use logger.warning_once to avoid massive outputs (huggingface#27428)

* use logger.warning_once to avoid massive outputs when training/finetuning longformer

* update more

src/transformers/models/big_bird/modeling_big_bird.py

@@ -2223,7 +2223,7 @@ def _pad_to_block_size(
 
         padding_len = (block_size - seq_len % block_size) % block_size
         if padding_len > 0:
-            logger.info(
+            logger.warning_once(
                 f"Input ids are automatically padded from {seq_len} to {seq_len + padding_len} to be a multiple of "
                 f"`config.block_size`: {block_size}"
             )

src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -2014,7 +2014,7 @@ def _pad_to_block_size(self, hidden_states: torch.Tensor, attention_mask: torch.
 
         padding_len = (block_size - seq_len % block_size) % block_size
         if padding_len > 0:
-            logger.info(
+            logger.warning_once(
                 f"Input ids are automatically padded from {seq_len} to {seq_len + padding_len} to be a multiple of "
                 f"`config.block_size`: {block_size}"
             )

src/transformers/models/led/modeling_led.py

@@ -1706,7 +1706,7 @@ def _pad_to_window_size(
 
         padding_len = (attention_window - seq_len % attention_window) % attention_window
         if padding_len > 0:
-            logger.info(
+            logger.warning_once(
                 f"Input ids are automatically padded from {seq_len} to {seq_len + padding_len} to be a multiple of "
                 f"`config.attention_window`: {attention_window}"
             )

src/transformers/models/led/modeling_tf_led.py

@@ -1859,7 +1859,7 @@ def _pad_to_window_size(
         padding_len = (attention_window - seq_len % attention_window) % attention_window
 
         if padding_len > 0:
-            logger.info(
+            logger.warning_once(
                 f"Input ids are automatically padded from {seq_len} to {seq_len + padding_len} to be a multiple of "
                 f"`config.attention_window`: {attention_window}"
             )

src/transformers/models/longformer/modeling_longformer.py

@@ -1598,7 +1598,7 @@ def _pad_to_window_size(
 
         # this path should be recorded in the ONNX export, it is fine with padding_len == 0 as well
         if padding_len > 0:
-            logger.info(
+            logger.warning_once(
                 f"Input ids are automatically padded from {seq_len} to {seq_len + padding_len} to be a multiple of "
                 f"`config.attention_window`: {attention_window}"
             )
@@ -1917,7 +1917,7 @@ def forward(
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         if global_attention_mask is None:
-            logger.info("Initializing global attention on CLS token...")
+            logger.warning_once("Initializing global attention on CLS token...")
             global_attention_mask = torch.zeros_like(input_ids)
             # global attention on cls token
             global_attention_mask[:, 0] = 1
@@ -2270,7 +2270,7 @@ def forward(
 
         # set global attention on question tokens
         if global_attention_mask is None and input_ids is not None:
-            logger.info("Initializing global attention on multiple choice...")
+            logger.warning_once("Initializing global attention on multiple choice...")
             # put global attention on all tokens after `config.sep_token_id`
             global_attention_mask = torch.stack(
                 [

src/transformers/models/longformer/modeling_tf_longformer.py

@@ -2213,7 +2213,7 @@ def call(
                 )
                 global_attention_mask = tf.cast(tf.fill(shape_list(input_ids), value=0), tf.int64)
             else:
-                logger.info("Initializing global attention on question tokens...")
+                logger.warning_once("Initializing global attention on question tokens...")
                 # put global attention on all tokens until `config.sep_token_id` is reached
                 sep_token_indices = tf.where(input_ids == self.config.sep_token_id)
                 sep_token_indices = tf.cast(sep_token_indices, dtype=tf.int64)
@@ -2341,7 +2341,7 @@ def call(
             global_attention_mask = tf.cast(global_attention_mask, tf.int64)
 
         if global_attention_mask is None and input_ids is not None:
-            logger.info("Initializing global attention on CLS token...")
+            logger.warning_once("Initializing global attention on CLS token...")
             # global attention on cls token
             global_attention_mask = tf.zeros_like(input_ids)
             updates = tf.ones(shape_list(input_ids)[0], dtype=tf.int64)

src/transformers/models/reformer/modeling_reformer.py

@@ -2139,7 +2139,7 @@ def _pad_to_mult_of_chunk_length(
         padded_seq_length=None,
         device=None,
     ):
-        logger.info(
+        logger.warning_once(
             f"Input ids are automatically padded from {input_shape[-1]} to {input_shape[-1] + padding_length} to be a "
             f"multiple of `config.chunk_length`: {padded_seq_length}"
         )