35% faster packing + rename bfd-requeue to bfd_split

tests/test_data_utils.py

@@ -23,6 +23,7 @@
 from transformers import AutoProcessor, AutoTokenizer, is_vision_available
 
 from trl.data_utils import (
+    PackingStrategy,
     apply_chat_template,
     extract_prompt,
     is_conversational,
@@ -1056,57 +1057,82 @@ def test_with_dataset(self):
             "attention_mask": [[0, 1, 1], [0, 0, 1, 1], [1]],
         }
         dataset = Dataset.from_dict(examples)
+        dataset = dataset.with_format("numpy", dtype="float32")
+        format = dataset.format
         seq_length = 3
         expected_output = {
             "input_ids": [[1, 2, 3], [4, 5, 6], [7, 8]],
             "attention_mask": [[0, 1, 1], [0, 0, 1], [1, 1]],
         }
         dataset = pack_dataset(dataset, seq_length, strategy="wrapped")
         assert dataset.to_dict() == expected_output
+        assert format == dataset.format
 
     def test_with_iterable_dataset(self):
         examples = {
             "input_ids": [[1, 2, 3], [4, 5, 6, 7], [8]],
             "attention_mask": [[0, 1, 1], [0, 0, 1, 1], [1]],
         }
         dataset = Dataset.from_dict(examples).to_iterable_dataset()
+        dataset = dataset.with_format("numpy")
+        formatting = dataset._formatting
         seq_length = 3
         expected_output = {
             "input_ids": [[1, 2, 3], [4, 5, 6], [7, 8]],
             "attention_mask": [[0, 1, 1], [0, 0, 1], [1, 1]],
         }
         dataset = pack_dataset(dataset, seq_length, strategy="wrapped")
         num_examples = len(examples[next(iter(examples))])
-        assert next(iter(dataset.batch(batch_size=num_examples))) == expected_output
+        assert next(iter(dataset.with_format(None).batch(batch_size=num_examples))) == expected_output
+        assert formatting == dataset._formatting
+
+
+class TestPackingStrategy(TrlTestCase):
+    def test_aliases(self):
+        assert PackingStrategy("bfd-split") is PackingStrategy.BFD_SPLIT
+        assert PackingStrategy("bfd-truncate") is PackingStrategy.BFD
+
+    def test_missing_value_raises_value_error(self):
+        with pytest.raises(ValueError, match="not a valid PackingStrategy"):
+            PackingStrategy("missing")
 
 
 class TestPackDatasetBfd(TrlTestCase):
-    def test_simple(self):
+    def test_with_dataset(self):
         examples = {
             "input_ids": [[1, 2, 3], [4, 5, 6, 7], [8]],
         }
         dataset = Dataset.from_dict(examples)
+        dataset = dataset.with_format("numpy", dtype="float32")
+        format = dataset.format
         seq_length = 4
         expected_output = {
             "input_ids": [[4, 5, 6, 7], [1, 2, 3, 8]],
             "seq_lengths": [[4], [3, 1]],
         }
         dataset = pack_dataset(dataset, seq_length, strategy="bfd")
+        expected_format = dataset.format
         assert dataset.to_dict() == expected_output
+        assert "seq_lengths" in expected_format["columns"]
+        expected_format["columns"].remove("seq_lengths")
+        assert format == dataset.format
 
     def test_with_iterable_dataset(self):
         examples = {
             "input_ids": [[1, 2, 3], [4, 5, 6, 7], [8]],
         }
         dataset = Dataset.from_dict(examples).to_iterable_dataset()
+        dataset = dataset.with_format("numpy")
+        formatting = dataset._formatting
         seq_length = 4
         expected_output = {
             "input_ids": [[4, 5, 6, 7], [1, 2, 3, 8]],
             "seq_lengths": [[4], [3, 1]],
         }
         dataset = pack_dataset(dataset, seq_length, strategy="bfd")
         num_examples = len(examples[next(iter(examples))])
-        assert next(iter(dataset.batch(batch_size=num_examples))) == expected_output
+        assert next(iter(dataset.with_format(None).batch(batch_size=num_examples))) == expected_output
+        assert formatting == dataset._formatting
 
     def test_with_overlong_0(self):
         examples = {
@@ -1118,7 +1144,7 @@ def test_with_overlong_0(self):
             "input_ids": [[1, 2, 3, 4], [8, 9, 10, 11], [6, 7, 5, 12]],
             "seq_lengths": [[4], [4], [2, 1, 1]],
         }
-        dataset = pack_dataset(dataset, seq_length, strategy="bfd-requeue")
+        dataset = pack_dataset(dataset, seq_length, strategy="bfd_split")
         assert dataset.to_dict() == expected_output
 
     def test_with_overlong_two_coluns(self):
@@ -1133,7 +1159,7 @@ def test_with_overlong_two_coluns(self):
             "col2": [[-1, 2, -3, 4], [-13, 14, -15, 16], [-7, 8, -9], [10, -11, 12], [-5, 6]],
             "seq_lengths": [[4], [4], [3], [3], [2]],
         }
-        dataset = pack_dataset(dataset, seq_length, strategy="bfd-requeue")
+        dataset = pack_dataset(dataset, seq_length, strategy="bfd_split")
         assert dataset.to_dict() == expected_output
 
     def test_with_non_power_of_2(self):
@@ -1146,10 +1172,10 @@ def test_with_non_power_of_2(self):
             "input_ids": [[1, 2, 3, 4, 5], [7, 8, 9, 10, 6], [11, 12, 13]],
             "seq_lengths": [[5], [4, 1], [3]],
         }
-        dataset = pack_dataset(dataset, seq_length, strategy="bfd-requeue")
+        dataset = pack_dataset(dataset, seq_length, strategy="bfd_split")
         assert dataset.to_dict() == expected_output
 
-    def test_default_no_requeue(self):
+    def test_default_no_split(self):
         """Test default 'bfd' strategy for SFT datasets (truncates overflow)."""
         examples = {
             "input_ids": [[1, 2, 3, 4, 5], [6, 7], [8, 9, 10, 11], [12]],
@@ -1164,6 +1190,19 @@ def test_default_no_requeue(self):
         dataset = pack_dataset(dataset, seq_length, strategy="bfd")
         assert dataset.to_dict() == expected_output
 
+    def test_with_empty_sequences(self):
+        examples = {
+            "input_ids": [[1, 2], [], [3, 4, 5], [], [6]],
+        }
+        dataset = Dataset.from_dict(examples)
+        seq_length = 4
+        expected_output = {
+            "input_ids": [[3, 4, 5, 6], [1, 2]],
+            "seq_lengths": [[3, 1], [2]],
+        }
+        dataset = pack_dataset(dataset, seq_length, strategy="bfd_split")
+        assert dataset.to_dict() == expected_output
+
 
 class TestTruncateExamples(TrlTestCase):
     def test_with_dataset(self):
@@ -1172,28 +1211,34 @@ def test_with_dataset(self):
             "attention_mask": [[0, 1, 1], [0, 0, 1, 1], [1]],
         }
         dataset = Dataset.from_dict(examples)
+        dataset = dataset.with_format("numpy", dtype="float32")
+        format = dataset.format
         max_length = 2
         expected_output = {
             "input_ids": [[1, 2], [4, 5], [8]],
             "attention_mask": [[0, 1], [0, 0], [1]],
         }
         dataset = truncate_dataset(dataset, max_length)
         assert dataset.to_dict() == expected_output
+        assert format == dataset.format
 
     def test_with_iterable_dataset(self):
         examples = {
             "input_ids": [[1, 2, 3], [4, 5, 6, 7], [8]],
             "attention_mask": [[0, 1, 1], [0, 0, 1, 1], [1]],
         }
         dataset = Dataset.from_dict(examples).to_iterable_dataset()
+        dataset = dataset.with_format("numpy")
+        formatting = dataset._formatting
         max_length = 2
         expected_output = {
             "input_ids": [[1, 2], [4, 5], [8]],
             "attention_mask": [[0, 1], [0, 0], [1]],
         }
         dataset = truncate_dataset(dataset, max_length)
         num_examples = len(examples[next(iter(examples))])
-        assert next(iter(dataset.batch(batch_size=num_examples))) == expected_output
+        assert next(iter(dataset.with_format(None).batch(batch_size=num_examples))) == expected_output
+        assert formatting == dataset._formatting
 
     def test_with_extra_column(self):
         examples = {

trl/__init__.py

@@ -28,6 +28,7 @@
 _import_structure = {
     "chat_template_utils": ["add_response_schema", "clone_chat_template", "get_training_chat_template"],
     "data_utils": [
+        "PackingStrategy",
         "apply_chat_template",
         "extract_prompt",
         "is_conversational",
@@ -72,6 +73,7 @@
 if TYPE_CHECKING:
     from .chat_template_utils import add_response_schema, clone_chat_template, get_training_chat_template
     from .data_utils import (
+        PackingStrategy,
         apply_chat_template,
         extract_prompt,
         is_conversational,