[Performance improvement] "Bad tokens ids" optimization (#6064)

* Optimized banned token masking

* Avoid duplicate EOS masking if in bad_words_id

* Updated mask generation to handle empty banned token list

* Addition of unit tests for the updated bad_words_ids masking

* Updated timeout handling in `test_postprocess_next_token_scores_large_bad_words_list` unit test

* Updated timeout handling in `test_postprocess_next_token_scores_large_bad_words_list` unit test (timeout does not work on Windows)

* Moving Marian import to the test context to allow TF only environments to run

* Moving imports to torch_available test

* Updated operations device and test

* Updated operations device and test

* Added docstring and comment for in-place scores modification

* Moving test to own test_generation_utils, use of lighter models for testing

* removed unneded imports in test_modeling_common

* revert formatting change for ModelTesterMixin

* Updated caching, simplified eos token id test, removed unnecessary @require_torch

* formatting compliance

src/transformers/data/test_generation_utils.py

@@ -0,0 +1,90 @@
+import random
+import unittest
+
+import timeout_decorator
+
+from transformers import is_torch_available
+from transformers.file_utils import cached_property
+from transformers.testing_utils import require_torch
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        MarianConfig,
+        MarianMTModel,
+    )
+
+
+@require_torch
+class GenerationUtilsTest(unittest.TestCase):
+    @cached_property
+    def config(self):
+        config = MarianConfig.from_pretrained("sshleifer/tiny-marian-en-de")
+        return config
+
+    @cached_property
+    def model(self):
+        return MarianMTModel(self.config)
+
+    def test_postprocess_next_token_scores(self):
+        config = self.config
+        model = self.model
+        # Initialize an input id tensor with batch size 8 and sequence length 12
+        input_ids = torch.arange(0, 96, 1).view((8, 12))
+        eos = config.eos_token_id
+        bad_words_ids_test_cases = [[[299]], [[23, 24], [54]], [[config.eos_token_id]], []]
+        masked_scores = [
+            [(0, 299), (1, 299), (2, 299), (3, 299), (4, 299), (5, 299), (6, 299), (7, 299)],
+            [(1, 24), (0, 54), (1, 54), (2, 54), (3, 54), (4, 54), (5, 54), (6, 54), (7, 54)],
+            [(0, eos), (1, eos), (2, eos), (3, eos), (4, eos), (5, eos), (6, eos), (7, eos)],
+            [],
+        ]
+
+        for test_case_index, bad_words_ids in enumerate(bad_words_ids_test_cases):
+            # Initialize a scores tensor with batch size 8 and vocabulary size 300
+            scores = torch.rand((8, 300))
+            output = model.postprocess_next_token_scores(
+                scores,
+                input_ids,
+                0,
+                bad_words_ids,
+                13,
+                15,
+                config.max_length,
+                config.eos_token_id,
+                config.repetition_penalty,
+                32,
+                5,
+            )
+            for masked_score in masked_scores[test_case_index]:
+                self.assertTrue(output[masked_score[0], masked_score[1]] == -float("inf"))
+
+    @timeout_decorator.timeout(10)
+    def test_postprocess_next_token_scores_large_bad_words_list(self):
+
+        config = self.config
+        model = self.model
+        # Initialize an input id tensor with batch size 8 and sequence length 12
+        input_ids = torch.arange(0, 96, 1).view((8, 12))
+
+        bad_words_ids = []
+        for _ in range(100):
+            length_bad_word = random.randint(1, 4)
+            bad_words_ids.append(random.sample(range(1, 300), length_bad_word))
+
+        scores = torch.rand((8, 300))
+        _ = model.postprocess_next_token_scores(
+            scores,
+            input_ids,
+            0,
+            bad_words_ids,
+            13,
+            15,
+            config.max_length,
+            config.eos_token_id,
+            config.repetition_penalty,
+            32,
+            5,
+        )

src/transformers/generation_utils.py

@@ -15,7 +15,7 @@
 # limitations under the License.
 
 import logging
-from typing import Iterable, Optional, Tuple
+from typing import Iterable, List, Optional, Tuple
 
 import torch
 from torch import Tensor
@@ -89,11 +89,12 @@ class GenerationMixin:
                 scores[i, banned_tokens] = -float("inf")
 
         if bad_words_ids is not None:
+            # Exclude EOS token (already processed)
+            bad_words_ids = list(filter(lambda bad_token_seq: bad_token_seq != [eos_token_id], bad_words_ids))
             # calculate a list of banned tokens according to bad words
-            banned_tokens = calc_banned_bad_words_ids(input_ids, bad_words_ids)
-
-            for i, banned_tokens in enumerate(banned_tokens):
-                scores[i, banned_tokens] = -float("inf")
+            banned_tokens = calc_banned_bad_words_ids(input_ids.tolist(), bad_words_ids)
+            # Modify the scores in place by setting the banned tokens logits to `-inf`
+            set_scores_to_inf_for_banned_tokens(scores, banned_tokens)
 
         return scores
 
@@ -893,7 +894,7 @@ def calc_banned_bad_words_ids(prev_input_ids: Iterable[int], bad_words_ids: Iter
                 bad_words_ids
             )
 
-            if _tokens_match(prev_input_ids_slice.tolist(), banned_token_seq[:-1]) is False:
+            if _tokens_match(prev_input_ids_slice, banned_token_seq[:-1]) is False:
                 # if tokens do not match continue
                 continue
 
@@ -904,6 +905,30 @@ def calc_banned_bad_words_ids(prev_input_ids: Iterable[int], bad_words_ids: Iter
     return banned_tokens
 
 
+def set_scores_to_inf_for_banned_tokens(scores: torch.Tensor, banned_tokens: List[List[int]]) -> None:
+    """ Modifies the scores in place by setting the banned token positions to `-inf`. Banned token is expected to be
+    a list of list of banned tokens to ban in the format [[batch index, vocabulary position],...]
+        Args:
+            scores: logits distribution of shape (batch size, vocabulary size)
+            banned_tokens: list of list of tokens to ban of length (batch_size)
+    """
+    banned_mask_list = []
+    for idx, batch_banned_tokens in enumerate(banned_tokens):
+        for token in batch_banned_tokens:
+            banned_mask_list.append([idx, token])
+    if not banned_mask_list:
+        return
+    banned_mask = torch.LongTensor(banned_mask_list)
+    indices = torch.ones(len(banned_mask))
+    # A sparse tensor is generated from a list of coordinates: [[0, 1], [0, 2], [2, 0]]. A conversion to dense tensor generates:
+    # [ 0  1  1 ]
+    # [ 0  0  0 ]
+    # [ 1  0  0 ]
+
+    banned_mask = torch.sparse.LongTensor(banned_mask.t(), indices, scores.size()).to(scores.device).to_dense().bool()
+    scores.masked_fill_(banned_mask, -float("inf"))
+
+
 def top_k_top_p_filtering(
     logits: Tensor,
     top_k: int = 0,