[feature] new arg use_tril_attention_mask

generate.py

@@ -32,16 +32,20 @@ def load_tokenizer(model_name_or_path: Union[str, os.PathLike]) -> PreTrainedTok
 
 
 def eval_prompts(
-    model: PreTrainedModel, tokenizer: PreTrainedTokenizer, prompts: List[str]
+    model: PreTrainedModel,
+    tokenizer: PreTrainedTokenizer,
+    prompts: List[str],
+    use_tril_attention_mask: bool = False,
 ) -> List[str]:
     inputs = tokenizer(
         prompts, padding=True, return_tensors='pt', return_attention_mask=True
     )
     inputs['position_ids'] = inputs.attention_mask.cumsum(-1) - 1
     inputs['position_ids'].masked_fill_(inputs.attention_mask == 0, 1)
-    inputs['attention_mask'] = (
-        inputs.attention_mask.unsqueeze(1) * inputs.attention_mask.unsqueeze(2)
-    ).tril()
+    if use_tril_attention_mask:
+        inputs['attention_mask'] = (
+            inputs.attention_mask.unsqueeze(1) * inputs.attention_mask.unsqueeze(2)
+        ).tril()
     inputs = inputs.to(model.device)
     with torch.inference_mode():
         output_ids = model.generate(
@@ -66,6 +70,17 @@ def parse_args():
         help="Name of or path to model",
         default='gpt2',
     )
+    parser.add_argument(
+        "--use_tril_attention_mask",
+        help="Use tril attention mask during training",
+        action="store_true",
+    )
+    parser.add_argument(
+        "--tokenizer_name_or_path",
+        type=str,
+        help="Name of or path to tokenizer",
+        default=None,
+    )
     args = parser.parse_args()
     return args
 
@@ -73,10 +88,13 @@ def parse_args():
 if __name__ == '__main__':
     args = parse_args()
 
+    if args.tokenizer_name_or_path is None:
+        args.tokenizer_name_or_path = args.model_name_or_path
+
     device = torch.device(0)
 
     model = load_model(args.model_name_or_path)
-    tokenizer = load_tokenizer(args.model_name_or_path)
+    tokenizer = load_tokenizer(args.tokenizer_name_or_path)
 
     model = model.to(device)
     prompts = [
@@ -87,7 +105,9 @@ if __name__ == '__main__':
         "这是一个最好的时代，这是一个最坏的时代。",
         "这是一个最好的时代，这是一个最坏的",
     ]
-    completes = eval_prompts(model, tokenizer, prompts)
+    completes = eval_prompts(
+        model, tokenizer, prompts, use_tril_attention_mask=args.use_tril_attention_mask
+    )
 
     for prompt, complete in zip(prompts, completes):
         print("[p]", prompt)

train.py

@@ -103,6 +103,11 @@ def parse_args():
         help="Name of or path to model",
         default='gpt2',
     )
+    parser.add_argument(
+        "--use_tril_attention_mask",
+        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(
@@ -165,10 +170,11 @@ def parse_args():
 
 
 class LitModule(pl.LightningModule):
-    def __init__(self, model_name: str):
+    def __init__(self, model_name: str, use_tril_attention_mask: str = False):
         super().__init__()
         self.save_hyperparameters()
         self.llm = self.register_core_module(init_model(model_name))
+        self.use_tril_attention_mask = use_tril_attention_mask
         self.metric_loss = torchmetrics.MeanMetric()
         self.metric_accuracy = torchmetrics.Accuracy(
             task='multiclass',
@@ -176,7 +182,7 @@ class LitModule(pl.LightningModule):
         )
 
     @cache
-    def get_tril_matrix(
+    def get_batch_tril_matrix(
         self, block_size: int, batch_size: Optional[int] = None
     ) -> torch.Tensor:
         matrix = torch.ones(block_size, block_size).tril()
@@ -190,9 +196,10 @@ class LitModule(pl.LightningModule):
 
     def training_step(self, batch: Dict[str, torch.Tensor], batch_idx):
         batch_size, block_size = batch['input_ids'].shape
-        batch['attention_mask'] = self.get_tril_matrix(
-            block_size, batch_size=batch_size
-        ).to(self.device)
+        if self.use_tril_attention_mask:
+            batch['attention_mask'] = self.get_batch_tril_matrix(
+                block_size, batch_size=batch_size
+            ).to(self.device)
         outputs = self.llm(**batch, return_dict=True)
         loss = outputs.loss
 
@@ -244,7 +251,7 @@ if __name__ == '__main__':
     set_seed(args.seed)
 
     # lightning module
-    lit_module = LitModule(args.model_name)
+    lit_module = LitModule(args.model_name, args.use_tril_attention_mask)
 
     # datasets
     tokenizer = load_tokenizer(args.tokenizer_name_or_path)