feat: allow for padding free plugin to be used without response template (#430)

* fix: allow for padding free + pretraining

Signed-off-by: Harikrishnan Balagopal <[email protected]>

* add data collator for padding free plugin scenario to be used for extended pretraining

Signed-off-by: Dushyant Behl <[email protected]>

* fix: update value error

Signed-off-by: Mehant Kammakomati <[email protected]>

* fix: delete images only when exists

Signed-off-by: Mehant Kammakomati <[email protected]>

---------

Signed-off-by: Harikrishnan Balagopal <[email protected]>
Signed-off-by: Dushyant Behl <[email protected]>
Signed-off-by: Mehant Kammakomati <[email protected]>
Co-authored-by: Harikrishnan Balagopal <[email protected]>
Co-authored-by: Mehant Kammakomati <[email protected]>

Author: 3 people

.github/workflows/image.yaml

@@ -15,9 +15,8 @@ jobs:
           sudo swapoff -a
           sudo rm -f /swapfile
           sudo apt clean
-          docker rmi $(docker image ls -aq)
+          if [ "$(docker image ls -q)" ]; then docker rmi $(docker image ls -aq); fi
           df -h
       - name: Build image
         run: |
           docker build -t fms-hf-tuning:dev . -f build/Dockerfile
-          

tests/data/test_data_preprocessing_utils.py

@@ -489,7 +489,7 @@ def test_is_pretokenized_data(data, result):
 
 @pytest.mark.parametrize(
     "packing, response_template, formatted_train_dataset,\
-     max_seq_length, instruction_template, expected_collator",
+     max_seq_length, instruction_template, is_padding_free, expected_collator",
     [
         (
             False,
@@ -501,6 +501,7 @@ def test_is_pretokenized_data(data, result):
             ),
             1024,
             None,
+            False,
             DataCollatorForCompletionOnlyLM,
         ),
         (
@@ -517,6 +518,7 @@ def test_is_pretokenized_data(data, result):
             ),
             1024,
             None,
+            False,
             DataCollatorForSeq2Seq,
         ),
         (
@@ -529,6 +531,7 @@ def test_is_pretokenized_data(data, result):
             ),
             1024,
             "\n### Text:",
+            False,
             DataCollatorForCompletionOnlyLM,
         ),
         (
@@ -545,6 +548,20 @@ def test_is_pretokenized_data(data, result):
             ),
             1024,
             "\n### Text:",
+            False,
+            DataCollatorForSeq2Seq,
+        ),
+        (
+            False,
+            None,
+            datasets.load_dataset(
+                "json",
+                data_files=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_JSON,
+                split="train",
+            ),
+            1024,
+            None,
+            True,
             DataCollatorForSeq2Seq,
         ),
     ],
@@ -555,6 +572,7 @@ def test_get_data_collator(
     formatted_train_dataset,
     max_seq_length,
     instruction_template,
+    is_padding_free,
     expected_collator,
 ):
     """Ensure that the correct collator type is fetched based on the data args"""
@@ -565,6 +583,7 @@ def test_get_data_collator(
         is_pretokenized_dataset(formatted_train_dataset),
         max_seq_length,
         instruction_template,
+        is_padding_free,
     )
     assert isinstance(collator, expected_collator)
 
@@ -1044,7 +1063,7 @@ def test_process_dataconfig_multiple_datasets_datafiles_sampling(
 
 
 @pytest.mark.parametrize(
-    "data_args",
+    "data_args, is_padding_free",
     [
         # single sequence JSON and response template
         (
@@ -1053,7 +1072,8 @@ def test_process_dataconfig_multiple_datasets_datafiles_sampling(
                 validation_data_path=TWITTER_COMPLAINTS_DATA_JSON,
                 dataset_text_field="output",
                 response_template="\n### Label:",
-            )
+            ),
+            False,
         ),
         # single sequence JSONL and response template
         (
@@ -1062,7 +1082,8 @@ def test_process_dataconfig_multiple_datasets_datafiles_sampling(
                 validation_data_path=TWITTER_COMPLAINTS_DATA_JSONL,
                 dataset_text_field="output",
                 response_template="\n### Label:",
-            )
+            ),
+            False,
         ),
         # single sequence PARQUET and response template
         (
@@ -1071,7 +1092,8 @@ def test_process_dataconfig_multiple_datasets_datafiles_sampling(
                 validation_data_path=TWITTER_COMPLAINTS_DATA_PARQUET,
                 dataset_text_field="output",
                 response_template="\n### Label:",
-            )
+            ),
+            False,
         ),
         # data formatter template with input/output JSON
         (
@@ -1080,7 +1102,8 @@ def test_process_dataconfig_multiple_datasets_datafiles_sampling(
                 validation_data_path=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_JSON,
                 data_formatter_template="### Text:{{input}} \n\n### Label: {{output}}",
                 response_template="\n### Label:",
-            )
+            ),
+            False,
         ),
         # data formatter template with input/output JSONL
         (
@@ -1089,7 +1112,8 @@ def test_process_dataconfig_multiple_datasets_datafiles_sampling(
                 validation_data_path=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_JSONL,
                 data_formatter_template="### Text:{{input}} \n\n### Label: {{output}}",
                 response_template="\n### Label:",
-            )
+            ),
+            False,
         ),
         # data formatter template with input/output PARQUET
         (
@@ -1098,32 +1122,44 @@ def test_process_dataconfig_multiple_datasets_datafiles_sampling(
                 validation_data_path=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_PARQUET,
                 data_formatter_template="### Text:{{input}} \n\n### Label: {{output}}",
                 response_template="\n### Label:",
-            )
+            ),
+            False,
         ),
         # input/output JSON with masking on input
         (
             configs.DataArguments(
                 training_data_path=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_JSON,
                 validation_data_path=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_JSON,
-            )
+            ),
+            False,
         ),
         # input/output JSONL with masking on input
         (
             configs.DataArguments(
                 training_data_path=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_JSONL,
                 validation_data_path=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_JSONL,
-            )
+            ),
+            False,
         ),
         # input/output PARQUET with masking on input
         (
             configs.DataArguments(
                 training_data_path=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_PARQUET,
                 validation_data_path=TWITTER_COMPLAINTS_DATA_INPUT_OUTPUT_PARQUET,
-            )
+            ),
+            False,
+        ),
+        (
+            configs.DataArguments(
+                training_data_path=TWITTER_COMPLAINTS_DATA_JSON,
+                validation_data_path=TWITTER_COMPLAINTS_DATA_JSON,
+                dataset_text_field="output",
+            ),
+            True,
         ),
     ],
 )
-def test_process_dataargs(data_args):
+def test_process_dataargs(data_args, is_padding_free):
     """Ensure that the train/eval data are properly formatted based on the data args / text field"""
     tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
     TRAIN_ARGS = configs.TrainingArguments(
@@ -1132,7 +1168,7 @@ def test_process_dataargs(data_args):
         output_dir="tmp",  # Not needed but positional
     )
     (train_set, eval_set, dataset_text_field, _, _, _) = process_dataargs(
-        data_args, tokenizer, TRAIN_ARGS
+        data_args, tokenizer, TRAIN_ARGS, is_padding_free=is_padding_free
     )
     assert isinstance(train_set, Dataset)
     assert isinstance(eval_set, Dataset)

tuning/config/acceleration_configs/attention_and_distributed_packing.py

@@ -47,3 +47,7 @@ class AttentionAndDistributedPackingConfig:
     def __post_init__(self):
         # ensure nested dataclasses initialized
         ensure_nested_dataclasses_initialized(self)
+
+    @property
+    def is_padding_free(self):
+        return self.padding_free is not None

tuning/data/data_preprocessing_utils.py

@@ -29,6 +29,7 @@ def get_data_collator(
     is_traindata_tokenized: bool,
     max_seq_length: int,
     instruction_template: Optional[str],
+    is_padding_free: bool = False,
 ) -> Callable:
     """Create and return the the appropriate collator type based on the configuration for packing,
     response_template, and dataset_text_field.
@@ -46,6 +47,8 @@ def get_data_collator(
             Max sequence length expected
         instruction_template: str
             str representing the human response in a chat template
+        is_padding_free: bool
+            if padding free plugin is used or not
 
     Returns:
         Callable
@@ -74,6 +77,16 @@ def get_data_collator(
                 tokenizer=tokenizer,
                 ignore_index=configs.IGNORE_INDEX,
             )
+
+        if is_padding_free:
+            # when packing is false but padding_free is used and
+            # no response template is used then its a pretrained scenario.
+            # Current plugin in fms-acceleration is compatible with
+            # `DataCollatorForSeq2Seq` collator hence we use this.
+            return DataCollatorForSeq2Seq(
+                tokenizer=tokenizer, padding=False, max_length=max_seq_length
+            )
+
         # Note that this automatically pads labels with -100
         # TODO check if this is sufficient for preprocessed
         if is_traindata_tokenized:

tuning/data/setup_dataprocessor.py

@@ -107,15 +107,22 @@ def _get_pretokenized_dataset_handlers(data_args, packing, is_eval_tokenized):
 
 
 ### Data format 2
-def _get_dataset_formatting_handlers(data_args, packing):
+def _get_dataset_formatting_handlers(data_args, packing, is_padding_free=False):
 
     if data_args.response_template is None:
         if packing is False:
-            raise ValueError(
-                "Since dataset_text_field or data_formatter_template \
-                is provided and packing is disabled, \
-                needs a corresponding response template for masking"
-            )
+            if is_padding_free:
+                logger.debug(
+                    "Assuming pretraining scenario (loss over all tokens) "
+                    + "because, packing is false,"
+                    + " padding_free plugin is used and no response template was provided."
+                )
+            else:
+                raise ValueError(
+                    "Since response_template is not provided for masking, \
+                    either use packing or padding_free to enable \
+                    pretraining scenario (loss over all tokens)."
+                )
 
     if data_args.response_template:
         # To use Response template, pass datasets with single sequence instances \
@@ -209,6 +216,7 @@ def _process_raw_data_args(
     packing: bool,
     max_seq_length: int,
     additional_data_handlers: Dict[str, Callable] = None,
+    is_padding_free: bool = False,
 ):
 
     # Create a data processor with default processor config
@@ -248,6 +256,7 @@ def _process_raw_data_args(
     tokenizer_kwargs = {}
     tokenizer_kwargs["max_length"] = max_seq_length
     tokenizer_kwargs["truncation"] = True
+    # Lets not pad in tokenizer...we can handle that in the collator
     tokenizer_kwargs["padding"] = False
 
     handlers = None
@@ -266,7 +275,7 @@ def _process_raw_data_args(
     elif data_args.data_formatter_template or data_args.dataset_text_field:
         # Data Format 3: Single Sequence Dataset
         handlers, dataset_text_field = _get_dataset_formatting_handlers(
-            data_args, packing
+            data_args, packing, is_padding_free
         )
     else:
         # Default Data Format: Dataset with Input/Output Fields
@@ -300,6 +309,7 @@ def process_dataargs(
     tokenizer: AutoTokenizer,
     train_args: TrainingArguments,
     additional_data_handlers: Dict[str, Callable] = None,
+    is_padding_free: bool = False,
 ):
     """
     Args:
@@ -310,6 +320,8 @@ def process_dataargs(
             Used for packing and max_seq_length
         additional_data_handlers: A Dict of [str, callable] data handlers
             which need to be registered with the data preprocessor
+        is_padding_free: A bool representing if Padding free plugin is enabled.
+                         Defaults to False.
     Returns:
         Tuple(Dataset, Dataset, str, DataCollator, int, Dict)
             tuple containing
@@ -345,6 +357,7 @@ def process_dataargs(
             train_args.packing,
             max_seq_length,
             additional_data_handlers,
+            is_padding_free,
         )
 
     # Note: This check should not be removed.
@@ -359,6 +372,7 @@ def process_dataargs(
         is_tokenized_dataset,
         max_seq_length,
         data_args.instruction_template,
+        is_padding_free=is_padding_free,
     )
 
     dataset_kwargs = {}

tuning/sft_trainer.py

@@ -306,6 +306,10 @@ def train(
     data_collator = None
     logger.info("Packing is set to %s ", train_args.packing)
 
+    is_padding_free = False
+    if attention_and_distributed_packing_config is not None:
+        is_padding_free = attention_and_distributed_packing_config.is_padding_free
+
     data_preprocessing_time = time.time()
     (
         formatted_train_dataset,
@@ -314,7 +318,13 @@ def train(
         data_collator,
         train_args.max_seq_length,
         dataset_kwargs,
-    ) = process_dataargs(data_args, tokenizer, train_args, additional_data_handlers)
+    ) = process_dataargs(
+        data_args,
+        tokenizer,
+        train_args,
+        additional_data_handlers,
+        is_padding_free=is_padding_free,
+    )
     additional_metrics["data_preprocessing_time"] = (
         time.time() - data_preprocessing_time
     )