Fix contrastive search to correctly handle input with padding (huggin…

…gface#33507)

* fix: handle padding in contrastive search for decoder-only models

* fix: handle padding in contrastive search for encoder-decoder models

* tests: move padding contrastive test to test_util, add t5 test

* fix: handle if model_kwargs["decoder_attention_mask"] is None

* refactor: improve padding input contrastive search generation tests

* chore: _ranking_fast to use LongTensor for cosine_matrix_mask

src/transformers/generation/utils.py

@@ -2604,6 +2604,15 @@ def _contrastive_search(
         unfinished_sequences = torch.ones(batch_size, dtype=torch.long, device=input_ids.device)
         model_kwargs = self._get_initial_cache_position(input_ids, model_kwargs)
 
+        # Create cosine_matrix_mask based on the attention_mask
+        cosine_matrix_mask = torch.ones_like(input_ids, dtype=torch.long)
+        if self.config.is_encoder_decoder:
+            if "decoder_attention_mask" in model_kwargs and model_kwargs["decoder_attention_mask"] is not None:
+                cosine_matrix_mask = model_kwargs["decoder_attention_mask"]
+        else:
+            cosine_matrix_mask = model_kwargs["attention_mask"]
+        cosine_matrix_mask = cosine_matrix_mask.repeat_interleave(top_k, dim=0)
+
         this_peer_finished = False
 
         while self._has_unfinished_sequences(this_peer_finished, synced_gpus, device=input_ids.device):
@@ -2771,7 +2780,12 @@ def _contrastive_search(
             # compute the degeneration penalty and re-rank the candidates based on the degeneration penalty and the
             # model confidence. Keeping `selected_idx` on CPU enables multi-device contrastive search and doesn't
             # introduce (noticeable) slowdowns on single-device runs.
-            selected_idx = _ranking_fast(context_hidden, next_hidden, top_k_probs, penalty_alpha, top_k)
+            selected_idx = _ranking_fast(
+                context_hidden, next_hidden, top_k_probs, cosine_matrix_mask, penalty_alpha, top_k
+            )
+            cosine_matrix_mask = torch.cat(
+                [cosine_matrix_mask, cosine_matrix_mask.new_ones((cosine_matrix_mask.shape[0], 1))], dim=-1
+            )
             selected_idx = selected_idx.to("cpu")
 
             # This will be used instead of the previous inneficient torch.stack(torch.split())
@@ -4283,6 +4297,7 @@ def _ranking_fast(
     context_hidden: torch.FloatTensor,
     next_hidden: torch.FloatTensor,
     next_top_k_probs: torch.FloatTensor,
+    cosine_matrix_mask: torch.LongTensor,
     alpha: float,
     beam_width: int,
 ) -> torch.FloatTensor:
@@ -4294,6 +4309,13 @@ def _ranking_fast(
     norm_context_hidden = context_hidden / context_hidden.norm(dim=2, keepdim=True)
     norm_next_hidden = next_hidden / next_hidden.norm(dim=2, keepdim=True)
     cosine_matrix = torch.matmul(norm_context_hidden, norm_next_hidden.transpose(1, 2)).squeeze(-1)  # [B*K, S]
+
+    # Penalize cosine_matrix based on the cosine_matrix_mask (ignore padding positions)
+    # Using a large negative value for masked positions
+    cosine_matrix_mask = cosine_matrix_mask.to(dtype=cosine_matrix.dtype)
+    cosine_matrix_mask = (1 - cosine_matrix_mask) * torch.finfo(cosine_matrix.dtype).min
+    cosine_matrix = cosine_matrix + cosine_matrix_mask
+
     degeneration_penalty, _ = torch.max(cosine_matrix, dim=-1)  # [B*K]
     next_top_k_probs = next_top_k_probs.view(-1)  # [B*K]
     contrastive_score = (1.0 - alpha) * next_top_k_probs - alpha * degeneration_penalty

tests/generation/test_utils.py

@@ -44,6 +44,7 @@
 
 if is_torch_available():
     import torch
+    import torch.nn.functional as F
 
     from transformers import (
         AutoModelForCausalLM,
@@ -59,6 +60,7 @@
         GPT2Tokenizer,
         ImageGPTForCausalImageModeling,
         SpeechEncoderDecoderModel,
+        T5ForConditionalGeneration,
     )
     from transformers.cache_utils import DynamicCache, EncoderDecoderCache, QuantoQuantizedCache, StaticCache
     from transformers.generation import (
@@ -3644,6 +3646,139 @@ def test_init_static_cache_multi_gpu(self):
         value_cache_1 = results.past_key_values.value_cache[1]
         self.assertTrue(key_cache_1.device == value_cache_1.device == torch.device(1))
 
+    @slow
+    def test_padding_input_contrastive_search_gpt2(self):
+        # Load the pre-trained GPT-2 model and tokenizer
+        model = GPT2LMHeadModel.from_pretrained("openai-community/gpt2")
+        model.to(torch_device)
+        tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2", clean_up_tokenization_spaces=True)
+
+        # Set the tokenizer to left-pad the sequences
+        tokenizer.padding_side = "left"
+
+        # Define the PAD token as the EOS token
+        tokenizer.pad_token = tokenizer.eos_token
+        model.generation_config.pad_token_id = model.generation_config.eos_token_id
+
+        # Define the input prompt
+        prompt_text = "The whispered legends of the haunted mansion spoke"
+
+        # Tokenize the input prompt
+        encoded_prompt = tokenizer(prompt_text, return_tensors="pt", padding=True)
+        input_ids = encoded_prompt.input_ids.to(torch_device)
+        attention_mask = encoded_prompt.attention_mask.to(torch_device)
+
+        # Define the contrastive search params
+        penalty_alpha = 0.6
+        top_k = 4
+
+        # Define the padding length to add to the input IDs and attention mask
+        padding_length = 10
+
+        # Generate text without padding
+        outputs = model.generate(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            do_sample=False,
+            penalty_alpha=penalty_alpha,
+            top_k=top_k,
+            max_new_tokens=64,
+        )
+        generated_text_no_padding = tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+        # Pad the input IDs and attention mask on the left
+        padded_input_ids = F.pad(
+            input_ids, (padding_length, 0), "constant", value=model.generation_config.pad_token_id
+        )
+        padded_attention_mask = F.pad(attention_mask, (padding_length, 0), "constant", value=0)
+
+        # Generate text with padded inputs
+        outputs_with_padding = model.generate(
+            input_ids=padded_input_ids,
+            attention_mask=padded_attention_mask,
+            do_sample=False,
+            penalty_alpha=penalty_alpha,
+            top_k=top_k,
+            max_new_tokens=64,
+        )
+        generated_text_with_padding = tokenizer.decode(outputs_with_padding[0], skip_special_tokens=True)
+
+        # Assert that the generated texts are identical for padded and non-padded inputs
+        self.assertEqual(generated_text_no_padding, generated_text_with_padding)
+        self.assertEqual(
+            generated_text_with_padding,
+            'The whispered legends of the haunted mansion spoke of the "souls of the dead" who were "falling '
+            'out of the sky" and "falling into the sea."\n\nThe ghostly apparitions were said to have been '
+            'created by the spirits of the dead, who were "falling out of the sky" and "falling into the sea',
+        )
+
+    @slow
+    def test_padding_input_contrastive_search_t5(self):
+        # Load the pre-trained T5 model and tokenizer
+        model = T5ForConditionalGeneration.from_pretrained("google-t5/t5-small")
+        model.to(torch_device)
+        tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-small", clean_up_tokenization_spaces=True)
+
+        # Define the input prompt
+        prompt_text = "translate English to German: I need to finish this task before the end of the day."
+
+        # Tokenize the input prompt
+        encoded_prompt = tokenizer(prompt_text, return_tensors="pt")
+        input_ids = encoded_prompt.input_ids.to(torch_device)
+        attention_mask = encoded_prompt.attention_mask.to(torch_device)
+
+        # Define the decoder prompt
+        decoder_prompt_text = "Ich muss diese Aufgabe"
+        encoded_decoder_prompt = tokenizer(decoder_prompt_text, add_special_tokens=False, return_tensors="pt")
+        decoder_input_ids = encoded_decoder_prompt.input_ids.to(torch_device)
+        decoder_attention_mask = encoded_decoder_prompt.attention_mask.to(torch_device)
+
+        # Define the contrastive search params
+        penalty_alpha = 0.6
+        top_k = 4
+
+        # Generate text without padding
+        outputs = model.generate(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            do_sample=False,
+            penalty_alpha=penalty_alpha,
+            top_k=top_k,
+            max_new_tokens=64,
+        )
+        generated_text_no_padding = tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+        # Define the padding length to add to the input IDs and attention mask
+        padding_length = 10
+
+        # Pad the decoder input IDs and attention mask on the left
+        padded_decoder_input_ids = F.pad(
+            decoder_input_ids, (padding_length, 0), "constant", value=model.generation_config.pad_token_id
+        )
+        padded_decoder_attention_mask = F.pad(decoder_attention_mask, (padding_length, 0), "constant", value=0)
+        # Since the decoder_start_token_id is the same as the pad_token_id,
+        # the last padded token represents the decoder start token.
+        # Set the attention mask for the decoder_start_token_id to True (1).
+        padded_decoder_attention_mask[:, padding_length - 1] = 1
+        # Generate text with padded inputs
+        outputs_with_padding = model.generate(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            decoder_input_ids=padded_decoder_input_ids,
+            decoder_attention_mask=padded_decoder_attention_mask,
+            do_sample=False,
+            penalty_alpha=penalty_alpha,
+            top_k=top_k,
+            max_new_tokens=64,
+        )
+        generated_text_with_padding = tokenizer.decode(outputs_with_padding[0], skip_special_tokens=True)
+
+        # Assert that the generated texts are identical for padded and non-padded inputs
+        self.assertEqual(generated_text_no_padding, generated_text_with_padding)
+        self.assertEqual(generated_text_no_padding, "Ich muss diese Aufgabe vor Ende des Tages beenden.")
+
 
 @require_torch
 class TokenHealingTestCase(unittest.TestCase):