feat: support padding-free + pretraining

Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com>

Author: kmehant

tuning/sft_trainer.py

@@ -341,6 +341,7 @@ def train(
         max_seq_length,
         data_args.tokens_field,
         data_args.instruction_template,
+        attention_and_distributed_packing_config.padding_free,
     )
 
     if framework is not None and framework.requires_agumentation:

tuning/utils/preprocessing_utils.py

@@ -167,6 +167,7 @@ def get_data_collator(
     max_seq_length: int,
     tokens_field: str = True,
     instruction_template: Optional[str] = None,
+    padding_free: str = None,
 ) -> Callable:
     """Create and return the the appropriate collator type based on the configuration for packing,
     response_template, and dataset_text_field.
@@ -186,6 +187,8 @@ def get_data_collator(
             feature having tokens
         instruction_template: Optional[str]
             start of user answer.
+        padding_free: str
+            padding free method
 
     Returns:
         Callable
@@ -240,6 +243,12 @@ def get_data_collator(
             return DataCollatorForSeq2Seq(
                 tokenizer=tokenizer, padding=True, max_length=max_seq_length
             )
+        if padding_free:
+            # when packing is false but padding_free is used and no response template is used
+            # then its a pretrained scenario.
+            return DataCollatorForSeq2Seq(
+                tokenizer=tokenizer, padding=False, max_length=max_seq_length
+            )
         raise ValueError(
             "Could not pick a data collator. Please refer to supported data formats"
         )