pipe neox_args into GPT2Dataset

megatron/data/data_utils.py

@@ -61,6 +61,7 @@ def build_the_dataset(
     skip_warmup,
     build_index_mappings=True,
     label_prefix=None,
+    neox_args=None,
 ):
     """Build train/valid/test datasets."""
 
@@ -85,6 +86,7 @@ def build_the_dataset(
         seed,
         build_index_mappings=build_index_mappings,
         label_dataset=label_dataset,
+        neox_args=neox_args
     )
     return dataset
 
@@ -98,6 +100,7 @@ def build_train_valid_test_datasets(
     seq_length,
     seed,
     skip_warmup,
+    neox_args=None
 ):
     """Build train, valid, and test datasets."""
 
@@ -139,6 +142,7 @@ def build_dataset(index, name):
                 seq_length,
                 seed,
                 use_shared_fs=use_shared_fs,
+                neox_args=neox_args
             )
         return dataset
 
@@ -224,6 +228,7 @@ def build_weighted_datasets(
                     skip_warmup=(not neox_args.mmap_warmup),
                     build_index_mappings=build_index_mappings,
                     label_prefix=label_path,
+                    neox_args=neox_args
                 )
             )
 
@@ -238,6 +243,7 @@ def build_weighted_datasets(
                     seed=neox_args.seed,
                     skip_warmup=(not neox_args.mmap_warmup),
                     build_index_mappings=build_index_mappings,
+                    neox_args=neox_args
                 )
             )
 
@@ -252,6 +258,7 @@ def build_weighted_datasets(
                     seed=neox_args.seed,
                     skip_warmup=(not neox_args.mmap_warmup),
                     build_index_mappings=build_index_mappings,
+                    neox_args=neox_args
                 )
             )
     return train_datasets, valid_datasets, test_datasets
@@ -394,6 +401,7 @@ def build_train_valid_test_data_iterators(neox_args):
                     train_weights,
                     valid_weights,
                     test_weights,
+                    neox_args=neox_args
                 )
 
             if train_datasets:
@@ -414,6 +422,7 @@ 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),
+                neox_args=neox_args
             )
 
         # Build dataloders.

megatron/data/gpt2_dataset.py

@@ -39,12 +39,15 @@ def __init__(
         build_index_mappings=True,
         use_shared_fs=True,
         label_dataset=None,
+        neox_args=None,
     ):
 
         self.name = name
         self.indexed_dataset = indexed_dataset
         self.label_dataset = label_dataset
 
+        self.neox_args = neox_args
+
         # Checks
         assert np.min(documents) >= 0
         assert np.max(documents) < indexed_dataset.sizes.shape[0]
@@ -116,21 +119,19 @@ def __getitem__(self, idx):
 
             conditional_training = True
             if conditional_training:
-                sample_text = self.neox_args.tokenizer.decode(sample[0])
+                sample_text = self.neox_args.tokenizer.detokenize(samples[0])
 
                 new_sample_text = ""
-                for i, (spacy_doc, meta) in enumerate(spacy_model.pipe((sample_text, "metadata"), n_process=8, as_tuples=True)):
+
+                for i, spacy_doc in enumerate(spacy_model.pipe([sample_text])):
                     for sent in spacy_doc.sents:
                         sent = sent.text_with_ws
                         new_sample_text += "<|endoftext|>"
                         new_sample_text += sent
-                encoded_new_text = self.neox_args.tokenizer.encode(new_sample_text)
-                print(len(encoded_new_text), len(samples[0])
+                encoded_new_text = self.neox_args.tokenizer.tokenize(new_sample_text)
+                # print(len(encoded_new_text), len(samples[0]))
 
 
-            for idx, (spacy_doc, meta) in enumerate(spacy_model.pipe(raw_texts, n_process=8, as_tuples=True)):
-                for sent in spacy_doc.sents:
-                    yield {'text': sent.text_with_ws, 'meta': meta, 'idx': idx}
             if len(datasets) == 1:
                 return {"text": np.array(samples[0], dtype=np.int64)}
             else: