi@users.noreply.github.com>
Signed-off-by: rjzamora <rzamora217@gmail.com>

* move compat code to _compat file

Signed-off-by: rjzamora <rzamora217@gmail.com>

* Only import PII constants during Curator import (NVIDIA#61)

* Move PII constants to a seperate file that does not import presidio/spacy and other GPU dependencies

Signed-off-by: Ayush Dattagupta <ayushdg95@gmail.com>

* Add comment around import, move constant import to global scope

Signed-off-by: Ayush Dattagupta <ayushdg95@gmail.com>

---------

Signed-off-by: Ayush Dattagupta <ayushdg95@gmail.com>

* add unit test

Signed-off-by: rjzamora <rzamora217@gmail.com>

* add pytest.mark.gpu

Signed-off-by: rjzamora <rzamora217@gmail.com>

* move extract_partitioning_index import for now

Signed-off-by: rjzamora <rzamora217@gmail.com>

* test both cudf and pandas

Signed-off-by: rjzamora <rzamora217@gmail.com>

* spelling

Signed-off-by: rjzamora <rzamora217@gmail.com>

---------

Signed-off-by: rjzamora <rzamora217@gmail.com>
Signed-off-by: Mehran Maghoumi <Maghoumi@users.noreply.github.com>
Signed-off-by: Ayush Dattagupta <ayushdg95@gmail.com>
Co-authored-by: Mehran Maghoumi <Maghoumi@users.noreply.github.com>
Co-authored-by: Ayush Dattagupta <ayushdg95@gmail.com>

Author: 3 people

nemo_curator/_compat.py

@@ -20,3 +20,4 @@
 # TODO: remove when dask min version gets bumped
 DASK_SHUFFLE_METHOD_ARG = _dask_version > parseVersion("2024.1.0")
 DASK_P2P_ERROR = _dask_version < parseVersion("2023.10.0")
+DASK_SHUFFLE_CAST_DTYPE = _dask_version > parseVersion("2023.12.0")

nemo_curator/utils/fuzzy_dedup_utils/merge_utils.py

@@ -16,13 +16,14 @@
 from operator import getitem
 
 import numpy as np
+import pandas as pd
 from dask.base import tokenize
 from dask.dataframe.core import new_dd_object
 from dask.dataframe.shuffle import partitioning_index
 from dask.highlevelgraph import HighLevelGraph
 from dask.utils import M
 
-from nemo_curator.utils.fuzzy_dedup_utils.shuffle_utils import rearange_by_column_direct
+from nemo_curator._compat import DASK_SHUFFLE_CAST_DTYPE
 
 
 def _split_part(part, nsplits):
@@ -129,6 +130,21 @@ def extract_partitioning_index(
     # a partition-wise merge between `left_df` and `right_df`.
     # We call this `global_partitioning_index`:
 
+    if DASK_SHUFFLE_CAST_DTYPE:
+        # Need to use the same type-casting logic as `shuffle`
+        dtypes = {}
+        if not isinstance(merge_on, list):
+            merge_on = [merge_on]
+        for col, dtype in left_df[merge_on].dtypes.items():
+            if pd.api.types.is_numeric_dtype(dtype):
+                dtypes[col] = np.float64
+        if not dtypes:
+            dtypes = None
+        cast_dtype = {"cast_dtype": dtypes}
+    else:
+        # `cast_dtype` argument doesn't exist yet
+        cast_dtype = {}
+
     num_bucket_files = bk_mapping.file_id.max() + 1
     global_partitioning_index = left_df[merge_on].map_partitions(
         partitioning_index,
@@ -137,6 +153,7 @@ def extract_partitioning_index(
         enforce_metadata=False,
         transform_divisions=False,
         align_dataframes=False,
+        **cast_dtype,
     )
 
     if total_bucket_partitions < num_bucket_files:
@@ -157,7 +174,7 @@ def extract_partitioning_index(
     # want to send the rows of `left_df` to the partition
     # indices encoded in `global_partitioning_index`. Instead, we
     # need to take a modulus with `parts_per_bucket_batch` to
-    # define a `"_partitoins"` column.
+    # define a `"_partitions"` column.
     left_df["_partitions"] = global_partitioning_index % parts_per_bucket_batch
 
     return left_df, global_partitioning_index
@@ -195,6 +212,10 @@ def merge_left_to_shuffled_right(
     subset_bucket_df,
     merge_on,
 ):
+    from nemo_curator.utils.fuzzy_dedup_utils.shuffle_utils import (
+        rearange_by_column_direct,
+    )
+
     # We are merging an unshuffled batch of "left" partitions
     # with a shuffled batch of "right" partitions. To minimize
     # data movement, we can manaully rerrange the "left" batch

tests/test_fuzzy_dedup.py

@@ -16,14 +16,17 @@
 from itertools import combinations
 from typing import Iterable
 
+import dask.dataframe as dd
 import numpy as np
 import pytest
 import yaml
+from dask import config
 from dask.dataframe.utils import assert_eq
 from distributed import Client
 
 from nemo_curator import LSH, FuzzyDuplicates, FuzzyDuplicatesConfig, MinHash
 from nemo_curator.datasets import DocumentDataset
+from nemo_curator.utils.fuzzy_dedup_utils.merge_utils import extract_partitioning_index
 from nemo_curator.utils.import_utils import gpu_only_import, gpu_only_import_from
 
 cudf = gpu_only_import("cudf")
@@ -367,3 +370,74 @@ def test_from_yaml(self, tmpdir):
         config = FuzzyDuplicatesConfig.from_yaml(tmpdir / "config.yaml")
         for param in yaml_params:
             assert getattr(config, param) == yaml_params[param]
+
+
+@pytest.mark.parametrize(
+    "backend",
+    [
+        "pandas",
+        pytest.param(
+            "cudf",
+            marks=pytest.mark.gpu,
+        ),
+    ],
+)
+def test_extract_partitioning_index(backend):
+
+    def add_partition_info(df, partition_info=None):
+        if partition_info is None:
+            df["file_id"] = -1
+        else:
+            df["file_id"] = partition_info["number"]
+        return df
+
+    with config.set({"dataframe.backend": backend}):
+
+        # Create a random `unshuffled` DataFrame with a
+        # "part_id" column to be used as the shuffle index
+        npartitions_left = 7
+        unshuffled = dd.from_dict(
+            {"part_id": np.random.randint(25, size=1000, dtype="int32")},
+            npartitions=npartitions_left,
+        )
+
+        # Create a `bk_mapping` DataFrame that defines
+        # the "correct" mapping beween "part_id" and
+        # the destination partition ("file_id")
+        npartitions_right = 5
+        bk_mapping = (
+            dd.from_dict(
+                {"part_id": np.arange(25, dtype="int32")},
+                npartitions=npartitions_right,
+            )
+            .shuffle("part_id")
+            .map_partitions(add_partition_info)
+            .compute()
+        )
+
+    # Use `extract_partitioning_index` to calculate
+    # the partitioning index and assign it as a new
+    # "_partitions" column
+    result, _ = extract_partitioning_index(
+        unshuffled,
+        "part_id",
+        bk_mapping,
+        npartitions_right,
+        npartitions_right,
+    )
+
+    # Rename the "_partitions" column, shuffle by "part_id",
+    # and then assign a "file_id" column to reflect the final
+    # partition of each row
+    check = (
+        result.rename(columns={"_partitions": "expected_file_id"})
+        .shuffle(
+            "part_id",
+            npartitions=npartitions_right,
+        )
+        .map_partitions(add_partition_info)
+        .compute()
+    )
+
+    # Check that the real and expected partitions match
+    assert (check["file_id"] == check["expected_file_id"]).all()