Train on wiki data.

lit_train.py

@@ -22,11 +22,11 @@ from utils import load_tokenizer
 def split_raw_dataset(
     raw_dataset: datasets.DatasetDict,
 ) -> Tuple[datasets.Dataset, datasets.Dataset]:
-    if 'validation' in raw_dataset:
-        train_dataset, val_dataset = raw_dataset['train'], raw_dataset['validation']
+    if "validation" in raw_dataset:
+        train_dataset, val_dataset = raw_dataset["train"], raw_dataset["validation"]
     else:
-        raw_dataset = raw_dataset['train'].train_test_split(test_size=0.05, seed=args.seed)
-        train_dataset, val_dataset = raw_dataset['train'], raw_dataset['test']
+        raw_dataset = raw_dataset["train"].train_test_split(test_size=0.05, seed=args.seed)
+        train_dataset, val_dataset = raw_dataset["train"], raw_dataset["test"]
     return train_dataset, val_dataset
 
 
@@ -39,32 +39,44 @@ def process_dataset(dataset: datasets.Dataset, tokenizer: PreTrainedTokenizer) -
             k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
             for k, t in concatenated_examples.items()
         }
-        result['labels'] = result['input_ids'].copy()
+        result["labels"] = result["input_ids"].copy()
         result = BatchEncoding(result)
         return result
 
+    def format_inputs(examples):
+        p = examples["段落"]
+        mergeLine = ""
+        for line in p:
+            mergeLine += line["内容"] + "\n"
+        return {"text": mergeLine}
+
     def tokenize_inputs(
         examples: Dict[str, list],
         tokenizer: PreTrainedTokenizer,
-        column_name: str = 'text',
+        column_name: str = "text",
     ) -> BatchEncoding:
         return tokenizer(examples[column_name], return_attention_mask=False)
 
     dataset_column_names = list(dataset.features)
     dataset = dataset.map(
-        partial(
-            tokenize_inputs,
-            tokenizer=tokenizer,
-            column_name=dataset_column_names[0],
-        ),
+        partial(format_inputs),
+        batched=False,
+        num_proc=args.num_proc,
+        remove_columns=dataset_column_names,
+    )
+    dataset_column_names = list(dataset.features)
+    dataset = dataset.map(
+        partial(tokenize_inputs, tokenizer=tokenizer),
         batched=True,
         num_proc=args.num_proc,
         remove_columns=dataset_column_names,
-    ).map(
+    )
+    dataset = dataset.map(
         partial(group_texts, block_size=tokenizer.model_max_length),
         batched=True,
         num_proc=args.num_proc,
     )
+
     return dataset
 
 
@@ -74,7 +86,7 @@ def parse_args():
         "--model_name",
         type=str,
         help="Name of or path to model",
-        default='gpt2',
+        default="gpt2",
     )
     parser.add_argument(
         "--learning_rate",
@@ -87,8 +99,12 @@ def parse_args():
         help="Use tril attention mask during training",
         action="store_true",
     )
-    parser.add_argument("--fp16", help="Enable fp16", action="store_true")
-    parser.add_argument("--bf16", help="Enable bf16", action="store_true")
+    parser.add_argument(
+        "--precision",
+        help="precision:bf16-mixed,16-mixed,32-true",
+        action="store_true",
+        default="16-mixed",
+    )
     parser.add_argument(
         "--tokenizer_name_or_path",
         type=str,
@@ -97,10 +113,10 @@ def parse_args():
     )
     parser.add_argument(
         "--dataset_name",
-        nargs='+',
+        nargs="+",
         type=str,
         help="Name(s) of dataset. To specify a config, pass a <dataset_name>:<dataset_config_name>",
-        default=["wikitext:wikitext-2-v1"],
+        default=["/home/colin/develop/dataset/liwu/MNBVC/wiki"],
     )
     parser.add_argument(
         "--train_batch_size",
@@ -124,7 +140,7 @@ def parse_args():
         "--num_proc",
         type=str,
         help="Number of data processes",
-        default=1,
+        default=12,
     )
     parser.add_argument(
         "--max_epochs",
@@ -136,7 +152,7 @@ def parse_args():
         "--strategy",
         type=str,
         help="Name of pytorch lightning distribution strategy",
-        default='ddp',
+        default="fsdp",
     )
     parser.add_argument(
         "--resume_from_ckpt_path",
@@ -154,7 +170,7 @@ def parse_args():
     return args
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     args = parse_args()
 
     if args.tokenizer_name_or_path is None:
@@ -170,8 +186,11 @@ if __name__ == '__main__':
     train_dataset_list = []
     val_dataset_list = []
     for dataset_name in args.dataset_name:
-        dataset_args = dataset_name.split(':')
-        raw_dataset = datasets.load_dataset(*dataset_args)
+        dataset_args = dataset_name.split(":")
+        raw_dataset = datasets.load_dataset(
+            "json", data_files="/home/colin/develop/dataset/liwu/MNBVC/wiki/20230197/0.jsonl.gz"
+        )
+        # raw_dataset = datasets.load_dataset(*dataset_args)
         train_dataset, val_dataset = split_raw_dataset(raw_dataset)
         train_dataset = process_dataset(train_dataset, tokenizer)
         val_dataset = process_dataset(val_dataset, tokenizer)
@@ -198,19 +217,13 @@ if __name__ == '__main__':
     )
 
     ne = next(train_dataloader._get_iterator())
-    print((ne["input_ids"]-ne["labels"]).numpy().tolist())
 
     # trainer
     # apply_all_patches()
-    torch.set_float32_matmul_precision('medium')
-    if args.bf16:
-        precision = 'bf16-mixed'
-    elif args.fp16:
-        precision = '16-mixed'
-    else:
-        precision = "32-true"
+    torch.set_float32_matmul_precision("medium")
+    precision = args.precision
     lit_trainer = pl.Trainer(
-        accelerator='gpu',
+        accelerator="gpu",
         precision=precision,
         log_every_n_steps=5,
         accumulate_grad_batches=args.accumulate_grad_batches,