adjust to accept different dataset type

megatron/data/data_utils.py

@@ -23,6 +23,7 @@
 from megatron.data.indexed_dataset import make_dataset as make_indexed_dataset
 from megatron.data.blendable_dataset import BlendableDataset
 from megatron.data.gpt2_dataset import GPT2Dataset
+from megatron.data.ptok_dataset import PTokDataset
 from megatron.data.samplers import DistributedBatchSampler
 
 
@@ -51,7 +52,18 @@ def make_data_loader(dataset, neox_args):
     )
 
 
+def get_dataset_obj(dataset_type):
+    print_rank_0("> building {} dataset ...".format(dataset_type))
+    # Select and instantiate the dataset.
+    if dataset_type.lower() == "default".lower():
+        dataset_obj = GPT2Dataset
+    elif dataset_type.lower() == "pause".lower():
+        dataset_obj = PTokDataset
+    return dataset_obj
+
+
 def build_the_dataset(
+    dataset_type,
     data_prefix,
     name,
     data_impl,
@@ -61,6 +73,7 @@ def build_the_dataset(
     skip_warmup,
     build_index_mappings=True,
     label_prefix=None,
+    dataset_configs=None,
 ):
     """Build train/valid/test datasets."""
 
@@ -75,7 +88,9 @@ def build_the_dataset(
     print_rank_0("     no. of documents:{}".format(total_num_of_documents))
     dataset = None
     documents = np.arange(start=0, stop=total_num_of_documents, step=1, dtype=np.int32)
-    dataset = GPT2Dataset(
+
+    dataset_obj = get_dataset_obj(dataset_type)
+    dataset = dataset_obj(
         name,
         data_prefix,
         documents,
@@ -85,11 +100,13 @@ def build_the_dataset(
         seed,
         build_index_mappings=build_index_mappings,
         label_dataset=label_dataset,
+        **dataset_configs if dataset_configs is not None else {},
     )
     return dataset
 
 
 def build_train_valid_test_datasets(
+    dataset_type,
     data_prefix,
     use_shared_fs,
     data_impl,
@@ -98,6 +115,7 @@ def build_train_valid_test_datasets(
     seq_length,
     seed,
     skip_warmup,
+    dataset_configs=None,
 ):
     """Build train, valid, and test datasets."""
 
@@ -130,7 +148,8 @@ def build_dataset(index, name):
                 start=splits[index], stop=splits[index + 1], step=1, dtype=np.int32
             )
 
-            dataset = GPT2Dataset(
+            dataset_obj = get_dataset_obj(dataset_type)
+            dataset = dataset_obj(
                 name,
                 data_prefix,
                 documents,
@@ -139,6 +158,7 @@ def build_dataset(index, name):
                 seq_length,
                 seed,
                 use_shared_fs=use_shared_fs,
+                **dataset_configs if dataset_configs is not None else {},
             )
         return dataset
 
@@ -215,6 +235,7 @@ def build_weighted_datasets(
         if train_path:
             train_datasets.append(
                 build_the_dataset(
+                    dataset_type=neox_args.dataset_type,
                     data_prefix=train_path,
                     name=f"train_{i}",
                     data_impl=neox_args.data_impl,
@@ -224,12 +245,14 @@ def build_weighted_datasets(
                     skip_warmup=(not neox_args.mmap_warmup),
                     build_index_mappings=build_index_mappings,
                     label_prefix=label_path,
+                    dataset_configs=neox_args.dataset_cfg,
                 )
             )
 
         if valid_path:
             valid_datasets.append(
                 build_the_dataset(
+                    dataset_type=neox_args.dataset_type,
                     data_prefix=valid_path,
                     name=f"valid_{i}",
                     data_impl=neox_args.data_impl,
@@ -238,12 +261,14 @@ def build_weighted_datasets(
                     seed=neox_args.seed,
                     skip_warmup=(not neox_args.mmap_warmup),
                     build_index_mappings=build_index_mappings,
+                    dataset_configs=neox_args.dataset_cfg,
                 )
             )
 
         if test_path:
             test_datasets.append(
                 build_the_dataset(
+                    dataset_type=neox_args.dataset_type,
                     data_prefix=test_path,
                     name=f"test_{i}",
                     data_impl=neox_args.data_impl,
@@ -252,6 +277,7 @@ def build_weighted_datasets(
                     seed=neox_args.seed,
                     skip_warmup=(not neox_args.mmap_warmup),
                     build_index_mappings=build_index_mappings,
+                    dataset_configs=neox_args.dataset_cfg,
                 )
             )
     return train_datasets, valid_datasets, test_datasets
@@ -406,6 +432,7 @@ def build_train_valid_test_data_iterators(neox_args):
             # when just data_path is provided
             # split dataset into train, valid and test from data_path
             train_ds, valid_ds, test_ds = build_train_valid_test_datasets(
+                dataset_type=neox_args.dataset_type,
                 data_prefix=neox_args.data_path,
                 use_shared_fs=neox_args.use_shared_fs,
                 data_impl=neox_args.data_impl,
@@ -414,6 +441,7 @@ def build_train_valid_test_data_iterators(neox_args):
                 seq_length=neox_args.seq_length,
                 seed=neox_args.seed,
                 skip_warmup=(not neox_args.mmap_warmup),
+                dataset_configs=neox_args.dataset_cfg,
             )
 
         # Build dataloders.