Merge pull request #222 from derekhiggins/checkpointing

Add data checkpointing capability

Author: markmc

src/instructlab/sdg/checkpointing.py

@@ -0,0 +1,86 @@
+# Standard
+import logging
+import uuid
+
+# Third Party
+from datasets import Dataset, concatenate_datasets, load_dataset
+from datasets.data_files import EmptyDatasetError
+
+# First Party
+from instructlab.sdg.utils import pandas
+
+logger = logging.getLogger(__name__)
+
+
+class Checkpointer:
+    def __init__(self, checkpoint_dir=None, save_freq=1):
+        self._checkpoint_dir = checkpoint_dir
+
+        self._save_freq = save_freq
+        self._cache = []
+
+    def checkpoint(self, dataset):
+        self._cache.append(dataset)
+        if len(self._cache) < self._save_freq:
+            return
+        self.save()
+        self._cache.clear()
+
+    def done(self):
+        if self._cache:
+            self.save()
+            self._cache.clear()
+
+    def save(self):
+        if self._checkpoint_dir is None:
+            return
+        checkpoint_id = uuid.uuid4().hex
+        checkpoint_file = (
+            f"{self._checkpoint_dir}/data_checkpoint_{checkpoint_id}.jsonl"
+        )
+        logger.info(f"Saving checkpoint to {checkpoint_file}")
+        # Saves all the current records to new file in the checkpoint dir
+        concatenate_datasets(self._cache).to_json(
+            checkpoint_file, orient="records", lines=True
+        )
+
+    def load(self, dataset: Dataset) -> Dataset:
+        if self._checkpoint_dir is None:
+            return dataset, None
+
+        try:
+            pre_generated_data = load_dataset(
+                "json", data_dir=self._checkpoint_dir, split="train"
+            )
+        except EmptyDatasetError:
+            logger.info(
+                f"No existing checkpoints found in {self._checkpoint_dir}, generating from scratch"
+            )
+            return dataset, None
+
+        logger.info(
+            f"Loading existing checkpoints from {self._checkpoint_dir}, with {pre_generated_data.num_rows} rows"
+        )
+        seed_data = self._get_missing_data(dataset, pre_generated_data)
+        logger.info(f"Found {seed_data.num_rows} missing rows in the dataset")
+        return seed_data, pre_generated_data
+
+    def _get_missing_data(self, seed_data, generated_data):
+        # Get the common columns between the two datasets
+        common_columns = list(
+            set(seed_data.column_names) & set(generated_data.column_names)
+        )
+
+        # Extract the relevant data based on common columns
+        seed_data_common = seed_data.select_columns(common_columns)
+        generated_data_common = generated_data.select_columns(common_columns)
+
+        # Convert to Pandas DataFrames for easier comparison
+        seed_df = seed_data_common.to_pandas()
+        generated_df = generated_data_common.to_pandas()
+
+        # Identify missing rows
+        missing_rows = ~seed_df.apply(tuple, 1).isin(generated_df.apply(tuple, 1))
+
+        missing_df = seed_data.to_pandas()[missing_rows]
+        return pandas.dataset_from_pandas_dataframe(missing_df)

src/instructlab/sdg/generate_data.py

@@ -5,6 +5,7 @@
 from importlib import resources
 from pathlib import Path
 from typing import Optional
+import dataclasses
 import json
 import os
 import time
@@ -181,6 +182,8 @@ def _context_init(
     model_family: str,
     model_id: str,
     num_instructions_to_generate: int,
+    checkpoint_dir: str,
+    save_freq: int,
     batch_num_workers: Optional[int],
     batch_size: Optional[int],
 ):
@@ -194,6 +197,8 @@ def _context_init(
         model_family=model_family,
         model_id=model_id,
         num_instructions_to_generate=num_instructions_to_generate,
+        checkpoint_dir=checkpoint_dir,
+        save_freq=save_freq,
         **extra_kwargs,
     )
 
@@ -284,6 +289,7 @@ def generate_data(
     client: Optional[openai.OpenAI] = None,
     pipeline: Optional[str] = "simple",
     batch_size: Optional[int] = None,
+    checkpoint_dir: Optional[str] = None,
 ) -> None:
     """Generate data for training and testing a model.
 
@@ -348,13 +354,17 @@ def generate_data(
         model_family,
         model_name,
         num_instructions_to_generate,
+        checkpoint_dir,
+        1,  # save_freq
         batch_size=batch_size,
         batch_num_workers=num_cpus,
     )
 
     sdg_knowledge, sdg_freeform_skill, sdg_grounded_skill = _sdg_init(ctx, pipeline)
 
-    mmlu_bench_pipe = mmlubench_pipe_init(ctx)
+    # Make sure checkpointing is disabled (we don't want this pipeline to load checkpoints from the main pipeline)
+    mmlu_ctx = dataclasses.replace(ctx, checkpoint_dir=None)
+    mmlu_bench_pipe = mmlubench_pipe_init(mmlu_ctx)
 
     mixer = _mixer_init(ctx, output_dir, date_suffix)
 

src/instructlab/sdg/pipeline.py

@@ -13,6 +13,7 @@
 import yaml
 
 # First Party
+from instructlab.sdg.checkpointing import Checkpointer
 from instructlab.sdg.utils import pandas
 
 # Local
@@ -61,6 +62,8 @@ class PipelineContext:  # pylint: disable=too-many-instance-attributes
     model_id: str
     num_instructions_to_generate: int
     dataset_num_procs: Optional[int] = DEFAULT_DATASET_NUM_PROCS
+    checkpoint_dir: Optional[str] = None
+    save_freq: Optional[int] = 1
     batch_size: int = DEFAULT_BATCH_SIZE
     batch_num_workers: Optional[int] = None
 
@@ -129,6 +132,12 @@ def generate(self, dataset) -> Dataset:
         Generate the dataset by running the pipeline steps.
         dataset: the input dataset
         """
+
+        # The checkpointer allows us to resume from where we left off
+        # Saving the output of pipe instances along the way
+        checkpointer = Checkpointer(self.ctx.checkpoint_dir, self.ctx.save_freq)
+        dataset, pre_generated_data = checkpointer.load(dataset)
+
         # If not batching, simply delegate to _generate_single
         if not self.ctx.batching_enabled:
             logger.info("Running pipeline single-threaded")
@@ -142,6 +151,7 @@ def generate(self, dataset) -> Dataset:
             self.ctx.batch_size,
         )
         input_splits = self._split_dataset(dataset)
+        output_splits = []
         with ThreadPoolExecutor(max_workers=self.ctx.batch_num_workers) as executor:
             futures = [
                 executor.submit(self._generate_single, input_split)
@@ -150,8 +160,13 @@ def generate(self, dataset) -> Dataset:
 
             # Collect the results of each batch as they finish. This needs to
             # wait for them all, so the order of waiting doesn't matter
-            output_splits = [future.result() for future in futures]
-
+            for future in futures:
+                ds = future.result()
+                output_splits.append(ds)
+                checkpointer.checkpoint(ds)
+        checkpointer.done()
+        if pre_generated_data:
+            output_splits.append(pre_generated_data)
         return concatenate_datasets(output_splits)
 
     ## Implementation Details ##

tests/test_checkpointing.py

@@ -0,0 +1,117 @@
+# Standard
+import json
+import os
+
+# Third Party
+from datasets import Dataset
+import pytest
+
+# First Party
+from instructlab.sdg.checkpointing import Checkpointer
+
+
+def _add_bar(sample, add_value=100):
+    sample["bar"] = sample["foo"] + add_value
+    return sample
+
+
+def _populate_checkpoints(tmpdir, dataset, checkpoints_count, remove_column):
+    for i in range(0, checkpoints_count):
+        checkpoint_dataset = dataset.select(range(i * 10, (i + 1) * 10))
+        checkpoint_dataset = checkpoint_dataset.map(
+            lambda x: _add_bar(x, add_value=100)
+        )
+        if remove_column:
+            checkpoint_dataset = checkpoint_dataset.remove_columns("foo")
+        checkpoint_dataset.to_json(
+            os.path.join(tmpdir, f"data_checkpoint_abcde{i}.jsonl"),
+            orient="records",
+            lines=True,
+        )
+
+
+def _validate_checkpoints(tmpdir, expected_files_count, expected_length, remove_column):
+    saved_files = os.listdir(tmpdir)
+    assert len(saved_files) == expected_files_count
+    assert all(f.startswith("data_checkpoint_") for f in saved_files)
+    assert all(f.endswith(".jsonl") for f in saved_files)
+
+    for f in saved_files:
+        with open(os.path.join(tmpdir, f), "r") as f:
+            l = list(f)
+            if isinstance(expected_length, list):
+                expected_length.remove(len(l))
+            else:
+                assert len(l) == expected_length
+            for s in l:
+                data = json.loads(s)
+                if remove_column:
+                    assert "foo" not in data and "bar" in data
+                else:
+                    assert "foo" in data and "bar" in data
+
+
+@pytest.mark.parametrize(
+    "save_freq, remove_column, dataset_size, init_checkpoints, splits, final_checkpoints, checkpoint_length",
+    [
+        (1, False, 10, 0, 0, 1, 10),
+        (1, True, 10, 0, 0, 1, 10),
+        (1, False, 100, 1, 9, 10, 10),
+        (1, True, 100, 1, 9, 10, 10),
+        (1, False, 100, 2, 8, 10, 10),
+        (3, False, 100, 2, 8, 5, [10, 10, 30, 30, 20]),
+    ],
+)
+def test_checkpointing(
+    tmpdir,
+    save_freq,
+    remove_column,
+    dataset_size,
+    init_checkpoints,
+    splits,
+    final_checkpoints,
+    checkpoint_length,
+):
+    # Our initial dataset
+    dataset = Dataset.from_list([{"idx": i, "foo": i} for i in range(dataset_size)])
+
+    # Generate and save some checkpoints to disk
+    _populate_checkpoints(tmpdir, dataset, init_checkpoints, remove_column)
+
+    # Load checkpoints, giving us the remaining dataset to process and
+    # the generated data loaded from the checkpoints
+    checkpointer = Checkpointer(checkpoint_dir=tmpdir, save_freq=save_freq)
+    dataset, pre_generated_data = checkpointer.load(dataset)
+
+    # Should be present, even if removed from the checkpoint (remove_column=True)
+    assert "foo" in dataset.features
+
+    # When testing save_freq, we will have checkpoints of different lengths
+    if isinstance(checkpoint_length, list):
+        checkpoints_total = sum(checkpoint_length[:init_checkpoints])
+    else:
+        checkpoints_total = checkpoint_length * init_checkpoints
+
+    # Validate pre-generated data loaded from the checkpoints
+    assert len(dataset) == (dataset_size - checkpoints_total)
+    if init_checkpoints > 0:
+        assert len(pre_generated_data) == checkpoints_total
+
+    # Apply pipeline to the remaining dataset and save checkpoints
+    if splits:
+        for i in range(0, splits):
+            split = dataset.select(range(i * 10, (i + 1) * 10))
+            split = split.map(lambda x: _add_bar(x, add_value=100))
+            if remove_column:
+                split = split.remove_columns("foo")
+            checkpointer.checkpoint(split)
+    else:
+        dataset = dataset.map(lambda x: _add_bar(x, add_value=10))
+        if remove_column:
+            dataset = dataset.remove_columns("foo")
+        checkpointer.checkpoint(dataset)
+
+    checkpointer.done()
+
+    # Validate that all checkpoints are now saved to disk
+    _validate_checkpoints(tmpdir, final_checkpoints, checkpoint_length, remove_column)

tests/test_generate_data.py

@@ -19,6 +19,8 @@ def test_context_init_batch_size_optional():
         "mixtral",
         "foo.bar",
         1,
+        "/checkpoint/dir",
+        1,
         batch_size=None,
         batch_num_workers=None,
     )
@@ -32,6 +34,8 @@ def test_context_init_batch_size_optional():
         "mixtral",
         "foo.bar",
         1,
+        "/checkpoint/dir",
+        1,
         batch_size=20,
         batch_num_workers=32,
     )