Add validation for maximum sequence length in modeling_whisper.py

src/transformers/models/whisper/modeling_whisper.py

@@ -1671,6 +1671,7 @@ def __init__(self, config: WhisperConfig):
         super().__init__(config)
         self.model = WhisperModel(config)
         self.proj_out = nn.Linear(config.d_model, config.vocab_size, bias=False)
+        self.max_target_positions = config.max_target_positions
 
         # Initialize weights and apply final processing
         self.post_init()
@@ -1723,7 +1724,7 @@ def forward(
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]`
             or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is
-            only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+            only computed for the tokens with labels in `[0, ..., config.vocab_size]`. `sequence_length` should be smaller than or equal to `config.max_target_positions`.
 
         Returns:
 
@@ -1751,6 +1752,10 @@ def forward(
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         if labels is not None:
+            if labels.shape[1] > self.max_target_positions:
+                raise ValueError(
+                    f"Labels' sequence length {labels.shape[1]} cannot exceed the maximum allowed length of {self.max_target_positions} tokens."
+                )
             if decoder_input_ids is None and decoder_inputs_embeds is None:
                 decoder_input_ids = shift_tokens_right(
                     labels, self.config.pad_token_id, self.config.decoder_start_token_id

tests/models/whisper/test_modeling_whisper.py

@@ -1676,6 +1676,63 @@ def test_flash_attn_2_generate_reuse_cache(self):
                     past_key_values=past_key_values,
                 )
 
+    def test_labels_sequence_max_length_correct(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_generative_model_classes:
+            input_features = input_dict["input_features"]
+
+            labels_length = config.max_target_positions
+            labels = torch.ones(1, labels_length, dtype=torch.int64)
+
+            model = model_class(config)
+            model(input_features=input_features, labels=labels)
+
+    def test_labels_sequence_max_length_correct_after_changing_config(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_generative_model_classes:
+            input_features = input_dict["input_features"]
+
+            config.max_target_positions += 100
+
+            labels_length = config.max_target_positions
+            labels = torch.ones(1, labels_length, dtype=torch.int64)
+
+            model = model_class(config)
+            model(input_features=input_features, labels=labels)
+
+    def test_labels_sequence_max_length_error(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_generative_model_classes:
+            input_features = input_dict["input_features"]
+
+            labels_length = config.max_target_positions + 1
+            labels = torch.ones(1, labels_length, dtype=torch.int64)
+
+            model = model_class(config)
+            with self.assertRaises(ValueError):
+                model(input_features=input_features, labels=labels)
+
+    def test_labels_sequence_max_length_error_after_changing_config(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+            input_features = input_dict["input_features"]
+
+            labels_length = config.max_target_positions + 1
+            labels = torch.ones(1, labels_length, dtype=torch.int64)
+
+            new_max_length = config.max_target_positions + 100
+            model.config.max_length = new_max_length
+            model.generation_config.max_length = new_max_length
+            config.max_target_positions = new_max_length
+
+            with self.assertRaises(ValueError):
+                model(input_features=input_features, labels=labels)
+
 
 @require_torch
 @require_torchaudio